-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrain.py
162 lines (142 loc) · 7.15 KB
/
train.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
import os
import torch
from torchvision import transforms
from torch.utils.data import DataLoader
from model.model_main import IQARegression
from model.backbone import inceptionresnetv2, Mixed_5b, Block35, SaveOutput
from option.config import Config
from trainer import train_epoch, eval_epoch
from utils.util import RandCrop, RandHorizontalFlip, RandRotation, Normalize, ToTensor, RandShuffle
# config file
config = Config({
# device
"GPU_ID": "0",
"num_workers": 8,
# model for LIVE/CSIQ/TID
# "n_enc_seq": 29*29, # feature map dimension (H x W) from backbone, this size is related to crop_size
# "n_dec_seq": 29*29, # feature map dimension (H x W) from backbone
# "n_layer": 2, # number of encoder/decoder layers
# "d_hidn": 256, # input channel (C) of encoder / decoder (input: C x N)
# "i_pad": 0,
# "d_ff": 1024, # feed forward hidden layer dimension
# "d_MLP_head": 512, # hidden layer of final MLP
# "n_head": 4, # number of head (in multi-head attention)
# "d_head": 256, # input channel (C) of each head (input: C x N) -> same as d_hidn
#"dropout": 0.1, # dropout ratio of transformer
# "emb_dropout": 0.1, # dropout ratio of input embedding
# "layer_norm_epsilon": 1e-12,
# "n_output": 1, # dimension of final prediction
# "crop_size": 256, # input image crop size
# model for PIPAL (NTIRE2021 Challenge)
"n_enc_seq": 21*21, # feature map dimension (H x W) from backbone, this size is related to crop_size
"n_dec_seq": 21*21, # feature map dimension (H x W) from backbone
"n_layer": 1, # number of encoder/decoder layers
"d_hidn": 128, # input channel (C) of encoder / decoder (input: C x N)
"i_pad": 0,
"d_ff": 1024, # feed forward hidden layer dimension
"d_MLP_head": 128, # hidden layer of final MLP
"n_head": 4, # number of head (in multi-head attention)
"d_head": 128, # input channel (C) of each head (input: C x N) -> same as d_hidn
"dropout": 0.1, # dropout ratio of transformer
"emb_dropout": 0.1, # dropout ratio of input embedding
"layer_norm_epsilon": 1e-12,
"n_output": 1, # dimension of final prediction
"crop_size": 192, # input image crop size
# data
"db_name": "PIPAL", # database name [ PIPAL | LIVE | CSIQ | TID2013 ]
"db_path": "/mnt/Dataset/anse_data/IQAdata/PIPAL/Train_Images", # root of dataset
"snap_path": "./weights/PIPAL", # path for saving weights
"txt_file_name": "./IQA_list/PIPAL_augment.txt", # image list file (.txt)
"train_size": 0.8,
"scenes": "all",
# ensemble in validation phase
"test_ensemble": True,
"n_ensemble": 5,
# optimization
"batch_size": 16,
"learning_rate": 2e-4,
"weight_decay": 1e-5,
"n_epoch": 300,
"val_freq": 1,
"save_freq": 5,
"checkpoint": None, # load pretrained weights
"T_max": 50, # cosine learning rate period (iteration)
"eta_min": 0 # mininum learning rate
})
# device setting
config.device = torch.device("cuda:%s" %config.GPU_ID if torch.cuda.is_available() else "cpu")
if torch.cuda.is_available():
print('Using GPU %s' % config.GPU_ID)
else:
print('Using CPU')
# data selection
if config.db_name == 'PIPAL':
from data.data_PIPAL import IQADataset
elif config.db_name == 'LIVE':
from data.data_LIVE import IQADataset
# data separation (8:2)
train_scene_list, test_scene_list = RandShuffle(config.scenes, config.train_size)
# for reproducing results
# train_scene_list = [*range(160)]
# test_scene_list = [*range(160, 200)]
# train_scene_list = [*range(23)]
# test_scene_list = [*range(23, 29)]
print('number of train scenes: %d' % len(train_scene_list))
print('number of test scenes: %d' % len(test_scene_list))
# data load
train_dataset = IQADataset(
db_path=config.db_path,
txt_file_name=config.txt_file_name,
transform=transforms.Compose([RandCrop(config.crop_size), Normalize(0.5, 0.5), RandHorizontalFlip(), RandRotation(), ToTensor()]),
train_mode=True,
scene_list=train_scene_list,
train_size=config.train_size
)
test_dataset = IQADataset(
db_path=config.db_path,
txt_file_name=config.txt_file_name,
transform= transforms.Compose([Normalize(0.5, 0.5), ToTensor()]) if config.test_ensemble else transforms.Compose([RandCrop(config.crop_size), Normalize(0.5, 0.5), ToTensor()]),
train_mode=False,
scene_list=test_scene_list,
train_size=config.train_size
)
train_loader = DataLoader(dataset=train_dataset, batch_size=config.batch_size, num_workers=config.num_workers, drop_last=True, shuffle=True)
test_loader = DataLoader(dataset=test_dataset, batch_size=config.batch_size, num_workers=config.num_workers, drop_last=True, shuffle=True)
# create model
model_transformer = IQARegression(config).to(config.device)
model_backbone = inceptionresnetv2(num_classes=1001, pretrained='imagenet+background').to(config.device)
# save intermediate layers
save_output = SaveOutput()
hook_handles = []
for layer in model_backbone.modules():
if isinstance(layer, Mixed_5b):
handle = layer.register_forward_hook(save_output)
hook_handles.append(handle)
elif isinstance(layer, Block35):
handle = layer.register_forward_hook(save_output)
hook_handles.append(handle)
# loss function
criterion = torch.nn.MSELoss()
optimizer = torch.optim.Adam(model_transformer.parameters(), lr=config.learning_rate, weight_decay=config.weight_decay)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=config.T_max, eta_min=config.eta_min)
# load weights & optimizer
if config.checkpoint is not None:
checkpoint = torch.load(config.checkpoint)
model_transformer.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
start_epoch = checkpoint['epoch']
loss = checkpoint['loss']
else:
start_epoch = 0
# make directory for saving weights
if not os.path.exists(config.snap_path):
os.mkdir(config.snap_path)
# train & validation
losses, scores = [], []
for epoch in range(start_epoch, config.n_epoch):
loss, rho_s, rho_p = train_epoch(config, epoch, model_transformer, model_backbone, save_output, criterion, optimizer, scheduler, train_loader)
# print('[train] epoch:%d / loss:%f / SROCC:%4f / PLCC:%4f' % (epoch+1, loss.item(), rho_s, rho_p))
if (epoch+1) % config.val_freq == 0:
loss, rho_s, rho_p = eval_epoch(config, epoch, model_transformer, model_backbone, save_output, criterion, test_loader)
# print('test epoch:%d / loss:%f /SROCC:%4f / PLCC:%4f' % (epoch+1, loss.item(), rho_s, rho_p))