-
Notifications
You must be signed in to change notification settings - Fork 16
/
utils.py
400 lines (347 loc) · 12.9 KB
/
utils.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
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
"""
Copyright (C) 2021 NVIDIA Corporation. All rights reserved.
Licensed under the NVIDIA Source Code License. See LICENSE at the main github page.
Authors: Seung Wook Kim, Jonah Philion, Antonio Torralba, Sanja Fidler
"""
import torch
import torch.nn as nn
from torch import autograd
from torch.autograd import Variable
import torch.nn.functional as F
from torch import distributions
import math
import random
import numpy as np
import os
def run_latent_decoder(latent_decoder, input, opts=None, return_all_outputs=False):
decoder_input = [input]
z_first, z_last = input[:, :opts.spatial_total_dim], input[:, opts.spatial_total_dim:]
z_first = z_first.view(input.size(0), -1, opts.spatial_h, opts.spatial_w)
bs = 4
num_chunk = int(np.ceil(z_first.size(0) / bs))
imgs = []
for i in range(num_chunk):
out = latent_decoder({'spatial_z':z_first[i*bs:(i+1)*bs],
'theme_z': z_last[i*bs:(i+1)*bs]}, decode_only=True)
imgs.append(out['image'])
output = torch.cat(imgs, dim=0)
return output
def get_latent_decoder(opts):
from latent_decoder_model.model.model import styleVAEGAN
img_size = opts.img_size[0]
latent_z_size = opts.latent_z_size
latent_decoder_model_path = opts.latent_decoder_model_path
if opts.gpu >= 0:
saved_ckpt = torch.load(latent_decoder_model_path)
else:
saved_ckpt = torch.load(latent_decoder_model_path, map_location=torch.device('cpu'))
opts.DO_PLAIN_GAN = False
saved_args = saved_ckpt['args']
vae_model = styleVAEGAN
model = vae_model(
saved_args.size, n_mlp=saved_args.n_mlp, channel_multiplier=saved_args.channel_multiplier, args=saved_args,
)
model.load_state_dict(saved_ckpt['vae_ema'], strict=False)
model.eval()
model = model.cpu()
return model
def save_model(fname, epoch, netG, netD, opts):
outdict = {'epoch': epoch, 'netG': netG.state_dict(), 'netD': netD.state_dict(), 'opts': opts}
torch.save(outdict, fname)
def save_optim(fname, epoch, optG_temporal, optG_graphic, optD):
outdict = {'epoch': epoch, 'optG_temporal': optG_temporal.state_dict(), 'optG_graphic': optG_graphic.state_dict(), 'optD': optD.state_dict()}
torch.save(outdict, fname)
def adjust_learning_rate(opt, lr):
for param_group in opt.param_groups:
param_group['lr'] = lr
def choose_optimizer(model, options, lr=None, exclude=None, include=None, model_name=''):
try:
wd = options.wd
except:
wd = 0.0
if lr == None:
lr = options.lr
if type(model) is list:
params = model
else:
params = model.parameters()
if exclude is not None:
params = []
for name, W in model.named_parameters():
if type(exclude) is list:
excluded = False
for exc in exclude:
if exc in name:
excluded = True
print(model_name + ', Exclude: ' + name)
break
if not excluded:
params.append(W)
print(model_name + ', Include: ' + name)
elif not exclude in name:
params.append(W)
print(model_name + ', Include: ' + name)
else:
print(model_name + ', Exclude: ' + name)
if include is not None:
params = []
for name, W in model.named_parameters():
if type(include) is list:
for inc in include:
if inc in name:
params.append(W)
print(model_name + ', Include: ' + name)
break
elif include in name:
params.append(W)
print(model_name + ', Include: ' + name)
if options.optimizer == 'sgd':
optimizer = torch.optim.SGD(params, lr=lr,
weight_decay=wd)
elif options.optimizer == 'adam':
optimizer = torch.optim.Adam(params, lr=lr,
weight_decay=wd, betas=(0.0, 0.9))
elif options.optimizer == 'rmsprop':
