-
Notifications
You must be signed in to change notification settings - Fork 6
/
train_WebVision.py
205 lines (164 loc) · 9.58 KB
/
train_WebVision.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
192
193
194
195
196
197
198
199
200
import numpy as np
import torch.utils.data as data
from tqdm import tqdm
from utils.ema import EMA
from utils.clip_wrapper import clip_img_wrap
from utils.webvision_data_utils import WebVision
import torch.optim as optim
from utils.learning import *
from model_diffusion import Diffusion
from utils.knn_utils import sample_knn_labels, knn, knn_labels, prepare_knn
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
import argparse
torch.manual_seed(123)
torch.cuda.manual_seed(123)
np.random.seed(123)
random.seed(123)
def train(diffusion_model, val_loader, device, model_save_dir, args, data_dir):
# device = diffusion_model.device
k = args.k
n_epochs = args.nepoch
n_class = 50
val_embed = np.load(os.path.join(data_dir, 'fp_embed_val_webvision.npy'))
train_embed = torch.tensor(np.load(os.path.join(data_dir, 'fp_embed_train_webvision.npy'))).to(device)
train_labels = torch.tensor(np.load(os.path.join(data_dir, 'train_labels_webvision.npy'))).to(device)
diffusion_model.fp_encoder.eval()
params = list(diffusion_model.model.parameters()) + list(diffusion_model.diffusion_encoder.parameters())
optimizer = optim.Adam(params, lr=0.0001, weight_decay=0.0, betas=(0.9, 0.999),
amsgrad=False, eps=1e-08)
diffusion_loss = nn.MSELoss(reduction='none')
# diffusion_loss = nn.MSELoss()
ema_helper = EMA(mu=0.9999)
ema_helper.register(diffusion_model.model)
max_accuracy = 0
print('Diffusion training start')
for epoch in range(n_epochs):
train_dataset = WebVision(data_root=data_dir, split='train', balance=True, randomize=True, cls_size=500, transform='train')
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.num_workers, worker_init_fn=init_fn, drop_last=True)
diffusion_model.diffusion_encoder.train()
diffusion_model.model.train()
with tqdm(enumerate(train_loader), total=len(train_loader), desc=f'train diffusion epoch {epoch}', ncols=120) as pbar:
for i, (x_batch, y_batch, _) in pbar:
with torch.no_grad():
fp_embd = diffusion_model.fp_encoder(x_batch.to(device))
y_labels_batch, sample_weight = sample_knn_labels(fp_embd, y_batch.to(device), train_embed,
train_labels, k=k, n_class=n_class)
y_one_hot_batch, y_logits_batch = cast_label_to_one_hot_and_prototype(y_labels_batch.to(torch.int64),
n_class=n_class)
y_0_batch = y_one_hot_batch.to(device)
# adjust_learning_rate
adjust_learning_rate(optimizer, i / len(train_loader) + epoch, warmup_epochs=1, n_epochs=n_epochs, lr_input=0.001)
n = x_batch.size(0)
# antithetic sampling
t = torch.randint(low=0, high=diffusion_model.num_timesteps, size=(n // 2 + 1,)).to(device)
t = torch.cat([t, diffusion_model.num_timesteps - 1 - t], dim=0)[:n]
# train with and without prior
output, e = diffusion_model.forward_t(y_0_batch, x_batch.to(device), t, fp_embd)
# compute loss
mse_loss = diffusion_loss(e, output)
weighted_mse_loss = torch.matmul(sample_weight.to(device), mse_loss)
loss = torch.mean(weighted_mse_loss)
# loss = diffusion_loss(e, output)
pbar.set_postfix({'loss': loss.item()})
# optimize diffusion model that predicts eps_theta
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(diffusion_model.model.parameters(), 1.0)
torch.nn.utils.clip_grad_norm_(diffusion_model.diffusion_encoder.parameters(), 1.0)
optimizer.step()
ema_helper.update(diffusion_model.model)
acc_val = test(diffusion_model, val_loader, val_embed)
print(f"epoch: {epoch}, val accuracy: {acc_val:.2f}%")
if acc_val > max_accuracy:
if args.device is None:
states = [diffusion_model.model.module.state_dict(),
diffusion_model.diffusion_encoder.module.state_dict()]
else:
states = [diffusion_model.model.state_dict(),
diffusion_model.diffusion_encoder.state_dict()]
torch.save(states, model_save_dir)
print("Model saved, best test accuracy at Epoch {}.".format(epoch))
max_accuracy = max(max_accuracy, acc_val)
def test(diffusion_model, test_loader, test_embed):
if not torch.is_tensor(test_embed):
test_embed = torch.tensor(test_embed).to(torch.float)
correct_cnt = 0
with torch.no_grad():
