-
Notifications
You must be signed in to change notification settings - Fork 0
/
multi_label_train.py
294 lines (256 loc) · 13.4 KB
/
multi_label_train.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
# Copyright (c) Facebook, Inc. and its affiliates.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os, sys
import argparse
import json
from pathlib import Path
import torch
from torch import nn
import torch.backends.cudnn as cudnn
from torchvision import transforms
from torchvision import models as torchvision_models
import utils
import vision_transformer as vits
from torch.utils.data import Dataset
from PIL import Image
def eval_linear(args):
utils.init_distributed_mode(args)
print("git:\n {}\n".format(utils.get_sha()))
print("\n".join("%s: %s" % (k, str(v)) for k, v in sorted(dict(vars(args)).items())))
cudnn.benchmark = True
# ============ building network ... ============
# if the network is a Vision Transformer (i.e. vit_tiny, vit_small, vit_base)
if args.arch in vits.__dict__.keys():
model = vits.__dict__[args.arch](patch_size=args.patch_size, num_classes=0)
embed_dim = model.embed_dim * (args.n_last_blocks + int(args.avgpool_patchtokens))
# otherwise, we check if the architecture is in torchvision models
elif args.arch in torchvision_models.__dict__.keys():
model = torchvision_models.__dict__[args.arch]()
embed_dim = model.fc.weight.shape[1]
model.fc = nn.Identity()
else:
print(f"Unknow architecture: {args.arch}")
sys.exit(1)
model.cuda()
model.eval()
# load weights to evaluate
utils.load_pretrained_weights(model, args.pretrained_weights, args.checkpoint_key, args.arch, args.patch_size)
print(f"Model {args.arch} built.")
# construct a linear classifier head
linear_classifier = LinearClassifier(embed_dim, num_labels=args.num_labels)
linear_classifier = linear_classifier.cuda()
linear_classifier = nn.parallel.DistributedDataParallel(linear_classifier, device_ids=[args.gpu])
# ============ preparing data ... ============
val_transform = transforms.Compose([
transforms.Resize(360, interpolation=3), # 3 is bicubic interpolation
# transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# dataset_val = datasets.ImageFolder(os.path.join(args.data_path, "val"), transform=val_transform)
dataset_val = myDataset(os.path.join(args.data_path, "val"), os.path.join(args.data_path, "val.json"), transform=val_transform)
val_loader = torch.utils.data.DataLoader(
dataset_val,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
)
if args.evaluate:
utils.load_pretrained_linear_weights(linear_classifier, args.arch, args.patch_size)
test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens)
print(f"Accuracy of the network on the {len(dataset_val)} test images: {test_stats['acc1']:.1f}%")
return
train_transform = transforms.Compose([
transforms.Resize(360, interpolation=3), # 3 is bicubic interpolation
# transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# dataset_train = datasets.ImageFolder(os.path.join(args.data_path, "train"), transform=train_transform)
dataset_train = myDataset(os.path.join(args.data_path, "train"), os.path.join(args.data_path, "train.json"), transform=train_transform)
sampler = torch.utils.data.distributed.DistributedSampler(dataset_train)
train_loader = torch.utils.data.DataLoader(
dataset_train,
sampler=sampler,
batch_size=args.batch_size_per_gpu,
num_workers=args.num_workers,
pin_memory=True,
)
print(f"Data loaded with {len(dataset_train)} train and {len(dataset_val)} val imgs.")
# set optimizer
optimizer = torch.optim.SGD(
linear_classifier.parameters(), # only updated params of linear classifier layer
args.lr * (args.batch_size_per_gpu * utils.get_world_size()) / 256., # linear scaling rule
momentum=0.9,
weight_decay=0, # we do not apply weight decay
)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, args.epochs, eta_min=0)
# Optionally resume from a checkpoint
to_restore = {"epoch": 0, "best_acc": 0.}
utils.restart_from_checkpoint(
os.path.join(args.output_dir, "360_checkpoint.pth.tar"),
run_variables=to_restore,
state_dict=linear_classifier,
optimizer=optimizer,
scheduler=scheduler,
)
start_epoch = to_restore["epoch"]
best_acc = to_restore["best_acc"]
for epoch in range(start_epoch, args.epochs):
train_loader.sampler.set_epoch(epoch)
train_stats = train(model, linear_classifier, optimizer, train_loader, epoch, args.n_last_blocks, args.avgpool_patchtokens)
scheduler.step()
log_stats = {**{f'train_{k}': v for k, v in train_stats.items()},
'epoch': epoch}
if epoch % args.val_freq == 0 or epoch == args.epochs - 1:
test_stats = validate_network(val_loader, model, linear_classifier, args.n_last_blocks, args.avgpool_patchtokens)
print(f"Accuracy at epoch {epoch} of the network on the {len(dataset_val)} test images: {test_stats['acc']:.1f}%")
best_acc = max(best_acc, test_stats["acc"])
print(f'Max accuracy so far: {best_acc:.2f}%')
log_stats = {**{k: v for k, v in log_stats.items()},
**{f'test_{k}': v for k, v in test_stats.items()}}
if utils.is_main_process():
with (Path(args.output_dir) / "360_log.txt").open("a") as f:
f.write(json.dumps(log_stats) + "\n")
save_dict = {
"epoch": epoch + 1,
"state_dict": linear_classifier.state_dict(),
"optimizer": optimizer.state_dict(),
"scheduler": scheduler.state_dict(),
"best_acc": best_acc,
}
torch.save(save_dict, os.path.join(args.output_dir, "checkpoint.pth.tar"))
print("Training of the supervised linear classifier on frozen features completed.\n"
"Test accuracy: {acc:.1f}".format(acc=best_acc))
def train(model, linear_classifier, optimizer, loader, epoch, n, avgpool):
