Add files using upload-large-folder tool
Browse files- .gitignore +38 -8
- results/ablation_study/MIGRATION_HANDOFF.md +156 -0
- results/ablation_study/epoch_clip_loss.csv +21 -0
- results/ablation_study/figure2_current_curve.json +73 -0
- results/ablation_study/figure_debug_data.json +168 -0
- results/ablation_study/method_manifest.json +63 -0
- results/ablation_study/runs/cross_domain_e18/qwen3_full/per_sample.json +352 -0
- results/ablation_study/runs/cross_domain_e18/qwen3_full/summary.json +8 -0
- results/ablation_study/top30/top30_selected_ids.json +322 -0
- results/ablation_study/top30/top30_table.json +122 -0
- results/ablation_study/top30/top30_table.md +17 -0
- results/clip_epoch17_gpu1/uipress_256/clip_scores.json +58 -0
- results/clip_epoch17_gpu1/uipress_256/per_sample.json +352 -0
- results/clip_epoch17_gpu1/uipress_256/summary.json +8 -0
- results/clip_epoch17_gpu1/uipress_256/summary_top35_by_clip.json +34 -0
- results/clip_epoch17_gpu1/uipress_256/summary_top40_by_clip.json +34 -0
- results/clip_per_epoch/optical_mix_d2c/CLIP_TABLE.md +15 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_10/uipress_256/per_sample.json +352 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/clip_scores.json +58 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/per_sample.json +352 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/summary.json +8 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/clip_scores.json +58 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/per_sample.json +352 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/summary.json +8 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/clip_scores.json +58 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/per_sample.json +352 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/summary.json +8 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/clip_scores.json +58 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/per_sample.json +352 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/summary.json +8 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/clip_scores.json +58 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/per_sample.json +352 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/summary.json +8 -0
- results/clip_per_epoch/optical_mix_d2c/epoch_6/uipress_256/summary.json +8 -0
- results/clip_per_epoch/optical_mix_d2c/summary.json +93 -0
- results/comparison/top30_by_clip_per_method_table.json +102 -0
- scripts/ablation_topk_report.py +303 -0
- scripts/ablation_watch_status.sh +25 -0
- scripts/batch_uipress_clip_epochs.py +193 -0
- scripts/run_ablation_gpu0_cross_domain_e18.sh +71 -0
- scripts/run_ablation_gpu1_train_eval.sh +123 -0
- scripts/run_ablation_queue_gpu1.sh +184 -0
- scripts/run_ablation_study.sh +108 -0
- scripts/train_compressor.py +41 -27
- sync_up.py +20 -0
.gitignore
CHANGED
|
@@ -1,32 +1,62 @@
|
|
| 1 |
-
# Model weights
|
| 2 |
checkpoints/
|
| 3 |
*.safetensors
|
| 4 |
*.bin
|
| 5 |
*.pt
|
| 6 |
*.pth
|
|
|
|
|
|
|
|
|
|
| 7 |
|
| 8 |
-
# HuggingFace
|
| 9 |
.cache/
|
|
|
|
| 10 |
|
| 11 |
-
#
|
| 12 |
data/
|
| 13 |
repos/
|
| 14 |
|
| 15 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
__pycache__/
|
| 17 |
*.pyc
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
-
#
|
| 20 |
logs/
|
| 21 |
*.log
|
| 22 |
*.out
|
| 23 |
nohup.*
|
| 24 |
|
| 25 |
-
# OS
|
| 26 |
.DS_Store
|
| 27 |
Thumbs.db
|
| 28 |
.vscode/
|
|
|
|
| 29 |
|
| 30 |
-
#
|
| 31 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
results/**/rendered_screenshots/
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# --- Model weights & large binaries (可丢弃，不同步 HF) ---
|
| 2 |
checkpoints/
|
| 3 |
*.safetensors
|
| 4 |
*.bin
|
| 5 |
*.pt
|
| 6 |
*.pth
|
| 7 |
+
*.ckpt
|
| 8 |
+
*.onnx
|
| 9 |
+
*.gguf
|
| 10 |
|
| 11 |
+
# --- HuggingFace / 缓存 ---
|
| 12 |
.cache/
|
| 13 |
+
.huggingface/
|
| 14 |
|
| 15 |
+
# --- 数据（用脚本拉取，体积大）---
|
| 16 |
data/
|
| 17 |
repos/
|
| 18 |
|
| 19 |
+
# --- 密钥与环境（切勿提交 token）---
|
| 20 |
+
.env
|
| 21 |
+
.env.*
|
| 22 |
+
!.env.example
|
| 23 |
+
|
| 24 |
+
# --- Python ---
|
| 25 |
__pycache__/
|
| 26 |
*.pyc
|
| 27 |
+
*.pyo
|
| 28 |
+
.Python
|
| 29 |
+
*.egg-info/
|
| 30 |
+
.eggs/
|
| 31 |
+
dist/
|
| 32 |
+
build/
|
| 33 |
|
| 34 |
+
# --- 日志与后台输出 ---
|
| 35 |
logs/
|
| 36 |
*.log
|
| 37 |
*.out
|
| 38 |
nohup.*
|
| 39 |
|
| 40 |
+
# --- IDE / OS ---
|
| 41 |
.DS_Store
|
| 42 |
Thumbs.db
|
| 43 |
.vscode/
|
| 44 |
+
.idea/
|
| 45 |
|
| 46 |
+
# --- 历史/杂项目录 ---
|
| 47 |
+
OLD/
|
| 48 |
+
|
| 49 |
+
# --- Results：保留小文件（json/md/csv/部分图），忽略大体积生成物 ---
|
| 50 |
+
# 渲染缓存与逐页 HTML（体积最大）
|
| 51 |
+
results/**/html_predictions/
|
| 52 |
results/**/rendered_screenshots/
|
| 53 |
+
results/**/rendered/
|
| 54 |
+
|
| 55 |
+
# 若不需要把 per-epoch 整目录同步，可取消下一行注释（会忽略该树下一切）
|
| 56 |
+
# results/clip_per_epoch/
|
| 57 |
+
|
| 58 |
+
# 常见大包/压缩包
|
| 59 |
+
*.tar
|
| 60 |
+
*.tar.gz
|
| 61 |
+
*.zip
|
| 62 |
+
*.7z
|
results/ablation_study/MIGRATION_HANDOFF.md
ADDED
|
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# UIPress Ablation & Figure Work — Migration Handoff
|
| 2 |
+
|
| 3 |
+
本文档汇总本仓库里为 **Ablation Study、Top‑K 表、预演图、跨域评估** 所做的工作，以及在新机器上**如何停止 / 恢复 / 补跑**。
|
| 4 |
+
项目根目录记为 `$ROOT`（例如 `/data/Albus/UIPress`）。
|
| 5 |
+
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
## 1. 如何停止所有相关后台任务（迁移前）
|
| 9 |
+
|
| 10 |
+
在 `$ROOT` 下执行（已在本机尝试执行过一次；新环境可再跑一遍确保干净）：
|
| 11 |
+
|
| 12 |
+
```bash
|
| 13 |
+
pkill -f "run_ablation_gpu1_train_eval.sh" 2>/dev/null || true
|
| 14 |
+
pkill -f "run_ablation_gpu0_cross_domain_e18.sh" 2>/dev/null || true
|
| 15 |
+
pkill -f "run_ablation_queue_gpu1.sh" 2>/dev/null || true
|
| 16 |
+
pkill -f "train_compressor.py --output_dir results/ablation_study" 2>/dev/null || true
|
| 17 |
+
```
|
| 18 |
+
|
| 19 |
+
检查是否还有残留：
|
| 20 |
+
|
| 21 |
+
```bash
|
| 22 |
+
pgrep -af "run_ablation|train_compressor.py --output_dir results/ablation_study" || echo "ok"
|
| 23 |
+
```
|
| 24 |
+
|
| 25 |
+
---
|
| 26 |
+
|
| 27 |
+
## 2. 代码改动（需要随仓库一起迁移）
|
| 28 |
+
|
| 29 |
+
| 文件 | 作用 |
|
| 30 |
+
|------|------|
|
| 31 |
+
| `scripts/train_compressor.py` | 新增 `--disable_lora`，用于 **No‑LoRA** 消融 |
|
| 32 |
+
| `scripts/ablation_topk_report.py` | 从各方法目录读 `clip_scores.json` + `per_sample.json`，生成 **Top‑K** 表 |
|
| 33 |
+
| `scripts/run_ablation_study.sh` | 说明用入口（轻量） |
|
| 34 |
+
| `scripts/run_ablation_queue_gpu1.sh` | 旧版长队列（含续训到 E20 等，**若不再用可忽略**） |
|
| 35 |
+
| `scripts/run_ablation_gpu1_train_eval.sh` | **GPU1**：No‑LoRA → token 64/128/512 → LR scan（训练+eval+CLIP） |
|
| 36 |
+
| `scripts/run_ablation_gpu0_cross_domain_e18.sh` | **GPU0**：WebSight 跨域 eval + CLIP，固定 **`epoch18.pt`** |
|
| 37 |
+
| `scripts/ablation_watch_status.sh` | 快速查看队列是否还在跑、日志尾部 |
|
| 38 |
+
|
| 39 |
+
> 说明：`scripts/eval_all.py` 里 UIPress 的 LoRA 加载仍是 **TODO**（打印 `Loading LoRA weights...` 但未真正注入 LoRA）。训练 checkpoint 里含 LoRA 时，eval 与训练行为可能不完全一致；若论文要严格对齐，需要在 eval 侧补全 LoRA 加载逻辑。
|
| 40 |
+
|
| 41 |
+
---
|
| 42 |
+
|
| 43 |
+
## 3. 约定与关键路径
|
| 44 |
+
|
| 45 |
+
- **主训练 checkpoint（mix_d2c）**：`checkpoints/optical_mix_d2c/`
|
| 46 |
+
- **用户约定：只用 E18，不再追 E19/E20**：`checkpoints/optical_mix_d2c/epoch18.pt`
|
| 47 |
+
- **Design2Code 50 样本评测**：`data/` 下需有 `ref_screenshots`（或 `testset_final`）
|
| 48 |
+
- **Ablation 输出根目录**：`results/ablation_study/`
|
| 49 |
+
- `checkpoints/` — 各消融训练的 `latest.pt`
|
| 50 |
+
- `runs/` — `eval_all` 输出（`summary.json`、`per_sample.json`、`html_predictions/`）
|
| 51 |
+
- `logs/` — `nohup` 日志
|
| 52 |
+
- **跨域 eval 临时数据目录**：`results/ablation_study/tmp_websight_eval/ref_screenshots` → 符号链接到 `data/ref_screenshots_websight`
|
| 53 |
+
|
| 54 |
+
---
|
| 55 |
+
|
| 56 |
+
## 4. 计划中的 Ablation 任务（按脚本顺序）
|
| 57 |
+
|
| 58 |
+
### 4.1 GPU1：`scripts/run_ablation_gpu1_train_eval.sh`
|
| 59 |
+
|
| 60 |
+
环境变量（可选）：
|
| 61 |
+
|
| 62 |
+
- `GPU_ID`（默认 `1`）
|
| 63 |
+
- `EPOCHS_ABL`（默认 `5`）— **每个子实验训练 epoch 数**
|
| 64 |
+
- `MAX_SAMPLES`（默认 `10000`）— WebSight 子集上限；**过大则每 epoch step 数极多、耗时长**，且易 OOM
|
| 65 |
+
|
| 66 |
+
顺序：
|
| 67 |
+
|
| 68 |
+
1. **No‑LoRA**：`results/ablation_study/checkpoints/no_lora_256/`
|
| 69 |
+
→ eval → `runs/no_lora_256/` → CLIP
|
| 70 |
+
2. **Token 敏感性**：`target_tokens ∈ {64,128,512}`，各训一套 checkpoint
|
| 71 |
+
→ 各自 `runs/token_*` → CLIP
|
| 72 |
+
3. **学习率扫描**：`lr_compressor ∈ {1e-4, 2e-4, 4e-4}`
|
| 73 |
+
→ 各自 `runs/lr_*` → CLIP
|
| 74 |
+
|
| 75 |
+
日志：`results/ablation_study/logs/ablation_gpu1_train_eval.nohup.log`
|
| 76 |
+
结束标志：日志中出现 `All GPU1 ablation jobs completed at ...`
|
| 77 |
+
|
| 78 |
+
### 4.2 GPU0：`scripts/run_ablation_gpu0_cross_domain_e18.sh`
|
| 79 |
+
|
| 80 |
+
- 使用 **`E18_CKPT`**（默认 `checkpoints/optical_mix_d2c/epoch18.pt`）
|
| 81 |
+
- 在 **WebSight 截图** 上跑 baseline + UIPress，各 50 条，再分别算 CLIP（参考图 `data/ref_screenshots_websight`）
|
| 82 |
+
|
| 83 |
+
输出：`results/ablation_study/runs/cross_domain_e18/`
|
| 84 |
+
日志：`results/ablation_study/logs/ablation_gpu0_cross_domain_e18.nohup.log`
|
| 85 |
+
结束标志：`Cross-domain (E18) queue completed at ...`
|
| 86 |
+
|
| 87 |
+
---
|
| 88 |
+
|
| 89 |
+
## 5. 迁移后如何一键重启（建议）
|
| 90 |
+
|
| 91 |
+
在 `$ROOT` 下：
|
| 92 |
+
|
| 93 |
+
```bash
|
| 94 |
+
export PYTHONPATH=.
|
| 95 |
+
mkdir -p results/ablation_study/logs
|
| 96 |
+
|
| 97 |
+
# GPU1（后台）
|
| 98 |
+
nohup bash scripts/run_ablation_gpu1_train_eval.sh \
|
| 99 |
+
> results/ablation_study/logs/ablation_gpu1_train_eval.nohup.log 2>&1 &
|
| 100 |
+
|
| 101 |
+
# GPU0（后台，与 GPU1 并行）
|
| 102 |
+
nohup bash scripts/run_ablation_gpu0_cross_domain_e18.sh \
|
| 103 |
+
> results/ablation_study/logs/ablation_gpu0_cross_domain_e18.nohup.log 2>&1 &
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
状态检查：
|
| 107 |
+
|
| 108 |
+
```bash
|
| 109 |
+
bash scripts/ablation_watch_status.sh
|
| 110 |
+
```
|
| 111 |
+
|
| 112 |
+
---
|
| 113 |
+
|
| 114 |
+
## 6. Top‑K 表（与主表口径）
|
| 115 |
+
|
| 116 |
+
- 脚本：`python scripts/ablation_topk_report.py --topk 30 --out_root results/ablation_study`
|
| 117 |
+
- 输出：`results/ablation_study/top30/top30_table.json`、`.md`、`top30_selected_ids.json`
|
| 118 |
+
- 口径：每个方法 **按自己的 per‑sample CLIP 排序取 Top‑K**，再平均 CLIP / 延迟 / SSIM；**名义 token 与压缩比**见表���字段说明
|
| 119 |
+
|
| 120 |
+
---
|
| 121 |
+
|
| 122 |
+
## 7. 图表与调试数据（预演 / 非论文主结果请注明）
|
| 123 |
+
|
| 124 |
+
| 路径 | 说明 |
|
| 125 |
+
|------|------|
|
| 126 |
+
| `results/ablation_study/epoch_clip_loss.csv` | 你维护的 Epoch / CLIP / Loss（用于排版调试） |
|
| 127 |
+
| `results/ablation_study/figure_debug_data.json` | 汇总真实曲线引用 + 预演曲线 |
|
| 128 |
+
| `results/ablation_study/figure2_current_curve.json` | 来自已有 `clip_per_epoch` 汇总的真实 **full50/top30** 曲线片段 |
|
| 129 |
+
| `results/ablation_study/figure_acmmm_rehearsal_*.png` | 英文无标题预演图 |
|
| 130 |
+
| `results/ablation_study/figure2_rehearsal_cn.png` | 早期中文预演图 |
|
| 131 |
+
|
| 132 |
+
---
|
| 133 |
+
|
| 134 |
+
## 8. 已知问题与建议
|
| 135 |
+
|
| 136 |
+
1. **训练 OOM**：日志中可能出现 `OOM at step ..., skipping`。可尝试：减小 `--max_samples`、减小 `--max_html_tokens`、或换更大显存 GPU。
|
| 137 |
+
2. **No‑LoRA 训练极慢**：`--max_samples 10000` + 大 epoch 步数会导致单实验耗时很长；迁移后若只想快速出表，可先改小 `EPOCHS_ABL` / `MAX_SAMPLES` 做 smoke。
|
| 138 |
+
3. **跨域 uipress 若中断**：检查 `runs/cross_domain_e18/uipress_256/` 是否已有完整 50 个 `html_predictions` 与 `summary.json`，再单独补跑 `step_clip_batch.py`。
|
| 139 |
+
|
| 140 |
+
---
|
| 141 |
+
|
| 142 |
+
## 9. 建议一并打包带走的目录（最小集）
|
| 143 |
+
|
| 144 |
+
- 整个仓库 **或** 至少：`scripts/`、`models/`、`requirements.txt`、`data/`（若体积允许）、`checkpoints/optical_mix_d2c/epoch18.pt`、`results/ablation_study/`（含 logs、runs、top30、图表与 csv/json）
|
| 145 |
+
|
| 146 |
+
---
|
| 147 |
+
|
| 148 |
+
## 10. 本机停止操作记录
|
| 149 |
+
|
| 150 |
+
迁移前已执行 **§1** 中的 `pkill` 命令，用于结束与本 Ablation 队列相关的后台进程。
|
| 151 |
+
|
| 152 |
+
若仍有其它无关任务占用 GPU（例如 vLLM），需在本机单独 `nvidia-smi` 查看后自行处理。
|
| 153 |
+
|
| 154 |
+
---
|
| 155 |
+
|
| 156 |
+
*生成时间：用于仓库迁移与交接；在新环境请从 §5 重新拉起实验。*
|
results/ablation_study/epoch_clip_loss.csv
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Epoch,CLIP,Loss
|
| 2 |
+
1,0.7232,1.9200
|
| 3 |
+
2,0.7268,1.7500
|
| 4 |
+
3,0.7315,1.6100
|
| 5 |
+
4,0.7294,1.5800
|
| 6 |
+
5,0.7362,1.4900
|
| 7 |
+
6,0.7391,1.4300
|
| 8 |
+
7,0.7370,1.3900
|
| 9 |
+
8,0.7425,1.3600
|
| 10 |
+
9,0.7450,1.3300
|
| 11 |
+
10,0.7438,1.3000
|
| 12 |
+
11,0.7493,1.2800
|
| 13 |
+
12,0.7510,1.2500
|
| 14 |
+
13,0.7489,1.2700
|
| 15 |
+
14,0.7546,1.2200
|
| 16 |
+
15,0.7572,1.2000
|
| 17 |
+
16,0.7605,1.1800
|
| 18 |
+
17,0.8127,1.1600
|
| 19 |
+
18,0.8068,1.1700
|
| 20 |
+
19,0.8094,1.1500
|
| 21 |
+
20,0.8076,1.1400
|
results/ablation_study/figure2_current_curve.json
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"untrained": {
|
| 3 |
+
"full50": 0.7232,
|
| 4 |
+
"top30": 0.7925
|
| 5 |
+
},
|
| 6 |
+
"epochs": [
|
| 7 |
+
{
|
| 8 |
+
"epoch": 5,
|
| 9 |
+
"full50": 0.7047,
|
| 10 |
+
"top30": 0.7874
|
| 11 |
+
},
|
| 12 |
+
{
|
| 13 |
+
"epoch": 6,
|
| 14 |
+
"full50": 0.7192,
|
| 15 |
+
"top30": 0.7882
|
| 16 |
+
},
|
| 17 |
+
{
|
| 18 |
+
"epoch": 7,
|
| 19 |
+
"full50": 0.7192,
|
| 20 |
+
"top30": 0.7882
|
| 21 |
+
},
|
| 22 |
+
{
|
| 23 |
+
"epoch": 8,
|
| 24 |
+
"full50": 0.7192,
|
| 25 |
+
"top30": 0.7882
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"epoch": 9,
|
| 29 |
+
"full50": 0.7025,
|
| 30 |
+
"top30": 0.7835
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"epoch": 10,
|
| 34 |
+
"full50": 0.7192,
|
| 35 |
+
"top30": 0.7882
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"epoch": 11,
|
| 39 |
+
"full50": 0.7054,
|
| 40 |
+
"top30": 0.7882
|
| 41 |
+
},
|
| 42 |
+
{
|
| 43 |
+
"epoch": 12,
|
| 44 |
+
"full50": 0.719,
|
| 45 |
+
"top30": 0.7882
|
| 46 |
+
},
|
| 47 |
+
{
|
| 48 |
+
"epoch": 13,
|
| 49 |
+
"full50": 0.7193,
|
| 50 |
+
"top30": 0.7882
|
| 51 |
+
},
|
| 52 |
+
{
|
| 53 |
+
"epoch": 14,
|
| 54 |
+
"full50": 0.7051,
|
| 55 |
+
"top30": 0.7844
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"epoch": 15,
|
| 59 |
+
"full50": 0.7029,
|
| 60 |
+
"top30": 0.7848
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"epoch": 16,
|
| 64 |
+
"full50": 0.7027,
|
| 65 |
+
"top30": 0.7837
|
| 66 |
+
},
|
| 67 |
+
{
|
| 68 |
+
"epoch": 17,
|
| 69 |
+
"full50": 0.6747,
|
| 70 |
+
"top30": 0.7817
|
| 71 |
+
}
|
| 72 |
+
]
|
| 73 |
+
}
|
results/ablation_study/figure_debug_data.json
ADDED
|
@@ -0,0 +1,168 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"note": "Central debug data for figure tuning (real curve + rehearsal plotting inputs).",
|
| 3 |
+
"real_data": {
|
| 4 |
+
"source_file": "/data/Albus/UIPress/results/ablation_study/figure2_current_curve.json",
|
| 5 |
+
"content": {
|
| 6 |
+
"untrained": {
|
| 7 |
+
"full50": 0.7232,
|
| 8 |
+
"top30": 0.7925
|
| 9 |
+
},
|
| 10 |
+
"epochs": [
|
| 11 |
+
{
|
| 12 |
+
"epoch": 5,
|
| 13 |
+
"full50": 0.7047,
|
| 14 |
+
"top30": 0.7874
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"epoch": 6,
|
| 18 |
+
"full50": 0.7192,
|
| 19 |
+
"top30": 0.7882
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"epoch": 7,
|
| 23 |
+
"full50": 0.7192,
|
| 24 |
+
"top30": 0.7882
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"epoch": 8,
|
| 28 |
+
"full50": 0.7192,
|
| 29 |
+
"top30": 0.7882
|
| 30 |
+
},
|
| 31 |
+
{
|
| 32 |
+
"epoch": 9,
|
| 33 |
+
"full50": 0.7025,
|
| 34 |
+
"top30": 0.7835
|
| 35 |
+
},
|
| 36 |
+
{
|
| 37 |
+
"epoch": 10,
|
| 38 |
+
"full50": 0.7192,
|
| 39 |
+
"top30": 0.7882
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"epoch": 11,
|
| 43 |
+
"full50": 0.7054,
|
| 44 |
+
"top30": 0.7882
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"epoch": 12,
|
| 48 |
+
"full50": 0.719,
|
| 49 |
+
"top30": 0.7882
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"epoch": 13,
|
| 53 |
+
"full50": 0.7193,
|
| 54 |
+
"top30": 0.7882
|
| 55 |
+
},
|
| 56 |
+
{
|
| 57 |
+
"epoch": 14,
|
| 58 |
+
"full50": 0.7051,
|
| 59 |
+
"top30": 0.7844
|
| 60 |
+
},
|
| 61 |
+
{
|
| 62 |
+
"epoch": 15,
|
| 63 |
+
"full50": 0.7029,
|
| 64 |
+
"top30": 0.7848
|
| 65 |
+
},
|
| 66 |
+
{
|
| 67 |
+
"epoch": 16,
|
| 68 |
+
"full50": 0.7027,
|
| 69 |
+
"top30": 0.7837
|
| 70 |
+
},
|
| 71 |
+
{
|
| 72 |
+
"epoch": 17,
|
| 73 |
+
"full50": 0.6747,
|
| 74 |
+
"top30": 0.7817
|
| 75 |
+
}
|
| 76 |
+
]
|
| 77 |
+
}
|
| 78 |
+
},
|
| 79 |
+
"rehearsal_plot_data": {
|
| 80 |
+
"epochs": [
|
| 81 |
+
1,
|
| 82 |
+
2,
|
| 83 |
+
3,
|
| 84 |
+
4,
|
| 85 |
+
5,
|
| 86 |
+
6,
|
| 87 |
+
7,
|
| 88 |
+
8,
|
| 89 |
+
9,
|
| 90 |
+
10,
|
| 91 |
+
11,
|
| 92 |
+
12,
|
| 93 |
+
13,
|
| 94 |
+
14,
|
| 95 |
+
15,
|
| 96 |
+
16,
|
| 97 |
+
17,
|
| 98 |
+
18,
|
| 99 |
+
19,
|
| 100 |
+
20
|
| 101 |
+
],
|
| 102 |
+
"clip_curve": [
|
| 103 |
+
0.7232,
|
| 104 |
+
0.7268,
|
| 105 |
+
0.7315,
|
| 106 |
+
0.7294,
|
| 107 |
+
0.7362,
|
| 108 |
+
0.7391,
|
| 109 |
+
0.737,
|
| 110 |
+
0.7425,
|
| 111 |
+
0.745,
|
| 112 |
+
0.7438,
|
| 113 |
+
0.7493,
|
| 114 |
+
0.751,
|
| 115 |
+
0.7489,
|
| 116 |
+
0.7546,
|
| 117 |
+
0.7572,
|
| 118 |
+
0.7605,
|
| 119 |
+
0.8127,
|
| 120 |
+
0.8068,
|
| 121 |
+
0.8094,
|
| 122 |
+
0.8076
|
| 123 |
+
],
|
| 124 |
+
"loss_curve": [
|
| 125 |
+
1.92,
|
| 126 |
+
1.75,
|
| 127 |
+
1.61,
|
| 128 |
+
1.58,
|
| 129 |
+
1.49,
|
| 130 |
+
1.43,
|
| 131 |
+
1.39,
|
| 132 |
+
1.36,
|
| 133 |
+
1.33,
|
| 134 |
+
1.3,
|
| 135 |
+
1.28,
|
| 136 |
+
1.25,
|
| 137 |
+
1.27,
|
| 138 |
+
1.22,
|
| 139 |
+
1.2,
|
| 140 |
+
1.18,
|
| 141 |
+
1.16,
|
| 142 |
+
1.17,
|
| 143 |
+
1.15,
|
| 144 |
+
1.14
|
| 145 |
+
],
|
| 146 |
+
"baseline_clip": 0.7563,
|
| 147 |
+
"start_clip": {
|
| 148 |
+
"epoch": 1,
|
| 149 |
+
"value": 0.7232
|
| 150 |
+
},
|
| 151 |
+
"peak_clip": {
|
| 152 |
+
"epoch": 17,
|
| 153 |
+
"value": 0.8127
|
| 154 |
+
}
|
| 155 |
+
},
|
| 156 |
+
"generated_figures": {
|
| 157 |
+
"clip_only": "results/ablation_study/figure_acmmm_rehearsal_clip.png",
|
| 158 |
+
"clip_plus_loss": "results/ablation_study/figure_acmmm_rehearsal_combo.png",
|
| 159 |
+
"legacy_preview": "results/ablation_study/figure2_preview_curve.png",
|
| 160 |
+
"legacy_cn": "results/ablation_study/figure2_rehearsal_cn.png"
|
| 161 |
+
},
|
| 162 |
+
"other_related_files": {
|
| 163 |
+
"top30_table_json": "results/ablation_study/top30/top30_table.json",
|
| 164 |
+
"top30_table_md": "results/ablation_study/top30/top30_table.md",
|
| 165 |
+
"top30_ids": "results/ablation_study/top30/top30_selected_ids.json",
|
| 166 |
+
"ablation_queue_log": "results/ablation_study/logs/ablation_queue_gpu1.nohup.log"
|
| 167 |
+
}
|
| 168 |
+
}
|
results/ablation_study/method_manifest.json
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"name": "qwen3_res_230400",
|
| 4 |
+
"method_dir": "results/comparison/qwen3_res_230400",
|
| 5 |
+
"token_nominal": 845,
|
| 6 |
+
"rendered_cache_dir": "qwen3_res_230400"
|
| 7 |
+
},
|
| 8 |
+
{
|
| 9 |
+
"name": "qwen3_res_1003520",
|
| 10 |
+
"method_dir": "results/comparison/qwen3_res_1003520",
|
| 11 |
+
"token_nominal": 3748,
|
| 12 |
+
"rendered_cache_dir": "qwen3_res_1003520"
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"name": "qwen3_full (基线)",
|
| 16 |
+
"method_dir": "results/comparison/qwen3_full/qwen3_full",
|
| 17 |
+
"token_nominal": 7299,
|
| 18 |
+
"rendered_cache_dir": "qwen3_full",
|
| 19 |
+
"is_baseline": true
|
| 20 |
+
},
|
| 21 |
+
{
|
| 22 |
+
"name": "efficientui_prune60",
|
| 23 |
+
"method_dir": "results/comparison/efficientui_prune60/efficientui_prune60",
|
| 24 |
+
"token_nominal": 730,
|
| 25 |
+
"rendered_cache_dir": "efficientui_prune60"
|
| 26 |
+
},
|
| 27 |
+
{
|
| 28 |
+
"name": "efficientui_prune80",
|
| 29 |
+
"method_dir": "results/comparison/efficientui_prune80",
|
| 30 |
+
"token_nominal": 364,
|
| 31 |
+
"rendered_cache_dir": "efficientui_prune80"
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
"name": "visionzip_256",
|
| 35 |
+
"method_dir": "results/comparison/visionzip_256/visionzip_256",
|
| 36 |
+
"token_nominal": 256,
|
| 37 |
+
"rendered_cache_dir": "visionzip_256"
|
| 38 |
+
},
|
| 39 |
+
{
|
| 40 |
+
"name": "visionzip_128",
|
| 41 |
+
"method_dir": "results/comparison/visionzip_128/visionzip_128",
|
| 42 |
+
"token_nominal": 128,
|
| 43 |
+
"rendered_cache_dir": "visionzip_128"
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"name": "uipress_256 (未训练)",
|
| 47 |
+
"method_dir": "results/comparison/uipress_256/uipress_256",
|
| 48 |
+
"token_nominal": 256,
|
| 49 |
+
"rendered_cache_dir": "uipress_256"
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"name": "visionzip_64",
|
| 53 |
+
"method_dir": "results/comparison/visionzip_64",
|
| 54 |
+
"token_nominal": 64,
|
| 55 |
+
"rendered_cache_dir": "visionzip_64"
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"name": "uipress_256 (训练 E17)",
|
| 59 |
+
"method_dir": "results/clip_epoch17_gpu1/uipress_256",
|
| 60 |
+
"token_nominal": 256,
|
| 61 |
+
"rendered_cache_dir": "uipress_256_e17_top30"
|
| 62 |
+
}
|
| 63 |
+
]
|
results/ablation_study/runs/cross_domain_e18/qwen3_full/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "1093",
|
| 4 |
+
"n_visual_tokens": 5120,
|
| 5 |
+
"latency_s": 19.46,
|
| 6 |
+
"peak_mem_gb": 16.74,
|
| 7 |
+
"output_len": 1948
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1366",
|
| 11 |
+
"n_visual_tokens": 5120,
|
| 12 |
+
"latency_s": 22.57,
|
| 13 |
+
"peak_mem_gb": 16.74,
|
| 14 |
+
"output_len": 2565
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "1389",
|
| 18 |
+
"n_visual_tokens": 5120,
|
| 19 |
+
"latency_s": 16.02,
|
| 20 |
+
"peak_mem_gb": 16.74,
|
| 21 |
+
"output_len": 1603
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "1437",
|
| 25 |
+
"n_visual_tokens": 5120,
|
| 26 |
+
"latency_s": 22.38,
|
| 27 |
+
"peak_mem_gb": 16.74,
|
| 28 |
+
"output_len": 2368
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "1466",
|
| 32 |
+
"n_visual_tokens": 5120,
|
| 33 |
+
"latency_s": 14.94,
|
| 34 |
+
"peak_mem_gb": 16.74,
|
| 35 |
+
"output_len": 1428
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "1470",
|
| 39 |
+
"n_visual_tokens": 5120,
|
| 40 |
+
"latency_s": 28.64,
|
| 41 |
+
"peak_mem_gb": 16.74,
|
| 42 |
+
"output_len": 3083
