-
Notifications
You must be signed in to change notification settings - Fork 3
/
utils.py
539 lines (434 loc) · 15.6 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
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
import random
import re
import time
import math
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
# from omegaconf import OmegaConf
from torch import distributions as pyd
from torch.distributions.utils import _standard_normal
from torch.distributions import Distribution, Normal
class eval_mode:
def __init__(self, *models):
self.models = models
def __enter__(self):
self.prev_states = []
for model in self.models:
self.prev_states.append(model.training)
model.train(False)
def __exit__(self, *args):
for model, state in zip(self.models, self.prev_states):
model.train(state)
return False
def set_seed_everywhere(seed):
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
def chain(*iterables):
for it in iterables:
yield from it
def soft_update_params(net, target_net, tau):
for param, target_param in zip(net.parameters(), target_net.parameters()):
target_param.data.copy_(tau * param.data +
(1 - tau) * target_param.data)
def hard_update_params(net, target_net):
for param, target_param in zip(net.parameters(), target_net.parameters()):
target_param.data.copy_(param.data)
def to_torch(xs, device):
return tuple(torch.as_tensor(x, device=device) for x in xs)
def weight_init(m):
"""Custom weight init for Conv2D and Linear layers."""
if isinstance(m, nn.Linear):
nn.init.orthogonal_(m.weight.data)
if hasattr(m.bias, 'data'):
m.bias.data.fill_(0.0)
elif isinstance(m, nn.Conv2d) or isinstance(m, nn.ConvTranspose2d):
gain = nn.init.calculate_gain('relu')
nn.init.orthogonal_(m.weight.data, gain)
if hasattr(m.bias, 'data'):
m.bias.data.fill_(0.0)
def param_norm(params, norm_type=2.0):
total_norm = torch.norm(
torch.stack([torch.norm(p.detach(), norm_type) for p in params]),
norm_type)
return total_norm.item()
class Until:
def __init__(self, until, action_repeat=1):
self._until = until
self._action_repeat = action_repeat
def __call__(self, step):
if self._until is None:
return True
until = self._until // self._action_repeat
return step < until
class Every:
def __init__(self, every, action_repeat=1):
self._every = every
self._action_repeat = action_repeat
def change_every(self, freq):
self._every *= freq
def __call__(self, step):
if self._every is None:
return False
every = self._every // self._action_repeat
if step % every == 0:
return True
return False
class Timer:
def __init__(self):
self._start_time = time.time()
self._last_time = time.time()
def reset(self):
elapsed_time = time.time() - self._last_time
self._last_time = time.time()
total_time = time.time() - self._start_time
return elapsed_time, total_time
def total_time(self):
return time.time() - self._start_time
class TruncatedNormal(pyd.Normal):
def __init__(self, loc, scale, low=-1.0, high=1.0, eps=1e-6):
super().__init__(loc, scale, validate_args=False)
self.low = low
self.high = high
self.eps = eps
def _clamp(self, x):
clamped_x = torch.clamp(x, self.low + self.eps, self.high - self.eps)
x = x - x.detach() + clamped_x.detach()
return x
def sample(self, clip=None, sample_shape=torch.Size()):
shape = self._extended_shape(sample_shape)
eps = _standard_normal(shape,
dtype=self.loc.dtype,
device=self.loc.device)
eps *= self.scale
if clip is not None:
eps = torch.clamp(eps, -clip, clip)
x = self.loc + eps
return self._clamp(x)
class TanhTransform(pyd.transforms.Transform):
domain = pyd.constraints.real
codomain = pyd.constraints.interval(-1.0, 1.0)
bijective = True
sign = +1
def __init__(self, cache_size=1):
super().__init__(cache_size=cache_size)
@staticmethod
def atanh(x):
return atanh(x)
# return 0.5 * (x.log1p() - (-x).log1p())
def __eq__(self, other):
return isinstance(other, TanhTransform)
def _call(self, x):
return x.tanh()
def _inverse(self, y):
