-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_pf.py
160 lines (119 loc) · 6.33 KB
/
train_pf.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
"""
Author: Haoran Chen
Date: 2024.07.07
"""
import os
import math
import numpy as np
import logging
from tqdm import tqdm
from PIL import Image
import clip
import torch
from torch.utils.data import DataLoader
from torch import nn
import torch.nn.functional as F
from torch.optim.lr_scheduler import CosineAnnealingLR
from prompt import clip_text_prompt_genarator
from utils import extract_features
def test(task_id, test_loader, prompt, tokenized_prompt, custom_clip_model, vpt_model, args):
tot_acc = 0
tot = 0
custom_clip_model.eval()
vpt_model.eval()
scale = custom_clip_model.logit_scale.exp()
with torch.no_grad():
for data, label, _ in test_loader:
tot += data.size()[0]
data = data.to(args["device"])
label = label.to(args["device"])
img_features, txt_features = custom_clip_model(data, prompt, tokenized_prompt)
logits_clip = scale * img_features @ txt_features.t()
logits_vpt, _ = vpt_model(data)
logits_vpt = logits_vpt[:, :logits_clip.shape[1]]
logits = (1 - nn.Sigmoid()(custom_clip_model.lambda_)) * logits_clip + nn.Sigmoid()(custom_clip_model.lambda_) * logits_vpt
softmaxed_logits = nn.Softmax(dim=1)(logits)
output = torch.argmax(softmaxed_logits, dim=1)
tot_acc += (output == label).sum().item()
return format(tot_acc / tot, '.3f')
def train_pf(clip_model, custom_clip_model, vpt_model, scenario_train, scenario_test, classnames, memory, class_mask, args):
acc_table = np.zeros([args["step"], args["step"]])
for task_id, dataset_train in enumerate(scenario_train):
logging.info("Start Training Task {}".format(task_id + 1))
custom_clip_model.train()
vpt_model.train()
task_classnames = classnames[task_id * args["increment"] : (task_id + 1) * args["increment"]]
custom_clip_model.update_parameters(task_id, task_classnames, args)
optimizer = torch.optim.AdamW(list(custom_clip_model.parameters()) + list(vpt_model.parameters()), lr=args["learning_rate"], weight_decay=args["weight_decay"])
clip_parameters = sum(p.numel() for p in custom_clip_model.parameters() if p.requires_grad)
vpt_parameters = sum(p.numel() for p in vpt_model.parameters() if p.requires_grad)
logging.info('number of trainable params: {}'.format(clip_parameters + vpt_parameters))
tot_clip_parameters = sum(p.numel() for p in custom_clip_model.parameters())
tot_vpt_parameters = sum(p.numel() for p in vpt_model.parameters())
logging.info('total number of params: {}'.format(tot_clip_parameters + tot_vpt_parameters))
scheduler = CosineAnnealingLR(optimizer, T_max=args["epoch"])
if task_id > 0:
mem_x, mem_y, mem_t = memory.get()
dataset_train.add_samples(mem_x, mem_y, mem_t)
train_loader = DataLoader(dataset_train, batch_size=args["batch_size"], drop_last=True, num_workers=4 * args["ngpus"], shuffle=True, pin_memory=True)
prog_bar = tqdm(range(args["epoch"]))
for _, epoch in enumerate(prog_bar):
losses = 0.0
for data, label, _ in train_loader:
data = data.to(args["device"])
label = label.to(args["device"])
label = label.to(torch.int64)
task_prompt, task_tokenized_prompt = clip_text_prompt_genarator(task_classnames, clip_model, custom_clip_model.task_raw_prompt)
if task_id > 0:
prompt = torch.cat([learned_prompt, task_prompt], dim=0)
tokenized_prompt = torch.cat([learned_tokenized_prompt, task_tokenized_prompt], dim=0)
else:
prompt = task_prompt
tokenized_prompt = task_tokenized_prompt
scale = custom_clip_model.logit_scale.exp()
image_feature, text_feature = custom_clip_model(data, prompt, tokenized_prompt)
logits_clip = scale * image_feature @ text_feature.t()
logits_vpt, _ = vpt_model(data)
logits_vpt = logits_vpt[:, :logits_clip.shape[1]]
logits = (1 - nn.Sigmoid()(custom_clip_model.lambda_)) * logits_clip + nn.Sigmoid()(custom_clip_model.lambda_) * logits_vpt
softmaxed_logits = F.log_softmax(logits, 1)
if task_id > 0:
weight = custom_clip_model.generate_weight_mask()
softmaxed_logits = softmaxed_logits * weight
loss = F.nll_loss(softmaxed_logits, label)
losses += loss.item()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(list(custom_clip_model.parameters()) + list(vpt_model.parameters()), 1.0)
optimizer.step()
scheduler.step()
info = "Task {}, Epoch {}/{} => Loss {:.3f}".format(
task_id + 1,
epoch + 1,
args["epoch"],
losses / len(train_loader)
)
prog_bar.set_description(info)
for task_id_test, dataset_test_ in enumerate(scenario_test):
if task_id_test > task_id:
break
else:
test_loader_ = DataLoader(dataset_test_, batch_size=1, drop_last=True,
num_workers=4 * args["ngpus"], shuffle=False)
acc = test(task_id, test_loader_, prompt, tokenized_prompt, custom_clip_model, vpt_model, args)
acc_table[task_id_test][task_id] = acc
cur_acc = acc_table[:, task_id]
logging.info("Task {} Average Accuracy is {:.3f}".format(task_id + 1, np.sum(cur_acc) / (task_id + 1)))
logging.info(acc_table)
learned_prompt = prompt
learned_tokenized_prompt = tokenized_prompt
if args["herding_method"] == 'barycenter':
features = extract_features(dataset_train, clip_model, args)
memory.add(*scenario_train[task_id].get_raw_samples(), features)
elif args["herding_method"] == 'closest':
pass
elif args["herding_method"] == 'random':
memory.add(*scenario_train[task_id].get_raw_samples(), None)
else:
raise Exception("Herding method not implemented")
return acc_table