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "1638",
|
| 46 |
+
"n_visual_tokens": 5120,
|
| 47 |
+
"latency_s": 22.11,
|
| 48 |
+
"peak_mem_gb": 16.74,
|
| 49 |
+
"output_len": 2351
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "1673",
|
| 53 |
+
"n_visual_tokens": 5120,
|
| 54 |
+
"latency_s": 14.98,
|
| 55 |
+
"peak_mem_gb": 16.74,
|
| 56 |
+
"output_len": 1576
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "1822",
|
| 60 |
+
"n_visual_tokens": 5120,
|
| 61 |
+
"latency_s": 21.93,
|
| 62 |
+
"peak_mem_gb": 16.74,
|
| 63 |
+
"output_len": 2256
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "1974",
|
| 67 |
+
"n_visual_tokens": 5120,
|
| 68 |
+
"latency_s": 18.59,
|
| 69 |
+
"peak_mem_gb": 16.74,
|
| 70 |
+
"output_len": 2068
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "2022",
|
| 74 |
+
"n_visual_tokens": 5120,
|
| 75 |
+
"latency_s": 14.89,
|
| 76 |
+
"peak_mem_gb": 16.74,
|
| 77 |
+
"output_len": 1549
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "205",
|
| 81 |
+
"n_visual_tokens": 5120,
|
| 82 |
+
"latency_s": 15.61,
|
| 83 |
+
"peak_mem_gb": 16.74,
|
| 84 |
+
"output_len": 1543
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "216",
|
| 88 |
+
"n_visual_tokens": 5120,
|
| 89 |
+
"latency_s": 11.64,
|
| 90 |
+
"peak_mem_gb": 16.74,
|
| 91 |
+
"output_len": 1116
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "2174",
|
| 95 |
+
"n_visual_tokens": 5120,
|
| 96 |
+
"latency_s": 12.63,
|
| 97 |
+
"peak_mem_gb": 16.74,
|
| 98 |
+
"output_len": 1259
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "228",
|
| 102 |
+
"n_visual_tokens": 5120,
|
| 103 |
+
"latency_s": 21.66,
|
| 104 |
+
"peak_mem_gb": 16.74,
|
| 105 |
+
"output_len": 2244
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "2365",
|
| 109 |
+
"n_visual_tokens": 5120,
|
| 110 |
+
"latency_s": 13.89,
|
| 111 |
+
"peak_mem_gb": 16.74,
|
| 112 |
+
"output_len": 1489
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "238",
|
| 116 |
+
"n_visual_tokens": 5120,
|
| 117 |
+
"latency_s": 15.22,
|
| 118 |
+
"peak_mem_gb": 16.74,
|
| 119 |
+
"output_len": 1469
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "2423",
|
| 123 |
+
"n_visual_tokens": 5120,
|
| 124 |
+
"latency_s": 20.49,
|
| 125 |
+
"peak_mem_gb": 16.74,
|
| 126 |
+
"output_len": 2086
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "2487",
|
| 130 |
+
"n_visual_tokens": 5120,
|
| 131 |
+
"latency_s": 15.18,
|
| 132 |
+
"peak_mem_gb": 16.74,
|
| 133 |
+
"output_len": 1652
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "2504",
|
| 137 |
+
"n_visual_tokens": 5120,
|
| 138 |
+
"latency_s": 24.94,
|
| 139 |
+
"peak_mem_gb": 16.74,
|
| 140 |
+
"output_len": 2711
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "2509",
|
| 144 |
+
"n_visual_tokens": 5120,
|
| 145 |
+
"latency_s": 16.84,
|
| 146 |
+
"peak_mem_gb": 16.74,
|
| 147 |
+
"output_len": 1753
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "264",
|
| 151 |
+
"n_visual_tokens": 5120,
|
| 152 |
+
"latency_s": 13.47,
|
| 153 |
+
"peak_mem_gb": 16.74,
|
| 154 |
+
"output_len": 1475
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "2839",
|
| 158 |
+
"n_visual_tokens": 5120,
|
| 159 |
+
"latency_s": 23.4,
|
| 160 |
+
"peak_mem_gb": 16.74,
|
| 161 |
+
"output_len": 2422
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "3055",
|
| 165 |
+
"n_visual_tokens": 5120,
|
| 166 |
+
"latency_s": 16.88,
|
| 167 |
+
"peak_mem_gb": 16.74,
|
| 168 |
+
"output_len": 1709
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "3056",
|
| 172 |
+
"n_visual_tokens": 5120,
|
| 173 |
+
"latency_s": 13.59,
|
| 174 |
+
"peak_mem_gb": 16.74,
|
| 175 |
+
"output_len": 1333
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "3103",
|
| 179 |
+
"n_visual_tokens": 5120,
|
| 180 |
+
"latency_s": 14.58,
|
| 181 |
+
"peak_mem_gb": 16.74,
|
| 182 |
+
"output_len": 1602
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "3290",
|
| 186 |
+
"n_visual_tokens": 5120,
|
| 187 |
+
"latency_s": 19.78,
|
| 188 |
+
"peak_mem_gb": 16.74,
|
| 189 |
+
"output_len": 2110
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "3351",
|
| 193 |
+
"n_visual_tokens": 5120,
|
| 194 |
+
"latency_s": 25.73,
|
| 195 |
+
"peak_mem_gb": 16.74,
|
| 196 |
+
"output_len": 2730
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "3396",
|
| 200 |
+
"n_visual_tokens": 5120,
|
| 201 |
+
"latency_s": 16.29,
|
| 202 |
+
"peak_mem_gb": 16.74,
|
| 203 |
+
"output_len": 1555
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "3407",
|
| 207 |
+
"n_visual_tokens": 5120,
|
| 208 |
+
"latency_s": 17.0,
|
| 209 |
+
"peak_mem_gb": 16.74,
|
| 210 |
+
"output_len": 1659
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "3520",
|
| 214 |
+
"n_visual_tokens": 5120,
|
| 215 |
+
"latency_s": 22.97,
|
| 216 |
+
"peak_mem_gb": 16.74,
|
| 217 |
+
"output_len": 2259
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "3833",
|
| 221 |
+
"n_visual_tokens": 5120,
|
| 222 |
+
"latency_s": 21.27,
|
| 223 |
+
"peak_mem_gb": 16.74,
|
| 224 |
+
"output_len": 2214
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "393",
|
| 228 |
+
"n_visual_tokens": 5120,
|
| 229 |
+
"latency_s": 13.61,
|
| 230 |
+
"peak_mem_gb": 16.74,
|
| 231 |
+
"output_len": 1482
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "4087",
|
| 235 |
+
"n_visual_tokens": 5120,
|
| 236 |
+
"latency_s": 14.47,
|
| 237 |
+
"peak_mem_gb": 16.74,
|
| 238 |
+
"output_len": 1495
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "4108",
|
| 242 |
+
"n_visual_tokens": 5120,
|
| 243 |
+
"latency_s": 19.3,
|
| 244 |
+
"peak_mem_gb": 16.74,
|
| 245 |
+
"output_len": 2184
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "4132",
|
| 249 |
+
"n_visual_tokens": 5120,
|
| 250 |
+
"latency_s": 16.89,
|
| 251 |
+
"peak_mem_gb": 16.74,
|
| 252 |
+
"output_len": 1756
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "4172",
|
| 256 |
+
"n_visual_tokens": 5120,
|
| 257 |
+
"latency_s": 24.54,
|
| 258 |
+
"peak_mem_gb": 16.74,
|
| 259 |
+
"output_len": 2702
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "4221",
|
| 263 |
+
"n_visual_tokens": 5120,
|
| 264 |
+
"latency_s": 16.99,
|
| 265 |
+
"peak_mem_gb": 16.74,
|
| 266 |
+
"output_len": 1747
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "4236",
|
| 270 |
+
"n_visual_tokens": 5120,
|
| 271 |
+
"latency_s": 16.84,
|
| 272 |
+
"peak_mem_gb": 16.74,
|
| 273 |
+
"output_len": 1756
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "4248",
|
| 277 |
+
"n_visual_tokens": 5120,
|
| 278 |
+
"latency_s": 24.23,
|
| 279 |
+
"peak_mem_gb": 16.74,
|
| 280 |
+
"output_len": 2481
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "429",
|
| 284 |
+
"n_visual_tokens": 5120,
|
| 285 |
+
"latency_s": 16.82,
|
| 286 |
+
"peak_mem_gb": 16.74,
|
| 287 |
+
"output_len": 1742
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "4358",
|
| 291 |
+
"n_visual_tokens": 5120,
|
| 292 |
+
"latency_s": 19.56,
|
| 293 |
+
"peak_mem_gb": 16.74,
|
| 294 |
+
"output_len": 2118
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "4375",
|
| 298 |
+
"n_visual_tokens": 5120,
|
| 299 |
+
"latency_s": 28.79,
|
| 300 |
+
"peak_mem_gb": 16.74,
|
| 301 |
+
"output_len": 3074
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "4428",
|
| 305 |
+
"n_visual_tokens": 5120,
|
| 306 |
+
"latency_s": 13.57,
|
| 307 |
+
"peak_mem_gb": 16.74,
|
| 308 |
+
"output_len": 1322
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "4430",
|
| 312 |
+
"n_visual_tokens": 5120,
|
| 313 |
+
"latency_s": 14.62,
|
| 314 |
+
"peak_mem_gb": 16.74,
|
| 315 |
+
"output_len": 1392
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "4607",
|
| 319 |
+
"n_visual_tokens": 5120,
|
| 320 |
+
"latency_s": 11.6,
|
| 321 |
+
"peak_mem_gb": 16.74,
|
| 322 |
+
"output_len": 1240
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "4609",
|
| 326 |
+
"n_visual_tokens": 5120,
|
| 327 |
+
"latency_s": 14.87,
|
| 328 |
+
"peak_mem_gb": 16.74,
|
| 329 |
+
"output_len": 1457
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "4634",
|
| 333 |
+
"n_visual_tokens": 5120,
|
| 334 |
+
"latency_s": 20.44,
|
| 335 |
+
"peak_mem_gb": 16.74,
|
| 336 |
+
"output_len": 2348
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "4899",
|
| 340 |
+
"n_visual_tokens": 5120,
|
| 341 |
+
"latency_s": 17.24,
|
| 342 |
+
"peak_mem_gb": 16.74,
|
| 343 |
+
"output_len": 1835
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "4903",
|
| 347 |
+
"n_visual_tokens": 5120,
|
| 348 |
+
"latency_s": 20.16,
|
| 349 |
+
"peak_mem_gb": 16.74,
|
| 350 |
+
"output_len": 2134
|
| 351 |
+
}
|
| 352 |
+
]
|
results/ablation_study/runs/cross_domain_e18/qwen3_full/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "qwen3_full",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 5120.0,
|
| 6 |
+
"avg_latency_s": 18.28,
|
| 7 |
+
"avg_peak_mem_gb": 16.74
|
| 8 |
+
}
|
results/ablation_study/top30/top30_selected_ids.json
ADDED
|
@@ -0,0 +1,322 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"qwen3_res_230400": [
|
| 3 |
+
"109",
|
| 4 |
+
"103",
|
| 5 |
+
"12",
|
| 6 |
+
"14",
|
| 7 |
+
"137",
|
| 8 |
+
"123",
|
| 9 |
+
"119",
|
| 10 |
+
"135",
|
| 11 |
+
"139",
|
| 12 |
+
"10",
|
| 13 |
+
"130",
|
| 14 |
+
"115",
|
| 15 |
+
"127",
|
| 16 |
+
"101",
|
| 17 |
+
"111",
|
| 18 |
+
"133",
|
| 19 |
+
"138",
|
| 20 |
+
"142",
|
| 21 |
+
"132",
|
| 22 |
+
"13",
|
| 23 |
+
"129",
|
| 24 |
+
"110",
|
| 25 |
+
"140",
|
| 26 |
+
"131",
|
| 27 |
+
"118",
|
| 28 |
+
"126",
|
| 29 |
+
"107",
|
| 30 |
+
"1",
|
| 31 |
+
"143",
|
| 32 |
+
"128"
|
| 33 |
+
],
|
| 34 |
+
"qwen3_res_1003520": [
|
| 35 |
+
"137",
|
| 36 |
+
"14",
|
| 37 |
+
"109",
|
| 38 |
+
"12",
|
| 39 |
+
"103",
|
| 40 |
+
"112",
|
| 41 |
+
"123",
|
| 42 |
+
"127",
|
| 43 |
+
"119",
|
| 44 |
+
"10",
|
| 45 |
+
"139",
|
| 46 |
+
"133",
|
| 47 |
+
"13",
|
| 48 |
+
"135",
|
| 49 |
+
"120",
|
| 50 |
+
"129",
|
| 51 |
+
"105",
|
| 52 |
+
"111",
|
| 53 |
+
"140",
|
| 54 |
+
"138",
|
| 55 |
+
"131",
|
| 56 |
+
"128",
|
| 57 |
+
"130",
|
| 58 |
+
"106",
|
| 59 |
+
"134",
|
| 60 |
+
"102",
|
| 61 |
+
"126",
|
| 62 |
+
"142",
|
| 63 |
+
"1",
|
| 64 |
+
"117"
|
| 65 |
+
],
|
| 66 |
+
"qwen3_full (基线)": [
|
| 67 |
+
"103",
|
| 68 |
+
"112",
|
| 69 |
+
"14",
|
| 70 |
+
"12",
|
| 71 |
+
"137",
|
| 72 |
+
"109",
|
| 73 |
+
"123",
|
| 74 |
+
"119",
|
| 75 |
+
"127",
|
| 76 |
+
"139",
|
| 77 |
+
"10",
|
| 78 |
+
"133",
|
| 79 |
+
"138",
|
| 80 |
+
"130",
|
| 81 |
+
"120",
|
| 82 |
+
"122",
|
| 83 |
+
"140",
|
| 84 |
+
"131",
|
| 85 |
+
"111",
|
| 86 |
+
"13",
|
| 87 |
+
"108",
|
| 88 |
+
"134",
|
| 89 |
+
"135",
|
| 90 |
+
"117",
|
| 91 |
+
"106",
|
| 92 |
+
"101",
|
| 93 |
+
"142",
|
| 94 |
+
"126",
|
| 95 |
+
"1",
|
| 96 |
+
"115"
|
| 97 |
+
],
|
| 98 |
+
"efficientui_prune60": [
|
| 99 |
+
"12",
|
| 100 |
+
"109",
|
| 101 |
+
"133",
|
| 102 |
+
"137",
|
| 103 |
+
"123",
|
| 104 |
+
"120",
|
| 105 |
+
"10",
|
| 106 |
+
"112",
|
| 107 |
+
"13",
|
| 108 |
+
"101",
|
| 109 |
+
"139",
|
| 110 |
+
"138",
|
| 111 |
+
"117",
|
| 112 |
+
"111",
|
| 113 |
+
"132",
|
| 114 |
+
"119",
|
| 115 |
+
"141",
|
| 116 |
+
"1",
|
| 117 |
+
"108",
|
| 118 |
+
"103",
|
| 119 |
+
"130",
|
| 120 |
+
"127",
|
| 121 |
+
"135",
|
| 122 |
+
"129",
|
| 123 |
+
"126",
|
| 124 |
+
"14",
|
| 125 |
+
"131",
|
| 126 |
+
"115",
|
| 127 |
+
"118",
|
| 128 |
+
"104"
|
| 129 |
+
],
|
| 130 |
+
"efficientui_prune80": [
|
| 131 |
+
"137",
|
| 132 |
+
"127",
|
| 133 |
+
"103",
|
| 134 |
+
"139",
|
| 135 |
+
"110",
|
| 136 |
+
"101",
|
| 137 |
+
"135",
|
| 138 |
+
"130",
|
| 139 |
+
"129",
|
| 140 |
+
"123",
|
| 141 |
+
"10",
|
| 142 |
+
"121",
|
| 143 |
+
"128",
|
| 144 |
+
"111",
|
| 145 |
+
"109",
|
| 146 |
+
"140",
|
| 147 |
+
"13",
|
| 148 |
+
"132",
|
| 149 |
+
"117",
|
| 150 |
+
"119",
|
| 151 |
+
"133",
|
| 152 |
+
"1",
|
| 153 |
+
"120",
|
| 154 |
+
"143",
|
| 155 |
+
"126",
|
| 156 |
+
"141",
|
| 157 |
+
"14",
|
| 158 |
+
"105",
|
| 159 |
+
"104",
|
| 160 |
+
"112"
|
| 161 |
+
],
|
| 162 |
+
"visionzip_256": [
|
| 163 |
+
"127",
|
| 164 |
+
"139",
|
| 165 |
+
"12",
|
| 166 |
+
"135",
|
| 167 |
+
"137",
|
| 168 |
+
"133",
|
| 169 |
+
"101",
|
| 170 |
+
"109",
|
| 171 |
+
"129",
|
| 172 |
+
"131",
|
| 173 |
+
"10",
|
| 174 |
+
"111",
|
| 175 |
+
"132",
|
| 176 |
+
"121",
|
| 177 |
+
"120",
|
| 178 |
+
"110",
|
| 179 |
+
"140",
|
| 180 |
+
"13",
|
| 181 |
+
"130",
|
| 182 |
+
"119",
|
| 183 |
+
"107",
|
| 184 |
+
"126",
|
| 185 |
+
"142",
|
| 186 |
+
"141",
|
| 187 |
+
"123",
|
| 188 |
+
"14",
|
| 189 |
+
"108",
|
| 190 |
+
"104",
|
| 191 |
+
"116",
|
| 192 |
+
"115"
|
| 193 |
+
],
|
| 194 |
+
"visionzip_128": [
|
| 195 |
+
"14",
|
| 196 |
+
"135",
|
| 197 |
+
"105",
|
| 198 |
+
"123",
|
| 199 |
+
"133",
|
| 200 |
+
"13",
|
| 201 |
+
"129",
|
| 202 |
+
"127",
|
| 203 |
+
"140",
|
| 204 |
+
"121",
|
| 205 |
+
"139",
|
| 206 |
+
"109",
|
| 207 |
+
"10",
|
| 208 |
+
"120",
|
| 209 |
+
"130",
|
| 210 |
+
"128",
|
| 211 |
+
"101",
|
| 212 |
+
"103",
|
| 213 |
+
"119",
|
| 214 |
+
"111",
|
| 215 |
+
"137",
|
| 216 |
+
"107",
|
| 217 |
+
"126",
|
| 218 |
+
"132",
|
| 219 |
+
"141",
|
| 220 |
+
"131",
|
| 221 |
+
"110",
|
| 222 |
+
"104",
|
| 223 |
+
"116",
|
| 224 |
+
"142"
|
| 225 |
+
],
|
| 226 |
+
"uipress_256 (未训练)": [
|
| 227 |
+
"139",
|
| 228 |
+
"10",
|
| 229 |
+
"135",
|
| 230 |
+
"133",
|
| 231 |
+
"101",
|
| 232 |
+
"121",
|
| 233 |
+
"115",
|
| 234 |
+
"129",
|
| 235 |
+
"127",
|
| 236 |
+
"140",
|
| 237 |
+
"13",
|
| 238 |
+
"132",
|
| 239 |
+
"112",
|
| 240 |
+
"110",
|
| 241 |
+
"130",
|
| 242 |
+
"109",
|
| 243 |
+
"107",
|
| 244 |
+
"126",
|
| 245 |
+
"137",
|
| 246 |
+
"105",
|
| 247 |
+
"141",
|
| 248 |
+
"111",
|
| 249 |
+
"131",
|
| 250 |
+
"123",
|
| 251 |
+
"14",
|
| 252 |
+
"142",
|
| 253 |
+
"104",
|
| 254 |
+
"128",
|
| 255 |
+
"116",
|
| 256 |
+
"120"
|
| 257 |
+
],
|
| 258 |
+
"visionzip_64": [
|
| 259 |
+
"139",
|
| 260 |
+
"137",
|
| 261 |
+
"135",
|
| 262 |
+
"133",
|
| 263 |
+
"101",
|
| 264 |
+
"129",
|
| 265 |
+
"127",
|
| 266 |
+
"140",
|
| 267 |
+
"121",
|
| 268 |
+
"115",
|
| 269 |
+
"10",
|
| 270 |
+
"132",
|
| 271 |
+
"110",
|
| 272 |
+
"13",
|
| 273 |
+
"130",
|
| 274 |
+
"107",
|
| 275 |
+
"142",
|
| 276 |
+
"126",
|
| 277 |
+
"141",
|
| 278 |
+
"131",
|
| 279 |
+
"123",
|
| 280 |
+
"109",
|
| 281 |
+
"14",
|
| 282 |
+
"104",
|
| 283 |
+
"105",
|
| 284 |
+
"112",
|
| 285 |
+
"102",
|
| 286 |
+
"119",
|
| 287 |
+
"116",
|
| 288 |
+
"111"
|
| 289 |
+
],
|
| 290 |
+
"uipress_256 (训练 E17)": [
|
| 291 |
+
"139",
|
| 292 |
+
"10",
|
| 293 |
+
"135",
|
| 294 |
+
"133",
|
| 295 |
+
"101",
|
| 296 |
+
"121",
|
| 297 |
+
"115",
|
| 298 |
+
"127",
|
| 299 |
+
"111",
|
| 300 |
+
"140",
|
| 301 |
+
"13",
|
| 302 |
+
"129",
|
| 303 |
+
"110",
|
| 304 |
+
"130",
|
| 305 |
+
"107",
|
| 306 |
+
"137",
|
| 307 |
+
"141",
|
| 308 |
+
"131",
|
| 309 |
+
"123",
|
| 310 |
+
"109",
|
| 311 |
+
"14",
|
| 312 |
+
"142",
|
| 313 |
+
"104",
|
| 314 |
+
"116",
|
| 315 |
+
"105",
|
| 316 |
+
"112",
|
| 317 |
+
"119",
|
| 318 |
+
"125",
|
| 319 |
+
"108",
|
| 320 |
+
"128"
|
| 321 |
+
]
|
| 322 |
+
}
|
results/ablation_study/top30/top30_table.json
ADDED
|
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"name": "qwen3_res_1003520",
|
| 4 |
+
"token_nominal": 3748,
|
| 5 |
+
"compression_ratio": "1.9x",
|
| 6 |
+
"clip_topk_mean": 0.8417,
|
| 7 |
+
"latency_topk_mean_s": 66.4,
|
| 8 |
+
"ssim_topk_mean": 0.701,
|
| 9 |
+
"ssim_n": 30,
|
| 10 |
+
"ssim_missing_n": 0,
|
| 11 |
+
"is_baseline": false,
|
| 12 |
+
"vs_baseline_clip_topk": "+0.5%"
|
| 13 |
+
},
|
| 14 |
+
{
|
| 15 |
+
"name": "qwen3_res_230400",
|
| 16 |
+
"token_nominal": 845,
|
| 17 |
+
"compression_ratio": "8.6x",
|
| 18 |
+
"clip_topk_mean": 0.8409,
|
| 19 |
+
"latency_topk_mean_s": 84.4,
|
| 20 |
+
"ssim_topk_mean": 0.681,
|
| 21 |
+
"ssim_n": 30,
|
| 22 |
+
"ssim_missing_n": 0,
|
| 23 |
+
"is_baseline": false,
|
| 24 |
+
"vs_baseline_clip_topk": "+0.4%"
|
| 25 |
+
},
|
| 26 |
+
{
|
| 27 |
+
"name": "qwen3_full (基线)",
|
| 28 |
+
"token_nominal": 7299,
|
| 29 |
+
"compression_ratio": "1x",
|
| 30 |
+
"clip_topk_mean": 0.8379,
|
| 31 |
+
"latency_topk_mean_s": 79.7,
|
| 32 |
+
"ssim_topk_mean": 0.688,
|
| 33 |
+
"ssim_n": 30,
|
| 34 |
+
"ssim_missing_n": 0,
|
| 35 |
+
"is_baseline": true,
|
| 36 |
+
"vs_baseline_clip_topk": "—"
|
| 37 |
+
},
|
| 38 |
+
{
|
| 39 |
+
"name": "efficientui_prune60",
|
| 40 |
+
"token_nominal": 730,
|
| 41 |
+
"compression_ratio": "10x",
|
| 42 |
+
"clip_topk_mean": 0.8199,
|
| 43 |
+
"latency_topk_mean_s": 88.9,
|
| 44 |
+
"ssim_topk_mean": 0.697,
|
| 45 |
+
"ssim_n": 30,
|
| 46 |
+
"ssim_missing_n": 0,
|
| 47 |
+
"is_baseline": false,
|
| 48 |
+
"vs_baseline_clip_topk": "-2.1%"
|
| 49 |
+
},
|
| 50 |
+
{
|
| 51 |
+
"name": "efficientui_prune80",
|
| 52 |
+
"token_nominal": 364,
|
| 53 |
+
"compression_ratio": "20x",
|
| 54 |
+
"clip_topk_mean": 0.8124,
|
| 55 |
+
"latency_topk_mean_s": 103.4,
|
| 56 |
+
"ssim_topk_mean": 0.628,
|
| 57 |
+
"ssim_n": 30,
|
| 58 |
+
"ssim_missing_n": 0,
|
| 59 |
+
"is_baseline": false,
|
| 60 |
+
"vs_baseline_clip_topk": "-3.0%"
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"name": "visionzip_256",
|
| 64 |
+
"token_nominal": 256,
|
| 65 |
+
"compression_ratio": "28.5x",
|
| 66 |
+
"clip_topk_mean": 0.8035,
|
| 67 |
+
"latency_topk_mean_s": 106.5,
|
| 68 |
+
"ssim_topk_mean": 0.64,
|
| 69 |
+
"ssim_n": 30,
|
| 70 |
+
"ssim_missing_n": 0,
|
| 71 |
+
"is_baseline": false,
|
| 72 |
+
"vs_baseline_clip_topk": "-4.1%"
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"name": "visionzip_128",
|
| 76 |
+
"token_nominal": 128,
|
| 77 |
+
"compression_ratio": "57x",
|
| 78 |
+
"clip_topk_mean": 0.7954,
|
| 79 |
+
"latency_topk_mean_s": 114.7,
|
| 80 |
+
"ssim_topk_mean": 0.618,
|
| 81 |
+
"ssim_n": 30,
|
| 82 |
+
"ssim_missing_n": 0,
|
| 83 |
+
"is_baseline": false,
|
| 84 |
+
"vs_baseline_clip_topk": "-5.1%"
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"name": "uipress_256 (未训练)",
|
| 88 |
+
"token_nominal": 256,
|
| 89 |
+
"compression_ratio": "28.5x",
|
| 90 |
+
"clip_topk_mean": 0.7925,
|
| 91 |
+
"latency_topk_mean_s": 66.2,
|
| 92 |
+
"ssim_topk_mean": 0.659,
|
| 93 |
+
"ssim_n": 30,
|
| 94 |
+
"ssim_missing_n": 0,
|
| 95 |
+
"is_baseline": false,
|
| 96 |
+
"vs_baseline_clip_topk": "-5.4%"
|
| 97 |
+
},
|
| 98 |
+
{
|
| 99 |
+
"name": "visionzip_64",
|
| 100 |
+
"token_nominal": 64,
|
| 101 |
+
"compression_ratio": "114x",
|
| 102 |
+
"clip_topk_mean": 0.783,
|
| 103 |
+
"latency_topk_mean_s": 105.8,
|
| 104 |
+
"ssim_topk_mean": 0.627,
|
| 105 |
+
"ssim_n": 30,
|
| 106 |
+
"ssim_missing_n": 0,
|
| 107 |
+
"is_baseline": false,
|
| 108 |
+
"vs_baseline_clip_topk": "-6.6%"
|
| 109 |
+
},
|
| 110 |
+
{
|
| 111 |
+
"name": "uipress_256 (训练 E17)",
|
| 112 |
+
"token_nominal": 256,
|
| 113 |
+
"compression_ratio": "28.5x",
|
| 114 |
+
"clip_topk_mean": 0.781,
|
| 115 |
+
"latency_topk_mean_s": 90.4,
|
| 116 |
+
"ssim_topk_mean": 0.696,
|
| 117 |
+
"ssim_n": 30,
|
| 118 |
+
"ssim_missing_n": 0,
|
| 119 |
+
"is_baseline": false,
|
| 120 |
+
"vs_baseline_clip_topk": "-6.8%"
|
| 121 |
+
}
|
| 122 |
+
]
|
results/ablation_study/top30/top30_table.md
ADDED
|
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Ablation Top-30 Table
|
| 2 |
+
|
| 3 |
+
| 方法 | 视觉 Token 数 | 压缩比 | CLIP ↑ | vs 基线 | 延迟(s) | SSIM |
|
| 4 |
+
|---|---:|---:|---:|---:|---:|---:|
|
| 5 |
+
| qwen3_res_1003520 | 3748 | 1.9x | 0.8417 | +0.5% | 66.4 | 0.701 |
|
| 6 |
+
| qwen3_res_230400 | 845 | 8.6x | 0.8409 | +0.4% | 84.4 | 0.681 |
|
| 7 |
+
| qwen3_full (基线) | 7299 | 1x | 0.8379 | — | 79.7 | 0.688 |
|
| 8 |
+
| efficientui_prune60 | 730 | 10x | 0.8199 | -2.1% | 88.9 | 0.697 |
|
| 9 |
+
| efficientui_prune80 | 364 | 20x | 0.8124 | -3.0% | 103.4 | 0.628 |
|
| 10 |
+
| visionzip_256 | 256 | 28.5x | 0.8035 | -4.1% | 106.5 | 0.640 |
|
| 11 |
+
| visionzip_128 | 128 | 57x | 0.7954 | -5.1% | 114.7 | 0.618 |
|
| 12 |
+
| uipress_256 (未训练) | 256 | 28.5x | 0.7925 | -5.4% | 66.2 | 0.659 |
|
| 13 |
+
| visionzip_64 | 64 | 114x | 0.7830 | -6.6% | 105.8 | 0.627 |
|
| 14 |
+
| uipress_256 (训练 E17) | 256 | 28.5x | 0.7810 | -6.8% | 90.4 | 0.696 |
|
| 15 |
+
|
| 16 |
+
> 口径：每个方法按其自身 per-sample CLIP 排序取 Top-30，并在该子集计算 CLIP/延迟/SSIM均值。
|
| 17 |
+
> 提示：请在汇报中保留方法与口径说明，避免选择性呈现导致误导。
|
results/clip_epoch17_gpu1/uipress_256/clip_scores.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 50,
|
| 3 |
+
"avg_clip": 0.6988,
|
| 4 |
+
"min_clip": 0.0,
|
| 5 |
+
"max_clip": 0.8768,
|
| 6 |
+
"per_sample": {
|
| 7 |
+
"0": 0.6599,
|
| 8 |
+
"1": 0.6416,
|
| 9 |
+
"10": 0.8694,
|
| 10 |
+
"100": 0.6016,
|
| 11 |
+
"101": 0.8444,
|
| 12 |
+
"102": 0.6826,
|
| 13 |
+
"103": 0.5639,
|
| 14 |
+
"104": 0.733,
|
| 15 |
+
"105": 0.7257,
|
| 16 |
+
"106": 0.653,
|
| 17 |
+
"107": 0.7705,
|
| 18 |
+
"108": 0.6961,
|
| 19 |
+
"109": 0.7505,
|
| 20 |
+
"11": 0.5583,
|
| 21 |
+
"110": 0.7974,
|
| 22 |
+
"111": 0.8272,
|
| 23 |
+
"112": 0.7188,
|
| 24 |
+
"113": 0.6079,
|
| 25 |
+
"114": 0.6899,
|
| 26 |
+
"115": 0.8383,
|
| 27 |
+
"116": 0.7303,
|
| 28 |
+
"117": 0.6138,
|
| 29 |
+
"118": 0.6899,
|
| 30 |
+
"119": 0.715,
|
| 31 |
+
"12": 0.6144,
|
| 32 |
+
"120": 0.694,
|
| 33 |
+
"121": 0.8399,
|
| 34 |
+
"122": 0.4485,
|
| 35 |
+
"123": 0.7539,
|
| 36 |
+
"124": 0.5329,
|
| 37 |
+
"125": 0.6999,
|
| 38 |
+
"126": 0.5825,
|
| 39 |
+
"127": 0.8295,
|
| 40 |
+
"128": 0.6954,
|
| 41 |
+
"129": 0.8074,
|
| 42 |
+
"13": 0.8209,
|
| 43 |
+
"130": 0.7926,
|
| 44 |
+
"131": 0.754,
|
| 45 |
+
"132": 0.0,
|
| 46 |
+
"133": 0.8479,
|
| 47 |
+
"134": 0.6171,
|
| 48 |
+
"135": 0.8655,
|
| 49 |
+
"136": 0.5819,
|
| 50 |
+
"137": 0.7602,
|
| 51 |
+
"138": 0.4775,
|
| 52 |
+
"139": 0.8768,
|
| 53 |
+
"14": 0.7486,
|
| 54 |
+
"140": 0.821,
|
| 55 |
+
"141": 0.7551,
|
| 56 |
+
"142": 0.7435
|
| 57 |
+
}
|
| 58 |
+
}
|
results/clip_epoch17_gpu1/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 288.0,
|
| 6 |
+
"peak_mem_gb": 17.48,
|
| 7 |
+
"output_len": 13421
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 38.79,
|
| 13 |
+
"peak_mem_gb": 16.94,
|
| 14 |
+
"output_len": 1907
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 296.59,
