-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
180 lines (142 loc) · 6.21 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
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
import os
import time
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from torch.autograd import Variable
import cv2
import numpy as np
import argparse
import tqdm
from PIL import Image
# Model
import CRAFT.craft_utils as craft_utils
import CRAFT.imgproc as imgproc
import CRAFT.file_utils as file_utils
from CRAFT.craft import CRAFT
from collections import OrderedDict
# Data loader
from loader import SynthTextLoader
def load_statedict(state_dict):
if list(state_dict.keys())[0].startswith("module"):
start_idx = 1
else:
start_idx = 0
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = ".".join(k.split(".")[start_idx:])
new_state_dict[name] = v
return new_state_dict
def str2bool(v):
return v.lower() in ("yes", "y", "true", "t", "1")
def hard_worst_loss(loss, groundtruth):
forceground_idx = groundtruth.nonzero().permute(1, 0)[0]
background_idx = (groundtruth == 0).nonzero().permute(1, 0)[0]
force_loss = loss[forceground_idx]
back_loss = loss[background_idx]
n_back_samples = min(forceground_idx.size(0) * 3, background_idx.size(0))
top_k_back_loss_idx = np.argsort(-back_loss.data.cpu().numpy())[:n_back_samples]
back_loss = back_loss[top_k_back_loss_idx]
return (force_loss.sum() + back_loss.sum()) / (force_loss.size(0) + back_loss.size(0))
def OHEMLoss(pd_char_masks, gt_char_masks, pd_aff_masks, gt_aff_masks):
b, h, w = gt_char_masks.size()
# Flatten tensor
pd_char_masks = pd_char_masks.contiguous().view(b * h * w)
gt_char_masks = gt_char_masks.contiguous().view(b * h * w)
pd_aff_masks = pd_aff_masks.contiguous().view(b * h * w)
gt_aff_masks = gt_aff_masks.contiguous().view(b * h * w)
# MSE loss for per pixel
char_loss = torch.nn.MSELoss(reduction='none')(pd_char_masks, gt_char_masks)
aff_loss = torch.nn.MSELoss(reduction='none')(pd_aff_masks, gt_aff_masks)
# Find the worst loss
hard_loss_char = hard_worst_loss(char_loss, gt_char_masks)
hard_loss_aff = hard_worst_loss(aff_loss, gt_aff_masks)
return hard_loss_char + hard_loss_aff
class ModelFit:
def __init__(self, model, optimizer, train_logger):
self.model = model
self.optimizer = optimizer
self.train_logger = train_logger
def train(self, data_loader, data_folder, save_img, save_char, save_aff,
num_epochs, batch_size, shuffle):
self.model.train()
loader = data_loader(batch_size, shuffle, data_folder, save_img, save_char, save_aff)
loader = loader.loader()
n_iter = len(loader)
for epoch in range(num_epochs):
print('Train model with epoch ', epoch + 1)
total_loss = 0
for batch_idx, (images, char_masks, aff_masks) in enumerate(loader):
images = images.cuda()
char_masks = char_masks.cuda()
char_masks = char_masks[:, 0, :, :]
aff_masks = aff_masks.cuda()
aff_masks = aff_masks[:, 0, :, :]
# Clear gradients
self.optimizer.zero_grad()
# Outputs
output, _ = self.model(images)
# Predict map for character
out_char = output[:, :, :, 0]
# Predict map for aff
out_aff = output[:, :, :, 1]
# Loss
loss = OHEMLoss(out_char, char_masks, out_aff, aff_masks)
loss.backward()
self.optimizer.step()
total_loss += loss.item()
torch.save(self.model.state_dict(), 'checkpoints/model.pth')
def valid(self, data_loader, data_folder, save_img, save_char, save_aff):
self.model.eval()
loader = data_loader(1, False, data_folder, save_img, save_char, save_aff)
loader = loader.loader()
n_iter = len(loader)
valid_pbar = tqdm.tqdm(enumerate(loader), total=n_iter)
with torch.no_grad():
for batch_idx, (images, char_masks, aff_masks) in valid_pbar:
images = images.cuda()
char_masks = char_masks.cuda()
char_masks = char_masks[:, 0, :, :]
aff_masks = aff_masks.cuda()
aff_masks = aff_masks[:, 0, :, :]
# Outputs
output, _ = self.model(images)
# Predict map for character
out_char = output[:, :, :, 0]
# Predict map for aff
out_aff = output[:, :, :, 1]
image = out_char.cpu()
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
show = transforms.ToPILImage()
image = show(image)
image.save('debug.png')
if __name__ == '__main__':
data_folder = '/mnt/data/hades/source/ocr_rnd/datasets/SynthText'
save_img = '/mnt/data/hades/source/sekiwa_rnd/dataset/synthtext/images'
save_char = '/mnt/data/hades/source/sekiwa_rnd/dataset/synthtext/char_mask'
save_aff = '/mnt/data/hades/source/sekiwa_rnd/dataset/synthtext/aff_mask'
# Load pre-trained model
model = CRAFT()
print('Loading weights from checkpoint')
parser = argparse.ArgumentParser(description='CRAFT Text Detection')
parser.add_argument('--cuda', default=True, type=str2bool, help='Use cuda to train model')
parser.add_argument('--trained_model', default='CRAFT/craft_mlt_25k.pth', type=str, help='Pretrained model')
args = parser.parse_args()
# if args.cuda:
# model.load_state_dict(load_statedict(torch.load(args.trained_model)))
# else:
# model.load_state_dict(load_statedict(torch.load(args.trained_model, map_location='cpu')))
if args.cuda:
model = model.cuda()
# Optimizer
optimizer = torch.optim.Adam(model.parameters(),
lr=0.001,
betas=(0.9, 0.98), eps=1e-9)
# Train model
trainer = ModelFit(model, optimizer, None)
num_epochs = 20
batch_size = 1
shuffle = True
trainer.train(SynthTextLoader, data_folder, save_img, save_char, save_aff, num_epochs, batch_size, shuffle)
trainer.valid(SynthTextLoader, data_folder, save_img, save_char, save_aff)