# We do not clamp to the boundary here as it may degrade the performance of certain algorithms.
# one should use `cache_size=1` instead
return self.atanh(y)
def log_abs_det_jacobian(self, x, y):
# We use a formula that is more numerically stable, see details in the following link
# https://github.com/tensorflow/probability/commit/ef6bb176e0ebd1cf6e25c6b5cecdd2428c22963f#diff-e120f70e92e6741bca649f04fcd907b7
return 2. * (math.log(2.) - x - F.softplus(-2. * x))
class SquashedNormal(pyd.transformed_distribution.TransformedDistribution):
def __init__(self, loc, scale):
self.loc = loc
self.scale = scale
self.base_dist = pyd.Normal(loc, scale)
transforms = [TanhTransform()]
super().__init__(self.base_dist, transforms)
@property
def mean(self):
mu = self.loc
for tr in self.transforms:
mu = tr(mu)
return mu
class SquashedNormal2:
def __init__(self, loc, scale):
self.loc = loc
self.scale = scale
self.normal = pyd.normal.Normal(loc, scale)
def sample(self, sample_shape=torch.Size()):
# Gradients will and should *not* pass through this operation.
z = self.normal.sample(sample_shape=sample_shape).detach()
return torch.tanh(z)
def rsample(self, sample_shape=torch.Size()):
# Gradients will and should pass through this operation.
z = self.normal.rsample(sample_shape=sample_shape)
return torch.tanh(z)
def log_prob(self, action):
# pre_tanh_value = torch.log((1 + action) / (1 - action)) / 2 # arc-tan 函数, 还原到取 tanh 之前时候
pre_tanh_value = atanh(action)
logp_pi = self.normal.log_prob(pre_tanh_value).sum(dim=-1)
if torch.isnan(logp_pi).any():
print("gaussian error:", logp_pi, self.loc, (self.scale >= 0).all())
# print("in log_prob 0:", action.shape, pre_tanh_value.shape)
# print("in log_prob 1:", self.normal.log_prob(pre_tanh_value).shape, logp_pi.shape)
logp_pi -= (2 * (np.log(2) - pre_tanh_value - F.softplus(-2 * pre_tanh_value))).sum(dim=-1)
# print("in log_prob 2:", logp_pi.shape)
# print("is nan 0:", torch.isnan(logp_pi).any())
# if torch.isnan(logp_pi).any():
# print("is nan 1:", torch.isnan(pre_tanh_value).any(), torch.isnan(action).any())
# print("is nan 2:", torch.isnan(F.softplus(-2 * pre_tanh_value)).any())
# print("is nan 3:", torch.isnan(self.normal.log_prob(pre_tanh_value)).any())
# print("is nan 4:", torch.isnan(self.loc).any(), torch.isnan(self.scale).any())
# print("is nan 5:", torch.isnan((2 * (np.log(2) - pre_tanh_value))).any())
# print("is nan 6:", torch.isnan(-F.softplus(-2 * pre_tanh_value)).any())
# print("is nan 7:", torch.isnan(-pre_tanh_value-F.softplus(-2 * pre_tanh_value)).any())
# print("is nan 8:", torch.isnan(np.log(2)-pre_tanh_value-F.softplus(-2 * pre_tanh_value)).any())
# print("is nan 9:", np.log(2), pre_tanh_value.shape, F.softplus(-2 * pre_tanh_value).shape)
# print("is nan 10:", action.mean(), pre_tanh_value.mean(), pre_tanh_value.min(), pre_tanh_value.max(), F.softplus(-2 * pre_tanh_value).mean())
# print("is nan 11:", torch.isnan(logp_pi).any())
# print("")
return logp_pi
@property
def mean(self):
return torch.tanh(self.loc)
def weight_init_he_relu(m, activation='relu'):
""" kaiming uniform init for linear layers."""
if isinstance(m, nn.Linear):
nn.init.kaiming_uniform_(m.weight.data, mode='fan_in', nonlinearity=activation)
if hasattr(m.bias, 'data'):
m.bias.data.fill_(0.01)
def weight_init_he_lrelu(m, activation='leaky_relu'):
""" kaiming uniform init for linear layers."""
if isinstance(m, nn.Linear):
nn.init.kaiming_uniform_(m.weight.data, mode='fan_in', nonlinearity=activation)
if hasattr(m.bias, 'data'):
m.bias.data.fill_(0.01)
def fanin_init(tensor):
size = tensor.size()
if len(size) == 2:
fan_in = size[0]
elif len(size) > 2:
fan_in = np.prod(size[1:])
else:
raise Exception("Shape must be have dimension at least 2.")
bound = 1. / np.sqrt(fan_in)
return tensor.data.uniform_(-bound, bound)
def weight_init_cql(m):
""" kaiming uniform init for linear layers."""