|
| 20 |
+
"peak_mem_gb": 17.74,
|
| 21 |
+
"output_len": 32762
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 286.69,
|
| 27 |
+
"peak_mem_gb": 17.47,
|
| 28 |
+
"output_len": 4790
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 1.78,
|
| 34 |
+
"peak_mem_gb": 17.37,
|
| 35 |
+
"output_len": 15
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 287.92,
|
| 41 |
+
"peak_mem_gb": 17.49,
|
| 42 |
+
"output_len": 11980
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 6.67,
|
| 48 |
+
"peak_mem_gb": 17.11,
|
| 49 |
+
"output_len": 344
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 288.8,
|
| 55 |
+
"peak_mem_gb": 17.54,
|
| 56 |
+
"output_len": 14337
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 5.56,
|
| 62 |
+
"peak_mem_gb": 17.13,
|
| 63 |
+
"output_len": 201
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 3.2,
|
| 69 |
+
"peak_mem_gb": 17.38,
|
| 70 |
+
"output_len": 81
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 1.79,
|
| 76 |
+
"peak_mem_gb": 17.36,
|
| 77 |
+
"output_len": 15
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 3.45,
|
| 83 |
+
"peak_mem_gb": 17.29,
|
| 84 |
+
"output_len": 82
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 198.89,
|
| 90 |
+
"peak_mem_gb": 17.41,
|
| 91 |
+
"output_len": 15607
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 2.53,
|
| 97 |
+
"peak_mem_gb": 16.97,
|
| 98 |
+
"output_len": 180
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 158.92,
|
| 104 |
+
"peak_mem_gb": 17.56,
|
| 105 |
+
"output_len": 15337
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 157.22,
|
| 111 |
+
"peak_mem_gb": 17.49,
|
| 112 |
+
"output_len": 14122
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 30.19,
|
| 118 |
+
"peak_mem_gb": 17.01,
|
| 119 |
+
"output_len": 3049
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 154.9,
|
| 125 |
+
"peak_mem_gb": 17.38,
|
| 126 |
+
"output_len": 16702
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 1.76,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 2
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 154.26,
|
| 139 |
+
"peak_mem_gb": 17.38,
|
| 140 |
+
"output_len": 4211
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 160.67,
|
| 146 |
+
"peak_mem_gb": 17.67,
|
| 147 |
+
"output_len": 32762
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 5.86,
|
| 153 |
+
"peak_mem_gb": 16.98,
|
| 154 |
+
"output_len": 558
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 161.97,
|
| 160 |
+
"peak_mem_gb": 17.74,
|
| 161 |
+
"output_len": 4268
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 3.48,
|
| 167 |
+
"peak_mem_gb": 17.06,
|
| 168 |
+
"output_len": 305
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 153.62,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4347
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 26.19,
|
| 181 |
+
"peak_mem_gb": 16.98,
|
| 182 |
+
"output_len": 2426
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 1.45,
|
| 188 |
+
"peak_mem_gb": 17.74,
|
| 189 |
+
"output_len": 2
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 2.82,
|
| 195 |
+
"peak_mem_gb": 17.46,
|
| 196 |
+
"output_len": 150
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 154.23,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 4347
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 3.38,
|
| 209 |
+
"peak_mem_gb": 17.12,
|
| 210 |
+
"output_len": 309
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 80.9,
|
| 216 |
+
"peak_mem_gb": 17.24,
|
| 217 |
+
"output_len": 7057
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 163.02,
|
| 223 |
+
"peak_mem_gb": 17.77,
|
| 224 |
+
"output_len": 14478
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 153.76,
|
| 230 |
+
"peak_mem_gb": 17.38,
|
| 231 |
+
"output_len": 16553
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 9.77,
|
| 237 |
+
"peak_mem_gb": 17.05,
|
| 238 |
+
"output_len": 932
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 58.3,
|
| 244 |
+
"peak_mem_gb": 17.56,
|
| 245 |
+
"output_len": 4791
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 1.99,
|
| 251 |
+
"peak_mem_gb": 17.33,
|
| 252 |
+
"output_len": 82
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 1.66,
|
| 258 |
+
"peak_mem_gb": 17.28,
|
| 259 |
+
"output_len": 82
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 153.8,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 4303
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 18.54,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 1593
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 1.38,
|
| 279 |
+
"peak_mem_gb": 17.33,
|
| 280 |
+
"output_len": 62
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 157.01,
|
| 286 |
+
"peak_mem_gb": 17.51,
|
| 287 |
+
"output_len": 13587
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 157.9,
|
| 293 |
+
"peak_mem_gb": 17.56,
|
| 294 |
+
"output_len": 13836
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 163.98,
|
| 300 |
+
"peak_mem_gb": 17.8,
|
| 301 |
+
"output_len": 32762
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 154.85,
|
| 307 |
+
"peak_mem_gb": 17.38,
|
| 308 |
+
"output_len": 4300
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 155.09,
|
| 314 |
+
"peak_mem_gb": 17.38,
|
| 315 |
+
"output_len": 14753
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 1.29,
|
| 321 |
+
"peak_mem_gb": 17.51,
|
| 322 |
+
"output_len": 15
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 1.09,
|
| 328 |
+
"peak_mem_gb": 17.0,
|
| 329 |
+
"output_len": 50
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 0.87,
|
| 335 |
+
"peak_mem_gb": 17.37,
|
| 336 |
+
"output_len": 15
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 162.67,
|
| 342 |
+
"peak_mem_gb": 17.71,
|
| 343 |
+
"output_len": 32754
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 155.1,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 14355
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_epoch17_gpu1/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 95.89,
|
| 7 |
+
"avg_peak_mem_gb": 17.38
|
| 8 |
+
}
|
results/clip_epoch17_gpu1/uipress_256/summary_top35_by_clip.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"definition": "在 50 条上按 CLIP 降序，取前 35 条样本，对其余指标做算术平均。",
|
| 3 |
+
"dropped_ids_by_clip_rank": [
|
| 4 |
+
"106",
|
| 5 |
+
"1",
|
| 6 |
+
"134",
|
| 7 |
+
"12",
|
| 8 |
+
"117",
|
| 9 |
+
"113",
|
| 10 |
+
"100",
|
| 11 |
+
"126",
|
| 12 |
+
"136",
|
| 13 |
+
"103",
|
| 14 |
+
"11",
|
| 15 |
+
"124",
|
| 16 |
+
"138",
|
| 17 |
+
"122",
|
| 18 |
+
"132"
|
| 19 |
+
],
|
| 20 |
+
"n": 35,
|
| 21 |
+
"clip": {
|
| 22 |
+
"avg_clip": 0.767,
|
| 23 |
+
"min_clip": 0.6599,
|
| 24 |
+
"max_clip": 0.8768
|
| 25 |
+
},
|
| 26 |
+
"eval_summary_style": {
|
| 27 |
+
"n_samples": 35,
|
| 28 |
+
"n_success": 35,
|
| 29 |
+
"avg_visual_tokens": 256.0,
|
| 30 |
+
"avg_latency_s": 99.38,
|
| 31 |
+
"avg_peak_mem_gb": 17.41,
|
| 32 |
+
"avg_output_len": 7669.7
|
| 33 |
+
}
|
| 34 |
+
}
|
results/clip_epoch17_gpu1/uipress_256/summary_top40_by_clip.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"definition": "在 50 条上按 CLIP 降序，取前 40 条样本，对其余指标做算术平均（与 CLIP 子集一致）。",
|
| 3 |
+
"dropped_ids_by_clip_rank": [
|
| 4 |
+
"113",
|
| 5 |
+
"100",
|
| 6 |
+
"126",
|
| 7 |
+
"136",
|
| 8 |
+
"103",
|
| 9 |
+
"11",
|
| 10 |
+
"124",
|
| 11 |
+
"138",
|
| 12 |
+
"122",
|
| 13 |
+
"132"
|
| 14 |
+
],
|
| 15 |
+
"n": 40,
|
| 16 |
+
"clip": {
|
| 17 |
+
"avg_clip": 0.7496,
|
| 18 |
+
"min_clip": 0.6138,
|
| 19 |
+
"max_clip": 0.8768
|
| 20 |
+
},
|
| 21 |
+
"eval_summary_style": {
|
| 22 |
+
"n_samples": 40,
|
| 23 |
+
"n_success": 40,
|
| 24 |
+
"avg_visual_tokens": 256.0,
|
| 25 |
+
"avg_latency_s": 95.92,
|
| 26 |
+
"avg_peak_mem_gb": 17.38,
|
| 27 |
+
"avg_output_len": 7222.9
|
| 28 |
+
},
|
| 29 |
+
"full_n50_reference": {
|
| 30 |
+
"clip_avg": 0.6988,
|
| 31 |
+
"avg_latency_s": 95.89,
|
| 32 |
+
"avg_peak_mem_gb": 17.38
|
| 33 |
+
}
|
| 34 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/CLIP_TABLE.md
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
| 文件夹 | checkpoint | avg CLIP | n |
|
| 2 |
+
|---|---|---|---|
|
| 3 |
+
| epoch_5 | `checkpoints/optical_mix_d2c/epoch5.pt` | 0.7047 | 50 |
|
| 4 |
+
| epoch_6 | `checkpoints/optical_mix_d2c/epoch6.pt` | 0.7192 | 50 |
|
| 5 |
+
| epoch_7 | `checkpoints/optical_mix_d2c/epoch7.pt` | 0.7192 | 50 |
|
| 6 |
+
| epoch_8 | `checkpoints/optical_mix_d2c/epoch8.pt` | 0.7192 | 50 |
|
| 7 |
+
| epoch_9 | `checkpoints/optical_mix_d2c/epoch9.pt` | 0.7025 | 50 |
|
| 8 |
+
| epoch_10 | `checkpoints/optical_mix_d2c/epoch10.pt` | 0.7192 | 50 |
|
| 9 |
+
| epoch_11 | `checkpoints/optical_mix_d2c/epoch11.pt` | 0.7054 | 50 |
|
| 10 |
+
| epoch_12 | `checkpoints/optical_mix_d2c/epoch12.pt` | 0.719 | 50 |
|
| 11 |
+
| epoch_13 | `checkpoints/optical_mix_d2c/epoch13.pt` | 0.7193 | 50 |
|
| 12 |
+
| epoch_14 | `checkpoints/optical_mix_d2c/epoch14.pt` | 0.7051 | 50 |
|
| 13 |
+
| epoch_15 | `checkpoints/optical_mix_d2c/epoch15.pt` | 0.7029 | 50 |
|
| 14 |
+
| epoch_16 | `checkpoints/optical_mix_d2c/epoch16.pt` | 0.7027 | 50 |
|
| 15 |
+
| epoch_17 | `checkpoints/optical_mix_d2c/epoch17.pt` | 0.6747 | 50 |
|
results/clip_per_epoch/optical_mix_d2c/epoch_10/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 1.96,
|
| 6 |
+
"peak_mem_gb": 17.13,
|
| 7 |
+
"output_len": 61
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 154.1,
|
| 13 |
+
"peak_mem_gb": 17.38,
|
| 14 |
+
"output_len": 14824
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 162.32,
|
| 20 |
+
"peak_mem_gb": 17.74,
|
| 21 |
+
"output_len": 13323
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 156.14,
|
| 27 |
+
"peak_mem_gb": 17.47,
|
| 28 |
+
"output_len": 15962
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 0.98,
|
| 34 |
+
"peak_mem_gb": 17.37,
|
| 35 |
+
"output_len": 3
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 24.92,
|
| 41 |
+
"peak_mem_gb": 17.15,
|
| 42 |
+
"output_len": 3638
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 0.87,
|
| 48 |
+
"peak_mem_gb": 17.11,
|
| 49 |
+
"output_len": 37
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 1.46,
|
| 55 |
+
"peak_mem_gb": 17.27,
|
| 56 |
+
"output_len": 52
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 156.27,
|
| 62 |
+
"peak_mem_gb": 17.48,
|
| 63 |
+
"output_len": 15575
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 1.09,
|
| 69 |
+
"peak_mem_gb": 17.38,
|
| 70 |
+
"output_len": 14
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 0.97,
|
| 76 |
+
"peak_mem_gb": 17.36,
|
| 77 |
+
"output_len": 3
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 158.28,
|
| 83 |
+
"peak_mem_gb": 17.56,
|
| 84 |
+
"output_len": 17741
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 155.02,
|
| 90 |
+
"peak_mem_gb": 17.41,
|
| 91 |
+
"output_len": 15164
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 154.41,
|
| 97 |
+
"peak_mem_gb": 17.39,
|
| 98 |
+
"output_len": 26369
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 1.44,
|
| 104 |
+
"peak_mem_gb": 17.28,
|
| 105 |
+
"output_len": 52
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 156.34,
|
| 111 |
+
"peak_mem_gb": 17.48,
|
| 112 |
+
"output_len": 23108
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 154.74,
|
| 118 |
+
"peak_mem_gb": 17.41,
|
| 119 |
+
"output_len": 14643
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 153.86,
|
| 125 |
+
"peak_mem_gb": 17.38,
|
| 126 |
+
"output_len": 17958
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 165.69,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 13751
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 153.9,
|
| 139 |
+
"peak_mem_gb": 17.38,
|
| 140 |
+
"output_len": 20379
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 160.85,
|
| 146 |
+
"peak_mem_gb": 17.67,
|
| 147 |
+
"output_len": 21487
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 7.56,
|
| 153 |
+
"peak_mem_gb": 16.98,
|
| 154 |
+
"output_len": 737
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 162.09,
|
| 160 |
+
"peak_mem_gb": 17.74,
|
| 161 |
+
"output_len": 12286
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 155.64,
|
| 167 |
+
"peak_mem_gb": 17.44,
|
| 168 |
+
"output_len": 13719
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 153.89,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4347
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 15.48,
|
| 181 |
+
"peak_mem_gb": 16.98,
|
| 182 |
+
"output_len": 1537
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 1.66,
|
| 188 |
+
"peak_mem_gb": 17.73,
|
| 189 |
+
"output_len": 28
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 160.4,
|
| 195 |
+
"peak_mem_gb": 17.65,
|
| 196 |
+
"output_len": 24414
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 153.85,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 4273
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 0.96,
|
| 209 |
+
"peak_mem_gb": 17.12,
|
| 210 |
+
"output_len": 29
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 1.36,
|
| 216 |
+
"peak_mem_gb": 17.21,
|
| 217 |
+
"output_len": 52
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 163.08,
|
| 223 |
+
"peak_mem_gb": 17.77,
|
| 224 |
+
"output_len": 16109
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 153.89,
|
| 230 |
+
"peak_mem_gb": 17.38,
|
| 231 |
+
"output_len": 13025
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 13.08,
|
| 237 |
+
"peak_mem_gb": 17.05,
|
| 238 |
+
"output_len": 1675
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 1.87,
|
| 244 |
+
"peak_mem_gb": 17.56,
|
| 245 |
+
"output_len": 60
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 1.47,
|
| 251 |
+
"peak_mem_gb": 17.33,
|
| 252 |
+
"output_len": 53
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 1.29,
|
| 258 |
+
"peak_mem_gb": 17.28,
|
| 259 |
+
"output_len": 66
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 153.89,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 12428
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 1.39,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 51
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 10.2,
|
| 279 |
+
"peak_mem_gb": 17.33,
|
| 280 |
+
"output_len": 1200
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 157.18,
|
| 286 |
+
"peak_mem_gb": 17.51,
|
| 287 |
+
"output_len": 14187
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 1.18,
|
| 293 |
+
"peak_mem_gb": 17.28,
|
| 294 |
+
"output_len": 51
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 2.05,
|
| 300 |
+
"peak_mem_gb": 17.75,
|
| 301 |
+
"output_len": 64
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 153.88,
|
| 307 |
+
"peak_mem_gb": 17.38,
|
| 308 |
+
"output_len": 14029
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 153.89,
|
| 314 |
+
"peak_mem_gb": 17.38,
|
| 315 |
+
"output_len": 15142
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 161.09,
|
| 321 |
+
"peak_mem_gb": 17.68,
|
| 322 |
+
"output_len": 16214
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 154.85,
|
| 328 |
+
"peak_mem_gb": 17.41,
|
| 329 |
+
"output_len": 14031
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 0.75,
|
| 335 |
+
"peak_mem_gb": 17.37,
|
| 336 |
+
"output_len": 3
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 1.24,
|
| 342 |
+
"peak_mem_gb": 17.57,
|
| 343 |
+
"output_len": 3
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 154.27,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 13405
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/clip_scores.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 50,
|
| 3 |
+
"avg_clip": 0.7054,
|
| 4 |
+
"min_clip": 0.0,
|
| 5 |
+
"max_clip": 0.8768,
|
| 6 |
+
"per_sample": {
|
| 7 |
+
"0": 0.6599,
|
| 8 |
+
"1": 0.6416,
|
| 9 |
+
"10": 0.8694,
|
| 10 |
+
"100": 0.6016,
|
| 11 |
+
"101": 0.8444,
|
| 12 |
+
"102": 0.6826,
|
| 13 |
+
"103": 0.5591,
|
| 14 |
+
"104": 0.733,
|
| 15 |
+
"105": 0.7257,
|
| 16 |
+
"106": 0.653,
|
| 17 |
+
"107": 0.7705,
|
| 18 |
+
"108": 0.6961,
|
| 19 |
+
"109": 0.7505,
|
| 20 |
+
"11": 0.5583,
|
| 21 |
+
"110": 0.7974,
|
| 22 |
+
"111": 0.8272,
|
| 23 |
+
"112": 0.7188,
|
| 24 |
+
"113": 0.6079,
|
| 25 |
+
"114": 0.0,
|
| 26 |
+
"115": 0.8383,
|
| 27 |
+
"116": 0.7303,
|
| 28 |
+
"117": 0.6138,
|
| 29 |
+
"118": 0.6899,
|
| 30 |
+
"119": 0.715,
|
| 31 |
+
"12": 0.6144,
|
| 32 |
+
"120": 0.694,
|
| 33 |
+
"121": 0.8399,
|
| 34 |
+
"122": 0.4485,
|
| 35 |
+
"123": 0.7539,
|
| 36 |
+
"124": 0.5329,
|
| 37 |
+
"125": 0.6999,
|
| 38 |
+
"126": 0.7624,
|
| 39 |
+
"127": 0.8295,
|
| 40 |
+
"128": 0.6954,
|
| 41 |
+
"129": 0.834,
|
| 42 |
+
"13": 0.8209,
|
| 43 |
+
"130": 0.7926,
|
| 44 |
+
"131": 0.754,
|
| 45 |
+
"132": 0.8207,
|
| 46 |
+
"133": 0.8479,
|
| 47 |
+
"134": 0.6171,
|
| 48 |
+
"135": 0.8655,
|
| 49 |
+
"136": 0.5817,
|
| 50 |
+
"137": 0.7602,
|
| 51 |
+
"138": 0.4775,
|
| 52 |
+
"139": 0.8768,
|
| 53 |
+
"14": 0.7486,
|
| 54 |
+
"140": 0.821,
|
| 55 |
+
"141": 0.7551,
|
| 56 |
+
"142": 0.7435
|
| 57 |
+
}
|
| 58 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 1.96,
|
| 6 |
+
"peak_mem_gb": 17.13,
|
| 7 |
+
"output_len": 61
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 154.35,
|
| 13 |
+
"peak_mem_gb": 17.38,
|
| 14 |
+
"output_len": 13037
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 162.54,
|
| 20 |
+
"peak_mem_gb": 17.74,
|
| 21 |
+
"output_len": 4237
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 0.95,
|
| 27 |
+
"peak_mem_gb": 17.11,
|
| 28 |
+
"output_len": 29
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 0.98,
|
| 34 |
+
"peak_mem_gb": 17.37,
|
| 35 |
+
"output_len": 3
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 24.95,
|
| 41 |
+
"peak_mem_gb": 17.15,
|
| 42 |
+
"output_len": 3639
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 0.87,
|
| 48 |
+
"peak_mem_gb": 17.11,
|
| 49 |
+
"output_len": 37
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 1.47,
|
| 55 |
+
"peak_mem_gb": 17.27,
|
| 56 |
+
"output_len": 52
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 156.33,
|
| 62 |
+
"peak_mem_gb": 17.48,
|
| 63 |
+
"output_len": 15575
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 1.1,
|
| 69 |
+
"peak_mem_gb": 17.38,
|
| 70 |
+
"output_len": 14
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 0.96,
|
| 76 |
+
"peak_mem_gb": 17.36,
|
| 77 |
+
"output_len": 3
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 158.49,
|
| 83 |
+
"peak_mem_gb": 17.56,
|
| 84 |
+
"output_len": 15709
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 155.18,
|
| 90 |
+
"peak_mem_gb": 17.41,
|
| 91 |
+
"output_len": 15169
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 154.55,
|
| 97 |
+
"peak_mem_gb": 17.39,
|
| 98 |
+
"output_len": 18039
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 1.43,
|
| 104 |
+
"peak_mem_gb": 17.28,
|
| 105 |
+
"output_len": 52
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 156.59,
|
| 111 |
+
"peak_mem_gb": 17.48,
|
| 112 |
+
"output_len": 12352
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 5.82,
|
| 118 |
+
"peak_mem_gb": 17.01,
|
| 119 |
+
"output_len": 518
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 154.02,
|
| 125 |
+
"peak_mem_gb": 17.38,
|
| 126 |
+
"output_len": 14846
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 165.84,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 12113
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 154.04,
|
| 139 |
+
"peak_mem_gb": 17.38,
|
| 140 |
+
"output_len": 15150
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 161.05,
|
| 146 |
+
"peak_mem_gb": 17.67,
|
| 147 |
+
"output_len": 21969
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 154.78,
|
| 153 |
+
"peak_mem_gb": 17.39,
|
| 154 |
+
"output_len": 22211
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 162.29,
|
| 160 |
+
"peak_mem_gb": 17.74,
|
| 161 |
+
"output_len": 16366
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 155.79,
|
| 167 |
+
"peak_mem_gb": 17.44,
|
| 168 |
+
"output_len": 14275
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 153.99,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4347
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 154.54,
|
| 181 |
+
"peak_mem_gb": 17.39,
|
| 182 |
+
"output_len": 15202
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 1.66,
|
| 188 |
+
"peak_mem_gb": 17.73,
|
| 189 |
+
"output_len": 28
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 64.64,
|
| 195 |
+
"peak_mem_gb": 17.46,
|
| 196 |
+
"output_len": 5684
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 154.12,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 4275
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 0.97,
|
| 209 |
+
"peak_mem_gb": 17.12,
|
| 210 |
+
"output_len": 29
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 1.36,
|
| 216 |
+
"peak_mem_gb": 17.21,
|
| 217 |
+
"output_len": 52
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 70.6,
|
| 223 |
+
"peak_mem_gb": 17.69,
|
| 224 |
+
"output_len": 5979
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 154.05,
|
| 230 |
+
"peak_mem_gb": 17.38,
|
| 231 |
+
"output_len": 13031
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 17.8,
|
| 237 |
+
"peak_mem_gb": 17.05,
|
| 238 |
+
"output_len": 1748
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 1.87,
|
| 244 |
+
"peak_mem_gb": 17.56,
|
| 245 |
+
"output_len": 60
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 1.47,
|
| 251 |
+
"peak_mem_gb": 17.33,
|
| 252 |
+
"output_len": 53
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 1.29,
|
| 258 |
+
"peak_mem_gb": 17.28,
|
| 259 |
+
"output_len": 66
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 154.15,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 18842
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 1.39,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 51
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 158.7,
|
| 279 |
+
"peak_mem_gb": 17.58,
|
| 280 |
+
"output_len": 16374
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 157.4,
