-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_inference_pyro.py
556 lines (457 loc) · 22.9 KB
/
run_inference_pyro.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
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
import torch
import pyro
import pyro.distributions as dist
import matplotlib
import matplotlib.pyplot as plt
from torch.distributions.constraints import positive
import logging
import os
import psutil
import numpy as np
import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import pickle
import scipy
import pyro
from pyro.infer import MCMC, NUTS
import ssms
import lanfactory
torch.set_default_dtype(torch.float32)
from lanfactory.trainers.torch_mlp import TorchMLP
import lanfactory
import ssms
import arviz as az
import cloudpickle as cpickle
#from bokeh.resources import INLINE
#import bokeh.io
#from bokeh import *
#bokeh.io.output_notebook(INLINE)
import argparse
import pathlib
from time import time
from copy import deepcopy
# Necessary?
pyro.enable_validation(True)
pyro.set_rng_seed(20)
logging.basicConfig(format='%(message)s', level=logging.INFO)
network_files = {'network_config': {'ddm': 'd27193a4153011ecb76ca0423f39a3e6_ddm_torch__network_config.pickle',
'angle': 'eba53550128911ec9fef3cecef056d26_angle_torch__network_config.pickle',
'weibull': '44deb16a127f11eca325a0423f39b436_weibull_torch__network_config.pickle',
'levy': '80dec298152e11ec88b8ac1f6bfea5a4_levy_torch__network_config.pickle'},
'network': {'ddm': 'd27193a4153011ecb76ca0423f39a3e6_ddm_torch_state_dict.pt',
'angle': 'eba53550128911ec9fef3cecef056d26_angle_torch_state_dict.pt',
'weibull': '44deb16a127f11eca325a0423f39b436_weibull_torch_state_dict.pt',
'levy': '80dec298152e11ec88b8ac1f6bfea5a4_levy_torch_state_dict.pt'}
}
initial_params_dict = {'ddm': {'v': torch.tensor(0.5),
'a': torch.tensor(1.5),
'z': torch.tensor(0.5),
't': torch.tensor(1.0)
},
'angle': {'v': torch.tensor(0.5),
'a': torch.tensor(1.5),
'z': torch.tensor(0.5),
't': torch.tensor(1.0),
'theta': torch.tensor(0.5)
},
'weibull': {'v': torch.tensor(0.5),
'a': torch.tensor(1.5),
'z': torch.tensor(0.5),
't': torch.tensor(1.0),
'theta': torch.tensor(0.5)
},
'levy': {'v': torch.tensor(0.5),
'a': torch.tensor(1.5),
'z': torch.tensor(0.5),
't': torch.tensor(1.0),
'alpha': torch.tensor(1.5),
},
}
def load_network(model = 'ddm', machine = 'gpu'):
network_config = pickle.load(open('nets/' + network_files['network_config'][model],
'rb'))
network = lanfactory.trainers.torch_mlp.TorchMLP(network_config = network_config,
input_shape = model_config['n_params'] + 2)
if machine == 'cpu':
network.load_state_dict(torch.load('nets/' + network_files['network'][model],
map_location = torch.device('cpu')))
else:
network.load_state_dict(torch.load('nets/' + network_files['network'][model]))
network.eval()
return network
import math
from numbers import Real
from numbers import Number
import torch
from torch.distributions import constraints
from torch.distributions.exp_family import ExponentialFamily
from torch.distributions.utils import _standard_normal, broadcast_all
from utils import save_traces
class SSMDist(dist.TorchDistribution):
def __init__(self, params, num_trials, network, ssm_model):
self.net = network
self.num_trials = num_trials
self.ssm_model = ssm_model
self.params = params
if self.params.dim() == 1:
batch_shape = torch.Size((1,))
elif self.params.dim() > 1:
batch_shape = self.params.size()[:-1]
super().__init__(batch_shape = batch_shape,
event_shape = torch.Size((2, )),
validate_args=False)
def sample(self):
return sim_wrap(theta = params,
model = self.ssm_model)
def log_prob(self, value):
if self.params.dim() == 1:
tmp_params = self.params.tile(dims = (self.num_trials, 1))
tmp_dat = value
if self.params.dim() == 3:
tmp_params = self.params.repeat_interleave(self.num_trials, -2)
tmp_dat = value.repeat((self.params.size()[0], 1, 1))
net_in = torch.cat([tmp_params, tmp_dat],
dim = -1)
# Old
logp = torch.clip(self.net(net_in), min = -16.11).squeeze(dim = -1)
return logp
class SSMDistH(dist.TorchDistribution):
# arg_constraints = {'loc': constraints.interval(-1, 1),
# 'scale': constraints.interval(0.0001, 10)
# }
def __init__(self, params, num_trials, network, ssm_name):