if isinstance(m, nn.Linear):
fanin_init(m.weight)
if hasattr(m.bias, 'data'):
m.bias.data.fill_(0.1)
def grad_norm(params, norm_type=2.0):
params = [p for p in params if p.grad is not None]
# pytorch_total_params = sum(p.numel() for p in params if p.requires_grad)
total_norm = torch.norm(
torch.stack([torch.norm(p.grad.detach(), norm_type) for p in params]),
norm_type)
return total_norm.item()
def schedule(schdl, step):
try:
return float(schdl)
except ValueError:
match = re.match(r'linear\((.+),(.+),(.+)\)', schdl)
if match:
init, final, duration = [float(g) for g in match.groups()]
mix = np.clip(step / duration, 0.0, 1.0)
return (1.0 - mix) * init + mix * final
match = re.match(r'step_linear\((.+),(.+),(.+),(.+),(.+)\)', schdl)
if match:
init, final1, duration1, final2, duration2 = [
float(g) for g in match.groups()
]
if step <= duration1:
mix = np.clip(step / duration1, 0.0, 1.0)
return (1.0 - mix) * init + mix * final1
else:
mix = np.clip((step - duration1) / duration2, 0.0, 1.0)
return (1.0 - mix) * final1 + mix * final2
raise NotImplementedError(schdl)
class RandomShiftsAug(nn.Module):
def __init__(self, pad):
super().__init__()
self.pad = pad
def forward(self, x):
x = x.float()
n, c, h, w = x.size()
assert h == w
padding = tuple([self.pad] * 4)
x = F.pad(x, padding, 'replicate') # (n, c, h+2*pad, w+2*pad) after padding
eps = 1.0 / (h + 2 * self.pad)
arange = torch.linspace(-1.0 + eps,
1.0 - eps,
h + 2 * self.pad,
device=x.device,
dtype=x.dtype)[:h]
arange = arange.unsqueeze(0).repeat(h, 1).unsqueeze(2)
base_grid = torch.cat([arange, arange.transpose(1, 0)], dim=2)
base_grid = base_grid.unsqueeze(0).repeat(n, 1, 1, 1)
shift = torch.randint(0,
2 * self.pad + 1,
size=(n, 1, 1, 2),
device=x.device,
dtype=x.dtype)
shift *= 2.0 / (h + 2 * self.pad)
grid = base_grid + shift
return F.grid_sample(x,
grid,
padding_mode='zeros',
align_corners=False)
class RMS(object):
"""running mean and std """
def __init__(self, device, epsilon=1e-4, shape=(1,)):
self.M = torch.zeros(shape).to(device)
self.S = torch.ones(shape).to(device)
self.n = epsilon
def __call__(self, x):
bs = x.size(0)
delta = torch.mean(x, dim=0) - self.M
new_M = self.M + delta * bs / (self.n + bs)
new_S = (self.S * self.n + torch.var(x, dim=0) * bs +
torch.square(delta) * self.n * bs /
(self.n + bs)) / (self.n + bs)
self.M = new_M
self.S = new_S
self.n += bs
return self.M, self.S
class PBE(object):
"""particle-based entropy based on knn normalized by running mean """
def __init__(self, rms, knn_clip, knn_k, knn_avg, knn_rms, device):
self.rms = rms
self.knn_rms = knn_rms
self.knn_k = knn_k
self.knn_avg = knn_avg
self.knn_clip = knn_clip
self.device = device
def __call__(self, rep):
source = target = rep
b1, b2 = source.size(0), target.size(0)
# (b1, 1, c) - (1, b2, c) -> (b1, 1, c) - (1, b2, c) -> (b1, b2, c) -> (b1, b2)
sim_matrix = torch.norm(source[:, None, :].view(b1, 1, -1) -
target[None, :, :].view(1, b2, -1),
dim=-1,
p=2)
# sim_matrix calculate the distance between two points. shape=(b1, b2)
# The topk returned after the topk is the topk element closest to each element in b1 in b2
reward, _ = sim_matrix.topk(self.knn_k,
dim=1,
largest=False, # False return the minimal distance
sorted=True) # return shape=(b1, knn_k)
if not self.knn_avg: # only keep k-th nearest neighbor
reward = reward[:, -1] # k-th minimal
reward = reward.reshape(-1, 1) # (b1, 1)
reward /= self.rms(reward)[0] if self.knn_rms else 1.0 # regularize
reward = torch.maximum(
reward - self.knn_clip,
torch.zeros_like(reward).to(self.device)
) if self.knn_clip >= 0.0 else reward # (b1, 1)
else: # average over all k nearest neighbors
reward = reward.reshape(-1, 1) # (b1 * k, 1)
reward /= self.rms(reward)[0] if self.knn_rms else 1.0
reward = torch.maximum(
reward - self.knn_clip,
torch.zeros_like(reward).to(
self.device)) if self.knn_clip >= 0.0 else reward
reward = reward.reshape((b1, self.knn_k)) # (b1, k)
reward = reward.mean(dim=1, keepdim=True) # (b1, 1) take average and return
reward = torch.log(reward + 1.0)
return reward
class ConstantSchedule(object):
def __init__(self, value):
"""Value remains constant over time.