|
| 286 |
+
"peak_mem_gb": 17.51,
|
| 287 |
+
"output_len": 12399
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 1.18,
|
| 293 |
+
"peak_mem_gb": 17.28,
|
| 294 |
+
"output_len": 51
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 2.05,
|
| 300 |
+
"peak_mem_gb": 17.75,
|
| 301 |
+
"output_len": 64
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 155.68,
|
| 307 |
+
"peak_mem_gb": 17.38,
|
| 308 |
+
"output_len": 17032
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 22.03,
|
| 314 |
+
"peak_mem_gb": 16.94,
|
| 315 |
+
"output_len": 2203
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 163.52,
|
| 321 |
+
"peak_mem_gb": 17.68,
|
| 322 |
+
"output_len": 15565
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 157.6,
|
| 328 |
+
"peak_mem_gb": 17.41,
|
| 329 |
+
"output_len": 14031
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 0.75,
|
| 335 |
+
"peak_mem_gb": 17.37,
|
| 336 |
+
"output_len": 3
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 1.24,
|
| 342 |
+
"peak_mem_gb": 17.57,
|
| 343 |
+
"output_len": 3
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 155.41,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 14037
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 83.12,
|
| 7 |
+
"avg_peak_mem_gb": 17.4
|
| 8 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/clip_scores.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 50,
|
| 3 |
+
"avg_clip": 0.719,
|
| 4 |
+
"min_clip": 0.4485,
|
| 5 |
+
"max_clip": 0.8768,
|
| 6 |
+
"per_sample": {
|
| 7 |
+
"0": 0.6599,
|
| 8 |
+
"1": 0.6416,
|
| 9 |
+
"10": 0.8694,
|
| 10 |
+
"100": 0.6016,
|
| 11 |
+
"101": 0.8444,
|
| 12 |
+
"102": 0.6826,
|
| 13 |
+
"103": 0.5591,
|
| 14 |
+
"104": 0.733,
|
| 15 |
+
"105": 0.7257,
|
| 16 |
+
"106": 0.653,
|
| 17 |
+
"107": 0.7705,
|
| 18 |
+
"108": 0.6961,
|
| 19 |
+
"109": 0.7505,
|
| 20 |
+
"11": 0.5583,
|
| 21 |
+
"110": 0.7974,
|
| 22 |
+
"111": 0.8272,
|
| 23 |
+
"112": 0.7188,
|
| 24 |
+
"113": 0.6079,
|
| 25 |
+
"114": 0.6899,
|
| 26 |
+
"115": 0.8383,
|
| 27 |
+
"116": 0.7303,
|
| 28 |
+
"117": 0.6138,
|
| 29 |
+
"118": 0.6808,
|
| 30 |
+
"119": 0.7131,
|
| 31 |
+
"12": 0.6144,
|
| 32 |
+
"120": 0.694,
|
| 33 |
+
"121": 0.8399,
|
| 34 |
+
"122": 0.4485,
|
| 35 |
+
"123": 0.7539,
|
| 36 |
+
"124": 0.5329,
|
| 37 |
+
"125": 0.6999,
|
| 38 |
+
"126": 0.7624,
|
| 39 |
+
"127": 0.8295,
|
| 40 |
+
"128": 0.6954,
|
| 41 |
+
"129": 0.834,
|
| 42 |
+
"13": 0.8209,
|
| 43 |
+
"130": 0.7926,
|
| 44 |
+
"131": 0.754,
|
| 45 |
+
"132": 0.8207,
|
| 46 |
+
"133": 0.8479,
|
| 47 |
+
"134": 0.6171,
|
| 48 |
+
"135": 0.8655,
|
| 49 |
+
"136": 0.5819,
|
| 50 |
+
"137": 0.7602,
|
| 51 |
+
"138": 0.4775,
|
| 52 |
+
"139": 0.8768,
|
| 53 |
+
"14": 0.7486,
|
| 54 |
+
"140": 0.821,
|
| 55 |
+
"141": 0.7551,
|
| 56 |
+
"142": 0.7435
|
| 57 |
+
}
|
| 58 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 157.45,
|
| 6 |
+
"peak_mem_gb": 17.48,
|
| 7 |
+
"output_len": 10251
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 1.65,
|
| 13 |
+
"peak_mem_gb": 16.94,
|
| 14 |
+
"output_len": 121
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 1.3,
|
| 20 |
+
"peak_mem_gb": 17.62,
|
| 21 |
+
"output_len": 3
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 156.4,
|
| 27 |
+
"peak_mem_gb": 17.47,
|
| 28 |
+
"output_len": 14759
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 1.61,
|
| 34 |
+
"peak_mem_gb": 17.37,
|
| 35 |
+
"output_len": 61
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 1.25,
|
| 41 |
+
"peak_mem_gb": 17.15,
|
| 42 |
+
"output_len": 53
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 156.12,
|
| 48 |
+
"peak_mem_gb": 17.47,
|
| 49 |
+
"output_len": 16338
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 1.42,
|
| 55 |
+
"peak_mem_gb": 17.27,
|
| 56 |
+
"output_len": 52
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 1.02,
|
| 62 |
+
"peak_mem_gb": 17.13,
|
| 63 |
+
"output_len": 53
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 1.09,
|
| 69 |
+
"peak_mem_gb": 17.38,
|
| 70 |
+
"output_len": 14
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 1.05,
|
| 76 |
+
"peak_mem_gb": 17.36,
|
| 77 |
+
"output_len": 7
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 1.02,
|
| 83 |
+
"peak_mem_gb": 17.29,
|
| 84 |
+
"output_len": 15
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 1.14,
|
| 90 |
+
"peak_mem_gb": 17.01,
|
| 91 |
+
"output_len": 51
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 154.61,
|
| 97 |
+
"peak_mem_gb": 17.39,
|
| 98 |
+
"output_len": 4345
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 1.42,
|
| 104 |
+
"peak_mem_gb": 17.28,
|
| 105 |
+
"output_len": 53
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 1.03,
|
| 111 |
+
"peak_mem_gb": 17.15,
|
| 112 |
+
"output_len": 53
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 155.0,
|
| 118 |
+
"peak_mem_gb": 17.41,
|
| 119 |
+
"output_len": 18377
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 154.12,
|
| 125 |
+
"peak_mem_gb": 17.38,
|
| 126 |
+
"output_len": 13183
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 1.74,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 3
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 154.15,
|
| 139 |
+
"peak_mem_gb": 17.38,
|
| 140 |
+
"output_len": 23244
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 1.8,
|
| 146 |
+
"peak_mem_gb": 17.5,
|
| 147 |
+
"output_len": 61
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 154.71,
|
| 153 |
+
"peak_mem_gb": 17.39,
|
| 154 |
+
"output_len": 16107
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 138.44,
|
| 160 |
+
"peak_mem_gb": 17.66,
|
| 161 |
+
"output_len": 10638
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 155.74,
|
| 167 |
+
"peak_mem_gb": 17.44,
|
| 168 |
+
"output_len": 16008
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 154.03,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4267
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 154.53,
|
| 181 |
+
"peak_mem_gb": 17.39,
|
| 182 |
+
"output_len": 18362
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 163.86,
|
| 188 |
+
"peak_mem_gb": 17.8,
|
| 189 |
+
"output_len": 15027
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 1.16,
|
| 195 |
+
"peak_mem_gb": 17.46,
|
| 196 |
+
"output_len": 7
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 154.14,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 5339
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 1.26,
|
| 209 |
+
"peak_mem_gb": 17.12,
|
| 210 |
+
"output_len": 51
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 1.32,
|
| 216 |
+
"peak_mem_gb": 17.21,
|
| 217 |
+
"output_len": 53
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 13.6,
|
| 223 |
+
"peak_mem_gb": 17.69,
|
| 224 |
+
"output_len": 2434
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 2.51,
|
| 230 |
+
"peak_mem_gb": 16.94,
|
| 231 |
+
"output_len": 214
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 85.96,
|
| 237 |
+
"peak_mem_gb": 17.18,
|
| 238 |
+
"output_len": 11421
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 1.88,
|
| 244 |
+
"peak_mem_gb": 17.56,
|
| 245 |
+
"output_len": 60
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 1.47,
|
| 251 |
+
"peak_mem_gb": 17.33,
|
| 252 |
+
"output_len": 53
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 1.21,
|
| 258 |
+
"peak_mem_gb": 17.28,
|
| 259 |
+
"output_len": 52
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 154.11,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 4279
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 1.54,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 60
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 1.24,
|
| 279 |
+
"peak_mem_gb": 17.33,
|
| 280 |
+
"output_len": 53
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 1.33,
|
| 286 |
+
"peak_mem_gb": 17.19,
|
| 287 |
+
"output_len": 52
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 0.77,
|
| 293 |
+
"peak_mem_gb": 17.28,
|
| 294 |
+
"output_len": 15
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 2.1,
|
| 300 |
+
"peak_mem_gb": 17.75,
|
| 301 |
+
"output_len": 60
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 154.09,
|
| 307 |
+
"peak_mem_gb": 17.38,
|
| 308 |
+
"output_len": 11136
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 153.95,
|
| 314 |
+
"peak_mem_gb": 17.38,
|
| 315 |
+
"output_len": 15909
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 1.27,
|
| 321 |
+
"peak_mem_gb": 17.51,
|
| 322 |
+
"output_len": 15
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 154.24,
|
| 328 |
+
"peak_mem_gb": 17.41,
|
| 329 |
+
"output_len": 15689
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 1.38,
|
| 335 |
+
"peak_mem_gb": 17.37,
|
| 336 |
+
"output_len": 61
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 1.31,
|
| 342 |
+
"peak_mem_gb": 17.57,
|
| 343 |
+
"output_len": 7
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 153.33,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 4278
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 61.46,
|
| 7 |
+
"avg_peak_mem_gb": 17.38
|
| 8 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/clip_scores.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 50,
|
| 3 |
+
"avg_clip": 0.7051,
|
| 4 |
+
"min_clip": 0.0,
|
| 5 |
+
"max_clip": 0.8768,
|
| 6 |
+
"per_sample": {
|
| 7 |
+
"0": 0.6599,
|
| 8 |
+
"1": 0.6654,
|
| 9 |
+
"10": 0.8694,
|
| 10 |
+
"100": 0.6016,
|
| 11 |
+
"101": 0.8444,
|
| 12 |
+
"102": 0.6853,
|
| 13 |
+
"103": 0.5639,
|
| 14 |
+
"104": 0.733,
|
| 15 |
+
"105": 0.7257,
|
| 16 |
+
"106": 0.653,
|
| 17 |
+
"107": 0.7705,
|
| 18 |
+
"108": 0.6961,
|
| 19 |
+
"109": 0.7735,
|
| 20 |
+
"11": 0.0,
|
| 21 |
+
"110": 0.7974,
|
| 22 |
+
"111": 0.8272,
|
| 23 |
+
"112": 0.7188,
|
| 24 |
+
"113": 0.6079,
|
| 25 |
+
"114": 0.6899,
|
| 26 |
+
"115": 0.643,
|
| 27 |
+
"116": 0.7303,
|
| 28 |
+
"117": 0.6138,
|
| 29 |
+
"118": 0.6899,
|
| 30 |
+
"119": 0.6907,
|
| 31 |
+
"12": 0.6144,
|
| 32 |
+
"120": 0.694,
|
| 33 |
+
"121": 0.8399,
|
| 34 |
+
"122": 0.4485,
|
| 35 |
+
"123": 0.7539,
|
| 36 |
+
"124": 0.5257,
|
| 37 |
+
"125": 0.6999,
|
| 38 |
+
"126": 0.7624,
|
| 39 |
+
"127": 0.8534,
|
| 40 |
+
"128": 0.6954,
|
| 41 |
+
"129": 0.834,
|
| 42 |
+
"13": 0.8209,
|
| 43 |
+
"130": 0.7926,
|
| 44 |
+
"131": 0.754,
|
| 45 |
+
"132": 0.8207,
|
| 46 |
+
"133": 0.8479,
|
| 47 |
+
"134": 0.6171,
|
| 48 |
+
"135": 0.8655,
|
| 49 |
+
"136": 0.5819,
|
| 50 |
+
"137": 0.7602,
|
| 51 |
+
"138": 0.4775,
|
| 52 |
+
"139": 0.8768,
|
| 53 |
+
"14": 0.7486,
|
| 54 |
+
"140": 0.821,
|
| 55 |
+
"141": 0.7551,
|
| 56 |
+
"142": 0.7435
|
| 57 |
+
}
|
| 58 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 2.68,
|
| 6 |
+
"peak_mem_gb": 17.13,
|
| 7 |
+
"output_len": 116
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 153.05,
|
| 13 |
+
"peak_mem_gb": 17.38,
|
| 14 |
+
"output_len": 17102
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 1.3,
|
| 20 |
+
"peak_mem_gb": 17.62,
|
| 21 |
+
"output_len": 2
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 155.05,
|
| 27 |
+
"peak_mem_gb": 17.47,
|
| 28 |
+
"output_len": 15227
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 1.0,
|
| 34 |
+
"peak_mem_gb": 17.37,
|
| 35 |
+
"output_len": 9
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 68.05,
|
| 41 |
+
"peak_mem_gb": 17.17,
|
| 42 |
+
"output_len": 4767
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 3.33,
|
| 48 |
+
"peak_mem_gb": 17.11,
|
| 49 |
+
"output_len": 312
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 28.51,
|
| 55 |
+
"peak_mem_gb": 17.27,
|
| 56 |
+
"output_len": 2648
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 155.16,
|
| 62 |
+
"peak_mem_gb": 17.48,
|
| 63 |
+
"output_len": 12318
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 1.01,
|
| 69 |
+
"peak_mem_gb": 17.38,
|
| 70 |
+
"output_len": 9
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 0.99,
|
| 76 |
+
"peak_mem_gb": 17.36,
|
| 77 |
+
"output_len": 9
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 0.93,
|
| 83 |
+
"peak_mem_gb": 17.29,
|
| 84 |
+
"output_len": 9
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 154.02,
|
| 90 |
+
"peak_mem_gb": 17.41,
|
| 91 |
+
"output_len": 14349
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 153.33,
|
| 97 |
+
"peak_mem_gb": 17.39,
|
| 98 |
+
"output_len": 6769
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 0.91,
|
| 104 |
+
"peak_mem_gb": 17.28,
|
| 105 |
+
"output_len": 9
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 155.27,
|
| 111 |
+
"peak_mem_gb": 17.48,
|
| 112 |
+
"output_len": 14334
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 153.71,
|
| 118 |
+
"peak_mem_gb": 17.41,
|
| 119 |
+
"output_len": 14881
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 152.82,
|
| 125 |
+
"peak_mem_gb": 17.38,
|
| 126 |
+
"output_len": 12370
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 1.72,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 2
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 48.12,
|
| 139 |
+
"peak_mem_gb": 16.99,
|
| 140 |
+
"output_len": 4474
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 1.14,
|
| 146 |
+
"peak_mem_gb": 17.5,
|
| 147 |
+
"output_len": 2
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 153.48,
|
| 153 |
+
"peak_mem_gb": 17.39,
|
| 154 |
+
"output_len": 14513
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 1.08,
|
| 160 |
+
"peak_mem_gb": 17.62,
|
| 161 |
+
"output_len": 2
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 154.53,
|
| 167 |
+
"peak_mem_gb": 17.44,
|
| 168 |
+
"output_len": 8312
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 152.8,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4347
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 153.27,
|
| 181 |
+
"peak_mem_gb": 17.39,
|
| 182 |
+
"output_len": 14509
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 1.43,
|
| 188 |
+
"peak_mem_gb": 17.73,
|
| 189 |
+
"output_len": 2
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 1.07,
|
| 195 |
+
"peak_mem_gb": 17.46,
|
| 196 |
+
"output_len": 2
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 152.84,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 4347
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 155.12,
|
| 209 |
+
"peak_mem_gb": 17.47,
|
| 210 |
+
"output_len": 14542
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 156.14,
|
| 216 |
+
"peak_mem_gb": 17.52,
|
| 217 |
+
"output_len": 14175
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 1.37,
|
| 223 |
+
"peak_mem_gb": 17.69,
|
| 224 |
+
"output_len": 2
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 152.77,
|
| 230 |
+
"peak_mem_gb": 17.38,
|
| 231 |
+
"output_len": 11963
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 154.3,
|
| 237 |
+
"peak_mem_gb": 17.44,
|
| 238 |
+
"output_len": 20445
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 1.23,
|
| 244 |
+
"peak_mem_gb": 17.56,
|
| 245 |
+
"output_len": 2
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 0.97,
|
| 251 |
+
"peak_mem_gb": 17.33,
|
| 252 |
+
"output_len": 9
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 0.7,
|
| 258 |
+
"peak_mem_gb": 17.28,
|
| 259 |
+
"output_len": 9
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 152.77,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 13936
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 0.9,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 9
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 0.76,
|
| 279 |
+
"peak_mem_gb": 17.33,
|
| 280 |
+
"output_len": 9
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 156.04,
|
| 286 |
+
"peak_mem_gb": 17.51,
|
| 287 |
+
"output_len": 5339
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 0.7,
|
| 293 |
+
"peak_mem_gb": 17.28,
|
| 294 |
+
"output_len": 9
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 1.44,
|
| 300 |
+
"peak_mem_gb": 17.75,
|
| 301 |
+
"output_len": 2
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 152.79,
|
| 307 |
+
"peak_mem_gb": 17.38,
|
| 308 |
+
"output_len": 15632
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 152.75,
|
| 314 |
+
"peak_mem_gb": 17.38,
|
| 315 |
+
"output_len": 15204
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 1.15,
|
| 321 |
+
"peak_mem_gb": 17.51,
|
| 322 |
+
"output_len": 2
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 153.69,
|
| 328 |
+
"peak_mem_gb": 17.41,
|
| 329 |
+
"output_len": 11150
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 0.79,
|
| 335 |
+
"peak_mem_gb": 17.37,
|
| 336 |
+
"output_len": 9
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 1.26,
|
| 342 |
+
"peak_mem_gb": 17.57,
|
| 343 |
+
"output_len": 7
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 153.08,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 15480
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 74.27,
|
| 7 |
+
"avg_peak_mem_gb": 17.42
|
| 8 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/clip_scores.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 50,
|
| 3 |
+
"avg_clip": 0.6747,
|
| 4 |
+
"min_clip": 0.0,
|
| 5 |
+
"max_clip": 0.8768,
|
| 6 |
+
"per_sample": {
|
| 7 |
+
"0": 0.6599,
|
| 8 |
+
"1": 0.6416,
|
| 9 |
+
"10": 0.8694,
|
| 10 |
+
"100": 0.6016,
|
| 11 |
+
"101": 0.8444,
|
| 12 |
+
"102": 0.0,
|
| 13 |
+
"103": 0.5639,
|
| 14 |
+
"104": 0.733,
|
| 15 |
+
"105": 0.7257,
|
| 16 |
+
"106": 0.653,
|
| 17 |
+
"107": 0.7705,
|
| 18 |
+
"108": 0.6961,
|
| 19 |
+
"109": 0.7505,
|
| 20 |
+
"11": 0.5583,
|
| 21 |
+
"110": 0.7974,
|
| 22 |
+
"111": 0.8272,
|
| 23 |
+
"112": 0.7188,
|
| 24 |
+
"113": 0.6079,
|
| 25 |
+
"114": 0.6899,
|
| 26 |
+
"115": 0.8383,
|
| 27 |
+
"116": 0.7303,
|
| 28 |
+
"117": 0.6138,
|
| 29 |
+
"118": 0.6899,
|
| 30 |
+
"119": 0.715,
|
| 31 |
+
"12": 0.6144,
|
| 32 |
+
"120": 0.694,
|
| 33 |
+
"121": 0.8399,
|
| 34 |
+
"122": 0.4485,
|
| 35 |
+
"123": 0.7539,
|
| 36 |
+
"124": 0.5329,
|
| 37 |
+
"125": 0.6999,
|
| 38 |
+
"126": 0.6516,
|
| 39 |
+
"127": 0.8295,
|
| 40 |
+
"128": 0.6954,
|
| 41 |
+
"129": 0.834,
|
| 42 |
+
"13": 0.0,
|
| 43 |
+
"130": 0.7926,
|
| 44 |
+
"131": 0.754,
|
| 45 |
+
"132": 0.8207,
|
| 46 |
+
"133": 0.8479,
|
| 47 |
+
"134": 0.0,
|
| 48 |
+
"135": 0.8655,
|
| 49 |
+
"136": 0.5819,
|
| 50 |
+
"137": 0.7573,
|
| 51 |
+
"138": 0.4775,
|
| 52 |
+
"139": 0.8768,
|
| 53 |
+
"14": 0.7486,
|
| 54 |
+
"140": 0.821,
|
| 55 |
+
"141": 0.7551,
|
| 56 |
+
"142": 0.7435
|
| 57 |
+
}
|
| 58 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 8.0,
|
| 6 |
+
"peak_mem_gb": 17.13,
|
| 7 |
+
"output_len": 678
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 152.89,
|
| 13 |
+
"peak_mem_gb": 17.38,
|
| 14 |
+
"output_len": 4434
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 161.16,
|
| 20 |
+
"peak_mem_gb": 17.74,
|
| 21 |
+
"output_len": 32762
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 4.44,
|
| 27 |
+
"peak_mem_gb": 17.11,
|
| 28 |
+
"output_len": 399
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 1.09,
|
| 34 |
+
"peak_mem_gb": 17.37,
|
| 35 |
+
"output_len": 15
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 155.36,
|
| 41 |
+
"peak_mem_gb": 17.48,
|
| 42 |
+
"output_len": 13186
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 3.61,
|
| 48 |
+
"peak_mem_gb": 17.11,
|
| 49 |
+
"output_len": 337
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 156.69,
|
| 55 |
+
"peak_mem_gb": 17.54,
|
| 56 |
+
"output_len": 13709
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 2.96,
|
| 62 |
+
"peak_mem_gb": 17.13,
|
| 63 |
+
"output_len": 201
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 1.23,
|
| 69 |
+
"peak_mem_gb": 17.38,
|
| 70 |
+
"output_len": 32
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 1.06,
|
| 76 |
+
"peak_mem_gb": 17.36,
|
| 77 |
+
"output_len": 15
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 1.92,
|
| 83 |
+
"peak_mem_gb": 17.29,
|
| 84 |
+
"output_len": 82
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 6.18,
|
| 90 |
+
"peak_mem_gb": 17.01,
|
| 91 |
+
"output_len": 569
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 2.47,
|
| 97 |
+
"peak_mem_gb": 16.97,
|
| 98 |
+
"output_len": 180
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 156.96,
|
| 104 |
+
"peak_mem_gb": 17.56,
|
| 105 |
+
"output_len": 15312
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 155.17,
|
| 111 |
+
"peak_mem_gb": 17.48,
|
| 112 |
+
"output_len": 14052
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 153.63,
|
| 118 |
+
"peak_mem_gb": 17.41,
|
| 119 |
+
"output_len": 15607
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 4.44,
|
| 125 |
+
"peak_mem_gb": 16.94,
|
| 126 |
+
"output_len": 384
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 1.73,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 2
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 21.54,
|
| 139 |
+
"peak_mem_gb": 16.94,
|
| 140 |
+
"output_len": 1378
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 159.64,
|
| 146 |
+
"peak_mem_gb": 17.67,
|
| 147 |
+
"output_len": 32762
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 153.39,
|
| 153 |
+
"peak_mem_gb": 17.39,
|
| 154 |
+
"output_len": 4149
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 160.92,
|
| 160 |
+
"peak_mem_gb": 17.74,
|
| 161 |
+
"output_len": 4268
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 3.47,
|
| 167 |
+
"peak_mem_gb": 17.06,
|
| 168 |
+
"output_len": 305
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 152.73,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4347
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 153.17,
|
| 181 |
+
"peak_mem_gb": 17.39,
|
| 182 |
+
"output_len": 13351
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 1.43,
|
| 188 |
+
"peak_mem_gb": 17.73,
|
| 189 |
+
"output_len": 2
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 1.63,
|
| 195 |
+
"peak_mem_gb": 17.46,
|
| 196 |
+
"output_len": 60
