-
Notifications
You must be signed in to change notification settings - Fork 7
/
ft_acti.py
219 lines (180 loc) · 9.02 KB
/
ft_acti.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
import os
import torch.nn as nn
from datasets import Breakfast_feat
from torch.utils.data import DataLoader
from tqdm import tqdm
import wandb
import argparse
import shutil
from pathlib import Path
import yaml
from dotmap import DotMap
import pprint
from modules.fusion_module import fusion_base
from test_acti import validate
from utils.KLLoss import KLLoss
from utils.Augmentation import *
from utils.solver import _optimizer, _lr_scheduler
from utils.tools import *
from utils.text_prompt import *
from utils.saving import *
class TextCLIP(nn.Module):
def __init__(self, model):
super(TextCLIP, self).__init__()
self.model = model
def forward(self, text):
return self.model.encode_text(text)
class ImageCLIP(nn.Module):
def __init__(self, model):
super(ImageCLIP, self).__init__()
self.model = model
def forward(self, image):
return self.model.encode_image(image)
def main():
global args, best_prec1
global global_step
parser = argparse.ArgumentParser()
parser.add_argument('--config', '-cfg', default='./configs/breakfast/breakfast_acti_ft.yaml')
parser.add_argument('--log_time', default='')
parser.add_argument('--name', default='Transcls_ls')
args = parser.parse_args()
with open(args.config, 'r') as f:
config = yaml.load(f)
working_dir = os.path.join('./exp', config['network']['type'], config['network']['arch'], config['data']['dataset'],
args.log_time)
wandb.init(project=config['network']['type'],
name='{}_{}_{}_{}_{}'.format(args.log_time, config['network']['type'], config['network']['arch'],
config['data']['dataset'], args.name))
print('-' * 80)
print(' ' * 20, "working dir: {}".format(working_dir))
print('-' * 80)
print('-' * 80)
print(' ' * 30, "Config")
pp = pprint.PrettyPrinter(indent=4)
pp.pprint(config)
print('-' * 80)
config = DotMap(config)
Path(working_dir).mkdir(parents=True, exist_ok=True)
shutil.copy(args.config, working_dir)
shutil.copy('ft_acti.py', working_dir)
device = "cuda" if torch.cuda.is_available() else "cpu" # If using GPU then use mixed precision training.
model, clip_state_dict = clip.load(config.network.arch, device=device, jit=False, tsm=config.network.tsm,
T=config.data.num_segments, dropout=config.network.drop_out,
emb_dropout=config.network.emb_dropout, pretrain=config.network.init,
joint=config.network.joint) # Must set jit=False for training ViT-B/32
fusion_model = fusion_base(config.network.sim_header, clip_state_dict, config.data.num_frames)
fusion_model_up = fusion_base("Transf_cls", clip_state_dict, config.data.num_segments)
model_text = TextCLIP(model)
model_text = torch.nn.DataParallel(model_text).cuda()
fusion_model = torch.nn.DataParallel(fusion_model).cuda()
fusion_model_up = torch.nn.DataParallel(fusion_model_up).cuda()
wandb.watch(model)
wandb.watch(fusion_model)
train_data = Breakfast_feat(mode='train', num_frames=config.data.num_frames,
n_split=config.data.n_split, n_seg=config.data.num_segments)
train_loader = DataLoader(train_data, batch_size=config.data.batch_size, num_workers=config.data.workers,
shuffle=True, pin_memory=True, drop_last=True)
val_data = Breakfast_feat(mode='val', num_frames=config.data.num_frames,
n_split=config.data.n_split, n_seg=config.data.num_segments)
val_loader = DataLoader(val_data, batch_size=config.data.batch_size, num_workers=config.data.workers, shuffle=False,
pin_memory=True, drop_last=False)
if device == "cpu":
model_text.float()
# model_image.float()
else:
clip.model.convert_weights(
model_text) # Actually this line is unnecessary since clip by default already on float16
# clip.model.convert_weights(model_image)
loss_img = KLLoss()
loss_txt = KLLoss()
start_epoch = config.solver.start_epoch
if config.pretrain:
if os.path.isfile(config.pretrain):
