-
Notifications
You must be signed in to change notification settings - Fork 200
/
latent_sde.py
417 lines (339 loc) · 16.4 KB
/
latent_sde.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
# Copyright 2020 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Latent SDE fit to a single time series with uncertainty quantification."""
import argparse
import logging
import math
import os
import random
from collections import namedtuple
from typing import Optional, Union
import matplotlib.pyplot as plt
import numpy as np
import torch
import tqdm
from torch import distributions, nn, optim
import torchsde
# w/ underscore -> numpy; w/o underscore -> torch.
Data = namedtuple('Data', ['ts_', 'ts_ext_', 'ts_vis_', 'ts', 'ts_ext', 'ts_vis', 'ys', 'ys_'])
class LinearScheduler(object):
def __init__(self, iters, maxval=1.0):
self._iters = max(1, iters)
self._val = maxval / self._iters
self._maxval = maxval
def step(self):
self._val = min(self._maxval, self._val + self._maxval / self._iters)
@property
def val(self):
return self._val
class EMAMetric(object):
def __init__(self, gamma: Optional[float] = .99):
super(EMAMetric, self).__init__()
self._val = 0.
self._gamma = gamma
def step(self, x: Union[torch.Tensor, np.ndarray]):
x = x.detach().cpu().numpy() if torch.is_tensor(x) else x
self._val = self._gamma * self._val + (1 - self._gamma) * x
return self._val
@property
def val(self):
return self._val
def str2bool(v):
"""Used for boolean arguments in argparse; avoiding `store_true` and `store_false`."""
if isinstance(v, bool): return v
if v.lower() in ('yes', 'true', 't', 'y', '1'): return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'): return False
else: raise argparse.ArgumentTypeError('Boolean value expected.')
def manual_seed(seed: int):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
def _stable_division(a, b, epsilon=1e-7):
b = torch.where(b.abs().detach() > epsilon, b, torch.full_like(b, fill_value=epsilon) * b.sign())
return a / b
class LatentSDE(torchsde.SDEIto):
def __init__(self, theta=1.0, mu=0.0, sigma=0.5):
super(LatentSDE, self).__init__(noise_type="diagonal")
logvar = math.log(sigma ** 2 / (2. * theta))
# Prior drift.
self.register_buffer("theta", torch.tensor([[theta]]))
self.register_buffer("mu", torch.tensor([[mu]]))
self.register_buffer("sigma", torch.tensor([[sigma]]))
# p(y0).
self.register_buffer("py0_mean", torch.tensor([[mu]]))
self.register_buffer("py0_logvar", torch.tensor([[logvar]]))
# Approximate posterior drift: Takes in 2 positional encodings and the state.
self.net = nn.Sequential(
nn.Linear(3, 200),
nn.Tanh(),
nn.Linear(200, 200),
nn.Tanh(),
nn.Linear(200, 1)
)
# Initialization trick from Glow.
self.net[-1].weight.data.fill_(0.)
self.net[-1].bias.data.fill_(0.)
# q(y0).
self.qy0_mean = nn.Parameter(torch.tensor([[mu]]), requires_grad=True)
self.qy0_logvar = nn.Parameter(torch.tensor([[logvar]]), requires_grad=True)
def f(self, t, y): # Approximate posterior drift.
if t.dim() == 0:
t = torch.full_like(y, fill_value=t)
# Positional encoding in transformers for time-inhomogeneous posterior.
return self.net(torch.cat((torch.sin(t), torch.cos(t), y), dim=-1))
def g(self, t, y): # Shared diffusion.
return self.sigma.repeat(y.size(0), 1)
def h(self, t, y): # Prior drift.
return self.theta * (self.mu - y)
def f_aug(self, t, y): # Drift for augmented dynamics with logqp term.
y = y[:, 0:1]
f, g, h = self.f(t, y), self.g(t, y), self.h(t, y)
u = _stable_division(f - h, g)
f_logqp = .5 * (u ** 2).sum(dim=1, keepdim=True)
return torch.cat([f, f_logqp], dim=1)
def g_aug(self, t, y): # Diffusion for augmented dynamics with logqp term.
