-
Notifications
You must be signed in to change notification settings - Fork 2
/
utils.py
320 lines (259 loc) · 12.2 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
import logging
import os
import numpy as np
import torch
import torch.nn.functional as F
from torch.distributed import init_process_group
from torch.optim.lr_scheduler import LambdaLR
from torch.optim import Optimizer
def apply_norm(x, norm, batch_norm=False):
if batch_norm:
return norm(x)
else:
return norm(x.transpose(-1, -2)).transpose(-1, -2)
def ddp_setup(rank, world_size, port):
"""
Args:
rank: Unique identifier of each process
world_size: Total number of processes
"""
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = port
init_process_group(backend="nccl", rank=rank, world_size=world_size)
def split_train_val(train, val_split):
train_len = int(len(train) * (1.0-val_split))
train, val = torch.utils.data.random_split(
train,
(train_len, len(train) - train_len),
generator=torch.Generator().manual_seed(42),
)
return train, val
class DistributedSamplerNoDuplicate(torch.utils.data.DistributedSampler):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
if not self.drop_last and len(self.dataset) % self.num_replicas != 0:
# some ranks may have less samples, that's fine
if self.rank >= len(self.dataset) % self.num_replicas:
self.num_samples -= 1
self.total_size = len(self.dataset)
def log_sum_exp(x):
""" numerically stable log_sum_exp implementation that prevents overflow """
# TF ordering
axis = len(x.size()) - 1
m, _ = torch.max(x, dim=axis)
m2, _ = torch.max(x, dim=axis, keepdim=True)
return m + torch.log(torch.sum(torch.exp(x - m2), dim=axis))
def log_prob_from_logits(x):
""" numerically stable log_softmax implementation that prevents overflow """
# TF ordering
axis = len(x.size()) - 1
m, _ = torch.max(x, dim=axis, keepdim=True)
return x - m - torch.log(torch.sum(torch.exp(x - m), dim=axis, keepdim=True))
# https://github.com/pclucas14/pixel-cnn-pp/blob/master/utils.py
def discretized_mix_logistic_loss(x, l):
""" log-likelihood for mixture of discretized logistics, assumes the data has been rescaled to [-1,1] interval """
# Pytorch ordering
x = x.permute(0, 2, 3, 1)
l = l.permute(0, 2, 3, 1)
xs = [int(y) for y in x.size()]
ls = [int(y) for y in l.size()]
# here and below: unpacking the params of the mixture of logistics
nr_mix = int(ls[-1] / 10)
logit_probs = l[:, :, :, :nr_mix]
l = l[:, :, :, nr_mix:].contiguous().view(xs + [nr_mix * 3]) # 3 for mean, scale, coef
means = l[:, :, :, :, :nr_mix]
# log_scales = torch.max(l[:, :, :, :, nr_mix:2 * nr_mix], -7.)
log_scales = torch.clamp(l[:, :, :, :, nr_mix:2 * nr_mix], min=-7.)
coeffs = torch.tanh(l[:, :, :, :, 2 * nr_mix:3 * nr_mix])
# here and below: getting the means and adjusting them based on preceding
# sub-pixels
x = x.contiguous()
x = x.unsqueeze(-1) + torch.zeros(xs + [nr_mix], device=x.device)
m2 = (means[:, :, :, 1, :] + coeffs[:, :, :, 0, :]
* x[:, :, :, 0, :]).view(xs[0], xs[1], xs[2], 1, nr_mix)
m3 = (means[:, :, :, 2, :] + coeffs[:, :, :, 1, :] * x[:, :, :, 0, :] +
coeffs[:, :, :, 2, :] * x[:, :, :, 1, :]).view(xs[0], xs[1], xs[2], 1, nr_mix)
means = torch.cat((means[:, :, :, 0, :].unsqueeze(3), m2, m3), dim=3)
centered_x = x - means
inv_stdv = torch.exp(-log_scales)
plus_in = inv_stdv * (centered_x + 1. / 255.)
cdf_plus = torch.sigmoid(plus_in)
min_in = inv_stdv * (centered_x - 1. / 255.)
