Commit
·
d09c0e5
1
Parent(s):
6cd6a16
Add inference script and main model execution logic
Browse files- inference.py +149 -0
- run_inference.sh +10 -0
inference.py
ADDED
@@ -0,0 +1,149 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import sys
|
2 |
+
|
3 |
+
sys.path.append(".")
|
4 |
+
import argparse
|
5 |
+
import os
|
6 |
+
import random
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
import pickle
|
10 |
+
from configs.configuration_mmdit import MMDiTConfig
|
11 |
+
from models.modeling_motif_vision import MotifVision
|
12 |
+
|
13 |
+
from safetensors.torch import load_file
|
14 |
+
# from tools.motif_api import PromptRewriter
|
15 |
+
# from tools.nsfw_filtering import ContentFilter
|
16 |
+
|
17 |
+
def load_sharded_model(model_index_path):
|
18 |
+
"""
|
19 |
+
Loads a sharded model from a safetensors index file.
|
20 |
+
|
21 |
+
Args:
|
22 |
+
model_index_path (str): Path to the model.safetensors.index.json file.
|
23 |
+
"""
|
24 |
+
with open(model_index_path, 'r') as f:
|
25 |
+
index = json.load(f)
|
26 |
+
|
27 |
+
sharded_state_dicts = {}
|
28 |
+
folder = os.path.dirname(model_index_path)
|
29 |
+
for weight_name, filename in index["weight_map"].items():
|
30 |
+
if filename not in sharded_state_dicts:
|
31 |
+
sharded_state_dicts[filename] = load_file(os.path.join(folder, filename), device="cpu")
|
32 |
+
|
33 |
+
merged_state_dict = {}
|
34 |
+
for weight_name, filename in index["weight_map"].items():
|
35 |
+
merged_state_dict[weight_name] = sharded_state_dicts[filename][weight_name]
|
36 |
+
|
37 |
+
merged_state_dict = {k: v for k, v in merged_state_dict.items() if 'dit' in k}
|
38 |
+
return merged_state_dict
|
39 |
+
|
40 |
+
|
41 |
+
def main(args):
|
42 |
+
|
43 |
+
# Check if the prompt file exists
|
44 |
+
if not os.path.isfile(args.prompt_file):
|
45 |
+
print(f"Error: The prompt file '{args.prompt_file}' does not exist.")
|
46 |
+
sys.exit(1)
|
47 |
+
|
48 |
+
# List of prompts
|
49 |
+
with open(args.prompt_file) as f:
|
50 |
+
prompts = [prompt.rstrip() for prompt in f.readlines()]
|
51 |
+
|
52 |
+
# Load model configuration and model
|
53 |
+
config = MMDiTConfig.from_json_file(args.model_config)
|
54 |
+
config.vae_type = args.vae_type # VAE overriding
|
55 |
+
config.height = args.resolution
|
56 |
+
config.width = args.resolution
|
57 |
+
|
58 |
+
model = MotifVision(config)
|
59 |
+
|
60 |
+
# Load checkpoint
|
61 |
+
try:
|
62 |
+
ema_instance = torch.load(args.model_ckpt, weights_only=False)
|
63 |
+
ema_instance = {k: v for k, v in ema_instance.items() if "dit" in k}
|
64 |
+
except pickle.UnpicklingError as e:
|
65 |
+
print(f"Error loading checkpoint: {e}")
|
66 |
+
ema_instance = load_file(args.model_ckpt)
|
67 |
+
ema_instance = {k: v for k, v in ema_instance.items() if "dit" in k}
|
68 |
+
|
69 |
+
if "ema_model.bin" in args.model_ckpt:
|
70 |
+
# EMA checkpoint loading
|
71 |
+
for param, ema_param in zip(model.parameters(), ema_instance["shadow_params"]):
|
72 |
+
param.data.copy_(ema_param.data)
|
73 |
+
else:
|
74 |
+
# Non-EMA checkpoint loading
|
75 |
+
model.load_state_dict(ema_instance)
|
76 |
+
|
77 |
+
model = model.cuda()
|
78 |
+
model = model.to(dtype=torch.bfloat16)
|
79 |
+
model.eval()
|
80 |
+
|
81 |
+
# Use guidance scales from args or set default
|
82 |
+
guidance_scales = args.guidance_scales if args.guidance_scales else [5.0]
|
83 |
+
|
84 |
+
# If a single seed is passed without nargs, wrap it in a list
|
85 |
+
if isinstance(args.seed, int):
|
86 |
+
seeds = [args.seed]
|
87 |
+
else:
|
88 |
+
seeds = args.seed
|
89 |
+
|
90 |
+
for seed in seeds:
|
91 |
+
for guidance_scale in guidance_scales:
|
92 |
+