y = y[:, 0:1]
g = self.g(t, y)
g_logqp = torch.zeros_like(y)
return torch.cat([g, g_logqp], dim=1)
def forward(self, ts, batch_size, eps=None):
eps = torch.randn(batch_size, 1).to(self.qy0_std) if eps is None else eps
y0 = self.qy0_mean + eps * self.qy0_std
qy0 = distributions.Normal(loc=self.qy0_mean, scale=self.qy0_std)
py0 = distributions.Normal(loc=self.py0_mean, scale=self.py0_std)
logqp0 = distributions.kl_divergence(qy0, py0).sum(dim=1) # KL(t=0).
aug_y0 = torch.cat([y0, torch.zeros(batch_size, 1).to(y0)], dim=1)
aug_ys = sdeint_fn(
sde=self,
y0=aug_y0,
ts=ts,
method=args.method,
dt=args.dt,
adaptive=args.adaptive,
rtol=args.rtol,
atol=args.atol,
names={'drift': 'f_aug', 'diffusion': 'g_aug'}
)
ys, logqp_path = aug_ys[:, :, 0:1], aug_ys[-1, :, 1]
logqp = (logqp0 + logqp_path).mean(dim=0) # KL(t=0) + KL(path).
return ys, logqp
def sample_p(self, ts, batch_size, eps=None, bm=None):
eps = torch.randn(batch_size, 1).to(self.py0_mean) if eps is None else eps
y0 = self.py0_mean + eps * self.py0_std
return sdeint_fn(self, y0, ts, bm=bm, method='srk', dt=args.dt, names={'drift': 'h'})
def sample_q(self, ts, batch_size, eps=None, bm=None):
eps = torch.randn(batch_size, 1).to(self.qy0_mean) if eps is None else eps
y0 = self.qy0_mean + eps * self.qy0_std
return sdeint_fn(self, y0, ts, bm=bm, method='srk', dt=args.dt)
@property
def py0_std(self):
return torch.exp(.5 * self.py0_logvar)
@property
def qy0_std(self):
return torch.exp(.5 * self.qy0_logvar)
def make_segmented_cosine_data():
ts_ = np.concatenate((np.linspace(0.3, 0.8, 10), np.linspace(1.2, 1.5, 10)), axis=0)
ts_ext_ = np.array([0.] + list(ts_) + [2.0])
ts_vis_ = np.linspace(0., 2.0, 300)
ys_ = np.cos(ts_ * (2. * math.pi))[:, None]
ts = torch.tensor(ts_).float()
ts_ext = torch.tensor(ts_ext_).float()
ts_vis = torch.tensor(ts_vis_).float()
ys = torch.tensor(ys_).float().to(device)
return Data(ts_, ts_ext_, ts_vis_, ts, ts_ext, ts_vis, ys, ys_)
def make_irregular_sine_data():
ts_ = np.sort(np.random.uniform(low=0.4, high=1.6, size=16))
ts_ext_ = np.array([0.] + list(ts_) + [2.0])
ts_vis_ = np.linspace(0., 2.0, 300)
ys_ = np.sin(ts_ * (2. * math.pi))[:, None] * 0.8
ts = torch.tensor(ts_).float()
ts_ext = torch.tensor(ts_ext_).float()
ts_vis = torch.tensor(ts_vis_).float()
ys = torch.tensor(ys_).float().to(device)
return Data(ts_, ts_ext_, ts_vis_, ts, ts_ext, ts_vis, ys, ys_)
def make_data():
data_constructor = {
'segmented_cosine': make_segmented_cosine_data,
'irregular_sine': make_irregular_sine_data
}[args.data]
return data_constructor()
def main():
# Dataset.
ts_, ts_ext_, ts_vis_, ts, ts_ext, ts_vis, ys, ys_ = make_data()
# Plotting parameters.
vis_batch_size = 1024
ylims = (-1.75, 1.75)
alphas = [0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55]
percentiles = [0.999, 0.99, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3, 0.2, 0.1]
vis_idx = np.random.permutation(vis_batch_size)
# From https://colorbrewer2.org/.
if args.color == "blue":
sample_colors = ('#8c96c6', '#8c6bb1', '#810f7c')
fill_color = '#9ebcda'
mean_color = '#4d004b'
num_samples = len(sample_colors)
else:
sample_colors = ('#fc4e2a', '#e31a1c', '#bd0026')
fill_color = '#fd8d3c'
mean_color = '#800026'
num_samples = len(sample_colors)
eps = torch.randn(vis_batch_size, 1).to(device) # Fix seed for the random draws used in the plots.