diffusion_model.model.eval()
diffusion_model.diffusion_encoder.eval()
diffusion_model.fp_encoder.eval()
for test_batch_idx, data_batch in tqdm(enumerate(test_loader), total=len(test_loader), desc=f'evaluating diff', ncols=100):
[x_batch, target, indicies] = data_batch[:3]
target = target.to(device)
fp_embed = test_embed[indicies, :].to(device)
label_t_0 = diffusion_model.reverse_ddim(x_batch, stochastic=False, fp_x=fp_embed).detach().cpu()
# acc_temp = accuracy(label_t_0.detach().cpu(), target.cpu())[0].item()
# acc_avg += acc_temp
correct = cnt_agree(label_t_0.detach().cpu(), target.cpu())
correct_cnt += correct
acc = 100 * correct_cnt / test_embed.shape[0]
return acc
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--nepoch", default=300, help="number of training epochs", type=int)
parser.add_argument("--batch_size", default=256, help="batch_size", type=int)
parser.add_argument("--num_workers", default=16, help="num_workers", type=int)
parser.add_argument("--warmup_epochs", default=1, help="warmup_epochs", type=int)
parser.add_argument("--feature_dim", default=1024, help="feature_dim", type=int)
parser.add_argument("--k", default=50, help="k neighbors for knn", type=int)
parser.add_argument("--ddim_n_step", default=10, help="number of steps in ddim", type=int)
parser.add_argument("--diff_encoder", default='resnet50_l', help="which encoder for diffusion", type=str)
parser.add_argument("--gpu_devices", default=[0, 1, 2, 3], type=int, nargs='+', help="")
parser.add_argument("--device", default=None, help="which cuda to use", type=str)
args = parser.parse_args()
if args.device is None:
gpu_devices = ','.join([str(id) for id in args.gpu_devices])
os.environ["CUDA_VISIBLE_DEVICES"] = gpu_devices
device = 'cuda' if torch.cuda.is_available() else 'cpu'
else:
device = args.device
n_class = 50
# load datasets WebVsison
webvision_dir = os.path.join(os.getcwd(), 'WebVision')
print('data_dir', webvision_dir)
train_dataset = WebVision(data_root=webvision_dir, split='train', balance=False, randomize=False, cls_size=500,
transform='val')
train_labels = torch.tensor(train_dataset.targets).to(torch.long)
np.save(os.path.join(webvision_dir, f'train_labels_webvision.npy'), train_labels)
val_dataset = WebVision(data_root=webvision_dir, split='val')
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=512, shuffle=False, num_workers=args.num_workers)
# initialize diffusion model
fp_encoder = clip_img_wrap('ViT-L/14', device, center=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))
fp_dim = fp_encoder.dim
model_path = './model/LRA-diffusion_WebVision.pt'
diffusion_model = Diffusion(fp_encoder, num_timesteps=1000, n_class=n_class, fp_dim=fp_dim, device=device,
feature_dim=args.feature_dim, encoder_type=args.diff_encoder,
ddim_num_steps=args.ddim_n_step, beta_schedule='cosine')
state_dict = torch.load(model_path, map_location=torch.device(device))
diffusion_model.load_diffusion_net(state_dict)
diffusion_model.fp_encoder.eval()
# DataParallel wrapper
if args.device is None:
print('using DataParallel')
diffusion_model.model = nn.DataParallel(diffusion_model.model).to(device)
diffusion_model.diffusion_encoder = nn.DataParallel(diffusion_model.diffusion_encoder).to(device)
diffusion_model.fp_encoder = nn.DataParallel(fp_encoder).to(device)
else:
print('using single gpu')
diffusion_model.to(device)
# # pre-compute for fp embeddings on training data
print('pre-computing fp embeddings for training data')
train_embed_dir = os.path.join(webvision_dir, f'fp_embed_train_webvision.npy')
train_embed = prepare_fp_x(diffusion_model.fp_encoder, train_dataset, train_embed_dir, device=device,
fp_dim=fp_dim, batch_size=200)
# for validation data
print('pre-computing fp embeddings for validation data for webvision')
val_embed_dir = os.path.join(webvision_dir, f'fp_embed_val_webvision.npy')
val_embed = prepare_fp_x(diffusion_model.fp_encoder, val_dataset, val_embed_dir, device=device,
fp_dim=fp_dim, batch_size=200)
max_accuracy = test(diffusion_model, val_loader, val_embed)
print('test webvision accuracy:', max_accuracy)
# train the diffusion model
train(diffusion_model, val_loader, device, model_path, args, data_dir=webvision_dir)