-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathgenerate_large.py
170 lines (143 loc) · 6.29 KB
/
generate_large.py
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
"""
diffusion model training for large datasets (ImageNet, LSUN, ...)
"""
import argparse
import os
import pickle
import random
import numpy as np
import torch
from omegaconf import OmegaConf
from hydra.utils import instantiate
from torchvision.utils import save_image
from torch.nn.parallel import DistributedDataParallel as DDP
from tqdm import tqdm, trange
from models.cm.script_util import create_model_and_diffusion
from models.DxMI.openai_diffusion import OpenAIDiffusion
from utils import print0, mkdir_p
def generate():
global output_path, local_rank
n_batch_to_generate = args.n_sample / args.batchsize / ngpus
l_sample = []
sampler.eval()
i_img = 0
for i_batch in trange(int(n_batch_to_generate), ncols=80):
if args.guidance_scale is not None:
d_sample = trainer.sample_guidance(n_sample=args.batchsize, device=device, guidance_scale=args.guidance_scale)
else:
d_sample = sampler.sample(args.batchsize, device=device,
i_class=None, enable_grad=False)
sample = d_sample['sample']
if args.skip_fid:
sample = ((sample + 1)/2).clamp(0, 1)
for s in sample:
save_image(s, os.path.join(output_path, f"{local_rank}_{i_img}.png"))
i_img += 1
l_sample.append(sample.detach().cpu())
else:
sample = ((sample + 1) * 127.5).clamp(0, 255).to(torch.uint8)
l_sample.append(sample)
samples = torch.cat(l_sample)
if not args.skip_fid:
all_samples = [torch.zeros_like(samples) for _ in range(ngpus)]
torch.distributed.all_gather(all_samples, samples)
all_samples = torch.cat([t.detach().cpu() for t in all_samples])
return all_samples
else:
return samples
def fid():
all_samples = generate()
print(f'generated {len(all_samples)}')
if not args.skip_fid:
split_samples = all_samples[local_rank::ngpus]
split_samples = (split_samples / 255.).float()
act = get_activations_from_tensor(split_samples, extractor, batch_size=50, dims=dims, device=device)
all_act = [torch.zeros_like(act) for _ in range(ngpus)]
torch.distributed.all_gather(all_act, act)
if local_rank <= 0:
act = torch.cat([a.detach().cpu() for a in all_act]).detach().numpy()
m1, s1 = np.mean(act, axis=0), np.cov(act, rowvar=False)
fid = calculate_frechet_distance(mu1=m1, sigma1=s1, mu2=m2, sigma2=s2)
print(f'FID from {len(all_samples)} samples: {fid}')
if args.guidance_scale is not None:
np.savez(os.path.join(args.log_dir, f'samples_{len(all_samples)}_{args.guidance_scale}.npz'),
all_samples.permute(0,2,3,1).cpu().numpy())
else:
np.savez(os.path.join(args.log_dir, f'samples_{len(all_samples)}.npz'),
all_samples.permute(0,2,3,1).cpu().numpy())
return all_samples
local_rank = int(os.environ.get("LOCAL_RANK", 0))
########### Parse arguments
parser = argparse.ArgumentParser()
parser.add_argument('--log_dir', type=str, required=True, help='path to logdir')
parser.add_argument('--n_sample', type=int, required=True)
parser.add_argument('--batchsize', type=int, default=100)
parser.add_argument('--guidance_scale', type=float, default=None, help='Value guidance scale, 0.0 for no guidance')
parser.add_argument('--skip_fid', action='store_true',
help='''Do not calculate FID. Instead, save individual images as they are generated.
This option avoids memory issues with large images, such as LSUN''')
args, unknown = parser.parse_known_args()
############ load configs
cfg = OmegaConf.load(os.path.join(args.log_dir, 'config.yaml'))
print(OmegaConf.to_yaml(cfg))
ngpus = torch.cuda.device_count()
print(f'using {ngpus} GPUs')
############# setting seeds
torch.backends.cudnn.deterministic = True
device = "cuda:{}".format(local_rank)
seed = cfg.training.seed
torch.cuda.set_device(device)
torch.manual_seed(seed + local_rank)
np.random.seed(seed + local_rank)
torch.backends.cudnn.benchmark = False
torch.cuda.manual_seed_all(seed + local_rank)
random.seed(seed + local_rank)
os.environ['PYTHONHASHSEED'] = str(seed + local_rank)
############# Model
unet, diffusion = create_model_and_diffusion(**cfg.diffusion)
ckpt_path = os.path.join(args.log_dir, 'sampler.pth')
output_path = os.path.join(args.log_dir, f'generated')
mkdir_p(output_path)
ckpt_dict = torch.load(ckpt_path)
print0(f'checkpoint loaded from {ckpt_path}')
print0(f'Best FID: {ckpt_dict["fid"]}')
print0(f'iter: {ckpt_dict["i_iter"]}')
sampler = OpenAIDiffusion(unet, diffusion, **cfg.sampler)
sampler.net.load_state_dict(ckpt_dict['state_dict'])
sampler.net.to(device)
if cfg.diffusion.use_fp16:
unet.convert_to_fp16()
if args.guidance_scale is not None:
print("Loading value function for guidance")
value_path = os.path.join(args.log_dir, 'value.pth')
if not os.path.exists(value_path):
raise ValueError("Value ftn not found at {}".format(value_path))
v = instantiate(cfg.value).to(device)
checkpoint = torch.load(value_path, map_location=device)
v.load_state_dict(checkpoint['state_dict'])
v.eval()
else:
v = None
if args.guidance_scale is not None:
trainer = instantiate(cfg.trainer, batchsize=args.batchsize)
trainer.set_models(f=None, v=v, sampler=sampler, optimizer=None,
optimizer_fstar=None, optimizer_v=None)
############# DDP
torch.distributed.init_process_group(backend="nccl", init_method="env://")
sampler.net = DDP(unet, device_ids=[local_rank], output_device=local_rank)
############# FID
if not args.skip_fid:
from pytorch_fid.inception import InceptionV3
from pytorch_fid.fid_score import get_activations_from_tensor, calculate_frechet_distance
dims = 2048
block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
extractor = InceptionV3([block_idx]).to(device)
extractor.eval()
if cfg.data.name == 'imagenet64':
d_fid_stats = np.load('datasets/VIRTUAL_imagenet64_labeled.npz')
elif cfg.data.name == 'lsun_bedroom':
d_fid_stats = np.load('datasets/VIRTUAL_lsun_bedroom256.npz')
else:
raise NotImplementedError
m2, s2 = torch.tensor(d_fid_stats['mu']), torch.tensor(d_fid_stats['sigma'])
samples = fid()