Upload calib.py
Browse files
calib.py
ADDED
@@ -0,0 +1,174 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""Collect calibration dataset."""
|
3 |
+
|
4 |
+
import os
|
5 |
+
from dataclasses import dataclass
|
6 |
+
|
7 |
+
import datasets
|
8 |
+
import torch
|
9 |
+
from omniconfig import configclass
|
10 |
+
from torch import nn
|
11 |
+
from tqdm import tqdm
|
12 |
+
|
13 |
+
from deepcompressor.app.diffusion.config import DiffusionPtqRunConfig
|
14 |
+
from deepcompressor.utils.common import hash_str_to_int, tree_map
|
15 |
+
|
16 |
+
from ...utils import get_control
|
17 |
+
from ..data import get_dataset
|
18 |
+
from .utils import CollectHook
|
19 |
+
|
20 |
+
|
21 |
+
def process(x: torch.Tensor) -> torch.Tensor:
|
22 |
+
dtype = x.dtype
|
23 |
+
return torch.from_numpy(x.float().numpy()).to(dtype)
|
24 |
+
|
25 |
+
|
26 |
+
def collect(config: DiffusionPtqRunConfig, dataset: datasets.Dataset):
|
27 |
+
samples_dirpath = os.path.join(config.output.root, "samples")
|
28 |
+
caches_dirpath = os.path.join(config.output.root, "caches")
|
29 |
+
os.makedirs(samples_dirpath, exist_ok=True)
|
30 |
+
os.makedirs(caches_dirpath, exist_ok=True)
|
31 |
+
caches = []
|
32 |
+
|
33 |
+
pipeline = config.pipeline.build()
|
34 |
+
model = pipeline.unet if hasattr(pipeline, "unet") else pipeline.transformer
|
35 |
+
assert isinstance(model, nn.Module)
|
36 |
+
model.register_forward_hook(CollectHook(caches=caches), with_kwargs=True)
|
37 |
+
|
38 |
+
batch_size = config.eval.batch_size
|
39 |
+
print(f"In total {len(dataset)} samples")
|
40 |
+
print(f"Evaluating with batch size {batch_size}")
|
41 |
+
pipeline.set_progress_bar_config(desc="Sampling", leave=False, dynamic_ncols=True, position=1)
|
42 |
+
for batch in tqdm(
|
43 |
+
dataset.iter(batch_size=batch_size, drop_last_batch=False),
|
44 |
+
desc="Data",
|
45 |
+
leave=False,
|
46 |
+
dynamic_ncols=True,
|
47 |
+
total=(len(dataset) + batch_size - 1) // batch_size,
|
48 |
+
):
|
49 |
+
filenames = batch["filename"]
|
50 |
+
prompts = batch["prompt"]
|
51 |
+
seeds = [hash_str_to_int(name) for name in filenames]
|
52 |
+
generators = [torch.Generator(device=pipeline.device).manual_seed(seed) for seed in seeds]
|
53 |
+
pipeline_kwargs = config.eval.get_pipeline_kwargs()
|
54 |
+
|
55 |
+
task = config.pipeline.task
|
56 |
+
control_root = config.eval.control_root
|
57 |
+
if task in ["canny-to-image", "depth-to-image", "inpainting"]:
|
58 |
+
controls = get_control(
|
59 |
+
task,
|
60 |
+
batch["image"],
|
61 |
+
names=batch["filename"],
|
62 |
+
data_root=os.path.join(
|
63 |
+
control_root, collect_config.dataset_name, f"{dataset.config_name}-{config.eval.num_samples}"
|
64 |
+
),
|
65 |
+
)
|
66 |
+
if task == "inpainting":
|
67 |
+
pipeline_kwargs["image"] = controls[0]
|
68 |
+
pipeline_kwargs["mask_image"] = controls[1]
|
69 |
+
else:
|
70 |
+
pipeline_kwargs["control_image"] = controls
|
71 |
+
|
72 |
+
# Handle meta tensors by moving individual components
|
73 |
+
try:
|
74 |
+
pipeline = pipeline.to("cuda")
|
75 |
+
except NotImplementedError:
|
76 |
+
# Move individual pipeline components that have to_empty method
|
77 |
+
if hasattr(pipeline, 'transformer') and pipeline.transformer is not None:
|
78 |
+
try:
|
79 |
+
pipeline.transformer = pipeline.transformer.to("cuda")
|
80 |
+
except NotImplementedError:
|
81 |
+
pipeline.transformer = pipeline.transformer.to_empty(device="cuda")
|
82 |
+
|
83 |
+
if hasattr(pipeline, 'text_encoder') and pipeline.text_encoder is not None:
|
84 |
+
try:
|
85 |
+
pipeline.text_encoder = pipeline.text_encoder.to("cuda")
|
86 |
+
except NotImplementedError:
|
87 |
+
pipeline.text_encoder = pipeline.text_encoder.to_empty(device="cuda")
|
88 |
+
|
89 |
+
if hasattr(pipeline, 'text_encoder_2') and pipeline.text_encoder_2 is not None:
