forked from ashwatc/DTP_sim2real_ROAR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathoptimizer.py
288 lines (236 loc) · 14.9 KB
/
optimizer.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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
import os
import cv2
import gc
import random
import time
from tqdm import tqdm
import numpy as np
import matplotlib.pyplot as plt
import argparse
from glob import glob
import torch
import torch.nn as nn
import torchvision.transforms as transforms
from PIL import Image, ImageFilter
from models.OEFT import OEFT
parser = argparse.ArgumentParser(description='Code to optimize')
parser.add_argument('--device', help='cuda | cuda:0 | cpu', default="cuda", type=str)
parser.add_argument('--device_num', help='which GPUs to use', default="0", type=str)
parser.add_argument('--sample_freq', help="sampling frequency of saving results", default=500, type=float)
""" optimizer.py setting """
parser.add_argument('--content_root', help='folder of content images', default="../OEFT/example_210905/input/", type=str)
parser.add_argument('--style_root', help='folder of style images', default="../OEFT/example_210905/style/", type=str)
parser.add_argument('--save_root', help='folder of saving results', default="../OEFT/example_210929/experiment", type=str)
parser.add_argument('--fileType', help='png|jpg', default="png", type=str)
parser.add_argument('--keys', help='vgg layer names', default=['r12', 'r22', 'r34', 'r44', 'r54'], nargs="+")
parser.add_argument('--iter', help="number of iteration for optimization", default=1000, type=int)
parser.add_argument('--img_size', help="size of input image", default=256, type=int)
parser.add_argument('--pretrained', help="use pre-trained network or not", action="store_false")
parser.add_argument('--denorm', help="size of input image", action="store_false")
parser.add_argument('--lr', help="learning rate", default=1e-4, type=float)
parser.add_argument('--beta1', help="optimizer parameter", default=0.5, type=float)
parser.add_argument('--beta2', help="optimizer parameter", default=0.999, type=float)
parser.add_argument('--weight_decay', help="weight_decay", default=1e-4, type=float)
""" OEPT.py setting """
parser.add_argument('--warpFeat', help="use warped feature as decoder input", action="store_true")
parser.add_argument('--warpMv', help="use warped feature as moving averaged feature with content feature", action="store_true")
parser.add_argument('--warpRes', help="use warped image as residual", action="store_true")
parser.add_argument('--cycle', help="use cycle consistency regularization", action="store_true")
parser.add_argument('--res_wt', help="weight between decoder output and residual warped img", default=8/9, type=float)
parser.add_argument('--cycle_wt', help="weight of cycle consistency regularization", default=1., type=float)
# 256, 128, 64, 32, 16
parser.add_argument('--nce_wt', help='nce loss weights from each layer[256-16]', default=[1/8*1/4, 1/4*1/4, 1/2*1/4, 1.*1/4, 1.*1/4], nargs="+")
parser.add_argument('--nns_wt', help='NN style loss weights from each layer[256-16]', default=[1/16*1/4, 1/8*1/4, 1/4*1/4, 1/2*1/4, 1.*1/4], nargs="+")
parser.add_argument('--nce_temp', help="temperature for nce", default=0.07, type=float)
parser.add_argument('--nns_temp', help="temperature for nns", default=0.05, type=float)
parser.add_argument('--content_style_wt', help="weight of between content and style", default=4/5, type=float)
""" corrnet.py setting """
parser.add_argument('--corr_temp', help="temperature of correlation module", default=0.01, type=float)
parser.add_argument('--mv_wt', help="weight of moving average", default=0.6, type=float)
mean = np.array([0.485, 0.456, 0.406]).reshape(1,1,3)
std = np.array([0.229, 0.224, 0.225]).reshape(1,1,3)
args = parser.parse_args()
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
GPU_list = args.device_num
len_GPU = len( GPU_list.split(","))
print("@@@@@@@@@@@@ len_GPU: ", len_GPU)
os.environ["CUDA_VISIBLE_DEVICES"] = GPU_list
# Load content and style
print(os.listdir('./'))
def img_load(path):
img = cv2.imread(path)[::,::,::-1] # BGR to RGB, [0-255]
return img
def toPIL(img):
# image range should be [0-255] for converting.
img_type = str(type(img))
if 'numpy' in img_type:
img = Image.fromarray(img)
elif 'torch' in img_type:
img = transforms.ToPILImage()(img).convert("RGB")
return img
if __name__ == "__main__":
# parse options
keys = args.keys
content_root = args.content_root
style_root = args.style_root
if not os.path.exists(args.content_root):
print("!!! args.content_root does not exist !!!")
