-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
177 lines (151 loc) · 7.73 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
from utils import *
import argparse
import numpy as np
import datetime
from trainer import SPMPGAN_Trainer
from dataset import Image_Editing_Dataset
import torch
from torch.utils.data import DataLoader
import os
import shutil
import cv2
from metrics.fid_score import calculate_fid_given_paths
parser = argparse.ArgumentParser()
parser.add_argument('--config', type=str, default='configs/config.yaml', help='Path to the config file.')
parser.add_argument('--output_path', type=str, default='.', help="outputs path")
parser.add_argument('--dataset_name', type=str, default='ADE20k-room', help="dataset name")
parser.add_argument("--resume", action="store_true")
parser.add_argument('--resume_dir', type=str, default='', help="outputs path")
opts = parser.parse_args()
print_options(opts)
# cudnn.benchmark = True
# GPU
os.environ["CUDA_VISIBLE_DEVICES"] = '2'
# Load experiment setting
cfg = get_config(opts.config)
# datasets setting
if opts.dataset_name == 'ADE20k-room':
cfg['lab_dim'] = 151
cfg['max_epoch'] = 500
cfg['test_freq'] = 20
elif opts.dataset_name == 'ADE20k-landscape':
cfg['lab_dim'] = 151
cfg['max_epoch'] = 500
cfg['test_freq'] = 20
elif opts.dataset_name == 'cityscapes':
cfg['lab_dim'] = 34
cfg['max_epoch'] = 500
cfg['test_freq'] = 20
trainer = SPMPGAN_Trainer(cfg)
trainer.cuda()
# print model information
trainer.print_networks()
# Setup dataset
dataset_root = os.path.join(cfg['dataset_dir'], opts.dataset_name)
train_dataset = Image_Editing_Dataset(cfg, dataset_root, split='train', dataset_name=opts.dataset_name)
train_loader = DataLoader(dataset=train_dataset, batch_size=cfg['batch_size'], shuffle=cfg['shuffle'], num_workers=cfg['worker'])
test_dataset = Image_Editing_Dataset(cfg, dataset_root, split='test', dataset_name=opts.dataset_name)
test_loader = DataLoader(dataset=test_dataset, batch_size=cfg['test_batch_size'], shuffle=False, num_workers=cfg['worker'])
print('train dataset containing ', len(train_loader), 'images')
print('test dataset containing ', len(test_loader), 'images')
# Setup logger and output folders
if opts.resume:
checkpoint_directory = opts.resume_dir + 'checkpoints/'
image_directory = opts.resume_dir + 'images/'
result_directory = opts.resume_dir + 'results/'
cur_epoch = trainer.resume(checkpoint_directory, ckpt_filename=None) + 1
shutil.copy(opts.config, os.path.join(opts.resume_dir, 'config_resume.yaml')) # copy config file to output folder
else:
cur_epoch = 0
output_directory = os.path.join(opts.output_path + "/outputs", opts.dataset_name,
datetime.datetime.today().strftime('%Y-%m-%d_%H-%M-%S'))
checkpoint_directory, image_directory, result_directory = prepare_sub_folder(output_directory)
shutil.copy(opts.config, os.path.join(output_directory, 'config.yaml')) # copy config file to output folder
cfg['checkpoints_dir'] = checkpoint_directory
best_fid = float("inf")
print('training start at %d th epoch' % (cur_epoch))
# Start training
for epoch in range(cur_epoch, cfg['max_epoch']):
for i, data in enumerate(train_loader): # inner loop within one epoch
trainer.train()
trainer.set_input(data) # unpack data from dataset and apply preprocessing
trainer.optimize_parameters() # calculate loss functions, get gradients, update network weights
