-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathengine.py
180 lines (141 loc) · 7.26 KB
/
engine.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
180
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Train and eval functions used in main.py
"""
import os
import sys
import cv2
import math
import json
import torch
import numpy as np
import util.misc as utils
from typing import Iterable
from tqdm import tqdm
from util.visualize import tensor_to_cv2image
from skimage import io
def save_img(path, img):
# img (H,W,C) or (H,W) np.uint8
io.imsave(path+'/'+name+'.png', img)
def train_one_epoch(model: torch.nn.Module, discriminator: torch.nn.Module,
criterion: torch.nn.Module, data_loader: Iterable, optimizer,
device: torch.device, epoch: int, max_norm: float = 0,
lr_scheduler: list = [0], print_freq: int = 10, debug: bool = False,
optim_with_mask: bool = False):
model.train()
criterion.train()
# criterion_adversial.train()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
# print_freq = 10
optimizer['D'].param_groups[0]['lr'] = lr_scheduler[epoch]
optimizer['G'].param_groups[0]['lr'] = lr_scheduler[epoch]
optimizer['G'].param_groups[1]['lr'] = lr_scheduler[epoch] * 0.1
if debug:
count = 0
save_folder = '../debug/20220112-vis-scutens-train-textmask'
os.makedirs(save_folder, exist_ok=True)
for data in tqdm(data_loader):
for image, label, mask, mask_gt in zip(data['image'], data['label'], data['mask'], data['mask_gt']):
image = tensor_to_cv2image(image, False)
label = tensor_to_cv2image(label, False)
mask = tensor_to_cv2image(mask, False)
mask_gt = tensor_to_cv2image(mask_gt, False)
cv2.imwrite(os.path.join(save_folder, f'{count:06d}-image.jpg'), image)
cv2.imwrite(os.path.join(save_folder, f'{count:06d}-label.jpg'), label)
cv2.imwrite(os.path.join(save_folder, f'{count:06d}-mask.jpg'), mask)
cv2.imwrite(os.path.join(save_folder, f'{count:06d}-mask_gt.jpg'), mask_gt)
count += 1
return
for data in metric_logger.log_every(data_loader, print_freq, header):
images = data['image'].to(device); labels = data['label'].to(device); mask_gts = data['mask_gt'].to(device)
structure_im = data['structure_im'].to(device)
structure_lbl = data['structure_lbl'].to(device)
soft_mask = data['soft_mask'].to(device)
# soft_mask = 1 - soft_mask ### for syn
# mask_gts = 1 - mask_gts
outputs = model(images, mask_gts, labels, structure_im, structure_lbl, soft_mask)
real_prob = discriminator(labels, mask_gts)
fake_prob_D = discriminator(outputs['output'][-1].contiguous().detach(), mask_gts)
D_loss = criterion.discriminator_loss(real_prob, fake_prob_D)
optimizer['D'].zero_grad()
D_loss.backward()
optimizer['D'].step()
fake_prob_G = discriminator(outputs['output'][-1], mask_gts)
outputs['real_prob'] = real_prob
outputs['fake_prob_D'] = fake_prob_D
outputs['fake_prob_G'] = fake_prob_G
loss_dict = criterion(outputs, mask_gts, labels, structure_lbl)
loss_dict['D_loss'] = D_loss
weight_dict = {'MSR_loss': 1.5, 'prc_loss': 0.1, 'style_loss': 120, 'D_fake': 1, 'D_loss': 1, 'FM_loss': 1, 'structure_loss': 1}
for k in loss_dict.keys():
loss_dict[k] *= weight_dict[k]
G_loss = sum([loss_dict[k] for k in loss_dict.keys() if k != 'D_loss'])
optimizer['G'].zero_grad()
G_loss.backward()
# for name, param in model.named_parameters():
# if param.grad is None:
# print(name)
optimizer['G'].step()
# optimizer['D'].step()
# D_loss = loss_dict['D_loss']
# G_loss = sum([loss_dict[k] for k in loss_dict.keys() if k != 'D_loss'])
loss_dict_reduced = utils.reduce_dict(loss_dict)
# loss_dict_reduced_unscaled = {f'{k}_unscaled': v for k, v in loss_dict_reduced.items()}
loss_dict_reduced_scaled = {k: v * weight_dict[k] for k, v in loss_dict_reduced.items()}
losses_reduced_scaled = sum(loss_dict_reduced_scaled.values())
loss_value = losses_reduced_scaled.item()
metric_logger.update(loss=loss_value, **loss_dict_reduced_scaled)
metric_logger.update(lr=optimizer['G'].param_groups[0]["lr"])
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate(model, criterion, data_loader, device, output_dir, chars, start_index, visualize=False):
model.eval()
criterion.eval()
chars = list(chars)
for data in tqdm(data_loader):
images = data['image'].to(device)
mask_gts = data['mask_gt'].to(device)
labels = data['label'].to(device)
soft_mask = data['soft_mask'].to(device)
# import pdb;pdb.set_trace()
soft_mask = torch.mean(soft_mask, 1, keepdim=True)
mask_gts = torch.mean(mask_gts, 1, keepdim=True)
structure_im = data['structure_im'].to(device)
structure_lbl = data['structure_lbl'].to(device)
# outputs = model(images, mask_gts, None)
# soft_mask = mask_gts
# import pdb;pdb.set_trace()
outputs = model(images, mask_gts, labels, structure_im, structure_lbl, soft_mask)
# import pdb;pdb.set_trace()
str_output = outputs[-1]
str_output = str_output.cpu().clamp(min=0, max=1)
output = outputs * (1 - soft_mask) + images * soft_mask
output = output[-1].cpu().clamp(min=0, max=1)
output = tensor_to_cv2image(output,False)
str_output = tensor_to_cv2image(str_output,False)
stroke_gt = mask_gts.cpu()
stroke_mask = torch.cat((stroke_gt[0], stroke_gt[0], stroke_gt[0]), 0)
mask = tensor_to_cv2image(stroke_mask,False)
image_path = data['image_path'][0]
if 'scut-syn' in image_path.lower():
dataset_name_str = 'SCUT-Syn'
dataset_name_paste = 'SCUT_Syn_com'
elif 'scut-ens' in image_path.lower():
dataset_name_str = 'SCUT-ENS'
dataset_name_paste = 'SCUT_ENS_com'
save_folder = os.path.join(output_dir, dataset_name_str)
os.makedirs(save_folder, exist_ok=True)
save_path = os.path.join(save_folder, os.path.basename(image_path).replace('jpg', 'png'))
# import pdb;pdb.set_trace()
cv2.imwrite(save_path, str_output)
save_folder_com = os.path.join(output_dir, dataset_name_paste)
os.makedirs(save_folder_com, exist_ok=True)
save_path_com = os.path.join(save_folder_com, os.path.basename(image_path).replace('jpg', 'png'))
cv2.imwrite(save_path_com, output)
# cv2.imwrite(os.path.join(save_folder, 'mask_' + os.path.basename(image_path).replace('jpg', 'png')), mask)
# save_img(save_path, output)