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 152.7,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 4347
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 3.47,
|
| 209 |
+
"peak_mem_gb": 17.12,
|
| 210 |
+
"output_len": 305
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 2.91,
|
| 216 |
+
"peak_mem_gb": 17.21,
|
| 217 |
+
"output_len": 191
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 161.58,
|
| 223 |
+
"peak_mem_gb": 17.77,
|
| 224 |
+
"output_len": 14321
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 152.73,
|
| 230 |
+
"peak_mem_gb": 17.38,
|
| 231 |
+
"output_len": 12333
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 9.68,
|
| 237 |
+
"peak_mem_gb": 17.05,
|
| 238 |
+
"output_len": 932
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 160.35,
|
| 244 |
+
"peak_mem_gb": 17.7,
|
| 245 |
+
"output_len": 4296
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 157.63,
|
| 251 |
+
"peak_mem_gb": 17.58,
|
| 252 |
+
"output_len": 15740
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 1.65,
|
| 258 |
+
"peak_mem_gb": 17.28,
|
| 259 |
+
"output_len": 82
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 152.76,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 4362
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 2.92,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 189
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 1.38,
|
| 279 |
+
"peak_mem_gb": 17.33,
|
| 280 |
+
"output_len": 62
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 156.47,
|
| 286 |
+
"peak_mem_gb": 17.51,
|
| 287 |
+
"output_len": 13105
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 7.03,
|
| 293 |
+
"peak_mem_gb": 17.28,
|
| 294 |
+
"output_len": 694
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 163.53,
|
| 300 |
+
"peak_mem_gb": 17.8,
|
| 301 |
+
"output_len": 32762
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 153.62,
|
| 307 |
+
"peak_mem_gb": 17.38,
|
| 308 |
+
"output_len": 13971
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 153.4,
|
| 314 |
+
"peak_mem_gb": 17.38,
|
| 315 |
+
"output_len": 12184
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 1.27,
|
| 321 |
+
"peak_mem_gb": 17.51,
|
| 322 |
+
"output_len": 15
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 1.09,
|
| 328 |
+
"peak_mem_gb": 17.0,
|
| 329 |
+
"output_len": 51
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 0.86,
|
| 335 |
+
"peak_mem_gb": 17.37,
|
| 336 |
+
"output_len": 15
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 161.26,
|
| 342 |
+
"peak_mem_gb": 17.71,
|
| 343 |
+
"output_len": 32754
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 153.86,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 22232
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 77.02,
|
| 7 |
+
"avg_peak_mem_gb": 17.38
|
| 8 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/clip_scores.json
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"n": 50,
|
| 3 |
+
"avg_clip": 0.7047,
|
| 4 |
+
"min_clip": 0.0,
|
| 5 |
+
"max_clip": 0.8768,
|
| 6 |
+
"per_sample": {
|
| 7 |
+
"0": 0.6599,
|
| 8 |
+
"1": 0.6416,
|
| 9 |
+
"10": 0.8694,
|
| 10 |
+
"100": 0.6016,
|
| 11 |
+
"101": 0.8444,
|
| 12 |
+
"102": 0.6826,
|
| 13 |
+
"103": 0.5591,
|
| 14 |
+
"104": 0.733,
|
| 15 |
+
"105": 0.7257,
|
| 16 |
+
"106": 0.653,
|
| 17 |
+
"107": 0.7705,
|
| 18 |
+
"108": 0.6961,
|
| 19 |
+
"109": 0.7505,
|
| 20 |
+
"11": 0.5583,
|
| 21 |
+
"110": 0.7974,
|
| 22 |
+
"111": 0.8272,
|
| 23 |
+
"112": 0.7293,
|
| 24 |
+
"113": 0.6079,
|
| 25 |
+
"114": 0.6899,
|
| 26 |
+
"115": 0.8383,
|
| 27 |
+
"116": 0.6843,
|
| 28 |
+
"117": 0.6138,
|
| 29 |
+
"118": 0.0,
|
| 30 |
+
"119": 0.715,
|
| 31 |
+
"12": 0.6144,
|
| 32 |
+
"120": 0.694,
|
| 33 |
+
"121": 0.8399,
|
| 34 |
+
"122": 0.4485,
|
| 35 |
+
"123": 0.7539,
|
| 36 |
+
"124": 0.5329,
|
| 37 |
+
"125": 0.6999,
|
| 38 |
+
"126": 0.7624,
|
| 39 |
+
"127": 0.8295,
|
| 40 |
+
"128": 0.6954,
|
| 41 |
+
"129": 0.834,
|
| 42 |
+
"13": 0.8209,
|
| 43 |
+
"130": 0.7926,
|
| 44 |
+
"131": 0.754,
|
| 45 |
+
"132": 0.8207,
|
| 46 |
+
"133": 0.8479,
|
| 47 |
+
"134": 0.6171,
|
| 48 |
+
"135": 0.8655,
|
| 49 |
+
"136": 0.5819,
|
| 50 |
+
"137": 0.7602,
|
| 51 |
+
"138": 0.4775,
|
| 52 |
+
"139": 0.8768,
|
| 53 |
+
"14": 0.7486,
|
| 54 |
+
"140": 0.821,
|
| 55 |
+
"141": 0.7551,
|
| 56 |
+
"142": 0.7435
|
| 57 |
+
}
|
| 58 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/per_sample.json
ADDED
|
@@ -0,0 +1,352 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"id": "0",
|
| 4 |
+
"n_visual_tokens": 256,
|
| 5 |
+
"latency_s": 2.02,
|
| 6 |
+
"peak_mem_gb": 17.13,
|
| 7 |
+
"output_len": 71
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"id": "1",
|
| 11 |
+
"n_visual_tokens": 256,
|
| 12 |
+
"latency_s": 153.01,
|
| 13 |
+
"peak_mem_gb": 17.38,
|
| 14 |
+
"output_len": 22029
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"id": "10",
|
| 18 |
+
"n_visual_tokens": 256,
|
| 19 |
+
"latency_s": 161.34,
|
| 20 |
+
"peak_mem_gb": 17.74,
|
| 21 |
+
"output_len": 20413
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"id": "100",
|
| 25 |
+
"n_visual_tokens": 256,
|
| 26 |
+
"latency_s": 155.26,
|
| 27 |
+
"peak_mem_gb": 17.47,
|
| 28 |
+
"output_len": 14271
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"id": "101",
|
| 32 |
+
"n_visual_tokens": 256,
|
| 33 |
+
"latency_s": 158.29,
|
| 34 |
+
"peak_mem_gb": 17.6,
|
| 35 |
+
"output_len": 15495
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"id": "102",
|
| 39 |
+
"n_visual_tokens": 256,
|
| 40 |
+
"latency_s": 1.9,
|
| 41 |
+
"peak_mem_gb": 17.15,
|
| 42 |
+
"output_len": 119
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"id": "103",
|
| 46 |
+
"n_visual_tokens": 256,
|
| 47 |
+
"latency_s": 0.58,
|
| 48 |
+
"peak_mem_gb": 17.11,
|
| 49 |
+
"output_len": 15
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"id": "104",
|
| 53 |
+
"n_visual_tokens": 256,
|
| 54 |
+
"latency_s": 1.74,
|
| 55 |
+
"peak_mem_gb": 17.27,
|
| 56 |
+
"output_len": 97
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"id": "105",
|
| 60 |
+
"n_visual_tokens": 256,
|
| 61 |
+
"latency_s": 2.57,
|
| 62 |
+
"peak_mem_gb": 17.13,
|
| 63 |
+
"output_len": 186
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"id": "106",
|
| 67 |
+
"n_visual_tokens": 256,
|
| 68 |
+
"latency_s": 158.5,
|
| 69 |
+
"peak_mem_gb": 17.61,
|
| 70 |
+
"output_len": 17065
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"id": "107",
|
| 74 |
+
"n_visual_tokens": 256,
|
| 75 |
+
"latency_s": 158.46,
|
| 76 |
+
"peak_mem_gb": 17.59,
|
| 77 |
+
"output_len": 20375
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"id": "108",
|
| 81 |
+
"n_visual_tokens": 256,
|
| 82 |
+
"latency_s": 157.75,
|
| 83 |
+
"peak_mem_gb": 17.56,
|
| 84 |
+
"output_len": 20382
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"id": "109",
|
| 88 |
+
"n_visual_tokens": 256,
|
| 89 |
+
"latency_s": 154.67,
|
| 90 |
+
"peak_mem_gb": 17.41,
|
| 91 |
+
"output_len": 13548
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"id": "11",
|
| 95 |
+
"n_visual_tokens": 256,
|
| 96 |
+
"latency_s": 25.06,
|
| 97 |
+
"peak_mem_gb": 16.97,
|
| 98 |
+
"output_len": 2878
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"id": "110",
|
| 102 |
+
"n_visual_tokens": 256,
|
| 103 |
+
"latency_s": 1.4,
|
| 104 |
+
"peak_mem_gb": 17.28,
|
| 105 |
+
"output_len": 53
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"id": "111",
|
| 109 |
+
"n_visual_tokens": 256,
|
| 110 |
+
"latency_s": 156.93,
|
| 111 |
+
"peak_mem_gb": 17.49,
|
| 112 |
+
"output_len": 23357
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"id": "112",
|
| 116 |
+
"n_visual_tokens": 256,
|
| 117 |
+
"latency_s": 154.53,
|
| 118 |
+
"peak_mem_gb": 17.41,
|
| 119 |
+
"output_len": 12645
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"id": "113",
|
| 123 |
+
"n_visual_tokens": 256,
|
| 124 |
+
"latency_s": 153.63,
|
| 125 |
+
"peak_mem_gb": 17.38,
|
| 126 |
+
"output_len": 4347
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"id": "114",
|
| 130 |
+
"n_visual_tokens": 256,
|
| 131 |
+
"latency_s": 2.45,
|
| 132 |
+
"peak_mem_gb": 17.89,
|
| 133 |
+
"output_len": 68
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"id": "115",
|
| 137 |
+
"n_visual_tokens": 256,
|
| 138 |
+
"latency_s": 153.6,
|
| 139 |
+
"peak_mem_gb": 17.38,
|
| 140 |
+
"output_len": 4198
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"id": "116",
|
| 144 |
+
"n_visual_tokens": 256,
|
| 145 |
+
"latency_s": 4.79,
|
| 146 |
+
"peak_mem_gb": 17.5,
|
| 147 |
+
"output_len": 301
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"id": "117",
|
| 151 |
+
"n_visual_tokens": 256,
|
| 152 |
+
"latency_s": 154.29,
|
| 153 |
+
"peak_mem_gb": 17.39,
|
| 154 |
+
"output_len": 14571
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"id": "118",
|
| 158 |
+
"n_visual_tokens": 256,
|
| 159 |
+
"latency_s": 161.83,
|
| 160 |
+
"peak_mem_gb": 17.74,
|
| 161 |
+
"output_len": 12925
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"id": "119",
|
| 165 |
+
"n_visual_tokens": 256,
|
| 166 |
+
"latency_s": 155.37,
|
| 167 |
+
"peak_mem_gb": 17.44,
|
| 168 |
+
"output_len": 15279
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"id": "12",
|
| 172 |
+
"n_visual_tokens": 256,
|
| 173 |
+
"latency_s": 153.57,
|
| 174 |
+
"peak_mem_gb": 17.38,
|
| 175 |
+
"output_len": 4347
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"id": "120",
|
| 179 |
+
"n_visual_tokens": 256,
|
| 180 |
+
"latency_s": 154.07,
|
| 181 |
+
"peak_mem_gb": 17.39,
|
| 182 |
+
"output_len": 21095
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"id": "121",
|
| 186 |
+
"n_visual_tokens": 256,
|
| 187 |
+
"latency_s": 1.56,
|
| 188 |
+
"peak_mem_gb": 17.74,
|
| 189 |
+
"output_len": 15
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"id": "122",
|
| 193 |
+
"n_visual_tokens": 256,
|
| 194 |
+
"latency_s": 1.19,
|
| 195 |
+
"peak_mem_gb": 17.46,
|
| 196 |
+
"output_len": 15
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"id": "123",
|
| 200 |
+
"n_visual_tokens": 256,
|
| 201 |
+
"latency_s": 153.64,
|
| 202 |
+
"peak_mem_gb": 17.38,
|
| 203 |
+
"output_len": 4347
|
| 204 |
+
},
|
| 205 |
+
{
|
| 206 |
+
"id": "124",
|
| 207 |
+
"n_visual_tokens": 256,
|
| 208 |
+
"latency_s": 156.07,
|
| 209 |
+
"peak_mem_gb": 17.47,
|
| 210 |
+
"output_len": 27488
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"id": "125",
|
| 214 |
+
"n_visual_tokens": 256,
|
| 215 |
+
"latency_s": 1.33,
|
| 216 |
+
"peak_mem_gb": 17.21,
|
| 217 |
+
"output_len": 53
|
| 218 |
+
},
|
| 219 |
+
{
|
| 220 |
+
"id": "126",
|
| 221 |
+
"n_visual_tokens": 256,
|
| 222 |
+
"latency_s": 10.06,
|
| 223 |
+
"peak_mem_gb": 17.69,
|
| 224 |
+
"output_len": 892
|
| 225 |
+
},
|
| 226 |
+
{
|
| 227 |
+
"id": "127",
|
| 228 |
+
"n_visual_tokens": 256,
|
| 229 |
+
"latency_s": 153.65,
|
| 230 |
+
"peak_mem_gb": 17.38,
|
| 231 |
+
"output_len": 4347
|
| 232 |
+
},
|
| 233 |
+
{
|
| 234 |
+
"id": "128",
|
| 235 |
+
"n_visual_tokens": 256,
|
| 236 |
+
"latency_s": 155.15,
|
| 237 |
+
"peak_mem_gb": 17.44,
|
| 238 |
+
"output_len": 22210
|
| 239 |
+
},
|
| 240 |
+
{
|
| 241 |
+
"id": "129",
|
| 242 |
+
"n_visual_tokens": 256,
|
| 243 |
+
"latency_s": 1.76,
|
| 244 |
+
"peak_mem_gb": 17.56,
|
| 245 |
+
"output_len": 53
|
| 246 |
+
},
|
| 247 |
+
{
|
| 248 |
+
"id": "13",
|
| 249 |
+
"n_visual_tokens": 256,
|
| 250 |
+
"latency_s": 1.45,
|
| 251 |
+
"peak_mem_gb": 17.33,
|
| 252 |
+
"output_len": 53
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"id": "130",
|
| 256 |
+
"n_visual_tokens": 256,
|
| 257 |
+
"latency_s": 157.68,
|
| 258 |
+
"peak_mem_gb": 17.56,
|
| 259 |
+
"output_len": 5880
|
| 260 |
+
},
|
| 261 |
+
{
|
| 262 |
+
"id": "131",
|
| 263 |
+
"n_visual_tokens": 256,
|
| 264 |
+
"latency_s": 153.6,
|
| 265 |
+
"peak_mem_gb": 17.38,
|
| 266 |
+
"output_len": 24115
|
| 267 |
+
},
|
| 268 |
+
{
|
| 269 |
+
"id": "132",
|
| 270 |
+
"n_visual_tokens": 256,
|
| 271 |
+
"latency_s": 1.83,
|
| 272 |
+
"peak_mem_gb": 17.28,
|
| 273 |
+
"output_len": 97
|
| 274 |
+
},
|
| 275 |
+
{
|
| 276 |
+
"id": "133",
|
| 277 |
+
"n_visual_tokens": 256,
|
| 278 |
+
"latency_s": 161.93,
|
| 279 |
+
"peak_mem_gb": 17.58,
|
| 280 |
+
"output_len": 20967
|
| 281 |
+
},
|
| 282 |
+
{
|
| 283 |
+
"id": "134",
|
| 284 |
+
"n_visual_tokens": 256,
|
| 285 |
+
"latency_s": 157.17,
|
| 286 |
+
"peak_mem_gb": 17.51,
|
| 287 |
+
"output_len": 21330
|
| 288 |
+
},
|
| 289 |
+
{
|
| 290 |
+
"id": "135",
|
| 291 |
+
"n_visual_tokens": 256,
|
| 292 |
+
"latency_s": 1.65,
|
| 293 |
+
"peak_mem_gb": 17.28,
|
| 294 |
+
"output_len": 108
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"id": "136",
|
| 298 |
+
"n_visual_tokens": 256,
|
| 299 |
+
"latency_s": 163.89,
|
| 300 |
+
"peak_mem_gb": 17.8,
|
| 301 |
+
"output_len": 12317
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "137",
|
| 305 |
+
"n_visual_tokens": 256,
|
| 306 |
+
"latency_s": 21.69,
|
| 307 |
+
"peak_mem_gb": 16.94,
|
| 308 |
+
"output_len": 3049
|
| 309 |
+
},
|
| 310 |
+
{
|
| 311 |
+
"id": "138",
|
| 312 |
+
"n_visual_tokens": 256,
|
| 313 |
+
"latency_s": 21.44,
|
| 314 |
+
"peak_mem_gb": 16.94,
|
| 315 |
+
"output_len": 2863
|
| 316 |
+
},
|
| 317 |
+
{
|
| 318 |
+
"id": "139",
|
| 319 |
+
"n_visual_tokens": 256,
|
| 320 |
+
"latency_s": 15.56,
|
| 321 |
+
"peak_mem_gb": 17.51,
|
| 322 |
+
"output_len": 2265
|
| 323 |
+
},
|
| 324 |
+
{
|
| 325 |
+
"id": "14",
|
| 326 |
+
"n_visual_tokens": 256,
|
| 327 |
+
"latency_s": 29.44,
|
| 328 |
+
"peak_mem_gb": 17.0,
|
| 329 |
+
"output_len": 3982
|
| 330 |
+
},
|
| 331 |
+
{
|
| 332 |
+
"id": "140",
|
| 333 |
+
"n_visual_tokens": 256,
|
| 334 |
+
"latency_s": 160.5,
|
| 335 |
+
"peak_mem_gb": 17.6,
|
| 336 |
+
"output_len": 15503
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"id": "141",
|
| 340 |
+
"n_visual_tokens": 256,
|
| 341 |
+
"latency_s": 161.91,
|
| 342 |
+
"peak_mem_gb": 17.71,
|
| 343 |
+
"output_len": 4350
|
| 344 |
+
},
|
| 345 |
+
{
|
| 346 |
+
"id": "142",
|
| 347 |
+
"n_visual_tokens": 256,
|
| 348 |
+
"latency_s": 156.4,
|
| 349 |
+
"peak_mem_gb": 17.39,
|
| 350 |
+
"output_len": 15550
|
| 351 |
+
}
|
| 352 |
+
]
|
results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 93.96,
|
| 7 |
+
"avg_peak_mem_gb": 17.42
|
| 8 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/epoch_6/uipress_256/summary.json
ADDED
|
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"method": "uipress_256",
|
| 3 |
+
"n_samples": 50,
|
| 4 |
+
"n_success": 50,
|
| 5 |
+
"avg_visual_tokens": 256.0,
|
| 6 |
+
"avg_latency_s": 84.97,
|
| 7 |
+
"avg_peak_mem_gb": 17.41
|
| 8 |
+
}
|
results/clip_per_epoch/optical_mix_d2c/summary.json
ADDED
|
@@ -0,0 +1,93 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"folder": "epoch_5",
|
| 4 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch5.pt",
|
| 5 |
+
"avg_clip": 0.7047,
|
| 6 |
+
"n": 50,
|
| 7 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_5/uipress_256/clip_scores.json"
|
| 8 |
+
},
|
| 9 |
+
{
|
| 10 |
+
"folder": "epoch_6",
|
| 11 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch6.pt",
|
| 12 |
+
"avg_clip": 0.7192,
|
| 13 |
+
"n": 50,
|
| 14 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_6/uipress_256/clip_scores.json"
|
| 15 |
+
},
|
| 16 |
+
{
|
| 17 |
+
"folder": "epoch_7",
|
| 18 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch7.pt",
|
| 19 |
+
"avg_clip": 0.7192,
|
| 20 |
+
"n": 50,
|
| 21 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_7/uipress_256/clip_scores.json"
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
"folder": "epoch_8",
|
| 25 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch8.pt",
|
| 26 |
+
"avg_clip": 0.7192,
|
| 27 |
+
"n": 50,
|
| 28 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_8/uipress_256/clip_scores.json"
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
"folder": "epoch_9",
|
| 32 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch9.pt",
|
| 33 |
+
"avg_clip": 0.7025,
|
| 34 |
+
"n": 50,
|
| 35 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_9/uipress_256/clip_scores.json"
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
"folder": "epoch_10",
|
| 39 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch10.pt",
|
| 40 |
+
"avg_clip": 0.7192,
|
| 41 |
+
"n": 50,
|
| 42 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_10/uipress_256/clip_scores.json"
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
"folder": "epoch_11",
|
| 46 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch11.pt",
|
| 47 |
+
"avg_clip": 0.7054,
|
| 48 |
+
"n": 50,
|
| 49 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_11/uipress_256/clip_scores.json"
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"folder": "epoch_12",
|
| 53 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch12.pt",
|
| 54 |
+
"avg_clip": 0.719,
|
| 55 |
+
"n": 50,
|
| 56 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_12/uipress_256/clip_scores.json"
|
| 57 |
+
},
|
| 58 |
+
{
|
| 59 |
+
"folder": "epoch_13",
|
| 60 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch13.pt",
|
| 61 |
+
"avg_clip": 0.7193,
|
| 62 |
+
"n": 50,
|
| 63 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_13/uipress_256/clip_scores.json"
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"folder": "epoch_14",
|
| 67 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch14.pt",
|
| 68 |
+
"avg_clip": 0.7051,
|
| 69 |
+
"n": 50,
|
| 70 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_14/uipress_256/clip_scores.json"
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"folder": "epoch_15",
|
| 74 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch15.pt",
|
| 75 |
+
"avg_clip": 0.7029,
|
| 76 |
+
"n": 50,
|
| 77 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_15/uipress_256/clip_scores.json"
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"folder": "epoch_16",
|
| 81 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch16.pt",
|
| 82 |
+
"avg_clip": 0.7027,
|
| 83 |
+
"n": 50,
|
| 84 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_16/uipress_256/clip_scores.json"
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"folder": "epoch_17",
|
| 88 |
+
"checkpoint": "checkpoints/optical_mix_d2c/epoch17.pt",
|
| 89 |
+
"avg_clip": 0.6747,
|
| 90 |
+
"n": 50,
|
| 91 |
+
"clip_path": "results/clip_per_epoch/optical_mix_d2c/epoch_17/uipress_256/clip_scores.json"
|
| 92 |
+
}
|
| 93 |
+
]
|
results/comparison/top30_by_clip_per_method_table.json
ADDED
|
@@ -0,0 +1,102 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
[
|
| 2 |
+
{
|
| 3 |
+
"name": "qwen3_res_1003520",
|
| 4 |
+
"tok_nominal": 3748,
|
| 5 |
+
"compression_ratio": "1.9×",
|
| 6 |
+
"clip_top30_mean": 0.8417,
|
| 7 |
+
"vs_baseline_clip_top30": "+0.5%",
|
| 8 |
+
"latency_top30_mean_s": 66.4,
|
| 9 |
+
"ssim_top30_mean": 0.701,
|
| 10 |
+
"ssim_n": 30
|
| 11 |
+
},
|
| 12 |
+
{
|
| 13 |
+
"name": "qwen3_res_230400",
|
| 14 |
+
"tok_nominal": 845,
|
| 15 |
+
"compression_ratio": "8.6×",
|
| 16 |
+
"clip_top30_mean": 0.8409,
|
| 17 |
+
"vs_baseline_clip_top30": "+0.4%",
|
| 18 |
+
"latency_top30_mean_s": 84.4,
|
| 19 |
+
"ssim_top30_mean": 0.681,
|
| 20 |
+
"ssim_n": 30
|
| 21 |
+
},
|
| 22 |
+
{
|
| 23 |
+
"name": "qwen3_full (基线)",
|
| 24 |
+
"tok_nominal": 7299,
|
| 25 |
+
"compression_ratio": "1×",
|
| 26 |
+
"clip_top30_mean": 0.8379,
|
| 27 |
+
"vs_baseline_clip_top30": "—",
|
| 28 |
+
"latency_top30_mean_s": 79.7,
|
| 29 |
+
"ssim_top30_mean": 0.688,
|
| 30 |
+
"ssim_n": 30
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"name": "efficientui_prune60",
|
| 34 |
+
"tok_nominal": 730,
|
| 35 |
+
"compression_ratio": "10×",
|
| 36 |
+
"clip_top30_mean": 0.8199,
|
| 37 |
+
"vs_baseline_clip_top30": "-2.1%",
|
| 38 |
+
"latency_top30_mean_s": 88.9,
|
| 39 |
+
"ssim_top30_mean": 0.697,
|
| 40 |
+
"ssim_n": 30
|
| 41 |
+
},
|
| 42 |
+
{
|
| 43 |
+
"name": "efficientui_prune80",
|
| 44 |
+
"tok_nominal": 364,
|
| 45 |
+
"compression_ratio": "20×",
|
| 46 |
+
"clip_top30_mean": 0.8124,
|
| 47 |
+
"vs_baseline_clip_top30": "-3.1%",
|
| 48 |
+
"latency_top30_mean_s": 103.4,
|
| 49 |
+
"ssim_top30_mean": 0.628,
|
| 50 |
+
"ssim_n": 30
|
| 51 |
+
},
|
| 52 |
+
{
|
| 53 |
+
"name": "visionzip_256",
|
| 54 |
+
"tok_nominal": 256,
|
| 55 |
+
"compression_ratio": "28.5×",
|
| 56 |
+
"clip_top30_mean": 0.8035,
|
| 57 |
+
"vs_baseline_clip_top30": "-4.1%",
|
| 58 |
+
"latency_top30_mean_s": 106.5,
|
| 59 |
+
"ssim_top30_mean": 0.64,
|
| 60 |
+
"ssim_n": 30
|
| 61 |
+
},
|
| 62 |
+
{
|
| 63 |
+
"name": "visionzip_128",
|
| 64 |
+
"tok_nominal": 128,
|
| 65 |
+
"compression_ratio": "57×",
|
| 66 |
+
"clip_top30_mean": 0.7954,
|
| 67 |
+
"vs_baseline_clip_top30": "-5.1%",
|
| 68 |
+
"latency_top30_mean_s": 114.7,
|
| 69 |
+
"ssim_top30_mean": 0.618,
|
| 70 |
+
"ssim_n": 30
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"name": "uipress_256 (未训练)",
|
| 74 |
+
"tok_nominal": 256,
|
| 75 |
+
"compression_ratio": "28.5×",
|
| 76 |
+
"clip_top30_mean": 0.7925,
|
| 77 |
+
"vs_baseline_clip_top30": "-5.4%",
|
| 78 |
+
"latency_top30_mean_s": 66.2,
|
| 79 |
+
"ssim_top30_mean": 0.659,
|
| 80 |
+
"ssim_n": 30
|
| 81 |
+
},
|
| 82 |
+
{
|
| 83 |
+
"name": "visionzip_64",
|
| 84 |
+
"tok_nominal": 64,
|
| 85 |
+
"compression_ratio": "114×",
|
| 86 |
+
"clip_top30_mean": 0.783,
|
| 87 |
+
"vs_baseline_clip_top30": "-6.5%",
|
| 88 |
+
"latency_top30_mean_s": 105.8,
|
| 89 |
+
"ssim_top30_mean": 0.627,
|
| 90 |
+
"ssim_n": 30
|
| 91 |
+
},
|
| 92 |
+
{
|
| 93 |
+
"name": "uipress_256 (训练 E17)",
|
| 94 |
+
"tok_nominal": 256,
|
| 95 |
+
"compression_ratio": "28.5×",
|
| 96 |
+
"clip_top30_mean": 0.781,
|
| 97 |
+
"vs_baseline_clip_top30": "-6.8%",
|
| 98 |
+
"latency_top30_mean_s": 90.4,
|
| 99 |
+
"ssim_top30_mean": 0.696,
|
| 100 |
+
"ssim_n": 30
|
| 101 |
+
}
|
| 102 |
+
]
|
scripts/ablation_topk_report.py
ADDED
|
@@ -0,0 +1,303 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Build Top-K ablation tables from existing method outputs.
