-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_grounding.py
189 lines (167 loc) · 9.16 KB
/
train_grounding.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
import torch.optim as optim
import torch.utils.data
import torch.nn as nn
from tqdm import tqdm
import os
import numpy as np
from model import *
from datasets.flicker import get_flicker1K_dataset, get_flicker_dataset
from datasets.visual_genome import get_VG_dataset
from datasets.coco import get_coco_dataset
from utils import interpret_batch, interpret_new
import CLIP.clip as clip
from inference_grounding import inference_bbox
def norm_z(z):
return z / z.norm(dim=1).unsqueeze(dim=1)
def open_folder(path):
if not os.path.exists(path):
os.mkdir(path)
a = os.listdir(path)
os.mkdir(path + '/gpu' + str(len(a)))
return str(len(a))
def get_logits(clip_model, real_imgs, text_pos, text_neg):
logits_pos, _ = clip_model(real_imgs, text_pos)
logits_neg, _ = clip_model(real_imgs, text_neg)
logits_fr = torch.cat((logits_pos.diag().unsqueeze(-1),
logits_neg.diag().unsqueeze(-1)),
dim=1)
return logits_fr
def gen_step(optimizer_G, clip_model, real_imgs, text, model, criterion, args):
bs = real_imgs.shape[0]
gt = torch.zeros(bs).long().to('cuda:' + str(real_imgs.get_device()))
optimizer_G.zero_grad()
clip_model.to('cuda:' + str(real_imgs.get_device()))
model.to('cuda:' + str(real_imgs.get_device()))
device = "cuda" if torch.cuda.is_available() else "cpu"
text_pos = text[:, :, 0]
z_t = norm_z(clip_model.encode_text(text_pos))
real_imgs_224 = F.interpolate(real_imgs, size=(224, 224), mode="bilinear", align_corners=True)
cam = interpret_new(real_imgs_224.detach(), text_pos.detach(), clip_model, device).detach().clone().float()
cam = F.interpolate(cam, size=(int(args['Isize']), int(args['Isize'])), mode="bilinear", align_corners=True)
M = model(real_imgs, z_t)
clip_cam_loss = F.mse_loss(M, cam)
M = F.interpolate(M, size=(224, 224), mode="bilinear", align_corners=True)
z_fr = norm_z(clip_model.encode_image(real_imgs_224 * M))
z_bg = norm_z(clip_model.encode_image(real_imgs_224 * (1 - M)))
regularization = M.mean()
fr_loss = (1 - (z_fr @ z_t.T)).mean()
bg_loss = torch.abs(z_bg @ z_t.T).mean()
loss = float(args['w3']) * fr_loss + \
float(args['w0']) * regularization +\
float(args['w1']) * clip_cam_loss +\
float(args['w2']) * bg_loss.mean()
loss.backward()
optimizer_G.step()
return loss.item(), 0
def logger(writer, loss_list, tplt_loss, step):
writer.add_scalar('Loss', loss_list, global_step=step)
writer.add_scalar('tplt_loss', tplt_loss, global_step=step)
def train(ds, model, clip_model, optimizer_G, args):
loss_list = []
pbar = tqdm(ds)
criterion = nn.CrossEntropyLoss()
for i, inputs in enumerate(pbar):
real_imgs = inputs[0].cuda()
text_pos = inputs[1]
text_pos_token = clip.tokenize(text_pos).to('cuda').unsqueeze(dim=2)
loss, tplt_loss = gen_step(optimizer_G, clip_model, real_imgs, text_pos_token, model, criterion, args)
loss_list.append(loss)
pbar.set_description('(train) :: loss {loss:.4f}'.format(loss=np.mean(loss_list)))
return np.mean(loss_list)
def main(args=None):
args['is_blip'] = False
gpu_num = torch.cuda.device_count()
model = MultiModel(args=args)
model = torch.nn.DataParallel(model, list(range(gpu_num))).cuda()
if bool(int(args['resume'])):
model1 = torch.load(args['resume_folder'])
model.load_state_dict(model1.state_dict())
optimizer_G = optim.SGD(model.parameters(),
lr=float(args['learning_rate']),
weight_decay=float(args['WD']),
momentum=float(args['M']))
if args['task'] == 'flicker':
trainset, testset = get_flicker_dataset(args=args)
elif args['task'] == 'vg_train':
trainset = get_VG_dataset(args=args)
testset = get_flicker1K_dataset(args=args)
elif args['task'] == 'coco':
trainset = get_coco_dataset(args=args)
testset = get_flicker1K_dataset(args=args)
ds = torch.utils.data.DataLoader(trainset,
batch_size=int(args['Batch_size']),
num_workers=int(args['nW']),
shuffle=True,
drop_last=True)
ds_test = torch.utils.data.DataLoader(testset,
batch_size=1,
num_workers=int(args['nW_eval']),
shuffle=False,
drop_last=False)
model.train()
results_path = os.path.join('results',