linear_classifier.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
for (inp, target) in metric_logger.log_every(loader, 20, header): # log every 20 iterations
# move to gpu
inp = inp.cuda(non_blocking=True)
target = target.cuda(non_blocking=True)
# forward
with torch.no_grad():
if "vit" in args.arch:
intermediate_output = model.get_intermediate_layers(inp, n)
output = torch.cat([x[:, 0] for x in intermediate_output], dim=-1)
if avgpool:
output = torch.cat((output.unsqueeze(-1), torch.mean(intermediate_output[-1][:, 1:], dim=1).unsqueeze(-1)), dim=-1)
output = output.reshape(output.shape[0], -1)
else:
output = model(inp)
output = linear_classifier(output)
# compute cross entropy loss
# loss = nn.CrossEntropyLoss()(output, target)
loss = nn.BCEWithLogitsLoss()(output, target.float())
# compute the gradients
optimizer.zero_grad()
loss.backward()
# step
optimizer.step()
# log
torch.cuda.synchronize()
metric_logger.update(loss=loss.item())
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def validate_network(val_loader, model, linear_classifier, n, avgpool):
linear_classifier.eval()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
for inp, target in metric_logger.log_every(val_loader, 2, header):
# move to gpu
inp = inp.cuda(non_blocking=True) # used when pin_memory=True
target = target.cuda(non_blocking=True)
# forward
with torch.no_grad():
if "vit" in args.arch:
intermediate_output = model.get_intermediate_layers(inp, n)
output = torch.cat([x[:, 0] for x in intermediate_output], dim=-1)
if avgpool:
output = torch.cat((output.unsqueeze(-1), torch.mean(intermediate_output[-1][:, 1:], dim=1).unsqueeze(-1)), dim=-1)
output = output.reshape(output.shape[0], -1)
else:
output = model(inp)
output = linear_classifier(output)
loss = nn.BCEWithLogitsLoss()(output, target.float())
acc = utils.multilabel_accuracy(output, target)
batch_size = inp.shape[0]
metric_logger.update(loss=loss.item())
metric_logger.meters['acc'].update(acc.item(), n=batch_size)
print('* Multi-label Acc {acc:.3f} loss {losses.global_avg:.3f}'
.format(acc=acc, losses=metric_logger.loss))
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
class LinearClassifier(nn.Module):
"""Linear layer to train on top of frozen features"""
def __init__(self, dim, num_labels=1000):
super(LinearClassifier, self).__init__()
self.num_labels = num_labels
self.linear = nn.Linear(dim, num_labels)
self.linear.weight.data.normal_(mean=0.0, std=0.01)
self.linear.bias.data.zero_()
def forward(self, x):
# flatten
x = x.view(x.size(0), -1)
# linear layer
return self.linear(x)
# below is my code
class myDataset(Dataset):
def __init__(self, data_dir, label_dir, transform=None):
self.label_dir = label_dir
self.transform = transform
self.images = [os.path.join(data_dir, f) for f in os.listdir(data_dir) if f.endswith('.jpg')]
self.labels = json.load(open(label_dir))
def __len__(self):
return len(self.images)
def __getitem__(self, idx):
image = Image.open(self.images[idx])
image_name = self.images[idx].split('/')[-1]
image = self.transform(image)
return image, torch.tensor(self.labels[image_name])
# above is my code
if __name__ == '__main__':
parser = argparse.ArgumentParser('Evaluation with linear classification on ImageNet')
parser.add_argument('--n_last_blocks', default=4, type=int, help="""Concatenate [CLS] tokens
for the `n` last blocks. We use `n=4` when evaluating ViT-Small and `n=1` with ViT-Base.""")
parser.add_argument('--avgpool_patchtokens', default=False, type=utils.bool_flag,
help="""Whether ot not to concatenate the global average pooled features to the [CLS] token.
We typically set this to False for ViT-Small and to True with ViT-Base.""")
parser.add_argument('--arch', default='vit_small', type=str, help='Architecture')
parser.add_argument('--patch_size', default=16, type=int, help='Patch resolution of the model.')
parser.add_argument('--pretrained_weights', default='', type=str, help="Path to pretrained weights to evaluate.")
parser.add_argument("--checkpoint_key", default="teacher", type=str, help='Key to use in the checkpoint (example: "teacher")')
parser.add_argument('--epochs', default=100, type=int, help='Number of epochs of training.')
parser.add_argument("--lr", default=0.001, type=float, help="""Learning rate at the beginning of
training (highest LR used during training). The learning rate is linearly scaled
with the batch size, and specified here for a reference batch size of 256.
We recommend tweaking the LR depending on the checkpoint evaluated.""")
parser.add_argument('--batch_size_per_gpu', default=128, type=int, help='Per-GPU batch-size')
parser.add_argument("--dist_url", default="env://", type=str, help="""url used to set up
distributed training; see https://pytorch.org/docs/stable/distributed.html""")
parser.add_argument("--local_rank", default=0, type=int, help="Please ignore and do not set this argument.")
parser.add_argument('--data_path', default='/path/to/imagenet/', type=str)
parser.add_argument('--num_workers', default=10, type=int, help='Number of data loading workers per GPU.')
parser.add_argument('--val_freq', default=1, type=int, help="Epoch frequency for validation.")
parser.add_argument('--output_dir', default=".", help='Path to save logs and checkpoints')
parser.add_argument('--num_labels', default=1000, type=int, help='Number of labels for linear classifier')
parser.add_argument('--evaluate', dest='evaluate', action='store_true', help='evaluate model on validation set')
args = parser.parse_args()
eval_linear(args)