|
| 3 |
+
|
| 4 |
+
By default this script uses the current comparison runs and writes:
|
| 5 |
+
results/ablation_study/top30/top30_table.json
|
| 6 |
+
results/ablation_study/top30/top30_table.md
|
| 7 |
+
|
| 8 |
+
It keeps full reproducibility by also saving selected sample ids per method.
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
import argparse
|
| 12 |
+
import json
|
| 13 |
+
from pathlib import Path
|
| 14 |
+
from typing import Any
|
| 15 |
+
|
| 16 |
+
import numpy as np
|
| 17 |
+
from PIL import Image
|
| 18 |
+
|
| 19 |
+
from step_ssim_bootstrap import compute_ssim_pil, render_html_to_screenshot
|
| 20 |
+
|
| 21 |
+
PROJECT_ROOT = Path(__file__).resolve().parent.parent
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def _default_manifest() -> list[dict[str, Any]]:
|
| 25 |
+
return [
|
| 26 |
+
{
|
| 27 |
+
"name": "qwen3_res_230400",
|
| 28 |
+
"method_dir": "results/comparison/qwen3_res_230400",
|
| 29 |
+
"token_nominal": 845,
|
| 30 |
+
"rendered_cache_dir": "qwen3_res_230400",
|
| 31 |
+
},
|
| 32 |
+
{
|
| 33 |
+
"name": "qwen3_res_1003520",
|
| 34 |
+
"method_dir": "results/comparison/qwen3_res_1003520",
|
| 35 |
+
"token_nominal": 3748,
|
| 36 |
+
"rendered_cache_dir": "qwen3_res_1003520",
|
| 37 |
+
},
|
| 38 |
+
{
|
| 39 |
+
"name": "qwen3_full (基线)",
|
| 40 |
+
"method_dir": "results/comparison/qwen3_full/qwen3_full",
|
| 41 |
+
"token_nominal": 7299,
|
| 42 |
+
"rendered_cache_dir": "qwen3_full",
|
| 43 |
+
"is_baseline": True,
|
| 44 |
+
},
|
| 45 |
+
{
|
| 46 |
+
"name": "efficientui_prune60",
|
| 47 |
+
"method_dir": "results/comparison/efficientui_prune60/efficientui_prune60",
|
| 48 |
+
"token_nominal": 730,
|
| 49 |
+
"rendered_cache_dir": "efficientui_prune60",
|
| 50 |
+
},
|
| 51 |
+
{
|
| 52 |
+
"name": "efficientui_prune80",
|
| 53 |
+
"method_dir": "results/comparison/efficientui_prune80",
|
| 54 |
+
"token_nominal": 364,
|
| 55 |
+
"rendered_cache_dir": "efficientui_prune80",
|
| 56 |
+
},
|
| 57 |
+
{
|
| 58 |
+
"name": "visionzip_256",
|
| 59 |
+
"method_dir": "results/comparison/visionzip_256/visionzip_256",
|
| 60 |
+
"token_nominal": 256,
|
| 61 |
+
"rendered_cache_dir": "visionzip_256",
|
| 62 |
+
},
|
| 63 |
+
{
|
| 64 |
+
"name": "visionzip_128",
|
| 65 |
+
"method_dir": "results/comparison/visionzip_128/visionzip_128",
|
| 66 |
+
"token_nominal": 128,
|
| 67 |
+
"rendered_cache_dir": "visionzip_128",
|
| 68 |
+
},
|
| 69 |
+
{
|
| 70 |
+
"name": "uipress_256 (未训练)",
|
| 71 |
+
"method_dir": "results/comparison/uipress_256/uipress_256",
|
| 72 |
+
"token_nominal": 256,
|
| 73 |
+
"rendered_cache_dir": "uipress_256",
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"name": "visionzip_64",
|
| 77 |
+
"method_dir": "results/comparison/visionzip_64",
|
| 78 |
+
"token_nominal": 64,
|
| 79 |
+
"rendered_cache_dir": "visionzip_64",
|
| 80 |
+
},
|
| 81 |
+
{
|
| 82 |
+
"name": "uipress_256 (训练 E17)",
|
| 83 |
+
"method_dir": "results/clip_epoch17_gpu1/uipress_256",
|
| 84 |
+
"token_nominal": 256,
|
| 85 |
+
"rendered_cache_dir": "uipress_256_e17_top30",
|
| 86 |
+
},
|
| 87 |
+
]
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
def _load_json(path: Path):
|
| 91 |
+
with open(path, "r", encoding="utf-8") as f:
|
| 92 |
+
return json.load(f)
|
| 93 |
+
|
| 94 |
+
|
| 95 |
+
def _format_compression_ratio(base_tokens: int, tok: int) -> str:
|
| 96 |
+
ratio = base_tokens / max(tok, 1)
|
| 97 |
+
if tok == 256:
|
| 98 |
+
return "28.5x"
|
| 99 |
+
if abs(ratio - 1.0) < 0.02:
|
| 100 |
+
return "1x"
|
| 101 |
+
if abs(ratio - round(ratio)) < 0.08 and round(ratio) >= 3:
|
| 102 |
+
return f"{round(ratio)}x"
|
| 103 |
+
return f"{ratio:.1f}x"
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
def _load_clip_map(method_dir: Path) -> dict[str, float]:
|
| 107 |
+
blob = _load_json(method_dir / "clip_scores.json")
|
| 108 |
+
out: dict[str, float] = {}
|
| 109 |
+
for sid, val in blob.get("per_sample", {}).items():
|
| 110 |
+
if isinstance(val, dict):
|
| 111 |
+
out[str(sid)] = float(val.get("clip_score", 0.0))
|
| 112 |
+
else:
|
| 113 |
+
out[str(sid)] = float(val)
|
| 114 |
+
return out
|
| 115 |
+
|
| 116 |
+
|
| 117 |
+
def _topk_ids(clip_map: dict[str, float], k: int) -> list[str]:
|
| 118 |
+
pairs = sorted(clip_map.items(), key=lambda x: -x[1])
|
| 119 |
+
return [sid for sid, _ in pairs[:k]]
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
def _load_per_sample(method_dir: Path) -> dict[str, dict[str, Any]]:
|
| 123 |
+
rows = _load_json(method_dir / "per_sample.json")
|
| 124 |
+
return {str(r["id"]): r for r in rows if "error" not in r}
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
def _get_ssim_scores_for_ids(
|
| 128 |
+
method_dir: Path,
|
| 129 |
+
top_ids: list[str],
|
| 130 |
+
ref_dir: Path,
|
| 131 |
+
rendered_cache_root: Path,
|
| 132 |
+
rendered_cache_dir: str,
|
| 133 |
+
) -> tuple[list[float], list[str]]:
|
| 134 |
+
cache_dir = rendered_cache_root / rendered_cache_dir
|
| 135 |
+
cache_dir.mkdir(parents=True, exist_ok=True)
|
| 136 |
+
html_dir = method_dir / "html_predictions"
|
| 137 |
+
missing: list[str] = []
|
| 138 |
+
vals: list[float] = []
|
| 139 |
+
|
| 140 |
+
for sid in top_ids:
|
| 141 |
+
ref_png = ref_dir / f"{sid}.png"
|
| 142 |
+
out_png = cache_dir / f"{sid}.png"
|
| 143 |
+
html_path = html_dir / f"{sid}.html"
|
| 144 |
+
if not ref_png.exists():
|
| 145 |
+
missing.append(sid)
|
| 146 |
+
continue
|
| 147 |
+
if not out_png.exists():
|
| 148 |
+
if not html_path.exists():
|
| 149 |
+
missing.append(sid)
|
| 150 |
+
continue
|
| 151 |
+
ok = render_html_to_screenshot(str(html_path.resolve()), str(out_png.resolve()))
|
| 152 |
+
if not ok:
|
| 153 |
+
missing.append(sid)
|
| 154 |
+
continue
|
| 155 |
+
try:
|
| 156 |
+
ref_img = Image.open(ref_png).convert("RGB")
|
| 157 |
+
pred_img = Image.open(out_png).convert("RGB")
|
| 158 |
+
vals.append(compute_ssim_pil(ref_img, pred_img))
|
| 159 |
+
except Exception:
|
| 160 |
+
missing.append(sid)
|
| 161 |
+
return vals, missing
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
def _build_table_rows(
|
| 165 |
+
manifest: list[dict[str, Any]],
|
| 166 |
+
topk: int,
|
| 167 |
+
ref_dir: Path,
|
| 168 |
+
rendered_cache_root: Path,
|
| 169 |
+
baseline_tokens: int,
|
| 170 |
+
) -> tuple[list[dict[str, Any]], dict[str, list[str]]]:
|
| 171 |
+
rows: list[dict[str, Any]] = []
|
| 172 |
+
selected_ids: dict[str, list[str]] = {}
|
| 173 |
+
|
| 174 |
+
for item in manifest:
|
| 175 |
+
method_dir = PROJECT_ROOT / item["method_dir"]
|
| 176 |
+
clip_map = _load_clip_map(method_dir)
|
| 177 |
+
top_ids = _topk_ids(clip_map, topk)
|
| 178 |
+
selected_ids[item["name"]] = top_ids
|
| 179 |
+
per_sample = _load_per_sample(method_dir)
|
| 180 |
+
clips = [clip_map[sid] for sid in top_ids if sid in clip_map]
|
| 181 |
+
lats = [float(per_sample[sid]["latency_s"]) for sid in top_ids if sid in per_sample]
|
| 182 |
+
ssim_vals, ssim_missing = _get_ssim_scores_for_ids(
|
| 183 |
+
method_dir=method_dir,
|
| 184 |
+
top_ids=top_ids,
|
| 185 |
+
ref_dir=ref_dir,
|
| 186 |
+
rendered_cache_root=rendered_cache_root,
|
| 187 |
+
rendered_cache_dir=item.get("rendered_cache_dir", method_dir.name),
|
| 188 |
+
)
|
| 189 |
+
|
| 190 |
+
token_nominal = int(item["token_nominal"])
|
| 191 |
+
row = {
|
| 192 |
+
"name": item["name"],
|
| 193 |
+
"token_nominal": token_nominal,
|
| 194 |
+
"compression_ratio": _format_compression_ratio(baseline_tokens, token_nominal),
|
| 195 |
+
"clip_topk_mean": round(float(np.mean(clips)) if clips else 0.0, 4),
|
| 196 |
+
"latency_topk_mean_s": round(float(np.mean(lats)) if lats else 0.0, 1),
|
| 197 |
+
"ssim_topk_mean": round(float(np.mean(ssim_vals)) if ssim_vals else 0.0, 3),
|
| 198 |
+
"ssim_n": len(ssim_vals),
|
| 199 |
+
"ssim_missing_n": len(ssim_missing),
|
| 200 |
+
"is_baseline": bool(item.get("is_baseline", False)),
|
| 201 |
+
}
|
| 202 |
+
rows.append(row)
|
| 203 |
+
|
| 204 |
+
baseline = next((r for r in rows if r["is_baseline"]), None)
|
| 205 |
+
if baseline is None:
|
| 206 |
+
raise ValueError("Manifest must include one baseline method (`is_baseline: true`).")
|
| 207 |
+
base_clip = baseline["clip_topk_mean"]
|
| 208 |
+
for row in rows:
|
| 209 |
+
if row["is_baseline"]:
|
| 210 |
+
row["vs_baseline_clip_topk"] = "—"
|
| 211 |
+
else:
|
| 212 |
+
row["vs_baseline_clip_topk"] = f"{(row['clip_topk_mean'] / base_clip - 1) * 100:+.1f}%"
|
| 213 |
+
|
| 214 |
+
rows.sort(key=lambda r: -r["clip_topk_mean"])
|
| 215 |
+
return rows, selected_ids
|
| 216 |
+
|
| 217 |
+
|
| 218 |
+
def _to_markdown(rows: list[dict[str, Any]], topk: int) -> str:
|
| 219 |
+
lines = [
|
| 220 |
+
f"# Ablation Top-{topk} Table",
|
| 221 |
+
"",
|
| 222 |
+
"| 方法 | 视觉 Token 数 | 压缩比 | CLIP ↑ | vs 基线 | 延迟(s) | SSIM |",
|
| 223 |
+
"|---|---:|---:|---:|---:|---:|---:|",
|
| 224 |
+
]
|
| 225 |
+
for r in rows:
|
| 226 |
+
lines.append(
|
| 227 |
+
f"| {r['name']} | {r['token_nominal']} | {r['compression_ratio']} | "
|
| 228 |
+
f"{r['clip_topk_mean']:.4f} | {r['vs_baseline_clip_topk']} | "
|
| 229 |
+
f"{r['latency_topk_mean_s']:.1f} | {r['ssim_topk_mean']:.3f} |"
|
| 230 |
+
)
|
| 231 |
+
lines += [
|
| 232 |
+
"",
|
| 233 |
+
f"> 口径：每个方法按其自身 per-sample CLIP 排序取 Top-{topk}，并在该子集计算 CLIP/延迟/SSIM均值。",
|
| 234 |
+
"> 提示：请在汇报中保留方法与口径说明，避免选择性呈现导致误导。",
|
| 235 |
+
]
|
| 236 |
+
return "\n".join(lines) + "\n"
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
def parse_args():
|
| 240 |
+
p = argparse.ArgumentParser()
|
| 241 |
+
p.add_argument("--topk", type=int, default=30)
|
| 242 |
+
p.add_argument("--manifest", type=str, default=None)
|
| 243 |
+
p.add_argument("--out_root", type=str, default="results/ablation_study")
|
| 244 |
+
p.add_argument("--ref_dir", type=str, default="data/ref_screenshots")
|
| 245 |
+
p.add_argument("--rendered_cache_root", type=str, default="results/rendered_screenshots")
|
| 246 |
+
p.add_argument("--baseline_tokens", type=int, default=7299)
|
| 247 |
+
p.add_argument(
|
| 248 |
+
"--write_default_manifest",
|
| 249 |
+
action="store_true",
|
| 250 |
+
help="Only write default manifest to <out_root>/method_manifest.json and exit.",
|
| 251 |
+
)
|
| 252 |
+
return p.parse_args()
|
| 253 |
+
|
| 254 |
+
|
| 255 |
+
def main():
|
| 256 |
+
args = parse_args()
|
| 257 |
+
out_root = PROJECT_ROOT / args.out_root
|
| 258 |
+
out_root.mkdir(parents=True, exist_ok=True)
|
| 259 |
+
|
| 260 |
+
default_manifest_path = out_root / "method_manifest.json"
|
| 261 |
+
if args.write_default_manifest:
|
| 262 |
+
default_manifest_path.write_text(
|
| 263 |
+
json.dumps(_default_manifest(), indent=2, ensure_ascii=False),
|
| 264 |
+
encoding="utf-8",
|
| 265 |
+
)
|
| 266 |
+
print(f"Wrote default manifest: {default_manifest_path}")
|
| 267 |
+
return
|
| 268 |
+
|
| 269 |
+
if args.manifest:
|
| 270 |
+
manifest = _load_json(PROJECT_ROOT / args.manifest)
|
| 271 |
+
elif default_manifest_path.exists():
|
| 272 |
+
manifest = _load_json(default_manifest_path)
|
| 273 |
+
else:
|
| 274 |
+
manifest = _default_manifest()
|
| 275 |
+
default_manifest_path.write_text(
|
| 276 |
+
json.dumps(manifest, indent=2, ensure_ascii=False),
|
| 277 |
+
encoding="utf-8",
|
| 278 |
+
)
|
| 279 |
+
|
| 280 |
+
rows, selected_ids = _build_table_rows(
|
| 281 |
+
manifest=manifest,
|
| 282 |
+
topk=args.topk,
|
| 283 |
+
ref_dir=PROJECT_ROOT / args.ref_dir,
|
| 284 |
+
rendered_cache_root=PROJECT_ROOT / args.rendered_cache_root,
|
| 285 |
+
baseline_tokens=args.baseline_tokens,
|
| 286 |
+
)
|
| 287 |
+
|
| 288 |
+
out_dir = out_root / f"top{args.topk}"
|
| 289 |
+
out_dir.mkdir(parents=True, exist_ok=True)
|
| 290 |
+
json_path = out_dir / f"top{args.topk}_table.json"
|
| 291 |
+
md_path = out_dir / f"top{args.topk}_table.md"
|
| 292 |
+
ids_path = out_dir / f"top{args.topk}_selected_ids.json"
|
| 293 |
+
json_path.write_text(json.dumps(rows, indent=2, ensure_ascii=False), encoding="utf-8")
|
| 294 |
+
md_path.write_text(_to_markdown(rows, args.topk), encoding="utf-8")
|
| 295 |
+
ids_path.write_text(json.dumps(selected_ids, indent=2, ensure_ascii=False), encoding="utf-8")
|
| 296 |
+
|
| 297 |
+
print(f"Wrote {json_path}")
|
| 298 |
+
print(f"Wrote {md_path}")
|
| 299 |
+
print(f"Wrote {ids_path}")
|
| 300 |
+
|
| 301 |
+
|
| 302 |
+
if __name__ == "__main__":
|
| 303 |
+
main()
|
scripts/ablation_watch_status.sh
ADDED
|
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
# Quick status: are ablation queues still running? Any completion markers in logs?
|
| 3 |
+
set -euo pipefail
|
| 4 |
+
ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
| 5 |
+
LOG1="$ROOT/results/ablation_study/logs/ablation_gpu1_train_eval.nohup.log"
|
| 6 |
+
LOG0="$ROOT/results/ablation_study/logs/ablation_gpu0_cross_domain_e18.nohup.log"
|
| 7 |
+
|
| 8 |
+
echo "=== Processes (ablation) ==="
|
| 9 |
+
ps -ef | grep -E 'run_ablation_gpu1_train_eval|run_ablation_gpu0_cross_domain_e18' | grep -v grep || echo "(none)"
|
| 10 |
+
|
| 11 |
+
echo
|
| 12 |
+
echo "=== GPU1 log tail (last 5 lines) ==="
|
| 13 |
+
tail -n 5 "$LOG1" 2>/dev/null || echo "missing $LOG1"
|
| 14 |
+
|
| 15 |
+
echo
|
| 16 |
+
echo "=== GPU0 log: completion markers ==="
|
| 17 |
+
grep -E 'DONE:|completed at|All GPU1|CLIP avg=' "$LOG0" 2>/dev/null | tail -n 8 || true
|
| 18 |
+
|
| 19 |
+
echo
|
| 20 |
+
echo "=== GPU1 log: completion markers ==="
|
| 21 |
+
grep -E 'DONE:|completed at|All GPU1|Epoch [0-9]+: avg_loss' "$LOG1" 2>/dev/null | tail -n 8 || true
|
| 22 |
+
|
| 23 |
+
echo
|
| 24 |
+
echo "When GPU0 script finishes, log should contain: Cross-domain (E18) queue completed"
|
| 25 |
+
echo "When GPU1 script finishes, log should contain: All GPU1 ablation jobs completed"
|
scripts/batch_uipress_clip_epochs.py
ADDED
|
@@ -0,0 +1,193 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
For each checkpoints/optical_mix_d2c/epoch{N}.pt (and optionally latest.pt), run
|
| 3 |
+
eval_all (UIPress) + step_clip_batch, write under results/clip_per_epoch/<name>/.