'gpu' + args['folder'],
'results.csv')
best_path = os.path.join('results',
'gpu' + args['folder'],
'best.csv')
f_all = open(results_path, 'w')
f_best = open(best_path, 'w')
f_all.write('epoches,label,acc\n')
f_best.write('epoches,acc\n')
best = 0
device = "cuda" if torch.cuda.is_available() else "cpu"
clip_model, _ = clip.load("ViT-B/32", device=device, jit=False)
for epoch in range(int(args['epoches'])):
train(ds, model.train(), clip_model.eval(), optimizer_G, args)
acc = inference_bbox(ds_test, model.eval(), clip_model.eval(), epoch, args)
f_all.write(str(epoch) + ',' + str('test') + ',' + str(acc) + '\n')
f_all.flush()
if acc > best:
torch.save(model, args['path_best'])
best = acc
f_best.write(str(epoch) + ',' + str(acc) + '\n')
f_best.flush()
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description='Description of your program')
parser.add_argument('-lr', '--learning_rate', default=0.0012, help='learning_rate', required=False)
parser.add_argument('-bs', '--Batch_size', default=32, help='batch_size', required=False)
parser.add_argument('-epoches', '--epoches', default=5000, help='number of epoches', required=False)
parser.add_argument('-nW', '--nW', default=0, help='number of workers', required=False)
parser.add_argument('-nW_eval', '--nW_eval', default=0, help='number of workers', required=False)
parser.add_argument('-WD', '--WD', default=1e-4, help='weight decay', required=False)
parser.add_argument('-order_ae', '--order_ae', default=16, help='order of the backbone - ae', required=False)
parser.add_argument('-backbone', '--backbone', default='vgg', help='order of the backbone - ae', required=False)
parser.add_argument('-task', '--task', default='vg_train', help='dataset task', required=False)
parser.add_argument('-dataset', '--dataset', default='flicker', help='dataset task', required=False)
parser.add_argument('-Isize', '--Isize', default=304, help='image size', required=False)
parser.add_argument('-nC', '--nC', default=200, help='number of classes', required=False)
parser.add_argument('-th', '--th', default=0.1, help='evaluation th', required=False)
parser.add_argument('-temp', '--temp', default=1, help='pretrined models', required=False)
parser.add_argument('-w0', '--w0', default=0.25, help='pretrined models', required=False)
parser.add_argument('-w1', '--w1', default=16, help='pretrined models', required=False)
parser.add_argument('-w2', '--w2', default=0.5, help='pretrined models', required=False)
parser.add_argument('-w3', '--w3', default=0.25, help='pretrined models', required=False)
parser.add_argument('-M', '--M', default=0.9, help='pretrined models', required=False)
parser.add_argument('-prob', '--prob', default=10, help='pretrined models', required=False)
parser.add_argument('-step_size', '--step_size', default=20, help='pretrined models', required=False)
parser.add_argument('-gamma', '--gamma', default=1, help='pretrined models', required=False)
parser.add_argument('-resume', '--resume', default=False, help='pretrined models', required=False)
parser.add_argument('-resume_folder', '--resume_folder', default='41', help='pretrined models', required=False)
parser.add_argument('-pretrained', '--pretrained', default=False, help='pretrined models', required=False)
parser.add_argument('-img_path', '--img_path', default=True, help='pretrined models', required=False)
parser.add_argument('-data_path', '--data_path',
default='/media/media1/talshah/coco/VG', help='data set path', required=False)
parser.add_argument('-val_path', '--val_path',
default=r'/media/media1/talshah/coco/flicker', help='data set path', required=False)
args = vars(parser.parse_args())
folder = open_folder('results')
Isize = str(args['Isize'])
args['folder'] = folder
args['path_best'] = os.path.join('results', 'gpu' + folder, 'net_best.pth')
args['resume_folder'] = os.path.join('results', 'gpu' + args['resume_folder'], 'net_best.pth')
args['path_save_init'] = os.path.join('results', 'gpu' + folder, 'net_init.pth')
args['path_init'] = os.path.join('results', 'init', 'cub',
str(args['backbone']) + str(args['order_ae']), 'net_init.pth')
main(args=args)