-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_ar.py
161 lines (141 loc) · 6.75 KB
/
train_ar.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
import argparse
import datetime
import json
import random
from sched import scheduler
import time
from pathlib import Path
import torch
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
from torch.utils.data import DataLoader, DistributedSampler
from torch.optim import Adam, AdamW
from torch.optim.lr_scheduler import StepLR, CosineAnnealingLR
import pytorch_warmup as warmup
from dataset.dataset_inshop import Inshop
from dataset.dataset_cub200 import CUB
from dataset.dataset_cars196 import Cars
from dataset.dataset_imagenet_val500 import ImageNet
from dataset.dataset_places365 import Places
from model.CLIP.clip import clip
from model.IRGen import IRGen
from model.DictTree import TreeNode
from utils.utils import LabelSmoothingCrossEntropy, train_transform, get_trainable_params
from utils.logger import get_logger
import os
import pickle
import scheduler
if __name__ == '__main__':
torch.distributed.init_process_group(backend="nccl")
local_rank = torch.distributed.get_rank()
torch.cuda.set_device(local_rank)
device = torch.device("cuda", local_rank)
world_size = dist.get_world_size()
parser = argparse.ArgumentParser(description='Train IRGen')
parser.add_argument('--data_dir', default='/mnt/default/data/isc/isc/Img', type=str, help='datasets path')
parser.add_argument('--data_name', default='car', type=str, choices=['car', 'cub', 'isc', 'imagenet', 'places'],
help='dataset name')
parser.add_argument('--file_name', default='in-shop_clothes_retrieval_trainval.pkl', type=str)
parser.add_argument('--codes', default='isc_256rq4_ids.pkl', type=str, help='image identifier')
parser.add_argument('--pretrained_model', default=None, type=str)
parser.add_argument('--output_dir', default='./', type=str)
parser.add_argument('--lr', default=8e-5, type=float, help='learning rate')
parser.add_argument('--batch_size', default=64, type=int, help='train batch size')
parser.add_argument('--num_epochs', default=400, type=int, help='train epoch number')
parser.add_argument('--smoothing', default=0.1, type=int, help='labelsmoothce smoothing')
parser.add_argument('--local_rank', default=0, type=int)
opt = parser.parse_args()
# args parse
data_name, data_dir = opt.data_name, opt.data_dir
file_name = opt.file_name
codes = opt.codes
pretrained_model = opt.pretrained_model
output_dir = opt.output_dir
num_epochs, batch_size, lr = opt.num_epochs, opt.batch_size, opt.lr
with open(os.path.join(data_dir, codes),'rb')as f:
gnd = pickle.load(f)
mapping = gnd['mapping']
id_length = mapping.shape[-1]
num_classes = np.unique(mapping).shape[0]
if data_name == 'isc':
Dataset = Inshop(data_dir, 'db',transform=train_transform(256,224))
elif data_name == 'cub':
Dataset = CUB(data_dir, 'db',transform=train_transform(256,224))
elif data_name == 'car':
Dataset = Cars(data_dir, 'db',transform=train_transform(256,224))
elif data_name == 'imagenet':
Dataset = ImageNet(data_dir, 'db',transform=train_transform(256,224))
elif data_name == 'places':
Dataset = Places(data_dir, 'db',transform=train_transform(256,224))
train_sampler = DistributedSampler(Dataset, shuffle=True)
if data_name in ['isc', 'cub', 'car']:
model = IRGen(dec_depth=12, num_classes=num_classes, id_len=id_length)
else:
model = IRGen(dec_depth=24, num_classes=num_classes, id_len=id_length)
mm, preprocess = clip.load('ViT-B-16.pt')
mm = mm.to('cpu')
mm=mm.type(torch.float32)
model.encoder = mm.visual
if pretrained_model:
checkpoint = torch.load(model_dir)
state_dict = checkpoint['net']
state_dict = {k.replace("module.", ""): v for k, v in state_dict.items()}
model.load_state_dict(state_dict, strict=False)
model = model.to(device)
model = DistributedDataParallel(model, device_ids=[local_rank],
output_device=local_rank)
train_loader = DataLoader(Dataset, batch_size=batch_size, sampler=train_sampler, num_workers=8)
optimizer = AdamW([{'params': get_trainable_params(model.module.decoder)},
{'params':get_trainable_params(model.module.encoder), 'lr': 0.01*lr}
], lr=lr, betas=(0.9, 0.96), eps=1e-08, weight_decay=0.05)
scheduler = scheduler.CosineLRScheduler(optimizer=optimizer,
t_initial=100,
lr_min=5e-7,
warmup_t=20,
warmup_lr_init=5e-8,
cycle_decay=1,
cycle_limit = 4,
)
# criterion = nn.CrossEntropyLoss()
criterion = LabelSmoothingCrossEntropy()
logger = get_logger('train.log')
with open(os.path.join(data_dir, file_name), 'rb') as fin:
gnd = pickle.load(fin)
class_idx = gnd['class_idx']
Loss = []
LR = []
for i in range(num_epochs):
train_loader.sampler.set_epoch(num_epochs)
if dist.get_rank() == 0:
print('start epoch',i)
for j,(img,clss,idx) in enumerate(train_loader):
idx = [np.random.rand(1)*len(class_idx[int(clss[k])]) for k in range(len(img))]
tgt = np.array([mapping[class_idx[int(clss[k])][int(idx[k])]] for k in range(len(img))])
tgt[np.where(tgt==-1)]=num_classes
target = torch.tensor(tgt,dtype=torch.int64)
img = img.cuda()
target = target.cuda()
output = model(img, tgt=target)
output = output[:,:-1,:]
output = output.reshape(-1,num_classes)
target = target.reshape(-1)
loss = criterion(output, target, args.smoothing)
running_loss = loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
if j%10 == 9:
if dist.get_rank() == 0:
logger.info('Epoch:[{}/{}]\t step={}\t loss={:.10f}\t lr={:.8f}'.format(i , num_epochshs, j+1, loss, optimizer.param_groups[0]['lr'] ))
percent=(j+1)/len(train_loader)
scheduler.step(i+percent)
scheduler.step(i+1)
if dist.get_rank() == 0:
print('epoch:', i, 'loss:', loss)
if i>90 and i%50 == 49:
state = {'net':model.state_dict(), 'optimizer':optimizer.state_dict(), 'epoch':i+1, 'loss':running_loss}
torch.save(state, os.path.join(output_dir,'isc_rq_{}.pkl').format(i+1))
dist.barrier()