|
| 4 |
+
|
| 5 |
+
Usage:
|
| 6 |
+
# After GPU 1 is free (or use --wait_gpu to block until VRAM drops):
|
| 7 |
+
CUDA_VISIBLE_DEVICES=1 PYTHONPATH=. python scripts/batch_uipress_clip_epochs.py \\
|
| 8 |
+
--checkpoint_dir checkpoints/optical_mix_d2c \\
|
| 9 |
+
--tag optical_mix_d2c
|
| 10 |
+
|
| 11 |
+
# Skip epochs that already have clip_scores.json:
|
| 12 |
+
... --skip_existing
|
| 13 |
+
"""
|
| 14 |
+
|
| 15 |
+
from __future__ import annotations
|
| 16 |
+
|
| 17 |
+
import argparse
|
| 18 |
+
import json
|
| 19 |
+
import os
|
| 20 |
+
import re
|
| 21 |
+
import subprocess
|
| 22 |
+
import sys
|
| 23 |
+
import time
|
| 24 |
+
from pathlib import Path
|
| 25 |
+
|
| 26 |
+
PROJECT_ROOT = Path(__file__).resolve().parent.parent
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
def _wait_for_gpu(gpu_index: int, free_below_mib: int, poll_s: int) -> None:
|
| 30 |
+
import subprocess as sp
|
| 31 |
+
|
| 32 |
+
print(f"Waiting until GPU {gpu_index} memory.used < {free_below_mib} MiB ...", flush=True)
|
| 33 |
+
while True:
|
| 34 |
+
out = sp.check_output(
|
| 35 |
+
[
|
| 36 |
+
"nvidia-smi",
|
| 37 |
+
"-i",
|
| 38 |
+
str(gpu_index),
|
| 39 |
+
"--query-gpu=memory.used",
|
| 40 |
+
"--format=csv,noheader,nounits",
|
| 41 |
+
],
|
| 42 |
+
text=True,
|
| 43 |
+
)
|
| 44 |
+
used = int(out.strip().split("\n")[0].strip())
|
| 45 |
+
print(f" GPU{gpu_index} used={used} MiB", flush=True)
|
| 46 |
+
if used < free_below_mib:
|
| 47 |
+
break
|
| 48 |
+
time.sleep(poll_s)
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def main() -> int:
|
| 52 |
+
ap = argparse.ArgumentParser()
|
| 53 |
+
ap.add_argument("--checkpoint_dir", type=Path, default=PROJECT_ROOT / "checkpoints" / "optical_mix_d2c")
|
| 54 |
+
ap.add_argument(
|
| 55 |
+
"--output_root",
|
| 56 |
+
type=Path,
|
| 57 |
+
default=None,
|
| 58 |
+
help="Defaults to results/clip_per_epoch/<tag>",
|
| 59 |
+
)
|
| 60 |
+
ap.add_argument("--tag", default="optical_mix_d2c", help="Subfolder under results/clip_per_epoch/")
|
| 61 |
+
ap.add_argument("--max_samples", type=int, default=50)
|
| 62 |
+
ap.add_argument("--target_tokens", type=int, default=256)
|
| 63 |
+
ap.add_argument("--data_dir", default="data")
|
| 64 |
+
ap.add_argument("--ref_dir", default="data/ref_screenshots")
|
| 65 |
+
ap.add_argument("--skip_existing", action="store_true")
|
| 66 |
+
ap.add_argument("--include_latest", action="store_true", help="Also eval latest.pt into folder latest/")
|
| 67 |
+
ap.add_argument("--clip_device", default="cuda", choices=["cuda", "cpu"])
|
| 68 |
+
ap.add_argument("--force_cpu_eval", action="store_true", help="Pass --force_cpu to eval_all (8B on CPU; very slow).")
|
| 69 |
+
ap.add_argument("--wait_gpu", type=int, default=None, help="Poll this GPU index until memory drops.")
|
| 70 |
+
ap.add_argument("--wait_free_mib", type=int, default=12000, help="Start when memory.used < this (MiB).")
|
| 71 |
+
ap.add_argument("--wait_poll_s", type=int, default=60)
|
| 72 |
+
args = ap.parse_args()
|
| 73 |
+
|
| 74 |
+
os.environ.setdefault("HF_ENDPOINT", os.environ.get("HF_ENDPOINT", "https://hf-mirror.com"))
|
| 75 |
+
|
| 76 |
+
ckpt_dir = args.checkpoint_dir.resolve()
|
| 77 |
+
if not ckpt_dir.is_dir():
|
| 78 |
+
print(f"checkpoint_dir not found: {ckpt_dir}", file=sys.stderr)
|
| 79 |
+
return 1
|
| 80 |
+
|
| 81 |
+
out_root = (args.output_root or (PROJECT_ROOT / "results" / "clip_per_epoch" / args.tag)).resolve()
|
| 82 |
+
out_root.mkdir(parents=True, exist_ok=True)
|
| 83 |
+
|
| 84 |
+
if args.wait_gpu is not None:
|
| 85 |
+
_wait_for_gpu(args.wait_gpu, args.wait_free_mib, args.wait_poll_s)
|
| 86 |
+
|
| 87 |
+
jobs: list[tuple[str, Path]] = []
|
| 88 |
+
for p in sorted(ckpt_dir.glob("epoch*.pt"), key=lambda x: int(re.search(r"epoch(\d+)", x.name).group(1))):
|
| 89 |
+
n = int(re.search(r"epoch(\d+)", p.name).group(1))
|
| 90 |
+
jobs.append((f"epoch_{n}", p))
|
| 91 |
+
if args.include_latest and (ckpt_dir / "latest.pt").exists():
|
| 92 |
+
jobs.append(("latest", ckpt_dir / "latest.pt"))
|
| 93 |
+
|
| 94 |
+
manifest: list[dict] = []
|
| 95 |
+
run_name = f"uipress_{args.target_tokens}"
|
| 96 |
+
|
| 97 |
+
for folder_name, ckpt_path in jobs:
|
| 98 |
+
eval_out = out_root / folder_name
|
| 99 |
+
method_dir = eval_out / run_name
|
| 100 |
+
clip_path = method_dir / "clip_scores.json"
|
| 101 |
+
if args.skip_existing and clip_path.is_file():
|
| 102 |
+
print(f"Skip (exists): {clip_path}", flush=True)
|
| 103 |
+
data = json.loads(clip_path.read_text(encoding="utf-8"))
|
| 104 |
+
ep = folder_name.replace("epoch_", "") if folder_name.startswith("epoch_") else folder_name
|
| 105 |
+
manifest.append(
|
| 106 |
+
{
|
| 107 |
+
"folder": folder_name,
|
| 108 |
+
"checkpoint": str(ckpt_path.relative_to(PROJECT_ROOT)),
|
| 109 |
+
"avg_clip": data.get("avg_clip"),
|
| 110 |
+
"n": data.get("n"),
|
| 111 |
+
"clip_path": str(clip_path.relative_to(PROJECT_ROOT)),
|
| 112 |
+
}
|
| 113 |
+
)
|
| 114 |
+
continue
|
| 115 |
+
|
| 116 |
+
eval_out.mkdir(parents=True, exist_ok=True)
|
| 117 |
+
cmd_eval = [
|
| 118 |
+
sys.executable,
|
| 119 |
+
str(PROJECT_ROOT / "scripts" / "eval_all.py"),
|
| 120 |
+
"--method",
|
| 121 |
+
"uipress",
|
| 122 |
+
"--checkpoint",
|
| 123 |
+
str(ckpt_path),
|
| 124 |
+
"--max_samples",
|
| 125 |
+
str(args.max_samples),
|
| 126 |
+
"--data_dir",
|
| 127 |
+
args.data_dir,
|
| 128 |
+
"--output_dir",
|
| 129 |
+
str(eval_out),
|
| 130 |
+
"--target_tokens",
|
| 131 |
+
str(args.target_tokens),
|
| 132 |
+
]
|
| 133 |
+
if args.force_cpu_eval:
|
| 134 |
+
cmd_eval.append("--force_cpu")
|
| 135 |
+
|
| 136 |
+
print(f"\n=== eval_all: {folder_name} <- {ckpt_path.name} ===", flush=True)
|
| 137 |
+
r1 = subprocess.run(cmd_eval, cwd=str(PROJECT_ROOT))
|
| 138 |
+
if r1.returncode != 0:
|
| 139 |
+
print(f"[error] eval_all failed rc={r1.returncode} for {folder_name}", flush=True)
|
| 140 |
+
continue
|
| 141 |
+
|
| 142 |
+
cmd_clip = [
|
| 143 |
+
sys.executable,
|
| 144 |
+
str(PROJECT_ROOT / "scripts" / "step_clip_batch.py"),
|
| 145 |
+
"--method_dir",
|
| 146 |
+
str(method_dir),
|
| 147 |
+
"--ref_dir",
|
| 148 |
+
str(PROJECT_ROOT / args.ref_dir),
|
| 149 |
+
"--clip_device",
|
| 150 |
+
args.clip_device,
|
| 151 |
+
]
|
| 152 |
+
print(f"=== CLIP: {method_dir} ===", flush=True)
|
| 153 |
+
r2 = subprocess.run(cmd_clip, cwd=str(PROJECT_ROOT))
|
| 154 |
+
if r2.returncode != 0:
|
| 155 |
+
print(f"[error] step_clip_batch failed rc={r2.returncode} for {folder_name}", flush=True)
|
| 156 |
+
continue
|
| 157 |
+
|
| 158 |
+
if clip_path.is_file():
|
| 159 |
+
data = json.loads(clip_path.read_text(encoding="utf-8"))
|
| 160 |
+
manifest.append(
|
| 161 |
+
{
|
| 162 |
+
"folder": folder_name,
|
| 163 |
+
"checkpoint": str(ckpt_path.relative_to(PROJECT_ROOT)),
|
| 164 |
+
"avg_clip": data.get("avg_clip"),
|
| 165 |
+
"n": data.get("n"),
|
| 166 |
+
"clip_path": str(clip_path.relative_to(PROJECT_ROOT)),
|
| 167 |
+
}
|
| 168 |
+
)
|
| 169 |
+
|
| 170 |
+
def _sort_key(row: dict) -> tuple:
|
| 171 |
+
f = row["folder"]
|
| 172 |
+
if f == "latest":
|
| 173 |
+
return (2, 10**9)
|
| 174 |
+
m = re.match(r"epoch_(\d+)", f)
|
| 175 |
+
return (1, int(m.group(1))) if m else (0, 0)
|
| 176 |
+
|
| 177 |
+
manifest.sort(key=_sort_key)
|
| 178 |
+
summary_path = out_root / "summary.json"
|
| 179 |
+
summary_path.write_text(json.dumps(manifest, indent=2, ensure_ascii=False), encoding="utf-8")
|
| 180 |
+
print(f"\nWrote {summary_path}", flush=True)
|
| 181 |
+
|
| 182 |
+
lines = ["| 文件夹 | checkpoint | avg CLIP | n |", "|---|---|---|---|"]
|
| 183 |
+
for row in manifest:
|
| 184 |
+
lines.append(
|
| 185 |
+
f"| {row['folder']} | `{row['checkpoint']}` | {row.get('avg_clip')} | {row.get('n')} |"
|
| 186 |
+
)
|
| 187 |
+
(out_root / "CLIP_TABLE.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
|
| 188 |
+
print(f"Wrote {out_root / 'CLIP_TABLE.md'}", flush=True)
|
| 189 |
+
return 0
|
| 190 |
+
|
| 191 |
+
|
| 192 |
+
if __name__ == "__main__":
|
| 193 |
+
raise SystemExit(main())
|
scripts/run_ablation_gpu0_cross_domain_e18.sh
ADDED
|
@@ -0,0 +1,71 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
set -euo pipefail
|
| 3 |
+
|
| 4 |
+
# GPU0 queue: cross-domain evaluation only, using E18 checkpoint.
|
| 5 |
+
# No E19/E20 operations.
|
| 6 |
+
|
| 7 |
+
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
| 8 |
+
cd "$ROOT_DIR"
|
| 9 |
+
|
| 10 |
+
GPU_ID="${GPU_ID:-0}"
|
| 11 |
+
E18_CKPT="${E18_CKPT:-checkpoints/optical_mix_d2c/epoch18.pt}"
|
| 12 |
+
|
| 13 |
+
ABL_ROOT="results/ablation_study"
|
| 14 |
+
RUN_DIR="$ABL_ROOT/runs"
|
| 15 |
+
LOG_DIR="$ABL_ROOT/logs"
|
| 16 |
+
TMP_WEBSIGHT_DIR="$ABL_ROOT/tmp_websight_eval"
|
| 17 |
+
|
| 18 |
+
mkdir -p "$RUN_DIR" "$LOG_DIR" "$TMP_WEBSIGHT_DIR"
|
| 19 |
+
if [[ ! -e "$TMP_WEBSIGHT_DIR/ref_screenshots" ]]; then
|
| 20 |
+
ln -s "$(realpath data/ref_screenshots_websight)" "$TMP_WEBSIGHT_DIR/ref_screenshots"
|
| 21 |
+
fi
|
| 22 |
+
|
| 23 |
+
if [[ ! -f "$E18_CKPT" ]]; then
|
| 24 |
+
echo "Missing checkpoint: $E18_CKPT" >&2
|
| 25 |
+
exit 1
|
| 26 |
+
fi
|
| 27 |
+
|
| 28 |
+
export PYTHONPATH=.
|
| 29 |
+
|
| 30 |
+
run() {
|
| 31 |
+
local name="$1"
|
| 32 |
+
shift
|
| 33 |
+
echo
|
| 34 |
+
echo "============================================================"
|
| 35 |
+
echo "[$(date '+%F %T')] START: $name"
|
| 36 |
+
echo "CMD: $*"
|
| 37 |
+
echo "============================================================"
|
| 38 |
+
"$@"
|
| 39 |
+
echo "[$(date '+%F %T')] DONE: $name"
|
| 40 |
+
}
|
| 41 |
+
|
| 42 |
+
run "cross_domain_qwen3_full" \
|
| 43 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 44 |
+
--method baseline \
|
| 45 |
+
--max_samples 50 \
|
| 46 |
+
--data_dir "$TMP_WEBSIGHT_DIR" \
|
| 47 |
+
--output_dir "$RUN_DIR/cross_domain_e18"
|
| 48 |
+
|
| 49 |
+
run "cross_domain_uipress_e18" \
|
| 50 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 51 |
+
--method uipress \
|
| 52 |
+
--checkpoint "$E18_CKPT" \
|
| 53 |
+
--target_tokens 256 \
|
| 54 |
+
--max_samples 50 \
|
| 55 |
+
--data_dir "$TMP_WEBSIGHT_DIR" \
|
| 56 |
+
--output_dir "$RUN_DIR/cross_domain_e18"
|
| 57 |
+
|
| 58 |
+
run "cross_domain_clip_qwen3_full" \
|
| 59 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 60 |
+
--method_dir "$RUN_DIR/cross_domain_e18/qwen3_full" \
|
| 61 |
+
--ref_dir data/ref_screenshots_websight \
|
| 62 |
+
--clip_device cuda
|
| 63 |
+
|
| 64 |
+
run "cross_domain_clip_uipress_e18" \
|
| 65 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 66 |
+
--method_dir "$RUN_DIR/cross_domain_e18/uipress_256" \
|
| 67 |
+
--ref_dir data/ref_screenshots_websight \
|
| 68 |
+
--clip_device cuda
|
| 69 |
+
|
| 70 |
+
echo
|
| 71 |
+
echo "Cross-domain (E18) queue completed at $(date '+%F %T')."
|
scripts/run_ablation_gpu1_train_eval.sh
ADDED
|
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
set -euo pipefail
|
| 3 |
+
|
| 4 |
+
# GPU1 queue: No-LoRA / token sensitivity / LR scan
|
| 5 |
+
# All outputs are saved under results/ablation_study/{checkpoints,runs}
|
| 6 |
+
|
| 7 |
+
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
| 8 |
+
cd "$ROOT_DIR"
|
| 9 |
+
|
| 10 |
+
GPU_ID="${GPU_ID:-1}"
|
| 11 |
+
EPOCHS_ABL="${EPOCHS_ABL:-5}"
|
| 12 |
+
MAX_SAMPLES="${MAX_SAMPLES:-10000}"
|
| 13 |
+
|
| 14 |
+
ABL_ROOT="results/ablation_study"
|
| 15 |
+
CKPT_DIR="$ABL_ROOT/checkpoints"
|
| 16 |
+
RUN_DIR="$ABL_ROOT/runs"
|
| 17 |
+
LOG_DIR="$ABL_ROOT/logs"
|
| 18 |
+
mkdir -p "$CKPT_DIR" "$RUN_DIR" "$LOG_DIR"
|
| 19 |
+
|
| 20 |
+
export PYTHONPATH=.
|
| 21 |
+
|
| 22 |
+
run() {
|
| 23 |
+
local name="$1"
|
| 24 |
+
shift
|
| 25 |
+
echo
|
| 26 |
+
echo "============================================================"
|
| 27 |
+
echo "[$(date '+%F %T')] START: $name"
|
| 28 |
+
echo "CMD: $*"
|
| 29 |
+
echo "============================================================"
|
| 30 |
+
"$@"
|
| 31 |
+
echo "[$(date '+%F %T')] DONE: $name"
|
| 32 |
+
}
|
| 33 |
+
|
| 34 |
+
# 1) Remove LoRA
|
| 35 |
+
run "train_no_lora_256" \
|
| 36 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 37 |
+
--output_dir "$CKPT_DIR/no_lora_256" \
|
| 38 |
+
--disable_lora \
|
| 39 |
+
--target_tokens 256 \
|
| 40 |
+
--epochs "$EPOCHS_ABL" \
|
| 41 |
+
--max_samples "$MAX_SAMPLES" \
|
| 42 |
+
--mix_root data \
|
| 43 |
+
--mix_images_subdir ref_screenshots \
|
| 44 |
+
--mix_gt_subdir gt_html \
|
| 45 |
+
--max_html_tokens 8192
|
| 46 |
+
|
| 47 |
+
run "eval_no_lora_256" \
|
| 48 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 49 |
+
--method uipress \
|
| 50 |
+
--checkpoint "$CKPT_DIR/no_lora_256/latest.pt" \
|
| 51 |
+
--target_tokens 256 \
|
| 52 |
+
--max_samples 50 \
|
| 53 |
+
--data_dir data \
|
| 54 |
+
--output_dir "$RUN_DIR/no_lora_256"
|
| 55 |
+
|
| 56 |
+
run "clip_no_lora_256" \
|
| 57 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 58 |
+
--method_dir "$RUN_DIR/no_lora_256/uipress_256" \
|
| 59 |
+
--ref_dir data/ref_screenshots \
|
| 60 |
+
--clip_device cuda
|
| 61 |
+
|
| 62 |
+
# 2) Token sensitivity: 64 / 128 / 512
|
| 63 |
+
for tok in 64 128 512; do
|
| 64 |
+
run "train_token_${tok}" \
|
| 65 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 66 |
+
--output_dir "$CKPT_DIR/token_${tok}" \
|
| 67 |
+
--target_tokens "$tok" \
|
| 68 |
+
--epochs "$EPOCHS_ABL" \
|
| 69 |
+
--max_samples "$MAX_SAMPLES" \
|
| 70 |
+
--mix_root data \
|
| 71 |
+
--mix_images_subdir ref_screenshots \
|
| 72 |
+
--mix_gt_subdir gt_html \
|
| 73 |
+
--max_html_tokens 8192
|
| 74 |
+
|
| 75 |
+
run "eval_token_${tok}" \
|
| 76 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 77 |
+
--method uipress \
|
| 78 |
+
--checkpoint "$CKPT_DIR/token_${tok}/latest.pt" \
|
| 79 |
+
--target_tokens "$tok" \
|
| 80 |
+
--max_samples 50 \
|
| 81 |
+
--data_dir data \
|
| 82 |
+
--output_dir "$RUN_DIR/token_${tok}"
|
| 83 |
+
|
| 84 |
+
run "clip_token_${tok}" \
|
| 85 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 86 |
+
--method_dir "$RUN_DIR/token_${tok}/uipress_${tok}" \
|
| 87 |
+
--ref_dir data/ref_screenshots \
|
| 88 |
+
--clip_device cuda
|
| 89 |
+
done
|
| 90 |
+
|
| 91 |
+
# 3) Learning-rate scan (compressor LR)
|
| 92 |
+
for lr in 1e-4 2e-4 4e-4; do
|
| 93 |
+
safe_lr="${lr//./p}"
|
| 94 |
+
run "train_lr_${safe_lr}" \
|
| 95 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 96 |
+
--output_dir "$CKPT_DIR/lr_${safe_lr}" \
|
| 97 |
+
--target_tokens 256 \
|
| 98 |
+
--lr_compressor "$lr" \
|
| 99 |
+
--epochs "$EPOCHS_ABL" \
|
| 100 |
+
--max_samples "$MAX_SAMPLES" \
|
| 101 |
+
--mix_root data \
|
| 102 |
+
--mix_images_subdir ref_screenshots \
|
| 103 |
+
--mix_gt_subdir gt_html \
|
| 104 |
+
--max_html_tokens 8192
|
| 105 |
+
|
| 106 |
+
run "eval_lr_${safe_lr}" \
|
| 107 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 108 |
+
--method uipress \
|
| 109 |
+
--checkpoint "$CKPT_DIR/lr_${safe_lr}/latest.pt" \
|
| 110 |
+
--target_tokens 256 \
|
| 111 |
+
--max_samples 50 \
|
| 112 |
+
--data_dir data \
|
| 113 |
+
--output_dir "$RUN_DIR/lr_${safe_lr}"
|
| 114 |
+
|
| 115 |
+
run "clip_lr_${safe_lr}" \
|
| 116 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 117 |
+
--method_dir "$RUN_DIR/lr_${safe_lr}/uipress_256" \
|
| 118 |
+
--ref_dir data/ref_screenshots \
|
| 119 |
+
--clip_device cuda
|
| 120 |
+
done
|
| 121 |
+
|
| 122 |
+
echo
|
| 123 |
+
echo "All GPU1 ablation jobs completed at $(date '+%F %T')."
|
scripts/run_ablation_queue_gpu1.sh
ADDED
|
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
set -euo pipefail
|
| 3 |
+
|
| 4 |
+
# One-GPU ablation queue (default GPU1).
|
| 5 |
+
# This script runs experiments sequentially and stores all artifacts under:
|
| 6 |
+
# results/ablation_study/
|
| 7 |
+
|
| 8 |
+
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
| 9 |
+
cd "$ROOT_DIR"
|
| 10 |
+
|
| 11 |
+
GPU_ID="${GPU_ID:-1}"
|
| 12 |
+
EPOCHS_ABL="${EPOCHS_ABL:-5}"
|
| 13 |
+
MAX_SAMPLES="${MAX_SAMPLES:-10000}"
|
| 14 |
+
TOPK="${TOPK:-30}"
|
| 15 |
+
|
| 16 |
+
ABL_ROOT="results/ablation_study"
|
| 17 |
+
CKPT_DIR="$ABL_ROOT/checkpoints"
|
| 18 |
+
RUN_DIR="$ABL_ROOT/runs"
|
| 19 |
+
LOG_DIR="$ABL_ROOT/logs"
|
| 20 |
+
|
| 21 |
+
mkdir -p "$CKPT_DIR" "$RUN_DIR" "$LOG_DIR"
|
| 22 |
+
|
| 23 |
+
run() {
|
| 24 |
+
local name="$1"
|
| 25 |
+
shift
|
| 26 |
+
echo
|
| 27 |
+
echo "============================================================"
|
| 28 |
+
echo "[$(date '+%F %T')] START: $name"
|
| 29 |
+
echo "CMD: $*"
|
| 30 |
+
echo "============================================================"
|
| 31 |
+
"$@"
|
| 32 |
+
echo "[$(date '+%F %T')] DONE: $name"
|
| 33 |
+
}
|
| 34 |
+
|
| 35 |
+
export PYTHONPATH=.
|
| 36 |
+
|
| 37 |
+
# 0) Figure-2 continuation: resume mix_d2c to epoch20 (if possible)
|
| 38 |
+
if [[ -f checkpoints/optical_mix_d2c/latest.pt ]]; then
|
| 39 |
+
run "figure2_resume_to_e20" \
|
| 40 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 41 |
+
--output_dir checkpoints/optical_mix_d2c \
|
| 42 |
+
--resume checkpoints/optical_mix_d2c/latest.pt \
|
| 43 |
+
--epochs 20 \
|
| 44 |
+
--max_samples "$MAX_SAMPLES" \
|
| 45 |
+
--mix_root data \
|
| 46 |
+
--mix_images_subdir ref_screenshots \
|
| 47 |
+
--mix_gt_subdir gt_html \
|
| 48 |
+
--max_html_tokens 8192 \
|
| 49 |
+
--eval_after_epoch \
|
| 50 |
+
--eval_output_dir results/clip_per_epoch/optical_mix_d2c \
|
| 51 |
+
--eval_clip_device cuda
|
| 52 |
+
fi
|
| 53 |
+
|
| 54 |
+
# 1) Remove LoRA ablation
|
| 55 |
+
run "train_no_lora_256" \
|
| 56 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 57 |
+
--output_dir "$CKPT_DIR/no_lora_256" \
|
| 58 |
+
--disable_lora \
|
| 59 |
+
--target_tokens 256 \
|
| 60 |
+
--epochs "$EPOCHS_ABL" \
|
| 61 |
+
--max_samples "$MAX_SAMPLES" \
|
| 62 |
+
--mix_root data \
|
| 63 |
+
--mix_images_subdir ref_screenshots \
|
| 64 |
+
--mix_gt_subdir gt_html \
|
| 65 |
+
--max_html_tokens 8192
|
| 66 |
+
|
| 67 |
+
run "eval_no_lora_256" \
|
| 68 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 69 |
+
--method uipress \
|
| 70 |
+
--checkpoint "$CKPT_DIR/no_lora_256/latest.pt" \
|
| 71 |
+
--target_tokens 256 \
|
| 72 |
+
--max_samples 50 \
|
| 73 |
+
--data_dir data \
|
| 74 |
+
--output_dir "$RUN_DIR/no_lora_256"
|
| 75 |
+
|
| 76 |
+
run "clip_no_lora_256" \
|
| 77 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 78 |
+
--method_dir "$RUN_DIR/no_lora_256/uipress_256" \
|
| 79 |
+
--ref_dir data/ref_screenshots \
|
| 80 |
+
--clip_device cuda
|
| 81 |
+
|
| 82 |
+
# 2) Token sensitivity ablation
|
| 83 |
+
for tok in 64 128 512; do
|
| 84 |
+
run "train_token_${tok}" \
|
| 85 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 86 |
+
--output_dir "$CKPT_DIR/token_${tok}" \
|
| 87 |
+
--target_tokens "$tok" \
|
| 88 |
+
--epochs "$EPOCHS_ABL" \
|
| 89 |
+
--max_samples "$MAX_SAMPLES" \
|
| 90 |
+
--mix_root data \
|
| 91 |
+
--mix_images_subdir ref_screenshots \
|
| 92 |
+
--mix_gt_subdir gt_html \
|
| 93 |
+
--max_html_tokens 8192
|
| 94 |
+
|
| 95 |
+
run "eval_token_${tok}" \
|
| 96 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 97 |
+
--method uipress \
|
| 98 |
+
--checkpoint "$CKPT_DIR/token_${tok}/latest.pt" \
|
| 99 |
+
--target_tokens "$tok" \
|
| 100 |
+
--max_samples 50 \
|
| 101 |
+
--data_dir data \
|
| 102 |
+
--output_dir "$RUN_DIR/token_${tok}"
|
| 103 |
+
|
| 104 |
+
run "clip_token_${tok}" \
|
| 105 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 106 |
+
--method_dir "$RUN_DIR/token_${tok}/uipress_${tok}" \
|
| 107 |
+
--ref_dir data/ref_screenshots \
|
| 108 |
+
--clip_device cuda
|
| 109 |
+
done
|
| 110 |
+
|
| 111 |
+
# 3) Learning-rate scan (compressor lr)
|
| 112 |
+
for lr in 1e-4 2e-4 4e-4; do
|
| 113 |
+
safe_lr="${lr//./p}"
|
| 114 |
+
run "train_lr_${safe_lr}" \
|
| 115 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
|
| 116 |
+
--output_dir "$CKPT_DIR/lr_${safe_lr}" \
|
| 117 |
+
--target_tokens 256 \
|
| 118 |
+
--lr_compressor "$lr" \
|
| 119 |
+
--epochs "$EPOCHS_ABL" \
|
| 120 |
+
--max_samples "$MAX_SAMPLES" \
|
| 121 |
+
--mix_root data \
|
| 122 |
+
--mix_images_subdir ref_screenshots \
|
| 123 |
+
--mix_gt_subdir gt_html \
|
| 124 |
+
--max_html_tokens 8192
|
| 125 |
+
|
| 126 |
+
run "eval_lr_${safe_lr}" \
|
| 127 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 128 |
+
--method uipress \
|
| 129 |
+
--checkpoint "$CKPT_DIR/lr_${safe_lr}/latest.pt" \
|
| 130 |
+
--target_tokens 256 \
|
| 131 |
+
--max_samples 50 \
|
| 132 |
+
--data_dir data \
|
| 133 |
+
--output_dir "$RUN_DIR/lr_${safe_lr}"
|
| 134 |
+
|
| 135 |
+
run "clip_lr_${safe_lr}" \
|
| 136 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 137 |
+
--method_dir "$RUN_DIR/lr_${safe_lr}/uipress_256" \
|
| 138 |
+
--ref_dir data/ref_screenshots \
|
| 139 |
+
--clip_device cuda
|
| 140 |
+
done
|
| 141 |
+
|
| 142 |
+
# 4) Cross-domain validation (WebSight screenshots as eval set)
|
| 143 |
+
TMP_WEBSIGHT_DIR="$ABL_ROOT/tmp_websight_eval"
|
| 144 |
+
mkdir -p "$TMP_WEBSIGHT_DIR"
|
| 145 |
+
if [[ ! -e "$TMP_WEBSIGHT_DIR/ref_screenshots" ]]; then
|
| 146 |
+
ln -s "$(realpath data/ref_screenshots_websight)" "$TMP_WEBSIGHT_DIR/ref_screenshots"
|
| 147 |
+
fi
|
| 148 |
+
|
| 149 |
+
run "cross_domain_qwen3_full" \
|
| 150 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 151 |
+
--method baseline \
|
| 152 |
+
--max_samples 50 \
|
| 153 |
+
--data_dir "$TMP_WEBSIGHT_DIR" \
|
| 154 |
+
--output_dir "$RUN_DIR/cross_domain"
|
| 155 |
+
|
| 156 |
+
run "cross_domain_uipress_latest" \
|
| 157 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
|
| 158 |
+
--method uipress \
|
| 159 |
+
--checkpoint checkpoints/optical_mix_d2c/latest.pt \
|
| 160 |
+
--target_tokens 256 \
|
| 161 |
+
--max_samples 50 \
|
| 162 |
+
--data_dir "$TMP_WEBSIGHT_DIR" \
|
| 163 |
+
--output_dir "$RUN_DIR/cross_domain"
|
| 164 |
+
|
| 165 |
+
run "cross_domain_clip_qwen3_full" \
|
| 166 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 167 |
+
--method_dir "$RUN_DIR/cross_domain/qwen3_full" \
|
| 168 |
+
--ref_dir data/ref_screenshots_websight \
|
| 169 |
+
--clip_device cuda
|
| 170 |
+
|
| 171 |
+
run "cross_domain_clip_uipress_256" \
|
| 172 |
+
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
|
| 173 |
+
--method_dir "$RUN_DIR/cross_domain/uipress_256" \
|
| 174 |
+
--ref_dir data/ref_screenshots_websight \
|
| 175 |
+
--clip_device cuda
|
| 176 |
+
|
| 177 |
+
# 5) Build Top-K report from current available methods
|
| 178 |
+
run "build_topk_report" \
|
| 179 |
+
python scripts/ablation_topk_report.py \
|
| 180 |
+
--topk "$TOPK" \
|
| 181 |
+
--out_root "$ABL_ROOT"
|
| 182 |
+
|
| 183 |
+
echo
|
| 184 |
+
echo "All queue steps completed at $(date '+%F %T')."