optimizer = torch.optim.RMSprop(params, lr=lr, alpha=0.95, eps=0.01, momentum=0.9)
else:
raise RuntimeError('invalid oprimizer type')
return optimizer
def build_models(opts, tmp_get_old=False):
from simulator_model.dynamics_engine import EngineGenerator as Generator
from simulator_model.discriminator import Discriminator
# Build models
generator = Generator(
opts
)
discriminator = Discriminator(
opts,
nfilter=opts.nfilterD
)
if opts.gpu is not None and not opts.gpu < 0 :
return generator.to(opts.device), discriminator.to(opts.device)
else:
return generator, discriminator
def weights_init(m):
if isinstance(m, MyConvo2d):
if m.conv.weight is not None:
if m.he_init:
init.kaiming_uniform_(m.conv.weight)
else:
init.xavier_uniform_(m.conv.weight)
if m.conv.bias is not None:
init.constant_(m.conv.bias, 0.0)
if isinstance(m, nn.Linear):
if m.weight is not None:
init.xavier_uniform_(m.weight)
if m.bias is not None:
init.constant_(m.bias, 0.0)
def copy_weights(source, target):
target.data = source.data
return
from termcolor import colored
def print_color(txt, color):
''' print <txt> to terminal using colors
'''
print(colored(txt, color))
return
def check_arg(opts, arg):
v = vars(opts)
if arg in v:
if type(v[arg]) == bool:
return v[arg]
else:
return True
else:
return False
def check_gpu(gpu, *args):
'''
'''
if gpu == None or gpu < 0:
if isinstance(args[0], dict):
d = args[0]
var_dict = {}
for key in d:
var_dict[key] = Variable(d[key])
if len(args) > 1:
return [var_dict] + check_gpu(gpu, *args[1:])
else:
return [var_dict]
if isinstance(args[0], list):
return [Variable(a) for a in args[0]]
# a list of arguments
if len(args) > 1:
return [Variable(a) for a in args]
else:
return Variable(args[0])
else:
if isinstance(args[0], dict):
d = args[0]
var_dict = {}
for key in d:
var_dict[key] = Variable(d[key]).to('cuda')
if len(args) > 1:
return [var_dict] + check_gpu(gpu, *args[1:])
else:
return [var_dict]
if isinstance(args[0], list):
return [Variable(a).to('cuda') for a in args[0]]
# a list of arguments
if len(args) > 1:
return [Variable(a).to('cuda') for a in args]
else:
return Variable(args[0]).to('cuda')
def get_data(data_iters, opts, get_rand=False):
tmp_states, tmp_actions, tmp_neg_actions = [], [], []
states, actions, neg_actions = [], [], []
if type(data_iters) is list:
for data_iter in data_iters:
s, a, na = data_iter.next()
tmp_states.append(s)
tmp_actions.append(a)
tmp_neg_actions.append(na)
else:
s, a, na = next(data_iters)
tmp_states.append(s)
tmp_actions.append(a)
tmp_neg_actions.append(na)
for j in range(len(tmp_states[0])): # over time steps
gs, ga, gna = [], [], []
for k in range(len(tmp_states[0][0])): # over batches
for i in range(len(tmp_states)): # over data type
gs.append(tmp_states[i][j][k])
ga.append(tmp_actions[i][j][k])
gna.append(tmp_neg_actions[i][j][k])
states.append(torch.stack(gs, dim=0))
actions.append(torch.stack(ga, dim=0))
neg_actions.append(torch.stack(gna, dim=0))
states = [check_gpu(opts.gpu, a) for a in states]
actions = [check_gpu(opts.gpu, a) for a in actions]
neg_actions = [check_gpu(opts.gpu, a) for a in neg_actions]
return states, actions, neg_actions
def load_state_dict(self, state_dict):
import torch.nn as nn
own_state = self.state_dict()
for name, param in state_dict.items():
if name not in own_state:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ ' + name + ' NOT LOADED')
continue
if isinstance(param, nn.Parameter):
# backwards compatibility for serialized parameters
param = param.data
try:
own_state[name].copy_(param)
print('++++++++++++++++++++++++++++++ ' + name + ' LOADED')
except:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ ' + name + ' NOT LOADED')
print(param.size())
print(own_state[name].size())
continue
def compute_grad2(d_out, x_in, allow_unused=False, batch_size=None, gpu=0, ns=1):
if d_out is None:
return check_gpu(gpu, torch.FloatTensor([0]))
if batch_size is None:
batch_size = x_in.size(0)
grad_dout = autograd.grad(
outputs=d_out.sum(), inputs=x_in,
create_graph=True, retain_graph=True, only_inputs=True,
allow_unused=allow_unused
)[0]
# import pdb; pdb.set_trace();
grad_dout2 = grad_dout.pow(2)
# xassert(grad_dout2.size() == x_in.size())
reg = grad_dout2.view(batch_size, -1).sum(1) * (ns * 1.0 / 6)
return reg
def toggle_grad(model, requires_grad):
for p in model.parameters():
p.requires_grad_(requires_grad)
def load_my_state_dict(self, state_dict):
import torch.nn as nn
own_state = self.state_dict()
print('now')
for name, param in own_state.items():
print(name)
if name not in state_dict:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ ' + name + ' NOT EXIST IN SAVED MODEL CKPT')
continue
print('load')
for name, param in state_dict.items():
print(name)
name = name.replace('module.', '')
if name not in own_state:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ ' + name + ' NOT EXIST IN CURRENT CODE MODEL FILE')
continue
print(name)
if isinstance(param, nn.Parameter):
# backwards compatibility for serialized parameters
param = param.data
try:
own_state[name].copy_(param)
except:
print('@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ ' + name + ' NOT LOADED')
print(param.size())
print(own_state[name].size())
continue
def init_config_model_for_play():
import config
import sys
import numpy as np
from trainer import Trainer
parser = config.init_parser()
opts, args = parser.parse_args(sys.argv)
force_play_from_data = opts.force_play_from_data
saved_model = opts.saved_model
initial_screen = opts.initial_screen
seed = opts.seed
port = opts.port
log_dir = opts.log_dir
gpu = opts.gpu
latent_decoder_model_path = opts.latent_decoder_model_path
recording_name = opts.recording_name
# create model
opts = torch.load(
opts.saved_model,
map_location='cpu')['opts']
opts.seed = seed
opts.port = port
opts.log_dir = log_dir
opts.play = True
opts.bs = 1
opts.gpu = gpu
opts.saved_model = saved_model
opts.initial_screen = initial_screen
opts.recording_name = recording_name
opts.latent_decoder_model_path= latent_decoder_model_path
opts.force_play_from_data = force_play_from_data
warm_up = opts.warm_up
torch.manual_seed(opts.seed)
np.random.seed(opts.seed)
opts.width_mul = 1.0 if not check_arg(opts, 'width_mul') else opts.width_mul
opts.spatial_dim = 4 if not check_arg(opts, 'spatial_dim') else opts.spatial_dim
if not check_arg(opts, 'spatial_h'):
opts.spatial_h = opts.spatial_dim
if not check_arg(opts, 'spatial_w'):
opts.spatial_w = int(opts.spatial_dim*opts.width_mul)
if not check_arg(opts, 'theme_d'):
opts.theme_d = opts.separate_holistic_style_dim
if not check_arg(opts, 'spatial_d'):
opts.spatial_d = (opts.latent_z_size - opts.theme_d) // int(opts.spatial_dim*opts.spatial_dim*opts.width_mul)
opts.device = 'cuda'
if opts.gpu != -1:
torch.cuda.manual_seed(opts.seed)
# create model
netG, _ = build_models(opts)
trainer = Trainer(opts,
netG, None,
None, None, None,
opts.LAMBDA)
assert opts.saved_model != None, 'Empty saved model'
# load the weights
print('loading netG')
load_my_state_dict(trainer.netG,
torch.load(
opts.saved_model,
map_location='cpu')['netG'] )
print('Models loaded')
latent_decoder = get_latent_decoder(opts)
latent_decoder = latent_decoder.cuda(0)
return opts, trainer, gpu, latent_decoder