|
90 |
+
try:
|
91 |
+
pipeline.text_encoder_2 = pipeline.text_encoder_2.to("cuda")
|
92 |
+
except NotImplementedError:
|
93 |
+
pipeline.text_encoder_2 = pipeline.text_encoder_2.to_empty(device="cuda")
|
94 |
+
|
95 |
+
if hasattr(pipeline, 'vae') and pipeline.vae is not None:
|
96 |
+
try:
|
97 |
+
pipeline.vae = pipeline.vae.to("cuda")
|
98 |
+
except NotImplementedError:
|
99 |
+
pipeline.vae = pipeline.vae.to_empty(device="cuda")
|
100 |
+
|
101 |
+
result_images = pipeline(prompt=prompts, generator=generators, **pipeline_kwargs).images
|
102 |
+
num_guidances = (len(caches) // batch_size) // config.eval.num_steps
|
103 |
+
num_steps = len(caches) // (batch_size * num_guidances)
|
104 |
+
assert (
|
105 |
+
len(caches) == batch_size * num_steps * num_guidances
|
106 |
+
), f"Unexpected number of caches: {len(caches)} != {batch_size} * {config.eval.num_steps} * {num_guidances}"
|
107 |
+
for j, (filename, image) in enumerate(zip(filenames, result_images, strict=True)):
|
108 |
+
image.save(os.path.join(samples_dirpath, f"{filename}.png"))
|
109 |
+
for s in range(num_steps):
|
110 |
+
for g in range(num_guidances):
|
111 |
+
c = caches[s * batch_size * num_guidances + g * batch_size + j]
|
112 |
+
c["filename"] = filename
|
113 |
+
c["step"] = s
|
114 |
+
c["guidance"] = g
|
115 |
+
c = tree_map(lambda x: process(x), c)
|
116 |
+
torch.save(c, os.path.join(caches_dirpath, f"{filename}-{s:05d}-{g}.pt"))
|
117 |
+
caches.clear()
|
118 |
+
|
119 |
+
|
120 |
+
@configclass
|
121 |
+
@dataclass
|
122 |
+
class CollectConfig:
|
123 |
+
"""Configuration for collecting calibration dataset.
|
124 |
+
|
125 |
+
Args:
|
126 |
+
root (`str`, *optional*, defaults to `"datasets"`):
|
127 |
+
Root directory to save the collected dataset.
|
128 |
+
dataset_name (`str`, *optional*, defaults to `"qdiff"`):
|
129 |
+
Name of the collected dataset.
|
130 |
+
prompt_path (`str`, *optional*, defaults to `"prompts/qdiff.yaml"`):
|
131 |
+
Path to the prompt file.
|
132 |
+
num_samples (`int`, *optional*, defaults to `128`):
|
133 |
+
Number of samples to collect.
|
134 |
+
"""
|
135 |
+
|
136 |
+
root: str = "datasets"
|
137 |
+
dataset_name: str = "qdiff"
|
138 |
+
data_path: str = "prompts/qdiff.yaml"
|
139 |
+
num_samples: int = 128
|
140 |
+
|
141 |
+
|
142 |
+
if __name__ == "__main__":
|
143 |
+
parser = DiffusionPtqRunConfig.get_parser()
|
144 |
+
parser.add_config(CollectConfig, scope="collect", prefix="collect")
|
145 |
+
configs, _, unused_cfgs, unused_args, unknown_args = parser.parse_known_args()
|
146 |
+
ptq_config, collect_config = configs[""], configs["collect"]
|
147 |
+
assert isinstance(ptq_config, DiffusionPtqRunConfig)
|
148 |
+
assert isinstance(collect_config, CollectConfig)
|
149 |
+
if len(unused_cfgs) > 0:
|
150 |
+
print(f"Warning: unused configurations {unused_cfgs}")
|
151 |
+
if unused_args is not None:
|
152 |
+
print(f"Warning: unused arguments {unused_args}")
|
153 |
+
assert len(unknown_args) == 0, f"Unknown arguments: {unknown_args}"
|
154 |
+
|
155 |
+
collect_dirpath = os.path.join(
|
156 |
+
collect_config.root,
|
157 |
+
str(ptq_config.pipeline.dtype),
|
158 |
+
ptq_config.pipeline.name,
|
159 |
+
ptq_config.eval.protocol,
|
160 |
+
collect_config.dataset_name,
|
161 |
+
f"s{collect_config.num_samples}",
|
162 |
+
)
|
163 |
+
print(f"Saving caches to {collect_dirpath}")
|
164 |
+
|
165 |
+
dataset = get_dataset(
|
166 |
+
collect_config.data_path,
|
167 |
+
max_dataset_size=collect_config.num_samples,
|
168 |
+
return_gt=ptq_config.pipeline.task in ["canny-to-image"],
|
169 |
+
repeat=1,
|
170 |
+
)
|
171 |
+
|
172 |
+
ptq_config.output.root = collect_dirpath
|
173 |
+
os.makedirs(ptq_config.output.root, exist_ok=True)
|
174 |
+
collect(ptq_config, dataset=dataset)
|