cdf_min = torch.sigmoid(min_in)
# log probability for edge case of 0 (before scaling)
log_cdf_plus = plus_in - F.softplus(plus_in)
# log probability for edge case of 255 (before scaling)
log_one_minus_cdf_min = -F.softplus(min_in)
cdf_delta = cdf_plus - cdf_min # probability for all other cases
mid_in = inv_stdv * centered_x
# log probability in the center of the bin, to be used in extreme cases
# (not actually used in our code)
log_pdf_mid = mid_in - log_scales - 2. * F.softplus(mid_in)
# now select the right output: left edge case, right edge case, normal
# case, extremely low prob case (doesn't actually happen for us)
# this is what we are really doing, but using the robust version below for extreme cases in other applications and to avoid NaN issue with tf.select()
# log_probs = tf.select(x < -0.999, log_cdf_plus, tf.select(x > 0.999, log_one_minus_cdf_min, tf.log(cdf_delta)))
# robust version, that still works if probabilities are below 1e-5 (which never happens in our code)
# tensorflow backpropagates through tf.select() by multiplying with zero instead of selecting: this requires use to use some ugly tricks to avoid potential NaNs
# the 1e-12 in tf.maximum(cdf_delta, 1e-12) is never actually used as output, it's purely there to get around the tf.select() gradient issue
# if the probability on a sub-pixel is below 1e-5, we use an approximation
# based on the assumption that the log-density is constant in the bin of
# the observed sub-pixel value
inner_inner_cond = (cdf_delta > 1e-5).float()
inner_inner_out = inner_inner_cond * torch.log(torch.clamp(cdf_delta, min=1e-12)) + (1. - inner_inner_cond) * (log_pdf_mid - np.log(127.5))
inner_cond = (x > 0.999).float()
inner_out = inner_cond * log_one_minus_cdf_min + (1. - inner_cond) * inner_inner_out
cond = (x < -0.999).float()
log_probs = cond * log_cdf_plus + (1. - cond) * inner_out
log_probs = torch.sum(log_probs, dim=3) + log_prob_from_logits(logit_probs)
return -torch.sum(log_sum_exp(log_probs))
def sample_from_discretized_mix_logistic(l, nr_mix):
# Pytorch ordering
l = l.permute(0, 2, 3, 1)
ls = [int(y) for y in l.size()]
xs = ls[:-1] + [3]
# unpack parameters
logit_probs = l[:, :, :, :nr_mix]
l = l[:, :, :, nr_mix:].contiguous().view(xs + [nr_mix * 3])
# sample mixture indicator from softmax
temp = torch.empty(logit_probs.size(), device=l.device)
temp.uniform_(1e-5, 1. - 1e-5)
temp = logit_probs.data - torch.log(- torch.log(temp))
_, argmax = temp.max(dim=3)
one_hot = to_one_hot(argmax, nr_mix)
sel = one_hot.view(xs[:-1] + [1, nr_mix])
# select logistic parameters
means = torch.sum(l[:, :, :, :, :nr_mix] * sel, dim=4)
log_scales = torch.clamp(torch.sum(
l[:, :, :, :, nr_mix:2 * nr_mix] * sel, dim=4), min=-7.)
coeffs = torch.sum(torch.tanh(
l[:, :, :, :, 2 * nr_mix:3 * nr_mix]) * sel, dim=4)
# sample from logistic & clip to interval
# we don't actually round to the nearest 8bit value when sampling
u = torch.empty(means.size(), device=means.device)
u.uniform_(1e-5, 1. - 1e-5)
x = means + torch.exp(log_scales) * (torch.log(u) - torch.log(1. - u))
x0 = torch.clamp(torch.clamp(x[:, :, :, 0], min=-1.), max=1.)
x1 = torch.clamp(torch.clamp(
x[:, :, :, 1] + coeffs[:, :, :, 0] * x0, min=-1.), max=1.)
x2 = torch.clamp(torch.clamp(
x[:, :, :, 2] + coeffs[:, :, :, 1] * x0 + coeffs[:, :, :, 2] * x1, min=-1.), max=1.)