self.net = network
self.n_samples = num_trials
self.boundaries = ssms.config.model_config[ssm_name]['param_bounds']
self.out_of_bounds_val = -66.1
self.params = params
self.ssm_name = ssm_name
self.params_shape = self.params.size()
batch_shape = self.params.size()[:-1]
super().__init__(batch_shape = batch_shape, #torch.Size((my_size,)),
event_shape = torch.Size((2,)),
validate_args = False) #torch.Size((2,))) # event_shape = (1,))
def sample(self):
return sim_wrap(theta = self.params, model = self.ssm_name, n_samples = self.n_samples)
def log_prob(self, value):
if self.params.dim() == 3:
tmp_dat = value.repeat((self.params_shape[0], 1, 1, 1))
tmp_params = self.params.unsqueeze(1).tile((1, self.n_samples, 1, 1))
elif self.params.dim() == 4:
tmp_dat = value.repeat(self.params_shape[0], self.params_shape[1], 1, 1, 1)
tmp_params = self.params.unsqueeze(2).tile((1, 1, self.n_samples, 1, 1))
else:
tmp_dat = value
tmp_params = self.params.tile((self.n_samples, 1, 1))
net_in = torch.cat([tmp_params, tmp_dat], dim = -1)
logp = torch.clip(self.net(net_in), min = -16.11)
logp_squeezed = torch.squeeze(logp, dim = -1)
for i in range(ssms.config.model_config[self.ssm_name]['n_params']):
logp_squeezed = torch.where(net_in[..., i] < torch.tensor(self.boundaries[1][i]),
logp_squeezed,
torch.tensor(self.out_of_bounds_val))
logp_squeezed = torch.where(net_in[..., i] > torch.tensor(self.boundaries[0][i]),
logp_squeezed,
torch.tensor(self.out_of_bounds_val))
logp_squeezed = torch.squeeze(logp_squeezed) #, dim = 0) #.unsqueeze(1)
return logp_squeezed
def model_maker(model = 'ddm'):
model_config = ssms.config.model_config[model]
def ssm_model(num_trials, data, network, model):
param_list = []
for param in model_config['params']:
idx = model_config['params'].index(param)
param_list.append(pyro.sample(param, dist.Uniform(model_config['param_bounds'][0][idx],
model_config['param_bounds'][1][idx])))
with pyro.plate("data", num_trials) as data_plate:
return pyro.sample("obs",
SSMDist(torch.stack(param_list, dim = -1),
num_trials,
network,
model),
obs = data)
return ssm_model
def model_maker_hierarchical(model = 'ddm'):
model_config = deepcopy(ssms.config.model_config[model])
model_config['mu_mu_std'] = {'ddm': {'v': 0.5,
'a': 0.5,
'z': 0.5,
't': 0.5
},
'angle':{'v': 0.5,
'a': 0.5,
'z': 0.5,
't': 0.5,
'theta': 0.5
},
'weibull':{'v': 0.5,
'a': 0.5,
'z': 0.5,
't': 0.5,
'alpha': 0.5,
'beta': 0.5
},
'levy': {'v': 0.5,
'a': 0.5,
'z': 0.5,
't': 0.5,
'alpha': 0.5
}
}
def ssm_model_hierarchical(num_subjects, num_trials, data, network, ssm_name):
mu_mu_list = []
mu_std_list = []
for param in model_config['params']:
idx = model_config['params'].index(param)
param_mean = (model_config['param_bounds'][1][idx] + model_config['param_bounds'][0][idx]) / 2
mu_mu_list.append(pyro.sample(param + "_mu_mu",
dist.Normal(param_mean,
model_config['mu_mu_std'][model][param])
)
)
mu_std_list.append(pyro.sample(param + "_mu_std",
dist.HalfNormal(100.)
)
)
with pyro.plate("subjects", num_subjects) as subjects_plate:
subj_list = []
for param in model_config['params']:
idx = model_config['params'].index(param)
subj_list.append(pyro.sample(param + "_subj", dist.Normal(mu_mu_list[idx], mu_std_list[idx])
)
)
with pyro.plate("data", num_trials) as data_plate:
return pyro.sample("obs",
SSMDistH(torch.stack(subj_list, axis = -1), num_trials, network, ssm_name),
obs = data)
return ssm_model_hierarchical
from torch.distributions import transform_to
from typing import Callable, Optional
def init_to_uniform(
site: Optional[dict] = None,
radius: float = 2.0,
):
"""
Initialize to a random point in the area ``(-radius, radius)`` of
unconstrained domain.
:param float radius: specifies the range to draw an initial point in the
unconstrained domain.
"""
if site is None:
return functools.partial(init_to_uniform, radius=radius)
value = site["fn"].sample().detach()
t = transform_to(site["fn"].support)
value = t(torch.rand_like(t.inv(value)) * (2 * radius) - radius)
value._pyro_custom_init = False
return value
def init_to_uniform_custom(
site: Optional[dict] = None,
radius: float = 0.1,
):
"""
Initialize to a random point in the area ``(-radius, radius)`` of
unconstrained domain.
