{{ message }}
Repository navigation
Expand file tree
/
Copy pathablations.py
More file actions
198 lines (173 loc) · 8.79 KB
/
Copy pathablations.py
File metadata and controls
198 lines (173 loc) · 8.79 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
import torch
from diffusers import StableDiffusionPipeline, DDPMScheduler
import argparse
from sampling import sample_user_images
from prior_models import RebecaDiffusionPrior
from tqdm import tqdm
from utils import save_generated_data
import gc
def main(args):
if args.experiment_type == "cross-cfgs":
device = "cuda"
model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id).to(device)
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
pipe.safety_checker = None
filedir = "./data/flickr/evaluation/diffusion_priors/models/weights/"
winner = "comprehensive_study_20250830_013540/modelrdp_num_layers6_num_heads8_hidden_dim128_tokens32_lr0.0001_optadamw_schreduce_on_plateau_bs64_nssquaredcos_cap_v2_ts1000_spu100_csFalse_objsample_normnone_uthr0"
rdp = RebecaDiffusionPrior(
img_embed_dim=1024,
num_users=210,
num_tokens=32,
hidden_dim=128,
n_heads=8,
num_layers=6,
score_classes=2,
).to(device)
rdp.load_state_dict(torch.load(f'{filedir}/{winner}.pth'))
noise_scheduler = DDPMScheduler(
num_train_timesteps=1000,
beta_schedule="squaredcos_cap_v2",
clip_sample=False,
prediction_type="sample"
)
cfgs = [3, 6, 9, 12, 15]
pipe_cfg = [1, 3, 5, 7, 10]
images_per_user = 10
savedir = "./data/flickr/evaluation/ablations/cross-cfgs/"
users = list(range(210))
seed = 2
torch.manual_seed(seed)
for cfg in tqdm(cfgs):
for sd_cfg in pipe_cfg:
gen_data = sample_user_images(
diffusion_prior_model=rdp,
diffusion_pipe=pipe,
users=users,
images_per_user=images_per_user, # Fewer images per user to save memory
noise_scheduler=noise_scheduler,
guidance_scale=cfg,
prompt=[""],
negative_prompt=[""],
img_embedding_size=1024,
pipe_cfg=sd_cfg,
device=device,
)
torch.save(gen_data, f"{savedir}/embcfg_{cfg}_pipecfg_{sd_cfg}_seed_{seed}.data")
del gen_data
torch.cuda.empty_cache()
gc.collect()
if args.experiment_type == "cross-prompts":
device = "cuda"
model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id).to(device)
pipe.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
pipe.safety_checker = None
filedir = "./data/flickr/evaluation/diffusion_priors/models/weights/"
winner = "comprehensive_study_20250830_013540/modelrdp_num_layers6_num_heads8_hidden_dim128_tokens32_lr0.0001_optadamw_schreduce_on_plateau_bs64_nssquaredcos_cap_v2_ts1000_spu100_csFalse_objsample_normnone_uthr0"
rdp = RebecaDiffusionPrior(
img_embed_dim=1024,
num_users=210,
num_tokens=32,
hidden_dim=128,
n_heads=8,
num_layers=6,
score_classes=2,
).to(device)
rdp.load_state_dict(torch.load(f'{filedir}/{winner}.pth'))
noise_scheduler = DDPMScheduler(
num_train_timesteps=1000,
beta_schedule="squaredcos_cap_v2",
clip_sample=False,
prediction_type="sample"
)
rebeca_cfg = 7.0
pipe_cfg = 5.0
images_per_user = 10
savedir = "./data/flickr/evaluation/ablations/cross-prompts/"
users = list(range(210))
seed = 3
torch.manual_seed(seed)
prompt_candidates = [
([""], [""]),
(
["high quality photo"],
["bad quality photo, letters"]
),
(
["Realistic image, finely detailed, with balanced composition and harmonious elements. "
"Dynamic yet subtle tones, versatile style adaptable to diverse themes and aesthetics, "
"prioritizing clarity and authenticity."],
["deformed, ugly, wrong proportion, frame, watermark, low res, bad anatomy, worst quality, low quality"]
)
]
for i, prompt_level in enumerate(prompt_candidates):
pos_prompt, neg_prompt = prompt_level[0], prompt_level[1]
gen_data = sample_user_images(
diffusion_prior_model=rdp,
diffusion_pipe=pipe,
users=users,
images_per_user=images_per_user, # Fewer images per user to save memory
noise_scheduler=noise_scheduler,
guidance_scale=rebeca_cfg,
prompt=pos_prompt,
negative_prompt=neg_prompt,
img_embedding_size=1024,
pipe_cfg=pipe_cfg,
device=device,
)
torch.save(gen_data, f"{savedir}/embcfg_{rebeca_cfg}_pipecfg_{pipe_cfg}_promptlevel_{i}_seed_{seed}.data")
del gen_data
torch.cuda.empty_cache()
gc.collect()
elif args.experiment_type == "baseline-prompts":
n_images = args.n_images
prompt_level = args.prompt_level
dst_dir = args.dst_dir
model_id = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(model_id).to("cuda")
pipe.safety_checker = None
prompt_candidates = [
([""], [""]),
(
["high quality photo"],
["bad quality photo, letters"]
),
(
["Realistic image, finely detailed, with balanced composition and harmonious elements. "
"Dynamic yet subtle tones, versatile style adaptable to diverse themes and aesthetics, "
"prioritizing clarity and authenticity."],
["deformed, ugly, wrong proportion, frame, watermark, low res, bad anatomy, worst quality, low quality"]
)
]
if prompt_level==0:
prompts = prompt_candidates[0]
elif prompt_level==1:
prompts = prompt_candidates[1]
elif prompt_level==2:
prompts = prompt_candidates[2]
else:
raise ValueError("Please select a prompt level from 0, 1 and 2.")
pos_prompt, neg_prompt = prompts[0], prompts[1]
save_images = list()
for _ in range(n_images):
with torch.no_grad():
gen_images = pipe(
prompt=pos_prompt,
negative_prompt=neg_prompt,
num_inference_steps=50,
).images
torch.cuda.empty_cache()
save_images.extend(gen_images)
torch.save(save_images, f"./data/flickr/evaluation/ablations/baseline-prompts/images_prompt_level_{prompt_level}.imgs")
#save_generated_data(save_images, dst_dir)
else:
raise NotImplementedError("{args.experiment_type} is not implemented yet.")
if __name__=="__main__":
parser = argparse.ArgumentParser(description='Generate baseline images')
parser.add_argument('--experiment_type', type=str, required=True, default="cross-cfg", help='Experiment type. Implemented cross-cfg, varying embedding generator CFG with Image generator CFG.')
parser.add_argument('--n_images', type=int, default=100, help='Number of images')
parser.add_argument('--prompt_level', type=int, help='Prompt level')
parser.add_argument('--dst_dir', type=str, help='Destination directory')
args = parser.parse_args()
main(args)
You can’t perform that action at this time.