Parameters
----------
value: float
Constant value of the schedule
"""
self._v = value
def value(self, t):
"""See Schedule.value"""
return self._v
def linear_interpolation(l, r, alpha):
return l + alpha * (r - l)
class PiecewiseSchedule(object):
def __init__(self, endpoints, interpolation=linear_interpolation, outside_value=None):
""" From OpenAI baselines
"""
idxes = [e[0] for e in endpoints]
assert idxes == sorted(idxes)
self._interpolation = interpolation
self._outside_value = outside_value
self._endpoints = endpoints
def value(self, t):
# See Schedule.value
for (l_t, l), (r_t, r) in zip(self._endpoints[:-1], self._endpoints[1:]):
if l_t <= t and t < r_t:
alpha = float(t - l_t) / (r_t - l_t)
return self._interpolation(l, r, alpha)
# t does not belong to any of the pieces, so doom.
assert self._outside_value is not None
return self._outside_value
class TanhNormal(Distribution):
"""
Represent distribution of X where
X ~ tanh(Z)
Z ~ N(mean, std)
Note: this is not very numerically stable.
"""
def __init__(self, normal_mean, normal_std, epsilon=1e-6):
"""
:param normal_mean: Mean of the normal distribution
:param normal_std: Std of the normal distribution
:param epsilon: Numerical stability epsilon when computing log-prob.
"""
self.normal_mean = normal_mean
self.normal_std = normal_std
self.normal = Normal(normal_mean, normal_std)
self.epsilon = epsilon
def sample_n(self, n, return_pre_tanh_value=False):
z = self.normal.sample_n(n)
if return_pre_tanh_value:
return torch.tanh(z), z
else:
return torch.tanh(z)
def log_prob(self, value, pre_tanh_value=None):
"""
:param value: some value, x
:param pre_tanh_value: arctanh(x)
:return:
"""
if pre_tanh_value is None:
pre_tanh_value = torch.log((1 + value) / (1 - value)) / 2 # arc-tan
# log of action - log(1-tanh(a)*tanh(a)+e)
# refer https://garage.readthedocs.io/en/v2020.06.1/_modules/garage/torch/distributions/tanh_normal.html#TanhNormal.log_prob
return self.normal.log_prob(pre_tanh_value) - torch.log(1 - value * value + self.epsilon)
def sample(self, return_pretanh_value=False):
"""
Gradients will and should *not* pass through this operation.
See https://github.com/pytorch/pytorch/issues/4620 for discussion.
"""
z = self.normal.sample().detach()
if return_pretanh_value:
return torch.tanh(z), z
else:
return torch.tanh(z)
def rsample(self, return_pretanh_value=False):
"""
Sampling in the reparameterization case.
"""
z = (self.normal_mean.cuda() + self.normal_std.cuda() *
Normal(torch.zeros(self.normal_mean.size()).cuda(), torch.ones(self.normal_std.size()).cuda()).sample())
z.requires_grad_()
if return_pretanh_value:
return torch.tanh(z), z
else:
return torch.tanh(z)
def atanh(x):
one_plus_x = (1 + x).clamp(min=1e-6)
one_minus_x = (1 - x).clamp(min=1e-6)
return 0.5 * torch.log(one_plus_x / one_minus_x)
def total_parameters(qfs):
if not isinstance(qfs, list):
qfs = [qfs]
n = 0
for qfs_member in qfs:
n += sum(p.numel() for p in qfs_member.parameters())
return n
def decay_linear(t, init, minimum, total_steps):
fraction = min(float(t) / total_steps, 1.0)
return init + fraction * (minimum - init)