-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathadapter_boosting.py
176 lines (148 loc) · 6.74 KB
/
adapter_boosting.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
import argparse
from model import Adapted_Clip
from jittor import optim
from jittor import nn
import jclip as clip
from utls import get_text_features, get_train_imgs, get_eval_imgs, get_test_imgs
import jittor as jt
import numpy as np
from PIL import Image
import logging
from tqdm import tqdm
import time
logging.basicConfig(
level=logging.INFO, # log messages that is higher or equal to {level}
format='%(asctime)s - %(levelname)s - %(name)s - %(message)s'
)
def evaluate(eval_imgs, eval_labels, model , preprocess, text_features, eval_batch_size, alphas):
top_1_list = []
top_5_list = []
for i in tqdm(range(0, len(eval_imgs), eval_batch_size)):
batch_imgs = []
batch_labels = []
for j in range(eval_batch_size):
if i + j >= len(eval_imgs):
break
img = Image.open(eval_imgs[i + j])
img = preprocess(img).unsqueeze(0)
batch_imgs.append(img)
batch_labels.append(eval_labels[i + j])
batch_imgs = jt.concat(batch_imgs, dim=0)
batch_labels = jt.concat(batch_labels)
logits = model.boost_execute(batch_imgs, text_features, alphas)
_, top_labels = logits.topk(5, dim=-1)
for j in range(len(batch_labels)):
if top_labels[j][0] == batch_labels[j]:
top_1_list.append(1)
else:
top_1_list.append(0)
if batch_labels[j] in top_labels[j]:
top_5_list.append(1)
else:
top_5_list.append(0)
eval_top_1 = sum(top_1_list) / len(top_1_list)
eval_top_5 = sum(top_5_list) / len(top_5_list)
return eval_top_1, eval_top_5
def test(test_imgs, model, preprocess, text_features, test_batch_size, alphas):
save_file = open('result.txt', 'w')
preds = []
imgs_dir = '/data/TestSetA/'
with jt.no_grad():
for i in tqdm(range(0, len(test_imgs), test_batch_size)):
batch_imgs = []
for j in range(test_batch_size):
if i + j >= len(test_imgs):
break
img = Image.open(imgs_dir + test_imgs[i + j])
img = preprocess(img).unsqueeze(0)
batch_imgs.append(img)
batch_imgs = jt.concat(batch_imgs, dim=0)
logits = model.boost_execute(batch_imgs, text_features, alphas)
_, top_labels = logits.topk(5, dim=-1)
preds.extend(top_labels.tolist())
for j in range(len(batch_imgs)):
save_file.write(test_imgs[i + j] + ' ' + ' '.join([str(p.item()) for p in top_labels[j]]) + '\n')
def boosting_trainer(model, adapter_index, preprocess, alphas, args):
optimizer = optim.SGD(model.adapters[adapter_index].parameters(), lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
train_imgs, train_labels = get_train_imgs()
text_features = get_text_features(model.clip_model)
epoch_losses = []
for epoch in range(args.num_epochs):
batch_size = args.training_batch_size
# Shuffle the images
indexes = np.random.permutation(len(train_imgs))
train_imgs = np.array(train_imgs)
train_labels = np.array(train_labels)
train_imgs = train_imgs[indexes]
train_labels = train_labels[indexes]
losses = []
for i in range(0, len(train_imgs), batch_size):
batch_imgs = []
batch_labels = []
for j in range(batch_size):
if i + j >= len(train_imgs):
break
img = Image.open(train_imgs[i + j])
img = preprocess(img).unsqueeze(0)
batch_imgs.append(img)
batch_labels.append(train_labels[i + j])
batch_imgs = jt.concat(batch_imgs, dim=0)
batch_labels = jt.concat(batch_labels)
logits = model.execute(batch_imgs, text_features, adapter_index)
loss = criterion(logits, batch_labels)
optimizer.step(loss)
losses.append(loss[0])
epoch_losses.append(jt.mean(losses).item())
logging.info(f'Epoch: {epoch} Loss: {round(epoch_losses[-1], 3)}')
# Draw the loss curve
import matplotlib.pyplot as plt
plt.plot(epoch_losses)
plt.xlabel('Epoch')
plt.ylabel('Loss')
plt.title(f'Adapter {adapter_index} - Loss Curve')
plt.savefig(f'./fig/{args.adapter_type}/adapter_{adapter_index}_loss_curve.png')
# Caculate the alpha
alpha = 1
alphas.append(alpha)
if args.eval:
print('#Evaluating the model')
eval_imgs, eval_labels = get_eval_imgs()
eval_batch_size = args.eval_batch_size
with jt.no_grad():
eval_top_1, eval_top_5 = evaluate(eval_imgs, eval_labels, model, preprocess, text_features, eval_batch_size, [1 for _ in range(adapter_index + 1)])
print(f'Adapter {adapter_index} - Top 1 Accuracy: {eval_top_1}')
print(f'Adapter {adapter_index} - Top 5 Accuracy: {eval_top_5}')
if __name__ == '__main__':
jt.flags.use_cuda = 1
parser = argparse.ArgumentParser()
parser.add_argument('--training_batch_size', type=int, default=32)
parser.add_argument('--eval_batch_size', type=int, default=256)
parser.add_argument('--test_batch_size', type=int, default=256)
parser.add_argument('--num_epochs', type=int, default=50)
parser.add_argument('--num_adapters', type=int, default=5)
parser.add_argument('--learning_rate', type=float, default=1)
parser.add_argument('--adapter_type', type=str, default='conv', help='mlp, attn or conv')
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--eval', type=bool, default=False)
parser.add_argument('--test', type=bool, default=False)
args = parser.parse_args()
jt.misc.set_global_seed(args.seed)
start_time = time.time()
print('Start Time:', time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(start_time)))
for arg in vars(args):
print(f'{arg}: {getattr(args, arg)}')
clip_model, preprocess = clip.load("/weights/ViT-B-32.pkl")
model = Adapted_Clip(clip_model, args.num_adapters, adapter_type=args.adapter_type)
alphas = []
for i in range(args.num_adapters):
boosting_trainer(model, i, preprocess, alphas, args)
if args.test:
print('#Testing the model')
test_imgs = get_test_imgs()
text_features = get_text_features(clip_model)
test_batch_size = args.test_batch_size
test(test_imgs, model, preprocess, text_features, test_batch_size, alphas)
end_time = time.time()
print('End Time:', time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(end_time)))
print('Time Elapsed:', round(end_time - start_time, 2), 's')