exit()
if not os.path.exists(args.style_root):
print("!!! args.style_root does not exist !!!")
exit()
content_list = glob( os.path.join(content_root, "*.{}".format(args.fileType)) )
style_list = glob( os.path.join(style_root, "*.{}".format(args.fileType) ) )
if len(content_list) < 1 or len(style_list) < 1:
print("!!! The number of content and style images should be more than 1 !!!")
exit()
content_list.sort()
style_list.sort()
print("@@@@@@@@@ len(content_list): ", len(content_list))
for z in range(len(content_list)):
for zz in range(len(style_list)):
random_seed = 1006
os.environ['PYTHONHASHSEED'] = str(random_seed)
torch.manual_seed(random_seed)
torch.cuda.manual_seed(random_seed)
torch.cuda.manual_seed_all(random_seed) # if use multi-GPU
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
np.random.seed(random_seed)
random.seed(random_seed)
""" start iteration """
torch.cuda.empty_cache() # remove all caches
try:
""" content, style image path indexing """
content_path = content_list[z] # './examples/input/in11.png'
style_path = style_list[zz] # './examples/style/tar11.png'
""" img load """
content = img_load(content_path)
style = img_load(style_path)
content_256 = content.copy()
style_256 = style.copy()
except Exception as e:
print("image loading error : ", e)
continue
""" Convert numpy array to PIL.Image format """
""" and modify the range (0-255) to [0-1] """
content = toPIL(content)
style = toPIL(style)
""" Make transform """
transform_list = []
img_size = (args.img_size, args.img_size)
transform_list.append(transforms.Resize(img_size, interpolation=2)) # @@@@ args.interpol-method = 2
transform_list += [transforms.ToTensor()]
transform_list += [transforms.Normalize( (0.485, 0.456, 0.406),
(0.229, 0.224, 0.225))]
transform = transforms.Compose(transform_list)
""" do transform """
content = transform(content)
style = transform(style)
content = torch.unsqueeze( content, dim=0 )
style = torch.unsqueeze( style, dim=0 )
""" Load model """
model = OEFT(args=args, pretrained=args.pretrained)
model = model.to(args.device)
""" Define optimizer """
e_optimizer = torch.optim.Adam(model.corrNet.parameters(), lr=args.lr,
betas=(args.beta1, args.beta2)
)
g_optimizer = torch.optim.Adam(model.decoder.parameters(), lr=args.lr,
betas=(args.beta1, args.beta2)
)
for param_group in e_optimizer.param_groups:
param_group['lr'] = 1e-4 #hparams.initial_learning_rate
for param_group in g_optimizer.param_groups:
param_group['lr'] = 1e-4 #hparams.initial_learning_rate
if "cuda" in args.device:
content = content.type(torch.cuda.FloatTensor).to(args.device).detach()
style = style.type(torch.cuda.FloatTensor).to(args.device).detach()
new_input = content.clone().type(torch.cuda.FloatTensor).to(args.device)
else:
content = content.type(torch.FloatTensor).to(args.device).detach()
style = style.type(torch.FloatTensor).to(args.device).detach()
new_input = content.clone().type(torch.FloatTensor).to(args.device)
warp_result = None
dec_result = None
cycle_result = None
count = 0
model.train()
start = time.time()
prog_bar = tqdm(range(args.iter))
for i in prog_bar:
if count == 0:
warped_s2c_feat, warped_s2c_imgs, tr_s2c_img_decs, loss, loss_dict, nce_dict, nns_dict = model(style, content, step=i)
else:
if args.warpMv:
warped_s2c_feat = warped_s2c_feat
for key in args.keys:
warped_s2c_feat[key] = warped_s2c_feat[key].detach()
warped_s2c_feat, warped_s2c_imgs, tr_s2c_img_decs, loss, loss_dict, nce_dict, nns_dict = model(style, content, warped_s2c_feat, step=i)