print(("=> loading checkpoint '{}'".format(config.pretrain)))
checkpoint = torch.load(config.pretrain)
model.load_state_dict(checkpoint['model_state_dict'])
fusion_model.load_state_dict(checkpoint['fusion_model_state_dict'])
del checkpoint
else:
print(("=> no checkpoint found at '{}'".format(config.resume)))
if config.resume:
if os.path.isfile(config.resume):
print(("=> loading checkpoint '{}'".format(config.resume)))
checkpoint = torch.load(config.resume)
model.load_state_dict(checkpoint['model_state_dict'])
fusion_model.load_state_dict(checkpoint['fusion_model_state_dict'])
start_epoch = checkpoint['epoch']
print(("=> loaded checkpoint '{}' (epoch {})"
.format(config.evaluate, start_epoch)))
del checkpoint
else:
print(("=> no checkpoint found at '{}'".format(config.pretrain)))
classes, num_text_aug, text_dict = text_prompt_single(train_data.classes)
optimizer = _optimizer(config, model, fusion_model)
lr_scheduler = _lr_scheduler(config, optimizer)
scale = 768 ** -0.5
proj = nn.Parameter(scale * torch.randn(768, 512)).half().to(device)
best_prec1 = 0.0
if config.solver.evaluate:
prec1 = validate(start_epoch, val_loader, classes, device, model, fusion_model, config, num_text_aug, proj)
return
for k, v in model.named_parameters():
print('{}: {}'.format(k, v.requires_grad))
for epoch in range(start_epoch, config.solver.epochs):
# model_image.train()
model_text.train()
fusion_model.train()
fusion_model_up.train()
for kkk, (image_embedding, list_id) in enumerate(tqdm(train_loader)):
if config.solver.type != 'monitor':
if (kkk + 1) == 1 or (kkk + 1) % 10 == 0:
lr_scheduler.step(epoch + kkk / len(train_loader))
optimizer.zero_grad()
# images = images.view((-1, config.data.num_segments, 3) + images.size()[-2:])
b, n, f, d = image_embedding.size()
text_id = numpy.random.randint(num_text_aug, size=len(list_id))
texts = torch.stack([text_dict[j][i, :] for i, j in zip(list_id, text_id)])
image_embedding = image_embedding.to(device, non_blocking=True)
# image_embedding = image_embedding.mean(dim=2, keepdim=False).half()
image_embedding = image_embedding.half() @ proj
# images = images.to(device).view(-1, c, h,
# w) # omit the Image.fromarray if the images already in PIL format,
# change this line to images=list_image if using preprocess inside the dataset class
texts = texts.to(device, non_blocking=True)
# image_embedding = model_image(images)
# image_embedding = image_embedding.view(b, t, -1)
image_embedding = image_embedding.view(-1, f, 512)
image_embedding = fusion_model(image_embedding)
image_embedding = image_embedding.view(b, n, 512)
image_embedding = fusion_model_up(image_embedding)
text_embedding = model_text(texts)
if config.network.fix_text:
text_embedding.detach_()
logit_scale = model.logit_scale.exp()
logits_per_image, logits_per_text = create_logits(image_embedding, text_embedding, logit_scale)
ground_truth = torch.tensor(gen_label(list_id), dtype=image_embedding.dtype, device=device)
loss_imgs = loss_img(logits_per_image, ground_truth)
loss_texts = loss_txt(logits_per_text, ground_truth)
total_loss = (loss_imgs + loss_texts) / 2
wandb.log({"train_total_loss": total_loss, "lr": optimizer.param_groups[0]['lr'],
"train_loss_imgs": loss_imgs, "train_loss_texts": loss_texts})
total_loss.backward()
if device == "cpu":
optimizer.step()
else:
convert_models_to_fp32(model)
optimizer.step()
clip.model.convert_weights(model)
# if epoch % config.logging.eval_freq == 0: # and epoch>0
prec1 = validate(epoch, val_loader, classes, device, model, fusion_model, fusion_model_up, config, num_text_aug, proj)
wandb.log({"val_acc": prec1})
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
print('Testing: {}/{}'.format(prec1, best_prec1))
print('Saving:')
filename = "{}/last_model.pt".format(working_dir)
epoch_saving(epoch, model, fusion_model, optimizer, filename)
if is_best:
best_saving(working_dir, epoch, model, fusion_model, optimizer)
if __name__ == '__main__':
main()