bm = torchsde.BrownianInterval(
t0=ts_vis[0],
t1=ts_vis[-1],
size=(vis_batch_size, 1),
device=device,
levy_area_approximation='space-time'
) # We need space-time Levy area to use the SRK solver
# Model.
model = LatentSDE().to(device)
optimizer = optim.Adam(model.parameters(), lr=1e-2)
scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=.999)
kl_scheduler = LinearScheduler(iters=args.kl_anneal_iters)
logpy_metric = EMAMetric()
kl_metric = EMAMetric()
loss_metric = EMAMetric()
if args.show_prior:
with torch.no_grad():
zs = model.sample_p(ts=ts_vis, batch_size=vis_batch_size, eps=eps, bm=bm).squeeze()
ts_vis_, zs_ = ts_vis.cpu().numpy(), zs.cpu().numpy()
zs_ = np.sort(zs_, axis=1)
img_dir = os.path.join(args.train_dir, 'prior.png')
plt.subplot(frameon=False)
for alpha, percentile in zip(alphas, percentiles):
idx = int((1 - percentile) / 2. * vis_batch_size)
zs_bot_ = zs_[:, idx]
zs_top_ = zs_[:, -idx]
plt.fill_between(ts_vis_, zs_bot_, zs_top_, alpha=alpha, color=fill_color)
# `zorder` determines who's on top; the larger the more at the top.
plt.scatter(ts_, ys_, marker='x', zorder=3, color='k', s=35) # Data.
plt.ylim(ylims)
plt.xlabel('$t$')
plt.ylabel('$Y_t$')
plt.tight_layout()
plt.savefig(img_dir, dpi=args.dpi)
plt.close()
logging.info(f'Saved prior figure at: {img_dir}')
for global_step in tqdm.tqdm(range(args.train_iters)):
# Plot and save.
if global_step % args.pause_iters == 0:
img_path = os.path.join(args.train_dir, f'global_step_{global_step}.png')
with torch.no_grad():
zs = model.sample_q(ts=ts_vis, batch_size=vis_batch_size, eps=eps, bm=bm).squeeze()
samples = zs[:, vis_idx]
ts_vis_, zs_, samples_ = ts_vis.cpu().numpy(), zs.cpu().numpy(), samples.cpu().numpy()
zs_ = np.sort(zs_, axis=1)
plt.subplot(frameon=False)
if args.show_percentiles:
for alpha, percentile in zip(alphas, percentiles):
idx = int((1 - percentile) / 2. * vis_batch_size)
zs_bot_, zs_top_ = zs_[:, idx], zs_[:, -idx]
plt.fill_between(ts_vis_, zs_bot_, zs_top_, alpha=alpha, color=fill_color)
if args.show_mean:
plt.plot(ts_vis_, zs_.mean(axis=1), color=mean_color)
if args.show_samples:
for j in range(num_samples):
plt.plot(ts_vis_, samples_[:, j], color=sample_colors[j], linewidth=1.0)
if args.show_arrows:
num, dt = 12, 0.12
t, y = torch.meshgrid(
[torch.linspace(0.2, 1.8, num).to(device), torch.linspace(-1.5, 1.5, num).to(device)]
)
t, y = t.reshape(-1, 1), y.reshape(-1, 1)
fty = model.f(t=t, y=y).reshape(num, num)
dt = torch.zeros(num, num).fill_(dt).to(device)
dy = fty * dt
dt_, dy_, t_, y_ = dt.cpu().numpy(), dy.cpu().numpy(), t.cpu().numpy(), y.cpu().numpy()
plt.quiver(t_, y_, dt_, dy_, alpha=0.3, edgecolors='k', width=0.0035, scale=50)
if args.hide_ticks:
plt.xticks([], [])
plt.yticks([], [])
plt.scatter(ts_, ys_, marker='x', zorder=3, color='k', s=35) # Data.