if i % cfg['visual_img_freq'] == 0:
results, img_name = trainer.visual_results()
cur_img_dir = os.path.join(image_directory, 'epoch-%d_iter-%d_%s'%(epoch, i, img_name[0]))
if not os.path.exists(cur_img_dir):
os.makedirs(cur_img_dir)
is_fg = ['mask', 'lab', 'mask_seam', 'edge_map']
for name, img in results.items():
no_fg = True
if name in is_fg:
no_fg = False
save_name = 'epoch-%d_iter-%d_%s_'%(epoch, i, name)+'.png'
if name == 'lab':
lab = lab2im(img)
cv2.imwrite(os.path.join(cur_img_dir, save_name), lab)
# print('lab mean: ', lab2im(img).mean())
label_dir = os.path.join(cur_img_dir, 'label')
if not os.path.exists(label_dir):
os.makedirs(label_dir)
set = np.unique(lab)
for l in set:
cur_lab = np.array(np.equal(lab, l).astype(np.uint8)) * 255
cv2.imwrite(os.path.join(label_dir, str(l)+'.png'), cur_lab)
elif name == 'att':
att_dir = os.path.join(cur_img_dir, 'att')
if not os.path.exists(att_dir):
os.makedirs(att_dir)
for i_att, att in enumerate(img):
hm = tensor2hm(att)
cv2.imwrite(os.path.join(att_dir, 'att_'+str(i_att)+'.png'), cv2.applyColorMap(hm, cv2.COLORMAP_JET))
elif name == 'middle_avg':
cv2.imwrite(os.path.join(cur_img_dir, save_name), tensor2im(img[:,:3,:,:], no_fg=no_fg))
elif name == 'masks':
masks_dir = os.path.join(cur_img_dir, 'masks')
if not os.path.exists(masks_dir):
os.makedirs(masks_dir)
for idx, m in enumerate(img):
hm = tensor2hm(m)
cv2.imwrite(os.path.join(masks_dir, 'att_'+str(idx)+'.png'), cv2.applyColorMap(hm, cv2.COLORMAP_JET))
elif name in ['middle_avg_encoder', 'middle_avg_decoder']:
cv2.imwrite(os.path.join(cur_img_dir, save_name), tensor2im(img[:,:3,:,:], no_fg=no_fg))
else:
cv2.imwrite(os.path.join(cur_img_dir, save_name), tensor2im(img, no_fg=no_fg))
if i % cfg['print_loss_freq'] == 0: # print training losses and save logging information to the disk
print('print losses at the {} epoch {} iter'.format(epoch, i))
trainer.print_losses()
if (epoch+1) % cfg['save_epoch_freq'] == 0: # cache our model every <save_epoch_freq> epochs
print('saving the model at the end of epoch %d' % (epoch))
trainer.save_nets(epoch, cfg)
print('saving the model')
trainer.save_latest_nets(epoch, cfg)
# updating learning rate
# trainer.update_learning_rate()
# compute FID score
if epoch % cfg['test_freq'] == 0:
with torch.no_grad():
print('testing at %d epoch' % (epoch))
trainer.eval()
cur_save_dir = os.path.join(result_directory, str(epoch))
if not os.path.exists(cur_save_dir):
os.makedirs(cur_save_dir)
for i, data in enumerate(test_loader):
trainer.set_input(data)
trainer.forward()
results, img_name = trainer.visual_results()
# if not os.path.exists(os.path.join(cur_save_dir, img_name[0]+'.png')):
# os.makedirs(os.path.join(cur_save_dir, img_name[0]+'.png'))
cv2.imwrite(os.path.join(cur_save_dir, img_name[0]+'.png'), tensor2im(results['mask_fake_G3']))
path_gt = os.path.join(cfg['dataset_dir'], opts.dataset_name, 'test', 'images')
path_test = cur_save_dir
print('path_gt: ', path_gt)
print('path_test: ', path_test)
# compute FID score
path = [path_gt, path_test]
fid_value = calculate_fid_given_paths(path, cfg['test_batch_size'])
print('========FID==========: ', fid_value)
if fid_value < best_fid:
print('saving the current best model at the end of epoch %d' % (epoch))
trainer.save_nets(epoch, cfg, suffix='_best_FID_'+str(fid_value))
best_fid = fid_value