-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutil.py
193 lines (170 loc) · 8.44 KB
/
util.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
import random
import torch
from torch.nn import init
from torch.optim import lr_scheduler
import numpy as np
def check_manual_seed(seed):
""" If manual seed is not specified, choose a
random one and communicate it to the user.
Args:
seed: seed to check
"""
seed = seed or random.randint(1, 10000)
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
print("Using manual seed: {seed}".format(seed=seed))
return
def get_scheduler(optimizer, opt):
"""Return a learning rate scheduler
Parameters:
optimizer -- the optimizer of the network
opt (option class) -- stores all the experiment flags; needs to be a subclass of BaseOptions．
opt.lr_policy is the name of learning rate policy: linear | step | plateau | cosine
For 'linear', we keep the same learning rate for the first <opt.n_epochs> epochs
and linearly decay the rate to zero over the next <opt.n_epochs_decay> epochs.
For other schedulers (step, plateau, and cosine), we use the default PyTorch schedulers.
See https://pytorch.org/docs/stable/optim.html for more details.
"""
if opt.lr_policy == 'linear':
def lambda_rule(epoch):
lr_l = 1.0 - max(0, epoch + opt.iter_count - opt.n_iters) / float(opt.n_iters_decay + 1)
return lr_l
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lambda_rule)
elif opt.lr_policy == 'step':
scheduler = lr_scheduler.StepLR(optimizer, step_size=opt.lr_decay_iters, gamma=0.1)
elif opt.lr_policy == 'plateau':
scheduler = lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.2, threshold=0.01, patience=5)
elif opt.lr_policy == 'cosine':
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, T_max=opt.n_epochs, eta_min=0)
else:
return NotImplementedError('learning rate policy [%s] is not implemented', opt.lr_policy)
return scheduler
def init_weights(net, init_type='normal', init_gain=0.02):
"""Initialize network weights.
Parameters:
net (network) -- network to be initialized
init_type (str) -- the name of an initialization method: normal | xavier | kaiming | orthogonal
init_gain (float) -- scaling factor for normal, xavier and orthogonal.
We use 'normal' in the original pix2pix and CycleGAN paper. But xavier and kaiming might
work better for some applications. Feel free to try yourself.
"""
def init_func(m): # define the initialization function
classname = m.__class__.__name__
if hasattr(m, 'weight') and (classname.find('Conv') != -1 or classname.find('Linear') != -1):
if init_type == 'normal':
init.normal_(m.weight.data, 0.0, init_gain)
elif init_type == 'xavier':
init.xavier_normal_(m.weight.data, gain=init_gain)
elif init_type == 'kaiming':
init.kaiming_normal_(m.weight.data, a=0, mode='fan_in')
elif init_type == 'orthogonal':
init.orthogonal_(m.weight.data, gain=init_gain)
else:
raise NotImplementedError('initialization method [%s] is not implemented' % init_type)
if hasattr(m, 'bias') and m.bias is not None:
init.constant_(m.bias.data, 0.0)
elif classname.find('BatchNorm2d') != -1: # BatchNorm Layer's weight is not a matrix; only normal distribution applies.
init.normal_(m.weight.data, 1.0, init_gain)
init.constant_(m.bias.data, 0.0)
# print('initialize network with %s' % init_type)
net.apply(init_func) # apply the initialization function <init_func>
class ImagePool():
"""This class implements an image buffer that stores previously generated images.
This buffer enables us to update discriminators using a history of generated images
rather than the ones produced by the latest generators.
"""
def __init__(self, pool_size):
"""Initialize the ImagePool class
Parameters:
pool_size (int) -- the size of image buffer, if pool_size=0, no buffer will be created
"""
self.pool_size = pool_size
if self.pool_size > 0: # create an empty pool
self.num_imgs = 0
self.images = []
def query(self, images):
"""Return an image from the pool.
Parameters:
images: the latest generated images from the generator
Returns images from the buffer.
By 50/100, the buffer will return input images.
By 50/100, the buffer will return images previously stored in the buffer,
and insert the current images to the buffer.
"""
if self.pool_size == 0: # if the buffer size is 0, do nothing
return images
return_images = []
for image in images:
image = torch.unsqueeze(image.data, 0)
if self.num_imgs < self.pool_size: # if the buffer is not full; keep inserting current images to the buffer
self.num_imgs = self.num_imgs + 1
self.images.append(image)
return_images.append(image)
else:
p = random.uniform(0, 1)
if p > 0.5: # by 50% chance, the buffer will return a previously stored image, and insert the current image into the buffer
random_id = random.randint(0, self.pool_size - 1) # randint is inclusive
tmp = self.images[random_id].clone()
self.images[random_id] = image
return_images.append(tmp)
else: # by another 50% chance, the buffer will return the current image
return_images.append(image)
return_images = torch.cat(return_images, 0) # collect all the images and return
return return_images
def to_cuda(samples, device):
for k in samples.keys():
samples[k] = samples[k].to(device, non_blocking=True)
return samples
class data_prefetcher():
def __init__(self, loader, device, prefetch=True):
self.dataloader = loader
self.loader = iter(loader)
self.prefetch = prefetch
self.device = device
if prefetch:
self.stream = torch.cuda.Stream()
self.preload()
def preload(self):
try:
self.next_samples = next(self.loader)
except StopIteration:
self.loader = iter(self.dataloader)
self.next_samples = next(self.loader)
# self.next_samples = None
# return
# if record_stream() doesn't work, another option is to make sure device inputs are created
# on the main stream.
# self.next_input_gpu = torch.empty_like(self.next_input, device='cuda')
# self.next_target_gpu = torch.empty_like(self.next_target, device='cuda')
# Need to make sure the memory allocated for next_* is not still in use by the main stream
# at the time we start copying to next_*:
# self.stream.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(self.stream):
self.next_samples = to_cuda(self.next_samples, self.device)
# more code for the alternative if record_stream() doesn't work:
# copy_ will record the use of the pinned source tensor in this side stream.
# self.next_input_gpu.copy_(self.next_input, non_blocking=True)
# self.next_target_gpu.copy_(self.next_target, non_blocking=True)
# self.next_input = self.next_input_gpu
# self.next_target = self.next_target_gpu
# With Amp, it isn't necessary to manually convert data to half.
# if args.fp16:
# self.next_input = self.next_input.half()
# else:
def next(self):
if self.prefetch:
torch.cuda.current_stream().wait_stream(self.stream)
samples = self.next_samples
if samples is not None:
for k, v in samples.items():
v.record_stream(torch.cuda.current_stream())
self.preload()
else:
try:
samples = next(self.loader)
samples= to_cuda(samples, self.device)
except StopIteration:
samples = None
return samples