plt.ylim(ylims)
plt.xlabel('$t$')
plt.ylabel('$Y_t$')
plt.tight_layout()
plt.savefig(img_path, dpi=args.dpi)
plt.close()
logging.info(f'Saved figure at: {img_path}')
if args.save_ckpt:
torch.save(
{'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'kl_scheduler': kl_scheduler},
os.path.join(ckpt_dir, f'global_step_{global_step}.ckpt')
)
# Train.
optimizer.zero_grad()
zs, kl = model(ts=ts_ext, batch_size=args.batch_size)
zs = zs.squeeze()
zs = zs[1:-1] # Drop first and last which are only used to penalize out-of-data region and spread uncertainty.
likelihood_constructor = {"laplace": distributions.Laplace, "normal": distributions.Normal}[args.likelihood]
likelihood = likelihood_constructor(loc=zs, scale=args.scale)
logpy = likelihood.log_prob(ys).sum(dim=0).mean(dim=0)
loss = -logpy + kl * kl_scheduler.val
loss.backward()
optimizer.step()
scheduler.step()
kl_scheduler.step()
logpy_metric.step(logpy)
kl_metric.step(kl)
loss_metric.step(loss)
logging.info(
f'global_step: {global_step}, '
f'logpy: {logpy_metric.val:.3f}, '
f'kl: {kl_metric.val:.3f}, '
f'loss: {loss_metric.val:.3f}'
)
if __name__ == '__main__':
# The argparse format supports both `--boolean-argument` and `--boolean-argument True`.
# Trick from https://stackoverflow.com/questions/15008758/parsing-boolean-values-with-argparse.
parser = argparse.ArgumentParser()
parser.add_argument('--no-gpu', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--debug', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--train-dir', type=str, required=True)
parser.add_argument('--save-ckpt', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--data', type=str, default='segmented_cosine', choices=['segmented_cosine', 'irregular_sine'])
parser.add_argument('--kl-anneal-iters', type=int, default=100, help='Number of iterations for linear KL schedule.')
parser.add_argument('--train-iters', type=int, default=5000, help='Number of iterations for training.')
parser.add_argument('--pause-iters', type=int, default=50, help='Number of iterations before pausing.')
parser.add_argument('--batch-size', type=int, default=512, help='Batch size for training.')
parser.add_argument('--likelihood', type=str, choices=['normal', 'laplace'], default='laplace')
parser.add_argument('--scale', type=float, default=0.05, help='Scale parameter of Normal and Laplace.')
parser.add_argument('--adjoint', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--adaptive', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--method', type=str, default='euler', choices=('euler', 'milstein', 'srk'),
help='Name of numerical solver.')
parser.add_argument('--dt', type=float, default=1e-2)
parser.add_argument('--rtol', type=float, default=1e-3)
parser.add_argument('--atol', type=float, default=1e-3)
parser.add_argument('--show-prior', type=str2bool, default=True, const=True, nargs="?")
parser.add_argument('--show-samples', type=str2bool, default=True, const=True, nargs="?")
parser.add_argument('--show-percentiles', type=str2bool, default=True, const=True, nargs="?")
parser.add_argument('--show-arrows', type=str2bool, default=True, const=True, nargs="?")
parser.add_argument('--show-mean', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--hide-ticks', type=str2bool, default=False, const=True, nargs="?")
parser.add_argument('--dpi', type=int, default=300)
parser.add_argument('--color', type=str, default='blue', choices=('blue', 'red'))
args = parser.parse_args()
device = torch.device('cuda' if torch.cuda.is_available() and not args.no_gpu else 'cpu')
manual_seed(args.seed)
if args.debug:
logging.getLogger().setLevel(logging.INFO)
ckpt_dir = os.path.join(args.train_dir, 'ckpts')
os.makedirs(ckpt_dir, exist_ok=True)
sdeint_fn = torchsde.sdeint_adjoint if args.adjoint else torchsde.sdeint
main()