# Output directory structure: base_dir/seed_xxx/guidance_yyy
|
93 |
+
output_dir = os.path.join(args.output_dir, f"seed_{seed}", f"scale_{guidance_scale}")
|
94 |
+
os.makedirs(output_dir, exist_ok=True)
|
95 |
+
# Using for_loop when generating high-resolution images
|
96 |
+
for i in range(0, len(prompts), args.batch_size): # Process 1s prompts at a time
|
97 |
+
# Set random seeds
|
98 |
+
torch.manual_seed(seed)
|
99 |
+
random.seed(seed)
|
100 |
+
np.random.seed(seed)
|
101 |
+
|
102 |
+
batch_prompts = prompts[i : i + args.batch_size]
|
103 |
+
imgs = model.sample(
|
104 |
+
batch_prompts,
|
105 |
+
args.steps,
|
106 |
+
resolution=[args.resolution, args.resolution],
|
107 |
+
guidance_scale=guidance_scale,
|
108 |
+
step_scaling=1.0,
|
109 |
+
use_linear_quadratic_schedule=True,
|
110 |
+
linear_quadratic_emulating_steps=250,
|
111 |
+
get_intermediate_steps=args.streaming,
|
112 |
+
noisy_pad=args.noisy_pad,
|
113 |
+
zero_masking=args.zero_masking,
|
114 |
+
)
|
115 |
+
if args.streaming:
|
116 |
+
imgs, intermediate_imgs = imgs
|
117 |
+
if isinstance(intermediate_imgs, list):
|
118 |
+
for j, intermediate_img in enumerate(intermediate_imgs):
|
119 |
+
for k, img in enumerate(intermediate_img):
|
120 |
+
img.save(os.path.join(output_dir, f"{i + k:03d}_{j:03d}_intermediate.png"))
|
121 |
+
else:
|
122 |
+
# If intermediate_imgs is a single Image, save it directly
|
123 |
+
intermediate_imgs.save(os.path.join(output_dir, f"{i:03d}_0_intermediate.png"))
|
124 |
+
for j, img in enumerate(imgs):
|
125 |
+
img.save(os.path.join(output_dir, f"{i + j:03d}_check.png"))
|
126 |
+
|
127 |
+
|
128 |
+
if __name__ == "__main__":
|
129 |
+
parser = argparse.ArgumentParser(description="Generate images with model")
|
130 |
+
parser.add_argument("--model-config", type=str, required=True, help="Path to the model configuration file")
|
131 |
+
parser.add_argument("--model-ckpt", type=str, required=True, help="Path to the model checkpoint file")
|
132 |
+
parser.add_argument(
|
133 |
+
"--seed", type=int, nargs="*", default=[7777], help="Random seed(s) for reproducibility (can provide multiple)"
|
134 |
+
)
|
135 |
+
# parser.add_argument("--slg", type=int, nargs="*", default=None, help="")
|
136 |
+
parser.add_argument("--steps", type=int, default=50, help="Number of steps for image generation")
|
137 |
+
parser.add_argument("--resolution", type=int, default=256, help="Resolution of output images")
|
138 |
+
parser.add_argument("--batch-size", type=int, default=32)
|
139 |
+
parser.add_argument("--streaming", action="store_true")
|
140 |
+
parser.add_argument("--noisy-pad", action="store_true")
|
141 |
+
parser.add_argument("--zero-masking", action="store_true")
|
142 |
+
parser.add_argument("--vae-type", type=str, default="SD3", help="Type of VAE")
|
143 |
+
parser.add_argument("--prompt-file", type=str, default="prompt_128.txt", help="Path to the prompt file")
|
144 |
+
parser.add_argument("--guidance-scales", type=float, nargs="*", default=None, help="List of guidance scales")
|
145 |
+
parser.add_argument("--output-dir", type=str, default="output", help="Base output directory for generated images")
|
146 |
+
parser.add_argument("--lora-ckpt", action="store_true")
|
147 |
+
args = parser.parse_args()
|
148 |
+
|
149 |
+
main(args)
|
run_inference.sh
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
python inference.py \
|
2 |
+
--model-config configs/mmdit_xlarge_hq.json \
|
3 |
+
--model-ckpt checkpoints/pytorch_model_fsdp.bin \
|
4 |
+
--seed 7777 \
|
5 |
+
--steps 30 \
|
6 |
+
--resolution 1024 \
|
7 |
+
--prompt-file prompts/sample_prompts.txt \
|
8 |
+
--guidance-scales 4.0 \
|
9 |
+
--output-dir outputs \
|
10 |
+
--batch-size 1
|