-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain_ddp.py
191 lines (162 loc) · 7.24 KB
/
main_ddp.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
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import numpy as np
import torchvision.datasets
from torch.utils.data import DataLoader
import argparse
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
import glob
import webdataset as wds
import torch.nn.functional as F
from torch.cuda.amp import autocast, GradScaler
from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
from models.model import CLIPEncoder, Decoder
from torch.utils.tensorboard import SummaryWriter
import torch.profiler
import time
from util import *
class DynamicInfoNCELoss(nn.Module):
def __init__(self, initial_temp=1, final_temp=0.07, total_batches=10000):
super().__init__()
self.initial_temp = initial_temp
self.final_temp = final_temp
self.total_batches = total_batches
self.current_temp = initial_temp
self.decay_rate = self.calculate_decay_rate()
def forward(self, embeddings1, embeddings2):
embeddings1 = F.normalize(embeddings1, p=2, dim=1)
embeddings2 = F.normalize(embeddings2, p=2, dim=1)
similarity_matrix = torch.matmul(embeddings1, embeddings2.T) / self.current_temp
labels = torch.arange(similarity_matrix.shape[0], device=similarity_matrix.device)
loss = F.cross_entropy(similarity_matrix, labels)
return loss
def update_temperature(self, batch_count):
self.current_temp = self.initial_temp * np.exp(-self.decay_rate * batch_count)
def calculate_decay_rate(self):
return -np.log(self.final_temp / self.initial_temp) / self.total_batches
def enumerate_report(seq, delta, growth=1.0):
last = 0
count = 0
for count, item in enumerate(seq):
now = time.time()
if now - last > delta:
last = now
yield count, item, True
else:
yield count, item, False
delta *= growth
def make_dataloader(tar_dir, batch_size, world_size, rank):
"""Create a DataLoader for training with WebDataset."""
transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(degrees=15),
transforms.ToTensor(),
])
def handle_sample(sample):
image, text = sample
return transform(image), text
tar_files = glob.glob(os.path.join(tar_dir, "*.tar"))
dataset = (wds.WebDataset(tar_files, resampled=True, shardshuffle=True)
.shuffle(5000)
.decode("pil")
.to_tuple("jpg", "txt")
.map(handle_sample)
.batched(batch_size)
.with_epoch(len(tar_files))
# .prepare(5)
# .cache(size=5000)
)
loader = wds.WebLoader(dataset, batch_size=None, num_workers=8, pin_memory=True)
# loader = loader.unbatched().shuffle(1000).batched(batch_size)
return loader
def main_worker(args):
args.local_rank = int(os.environ["LOCAL_RANK"])
dist.init_process_group(backend='gloo', init_method=args.dist_url, world_size=args.world_size, rank=args.local_rank)
torch.cuda.set_device(args.local_rank)
clip_encoder = CLIPEncoder('ViT-B/32').to(args.local_rank)
decoder = Decoder(embed_dim=512).to(args.local_rank)
clip_encoder = DistributedDataParallel(clip_encoder, device_ids=[args.local_rank])
decoder = DistributedDataParallel(decoder, device_ids=[args.local_rank])
for param in clip_encoder.parameters():
param.requires_grad = False
# optimizer = optim.Lamb(decoder.parameters(),
# lr=1e-4,
# betas=(0.9, 0.999))
optimizer = torch.optim.AdamW(decoder.parameters(), args.lr)
scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=5000, T_mult=1)
scaler = GradScaler()
# writer = SummaryWriter(log_dir='./runs') if args.local_rank == 0 else None
criterion = DynamicInfoNCELoss()
train_loader = make_dataloader(args.tar_dir, args.batch_size, args.world_size, args.local_rank)
global_step = 0
start_epoch = 0
if args.checkpoint:
checkpoint = torch.load(args.checkpoint, map_location=f'cuda:{args.local_rank}')
decoder.module.load_state_dict(checkpoint['decoder_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
# global_step = checkpoint.get('global_step', 0)
# start_epoch = checkpoint.get('epoch', 0) + 1
else:
global_step = 0
start_epoch = 0
for epoch in range(start_epoch, args.epoch):
total_loss = 0
total_observer = 0
count = 0
for batch_idx, (images, text) in enumerate(train_loader):
images = images.to(args.local_rank)
optimizer.zero_grad()
with torch.no_grad():
e1 = clip_encoder.module.encode_img(images)
with autocast():
noise = decoder(e1)
noise = torch.clamp(noise, -args.eps, args.eps)
images_adv = []
for _ in range(args.chunk):
images_adv.append(torch.clamp(noise + images[torch.randperm(images.size(0))], 0, 1))
images_adv = torch.cat(images_adv, dim=0)
e2 = clip_encoder.module.encode_img(images_adv)
e2_chunks = torch.chunk(e2, args.chunk, dim=0)
sum_tensor = torch.zeros_like(e2_chunks[0])
for chunk in e2_chunks:
sum_tensor += chunk
e2 = sum_tensor / args.chunk
loss = criterion(e1, e2)
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
total_loss += loss.item()
count += 1
global_step += 1
scheduler.step()
if batch_idx <= 10000:
criterion.update_temperature(batch_idx)
if batch_idx % 500 == 0 and args.local_rank == 0:
avg_loss = total_loss / count
current_lr = optimizer.param_groups[0]['lr']
print(f'Batch {batch_idx}, Loss: {total_loss / count:.6f}, lr: {current_lr}')
torch.save({
'global_step': global_step,
'decoder_state_dict': decoder.module.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict(),
}, f"checkpoints/pre-trained.pt")
# if args.local_rank == 0:
# writer.close()
dist.destroy_process_group()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--tar_dir", type=str)
parser.add_argument("--epoch", type=int, default=10)
parser.add_argument("--batch_size", type=int, default=600)
parser.add_argument("--dist_url", type=str, default="tcp://127.0.0.1:23456")
parser.add_argument("--chunk", type=int, default=5)
parser.add_argument("--eps", type=float, default=16 / 255)
parser.add_argument("--lr", type=float, default=1e-4)
parser.add_argument("--checkpoint", type=str, default=None, help="path to checkpoint to load")
args = parser.parse_args()
ngpus_per_node = torch.cuda.device_count()
args.world_size = ngpus_per_node
main_worker(args)