out = torch.cat([x0.view(xs[:-1] + [1]), x1.view(xs[:-1] + [1]), x2.view(xs[:-1] + [1])], dim=3)
# put back in Pytorch ordering
out = out.permute(0, 3, 1, 2)
return out
def to_one_hot(tensor, n, fill_with=1.):
# we perform one hot encore with respect to the last axis
one_hot = torch.FloatTensor(tensor.size() + (n,), device=tensor.device).zero_()
one_hot.scatter_(len(tensor.size()), tensor.unsqueeze(-1), fill_with)
return one_hot
def sample_from_discretized_mix_logistic_1d(l, nr_mix):
# Pytorch ordering
l = l.permute(0, 2, 3, 1)
ls = [int(y) for y in l.size()]
xs = ls[:-1] + [1] #[3]
# unpack parameters
logit_probs = l[:, :, :, :nr_mix]
l = l[:, :, :, nr_mix:].contiguous().view(xs + [nr_mix * 2]) # for mean, scale
# sample mixture indicator from softmax
temp = torch.FloatTensor(logit_probs.size(), device=l.device)
temp.uniform_(1e-5, 1. - 1e-5)
temp = logit_probs.data - torch.log(- torch.log(temp))
_, argmax = temp.max(dim=3)
one_hot = to_one_hot(argmax, nr_mix)
sel = one_hot.view(xs[:-1] + [1, nr_mix])
# select logistic parameters
means = torch.sum(l[:, :, :, :, :nr_mix] * sel, dim=4)
log_scales = torch.clamp(torch.sum(
l[:, :, :, :, nr_mix:2 * nr_mix] * sel, dim=4), min=-7.)
u = torch.FloatTensor(means.size(), device=l.device)
u.uniform_(1e-5, 1. - 1e-5)
x = means + torch.exp(log_scales) * (torch.log(u) - torch.log(1. - u))
x0 = torch.clamp(torch.clamp(x[:, :, :, 0], min=-1.), max=1.)
out = x0.unsqueeze(1)
return out
def setup_logger(name, src, result_path, filename="log"):
logger = logging.getLogger(name)
logger.setLevel(logging.DEBUG)
log_path = os.path.join(result_path, filename)
makedirs(log_path)
info_file_handler = logging.FileHandler(log_path)
info_file_handler.setLevel(logging.INFO)
logger.addHandler(info_file_handler)
logger.info(src)
with open(src) as f:
logger.info(f.read())
return logger
def makedirs(filename):
if not os.path.exists(os.path.dirname(filename)):
os.makedirs(os.path.dirname(filename))
def ensure_path(path, other):
if os.path.exists(path):
return path
return other
def count_parameters(model):
# for p in model.parameters():
# if p.requires_grad:
# print(p.shape)
return sum(p.numel() for p in model.parameters() if p.requires_grad)
class DotDict(dict):
"""dot.notation access to dictionary attributes
From https://stackoverflow.com/questions/2352181/how-to-use-a-dot-to-access-members-of-dictionary
Note that there are issues with updating values of nested DotDicts
"""
def __getattr__(*args):
# Allow nested dicts
val = dict.get(*args)
return DotDict(val) if type(val) is dict else val
__setattr__ = dict.__setitem__
__delattr__ = dict.__delitem__
__dir__ = dict.keys
class DummyWandb:
def __init__(self, *args, **kwargs):
self.args = args
self.kwargs = kwargs
self.config = {}
self.name = ""
self.id = ""
self.path = ""
self.dir = "./"
@staticmethod
def init(*args, **kwargs):
return DummyWandb(*args, **kwargs)
def log(self, *args, **kwargs):
return
def watch(self, *args, **kwargs):
return
def finish(self, *args, **kwargs):
return
def save(self, *args, **kwargs):
return
def get_cosine_schedule_with_warmup(
optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, num_cycles: float = 0.5, last_epoch: int = -1
):
""" From: https://github.com/huggingface/transformers/blob/main/src/transformers/optimization.py,
This way we don't have dependency on the `transformers` package.
Create a schedule with a learning rate that decreases following the values of the cosine function between the
initial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the
initial lr set in the optimizer.
Args:
optimizer ([`~torch.optim.Optimizer`]):
The optimizer for which to schedule the learning rate.
num_warmup_steps (`int`):
The number of steps for the warmup phase.
num_training_steps (`int`):
The total number of training steps.
num_cycles (`float`, *optional*, defaults to 0.5):
The number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0
following a half-cosine).
last_epoch (`int`, *optional*, defaults to -1):
The index of the last epoch when resuming training.
Return:
`torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
"""
def lr_lambda(current_step):
if current_step < num_warmup_steps:
return float(current_step) / float(max(1, num_warmup_steps))
progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
return LambdaLR(optimizer, lr_lambda, last_epoch)