else:
warped_s2c_feat, warped_s2c_imgs, tr_s2c_img_decs, loss, loss_dict, nce_dict, nns_dict = model(style, content, step=i)
# summary_writer.add_scalars( "Total NCE and NNS" , loss_dict, i)
# summary_writer.add_scalars( "NCEs" , nce_dict, i)
# summary_writer.add_scalars( "NNSs" , nns_dict, i)
# nce_dict.update(nns_dict)
# summary_writer.add_scalars( "NCEs and NNSs" , nce_dict, i)
prog_bar.set_description("Pair:{}, iter:{}, loss_style:{}, loss_cont:{}, loss_cycle:{}".format(
z+1,
i+1,
loss_dict["L_style"],
loss_dict["L_content"],
loss_dict["L_cycle"])
)
e_optimizer.zero_grad()
g_optimizer.zero_grad()
loss.backward()
e_optimizer.step()
g_optimizer.step()
count += 1
""" generation result """
dec_result = tr_s2c_img_decs.clone().detach()
""" save the results """
if (i + 1) % args.sample_freq == 0 or i == 0 or i == args.iter-1:
c_img = os.path.basename(content_path) # 'in11.png'
s_img = os.path.basename(style_path) # 'tar11.png'
c_name = c_img.split('.' + args.fileType)[0] # in11
s_name = s_img.split('.' + args.fileType)[0] # tar11
pair_dir = '{}_'.format(z) + c_name + '_and_' + s_name
pair_iter_dir = '{}_'.format(z) + "iter" + str(i) + "_" +c_name + '_and_' + s_name
""" making folder to save results """
root_iter_path = os.path.join(os.path.join(args.save_root, c_name), pair_dir)
# root_iter_path = os.path.join(args.save_root, pair_dir)
save_dir_path = os.path.join(root_iter_path, pair_iter_dir)
if not os.path.isdir(args.save_root):
os.makedirs(args.save_root)
if not os.path.isdir(save_dir_path):
os.makedirs(save_dir_path)
""" denormalization """
if args.denorm == True:
result = ( np.clip(( dec_result[0].permute(1,2,0).clone().detach().cpu().numpy()) *std + mean, 0.0, 1.0)*255.0).astype('uint8')[::,::,::-1]
else:
result = ( np.clip(( dec_result[0].permute(1,2,0).clone().detach().cpu().numpy()), 0.0, 1.0)*255.0).astype('uint8')[::,::,::-1]
""" change the form """
content_save = (cv2.resize(content_256, (512,384))).astype('uint8')[::,::,::-1]
style_save = (cv2.resize(style_256, (512,384))).astype('uint8')[::,::,::-1]
result = (cv2.resize(result, (512,384))).astype('uint8')
bundle_result = np.stack( (content_save, style_save, result), axis=1 )
bundle_result = bundle_result.reshape((384, 512*3, 3))
""" save the result """
cv2.imwrite( os.path.join(save_dir_path, c_name+'.png'), content_save)
cv2.imwrite( os.path.join(save_dir_path, s_name+'.png'), style_save)
cv2.imwrite( os.path.join(save_dir_path, 'result.png'), result)
cv2.imwrite( os.path.join(save_dir_path, c_name + '_' + s_name + '_' + 'result_bundle.png'), bundle_result)
if i == args.iter - 1:
cv2.imwrite( os.path.join(root_iter_path, c_name + '_' + s_name + '_' + 'result_bundle.png'), bundle_result)
cv2.imwrite( os.path.join(os.path.join(args.save_root, c_name), c_name + '_' + s_name + '_' + 'result_bundle.png'), bundle_result)
# if args.denorm == True:
# warped_s2c_imgs= (np.clip( ( warped_s2c_imgs[0].clone().permute(1,2,0).detach().cpu().numpy()) *std + mean, 0.0, 1.0)*255.0).astype('uint8')[::,::,::-1]
# else:
# warped_s2c_imgs = (np.clip( ( warped_s2c_imgs[0].clone().permute(1,2,0).detach().cpu().numpy()), 0.0, 1.0)*255.0).astype('uint8')[::,::,::-1]
# resolution = ['256', '128', '64', '32', '16']
# cv2.imwrite( os.path.join(save_dir_path, 'warp_{0}.png'.format(resolution[2])) , warped_s2c_imgs ) # k=2 (64)
print("time :", time.time() - start)
print("root path: ", root_iter_path)
del model
gc.collect()