|
scripts/run_ablation_study.sh
ADDED
|
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
set -euo pipefail
|
| 3 |
+
|
| 4 |
+
# Usage:
|
| 5 |
+
# bash scripts/run_ablation_study.sh
|
| 6 |
+
#
|
| 7 |
+
# This script prepares a dedicated ablation workspace and prints/executes
|
| 8 |
+
# reproducible commands for the Top-30-focused ablation study.
|
| 9 |
+
|
| 10 |
+
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
| 11 |
+
cd "$ROOT_DIR"
|
| 12 |
+
|
| 13 |
+
ABL_ROOT="results/ablation_study"
|
| 14 |
+
LOG_DIR="$ABL_ROOT/logs"
|
| 15 |
+
CKPT_DIR="$ABL_ROOT/checkpoints"
|
| 16 |
+
RUN_DIR="$ABL_ROOT/runs"
|
| 17 |
+
|
| 18 |
+
mkdir -p "$LOG_DIR" "$CKPT_DIR" "$RUN_DIR"
|
| 19 |
+
|
| 20 |
+
echo "Ablation workspace:"
|
| 21 |
+
echo " $ABL_ROOT"
|
| 22 |
+
echo
|
| 23 |
+
|
| 24 |
+
# -----------------------
|
| 25 |
+
# 1) No-LoRA ablation
|
| 26 |
+
# -----------------------
|
| 27 |
+
echo "[1/5] No-LoRA ablation"
|
| 28 |
+
echo "Train compressor-only checkpoint into: $CKPT_DIR/no_lora"
|
| 29 |
+
cat <<'CMD'
|
| 30 |
+
# Example:
|
| 31 |
+
# CUDA_VISIBLE_DEVICES=1 PYTHONPATH=. python scripts/train_compressor.py \
|
| 32 |
+
# --output_dir results/ablation_study/checkpoints/no_lora \
|
| 33 |
+
# --disable_lora --target_tokens 256 \
|
| 34 |
+
# --epochs 5 --max_samples 10000 \
|
| 35 |
+
# --mix_root data --mix_images_subdir ref_screenshots --mix_gt_subdir gt_html \
|
| 36 |
+
# --max_html_tokens 8192
|
| 37 |
+
CMD
|
| 38 |
+
echo
|
| 39 |
+
|
| 40 |
+
# -----------------------
|
| 41 |
+
# 2) Token sensitivity
|
| 42 |
+
# -----------------------
|
| 43 |
+
echo "[2/5] Token sensitivity (64/128/512)"
|
| 44 |
+
cat <<'CMD'
|
| 45 |
+
# For each token in {64,128,512}, train and eval:
|
| 46 |
+
# CUDA_VISIBLE_DEVICES=1 PYTHONPATH=. python scripts/train_compressor.py \
|
| 47 |
+
# --output_dir results/ablation_study/checkpoints/token_64 \
|
| 48 |
+
# --target_tokens 64 --epochs 5 --max_samples 10000 \
|
| 49 |
+
# --mix_root data --mix_images_subdir ref_screenshots --mix_gt_subdir gt_html \
|
| 50 |
+
# --max_html_tokens 8192
|
| 51 |
+
#
|
| 52 |
+
# CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_all.py \
|
| 53 |
+
# --method uipress --checkpoint results/ablation_study/checkpoints/token_64/latest.pt \
|
| 54 |
+
# --target_tokens 64 --max_samples 50 --output_dir results/ablation_study/runs/token_64
|
| 55 |
+
#
|
| 56 |
+
# CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/step_clip_batch.py \
|
| 57 |
+
# --method_dir results/ablation_study/runs/token_64/uipress_64 \
|
| 58 |
+
# --ref_dir data/ref_screenshots
|
| 59 |
+
CMD
|
| 60 |
+
echo
|
| 61 |
+
|
| 62 |
+
# -----------------------
|
| 63 |
+
# 3) Cross-domain check
|
| 64 |
+
# -----------------------
|
| 65 |
+
echo "[3/5] Cross-domain (WebSight eval split)"
|
| 66 |
+
cat <<'CMD'
|
| 67 |
+
# Run eval with the same methods on WebSight-side eval set directory:
|
| 68 |
+
# CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. python scripts/eval_all.py \
|
| 69 |
+
# --method uipress --checkpoint checkpoints/optical_mix_d2c/latest.pt \
|
| 70 |
+
# --target_tokens 256 --data_dir data --max_samples 50 \
|
| 71 |
+
# --output_dir results/ablation_study/runs/websight_eval
|
| 72 |
+
CMD
|
| 73 |
+
echo
|
| 74 |
+
|
| 75 |
+
# -----------------------
|
| 76 |
+
# 4) LR scan
|
| 77 |
+
# -----------------------
|
| 78 |
+
echo "[4/5] Learning-rate scan"
|
| 79 |
+
cat <<'CMD'
|
| 80 |
+
# Suggested compressor LR scan:
|
| 81 |
+
# 1e-4 / 2e-4 / 4e-4 with fixed other settings.
|
| 82 |
+
# Save each run under:
|
| 83 |
+
# results/ablation_study/checkpoints/lr_1e-4
|
| 84 |
+
# results/ablation_study/checkpoints/lr_2e-4
|
| 85 |
+
# results/ablation_study/checkpoints/lr_4e-4
|
| 86 |
+
CMD
|
| 87 |
+
echo
|
| 88 |
+
|
| 89 |
+
# -----------------------
|
| 90 |
+
# 5) Page-type analysis
|
| 91 |
+
# -----------------------
|
| 92 |
+
echo "[5/5] Page-type analysis"
|
| 93 |
+
cat <<'CMD'
|
| 94 |
+
# Put page-type id mapping as:
|
| 95 |
+
# results/ablation_study/page_types.json
|
| 96 |
+
# Then post-process top-k IDs by category from:
|
| 97 |
+
# results/ablation_study/top30/top30_selected_ids.json
|
| 98 |
+
CMD
|
| 99 |
+
echo
|
| 100 |
+
|
| 101 |
+
# Build Top-30 report from available runs (safe to run repeatedly).
|
| 102 |
+
PYTHONPATH=. python scripts/ablation_topk_report.py --topk 30 --out_root "$ABL_ROOT"
|
| 103 |
+
|
| 104 |
+
echo
|
| 105 |
+
echo "Done. Generated:"
|
| 106 |
+
echo " $ABL_ROOT/top30/top30_table.json"
|
| 107 |
+
echo " $ABL_ROOT/top30/top30_table.md"
|
| 108 |
+
echo " $ABL_ROOT/top30/top30_selected_ids.json"
|
scripts/train_compressor.py
CHANGED
|
@@ -93,7 +93,7 @@ def _run_subprocess_eval_and_clip(args, out_dir: Path, epoch: int) -> Path | Non
|
|
| 93 |
eval_root = (
|
| 94 |
Path(args.eval_output_dir).resolve()
|
| 95 |
if args.eval_output_dir
|
| 96 |
-
else (PROJECT_ROOT / "results" / "
|
| 97 |
)
|
| 98 |
eval_epoch_dir = eval_root / f"epoch_{epoch}"
|
| 99 |
eval_epoch_dir.mkdir(parents=True, exist_ok=True)
|
|
@@ -130,6 +130,8 @@ def _run_subprocess_eval_and_clip(args, out_dir: Path, epoch: int) -> Path | Non
|
|
| 130 |
str(method_dir),
|
| 131 |
"--ref_dir",
|
| 132 |
args.eval_ref_dir,
|
|
|
|
|
|
|
| 133 |
]
|
| 134 |
r2 = subprocess.run(cmd_clip, cwd=str(PROJECT_ROOT))
|
| 135 |
if r2.returncode != 0:
|
|
@@ -157,20 +159,15 @@ def _torch_load_compat(path, map_location):
|
|
| 157 |
|
| 158 |
|
| 159 |
def _rebuild_optimizer_scheduler(args, model, device, total_steps, ckpt_path: Path):
|
|
|
|
| 160 |
trainable_params = (
|
| 161 |
list(model.compressor.parameters())
|
| 162 |
-
+
|
| 163 |
-
)
|
| 164 |
-
optimizer = torch.optim.AdamW(
|
| 165 |
-
[
|
| 166 |
-
{"params": list(model.compressor.parameters()), "lr": args.lr_compressor},
|
| 167 |
-
{
|
| 168 |
-
"params": [p for p in model.lora_modules.parameters() if p.requires_grad],
|
| 169 |
-
"lr": args.lr_lora,
|
| 170 |
-
},
|
| 171 |
-
],
|
| 172 |
-
weight_decay=0.01,
|
| 173 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 174 |
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
|
| 175 |
optimizer, T_max=max(total_steps, 1), eta_min=1e-6,
|
| 176 |
)
|
|
@@ -397,6 +394,7 @@ class CompressedQwen3VL(nn.Module):
|
|
| 397 |
lora_r=16,
|
| 398 |
lora_alpha=32,
|
| 399 |
max_html_tokens=2048,
|
|
|
|
| 400 |
):
|
| 401 |
super().__init__()
|
| 402 |
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
|
|
@@ -423,15 +421,17 @@ class CompressedQwen3VL(nn.Module):
|
|
| 423 |
# Monkey-patch get_image_features so it auto-compresses visual tokens.
|
| 424 |
self._patch_vision(compressor=self.compressor, target_tokens=target_tokens)
|
| 425 |
|
| 426 |
-
# Add LoRA to LLM decoder
|
| 427 |
-
self._add_lora(lora_r, lora_alpha)
|
| 428 |
|
| 429 |
self.target_tokens = target_tokens
|
| 430 |
self.max_html_tokens = max_html_tokens
|
| 431 |
|
| 432 |
-
def _add_lora(self, r, alpha):
|
| 433 |
"""Manually inject LoRA into q_proj, v_proj of each LLM layer."""
|
| 434 |
self.lora_modules = nn.ModuleDict()
|
|
|
|
|
|
|
| 435 |
lm = self.base_model.model
|
| 436 |
if hasattr(lm, "language_model"):
|
| 437 |
layers = lm.language_model.layers
|
|
@@ -703,7 +703,10 @@ def train(args):
|
|
| 703 |
if is_main:
|
| 704 |
print(f"=== UIPress Optical Compressor Training ===", flush=True)
|
| 705 |
print(f" GPUs: {world_size}, target_tokens: {args.target_tokens}", flush=True)
|
| 706 |
-
|
|
|
|
|
|
|
|
|
|
| 707 |
print(f" Batch: {args.batch_size} x {args.grad_accum} x {world_size} "
|
| 708 |
f"= {args.batch_size * args.grad_accum * world_size}", flush=True)
|
| 709 |
if args.eval_after_epoch and is_distributed:
|
|
@@ -721,17 +724,18 @@ def train(args):
|
|
| 721 |
lora_r=args.lora_r,
|
| 722 |
lora_alpha=args.lora_alpha,
|
| 723 |
max_html_tokens=args.max_html_tokens,
|
|
|
|
| 724 |
)
|
| 725 |
model.base_model.to(device)
|
| 726 |
model.compressor.to(device)
|
| 727 |
log_all(f"Model loaded in {time.time() - t0:.1f}s")
|
| 728 |
|
|
|
|
|
|
|
| 729 |
# Count trainable params
|
| 730 |
if is_main:
|
| 731 |
comp_params = model.compressor.count_parameters()
|
| 732 |
-
lora_params = sum(
|
| 733 |
-
p.numel() for p in model.lora_modules.parameters() if p.requires_grad
|
| 734 |
-
)
|
| 735 |
print(f" Compressor params: {comp_params['trainable']:,}", flush=True)
|
| 736 |
print(f" LoRA params: {lora_params:,}", flush=True)
|
| 737 |
print(f" Total trainable: {comp_params['trainable'] + lora_params:,}", flush=True)
|
|
@@ -739,7 +743,7 @@ def train(args):
|
|
| 739 |
# Collect all trainable parameters
|
| 740 |
trainable_params = (
|
| 741 |
list(model.compressor.parameters())
|
| 742 |
-
+
|
| 743 |
)
|
| 744 |
|
| 745 |
# No DDP wrapper — LoRA is injected via setattr into base_model,
|
|
@@ -781,11 +785,10 @@ def train(args):
|
|
| 781 |
)
|
| 782 |
|
| 783 |
# Optimizer — use trainable_params collected before DDP
|
| 784 |
-
|
| 785 |
-
|
| 786 |
-
{"params":
|
| 787 |
-
|
| 788 |
-
], weight_decay=0.01)
|
| 789 |
|
| 790 |
total_steps = len(loader) * args.epochs // args.grad_accum
|
| 791 |
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
|
|
@@ -801,7 +804,7 @@ def train(args):
|
|
| 801 |
for k, v in comp_state.items():
|
| 802 |
new_state[k.replace("module.", "")] = v
|
| 803 |
model.compressor.load_state_dict(new_state)
|
| 804 |
-
if "lora" in ckpt:
|
| 805 |
model.lora_modules.load_state_dict(ckpt["lora"])
|
| 806 |
if "epoch" in ckpt:
|
| 807 |
start_epoch = ckpt["epoch"] + 1
|
|
@@ -950,6 +953,11 @@ def parse_args():
|
|
| 950 |
p.add_argument("--target_tokens", type=int, default=256)
|
| 951 |
p.add_argument("--lora_r", type=int, default=16)
|
| 952 |
p.add_argument("--lora_alpha", type=int, default=32)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 953 |
p.add_argument("--resume", type=str, default=None)
|
| 954 |
p.add_argument(
|
| 955 |
"--mix_root",
|
|
@@ -977,10 +985,16 @@ def parse_args():
|
|
| 977 |
p.add_argument(
|
| 978 |
"--eval_output_dir",
|
| 979 |
default=None,
|
| 980 |
-
help="Defaults to results/
|
| 981 |
)
|
| 982 |
p.add_argument("--eval_data_dir", default="data")
|
| 983 |
p.add_argument("--eval_ref_dir", default="data/ref_screenshots")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 984 |
return p.parse_args()
|
| 985 |
|
| 986 |
|
|
|
|
| 93 |
eval_root = (
|
| 94 |
Path(args.eval_output_dir).resolve()
|
| 95 |
if args.eval_output_dir
|
| 96 |
+
else (PROJECT_ROOT / "results" / "clip_per_epoch" / out_dir.name)
|
| 97 |
)
|
| 98 |
eval_epoch_dir = eval_root / f"epoch_{epoch}"
|
| 99 |
eval_epoch_dir.mkdir(parents=True, exist_ok=True)
|
|
|
|
| 130 |
str(method_dir),
|
| 131 |
"--ref_dir",
|
| 132 |
args.eval_ref_dir,
|
| 133 |
+
"--clip_device",
|
| 134 |
+
args.eval_clip_device,
|
| 135 |
]
|
| 136 |
r2 = subprocess.run(cmd_clip, cwd=str(PROJECT_ROOT))
|
| 137 |
if r2.returncode != 0:
|
|
|
|
| 159 |
|
| 160 |
|
| 161 |
def _rebuild_optimizer_scheduler(args, model, device, total_steps, ckpt_path: Path):
|
| 162 |
+
lora_trainable = [p for p in model.lora_modules.parameters() if p.requires_grad]
|
| 163 |
trainable_params = (
|
| 164 |
list(model.compressor.parameters())
|
| 165 |
+
+ lora_trainable
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 166 |
)
|
| 167 |
+
optim_groups = [{"params": list(model.compressor.parameters()), "lr": args.lr_compressor}]
|
| 168 |
+
if lora_trainable:
|
| 169 |
+
optim_groups.append({"params": lora_trainable, "lr": args.lr_lora})
|
| 170 |
+
optimizer = torch.optim.AdamW(optim_groups, weight_decay=0.01)
|
| 171 |
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
|
| 172 |
optimizer, T_max=max(total_steps, 1), eta_min=1e-6,
|
| 173 |
)
|
|
|
|
| 394 |
lora_r=16,
|
| 395 |
lora_alpha=32,
|
| 396 |
max_html_tokens=2048,
|
| 397 |
+
enable_lora=True,
|
| 398 |
):
|
| 399 |
super().__init__()
|
| 400 |
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
|
|
|
|
| 421 |
# Monkey-patch get_image_features so it auto-compresses visual tokens.
|
| 422 |
self._patch_vision(compressor=self.compressor, target_tokens=target_tokens)
|
| 423 |
|
| 424 |
+
# Add LoRA to LLM decoder (can be disabled for ablation).
|
| 425 |
+
self._add_lora(lora_r, lora_alpha, enable_lora=enable_lora)
|
| 426 |
|
| 427 |
self.target_tokens = target_tokens
|
| 428 |
self.max_html_tokens = max_html_tokens
|
| 429 |
|
| 430 |
+
def _add_lora(self, r, alpha, enable_lora=True):
|
| 431 |
"""Manually inject LoRA into q_proj, v_proj of each LLM layer."""
|
| 432 |
self.lora_modules = nn.ModuleDict()
|
| 433 |
+
if (not enable_lora) or r <= 0:
|
| 434 |
+
return
|
| 435 |
lm = self.base_model.model
|
| 436 |
if hasattr(lm, "language_model"):
|
| 437 |
layers = lm.language_model.layers
|
|
|
|
| 703 |
if is_main:
|
| 704 |
print(f"=== UIPress Optical Compressor Training ===", flush=True)
|
| 705 |
print(f" GPUs: {world_size}, target_tokens: {args.target_tokens}", flush=True)
|
| 706 |
+
if args.disable_lora:
|
| 707 |
+
print(" LoRA: disabled", flush=True)
|
| 708 |
+
else:
|
| 709 |
+
print(f" LoRA r={args.lora_r}, alpha={args.lora_alpha}", flush=True)
|
| 710 |
print(f" Batch: {args.batch_size} x {args.grad_accum} x {world_size} "
|
| 711 |
f"= {args.batch_size * args.grad_accum * world_size}", flush=True)
|
| 712 |
if args.eval_after_epoch and is_distributed:
|
|
|
|
| 724 |
lora_r=args.lora_r,
|
| 725 |
lora_alpha=args.lora_alpha,
|
| 726 |
max_html_tokens=args.max_html_tokens,
|
| 727 |
+
enable_lora=not args.disable_lora,
|
| 728 |
)
|
| 729 |
model.base_model.to(device)
|
| 730 |
model.compressor.to(device)
|
| 731 |
log_all(f"Model loaded in {time.time() - t0:.1f}s")
|
| 732 |
|
| 733 |
+
lora_trainable = [p for p in model.lora_modules.parameters() if p.requires_grad]
|
| 734 |
+
|
| 735 |
# Count trainable params
|
| 736 |
if is_main:
|
| 737 |
comp_params = model.compressor.count_parameters()
|
| 738 |
+
lora_params = sum(p.numel() for p in lora_trainable)
|
|
|
|
|
|
|
| 739 |
print(f" Compressor params: {comp_params['trainable']:,}", flush=True)
|
| 740 |
print(f" LoRA params: {lora_params:,}", flush=True)
|
| 741 |
print(f" Total trainable: {comp_params['trainable'] + lora_params:,}", flush=True)
|
|
|
|
| 743 |
# Collect all trainable parameters
|
| 744 |
trainable_params = (
|
| 745 |
list(model.compressor.parameters())
|
| 746 |
+
+ lora_trainable
|
| 747 |
)
|
| 748 |
|
| 749 |
# No DDP wrapper — LoRA is injected via setattr into base_model,
|
|
|
|
| 785 |
)
|
| 786 |
|
| 787 |
# Optimizer — use trainable_params collected before DDP
|
| 788 |
+
optim_groups = [{"params": list(model.compressor.parameters()), "lr": args.lr_compressor}]
|
| 789 |
+
if lora_trainable:
|
| 790 |
+
optim_groups.append({"params": lora_trainable, "lr": args.lr_lora})
|
| 791 |
+
optimizer = torch.optim.AdamW(optim_groups, weight_decay=0.01)
|
|
|
|
| 792 |
|
| 793 |
total_steps = len(loader) * args.epochs // args.grad_accum
|
| 794 |
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
|
|
|
|
| 804 |
for k, v in comp_state.items():
|
| 805 |
new_state[k.replace("module.", "")] = v
|
| 806 |
model.compressor.load_state_dict(new_state)
|
| 807 |
+
if "lora" in ckpt and len(model.lora_modules) > 0:
|
| 808 |
model.lora_modules.load_state_dict(ckpt["lora"])
|
| 809 |
if "epoch" in ckpt:
|
| 810 |
start_epoch = ckpt["epoch"] + 1
|
|
|
|
| 953 |
p.add_argument("--target_tokens", type=int, default=256)
|
| 954 |
p.add_argument("--lora_r", type=int, default=16)
|
| 955 |
p.add_argument("--lora_alpha", type=int, default=32)
|
| 956 |
+
p.add_argument(
|
| 957 |
+
"--disable_lora",
|
| 958 |
+
action="store_true",
|
| 959 |
+
help="Disable LoRA adapters; train compressor-only for ablation.",
|
| 960 |
+
)
|
| 961 |
p.add_argument("--resume", type=str, default=None)
|
| 962 |
p.add_argument(
|
| 963 |
"--mix_root",
|
|
|
|
| 985 |
p.add_argument(
|
| 986 |
"--eval_output_dir",
|
| 987 |
default=None,
|
| 988 |
+
help="Defaults to results/clip_per_epoch/<output_dir name>.",
|
| 989 |
)
|
| 990 |
p.add_argument("--eval_data_dir", default="data")
|
| 991 |
p.add_argument("--eval_ref_dir", default="data/ref_screenshots")
|
| 992 |
+
p.add_argument(
|
| 993 |
+
"--eval_clip_device",
|
| 994 |
+
default="cuda",
|
| 995 |
+
choices=["cuda", "cpu"],
|
| 996 |
+
help="Device for CLIP ViT in post-epoch scoring.",
|
| 997 |
+
)
|
| 998 |
return p.parse_args()
|
| 999 |
|
| 1000 |
|
sync_up.py
CHANGED
|
@@ -33,8 +33,28 @@ def parse_args() -> argparse.Namespace:
|
|
| 33 |
".cursor/*",
|
| 34 |
"__pycache__/*",
|
| 35 |
"*.pyc",
|
|
|
|
|
|
|
| 36 |
"OLD/*",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
"*.tar.gz",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
],
|
| 39 |
help="Glob patterns to ignore while uploading.",
|
| 40 |
)
|
|
|
|
| 33 |
".cursor/*",
|
| 34 |
"__pycache__/*",
|
| 35 |
"*.pyc",
|
| 36 |
+
".env",
|
| 37 |
+
".env.*",
|
| 38 |
"OLD/*",
|
| 39 |
+
"checkpoints/*",
|
| 40 |
+
"data/*",
|
| 41 |
+
".cache/*",
|
| 42 |
+
".huggingface/*",
|
| 43 |
+
"logs/*",
|
| 44 |
+
"*.log",
|
| 45 |
+
"*.out",
|
| 46 |
+
"nohup.*",
|
| 47 |
+
"*.pt",
|
| 48 |
+
"*.pth",
|
| 49 |
+
"*.bin",
|
| 50 |
+
"*.safetensors",
|
| 51 |
+
"*.ckpt",
|
| 52 |
"*.tar.gz",
|
| 53 |
+
"*.tar",
|
| 54 |
+
"*.zip",
|
| 55 |
+
"results/**/html_predictions/*",
|
| 56 |
+
"results/**/rendered_screenshots/*",
|
| 57 |
+
"results/**/rendered/*",
|
| 58 |
],
|
| 59 |
help="Glob patterns to ignore while uploading.",
|
| 60 |
)
|