:param float radius: specifies the range to draw an initial point in the
unconstrained domain.
"""
if site is None:
return functools.partial(init_to_uniform, radius=radius)
site_num = site['fn'].shape()
# group level locs -------
if site['name'] == 'v_mu_mu':
return torch.rand(1) - 0.5
if site['name'] == 'a_mu_mu':
return (torch.rand(1) - 0.5) + 1.5
if site['name'] == 'z_mu_mu':
return (torch.rand(1) / 5) + 0.3
if site['name'] == 't_mu_mu':
return torch.rand(1) / 5
if site['name'] == 'theta_mu_mu':
return (torch.rand(1) / 10)
if site['name'] == 'alpha_mu_mu':
return (torch.rand(1) + 1.5)
if site['name'] == 'beta_mu_mu':
return (torch.rand(1) + 1.5)
# -------
# subject level locs -------
if site['name'] == 'v_subj':
return torch.rand(site_num) - 0.5
if site['name'] == 'a_subj':
return (torch.rand(site_num) - 0.5) + 1.5
if site['name'] == 'z_subj':
return (torch.rand(site_num) / 5) + 0.3
if site['name'] == 't_subj':
return torch.rand(site_num) / 5
if site['name'] == 'theta_subj':
return (torch.rand(site_num) / 10)
if site['name'] == 'alpha_subj':
return (torch.rand(site_num) + 1.5)
if site['name'] == 'beta_subj':
return (torch.rand(site_num) + 1.5)
# -------
# std parameters keep the uniform random init
value = site["fn"].sample().detach()
t = transform_to(site["fn"].support)
value = t(torch.rand_like(t.inv(value)) * (2 * radius) - radius)
value._pyro_custom_init = False
return value
if __name__ == "__main__":
# Command line interface arguments
CLI = argparse.ArgumentParser()
CLI.add_argument("--model",
type = str,
default = "ddm")
CLI.add_argument("--modeltype",
type = str,
default = "singlesubject")
CLI.add_argument("--nsteps",
type = int,
default = 5000)
CLI.add_argument("--nparticles",
type = int,
default = 10)
CLI.add_argument("--guide",
type = str,
default = 'normal')
CLI.add_argument("--machine",
type = str,
default = 'gpu')
CLI.add_argument("--idmin",
type = int,
default = 0)
CLI.add_argument("--idmax",
type = int,
default = 100)
CLI.add_argument("--progressbar",
type = int,
default = 0)
args = CLI.parse_args()
print(args)
# Model
if args.modeltype == 'singlesubject':
model = args.model # for now only DDM (once we have choice probability models --> all models applicable)
model_config = ssms.config.model_config[model].copy() # convenience
central_data_single_subject = pickle.load(open('data/single_subject/' + model + \
'_nsamples_1000_nparams_200_stdfracdenom_6.pickle', 'rb'))
arviz_path = 'data/single_subject/' + \
model + '_nsamples_1000_nparams_200_stdfracdenom_6'
pathlib.Path(arviz_path).mkdir(parents = True,
exist_ok = True)
for data_idx in range(args.idmin, args.idmax, 1):
data = central_data_single_subject['data'][data_idx]['pyro']
gt_params = central_data_single_subject['data'][data_idx]['gt_params']
n_samples = central_data_single_subject['data'][data_idx]['pyro'].shape[0]
network = load_network(model = model,
machine = args.machine)
ssm_model = model_maker(model = model)
pyro.clear_param_store()
# These should be reset each training loop.
if args.guide == 'normal':
auto_guide = pyro.infer.autoguide.AutoNormal(ssm_model)
elif args.guide == 'mvnormal':
auto_guide = pyro.infer.autoguide.AutoMultivariateNormal(ssm_model)
adam = pyro.optim.Adam({"lr": 0.02})
elbo = pyro.infer.Trace_ELBO(num_particles = args.nparticles,
vectorize_particles = True)
svi = pyro.infer.SVI(ssm_model, auto_guide, adam, elbo)
losses = []
start_t = time()
for step in range(args.nsteps): # Consider running for more steps.
loss = svi.step(*(n_samples, data, network, model))
losses.append(loss)
if step % 100 == 0:
logging.info("Step {} out of {}, Elbo loss: {}".format(step, args.nsteps, loss))
end_t = time()
print('Inference took: ', end_t - start_t, ' seconds...')
predictive = pyro.infer.Predictive(ssm_model, guide = auto_guide, num_samples = 2000)
svi_samples = predictive(*(n_samples, data, network, model)) #torch.tensor(data))
svi_samples_new = {key_: svi_samples[key_].numpy().swapaxes(0, 1) for key_ in svi_samples.keys() if key_ != 'obs'}
az_svi = az.convert_to_inference_data(svi_samples_new)
az_svi.posterior.attrs['runtime'] = end_t - start_t
az_svi.posterior.attrs['svi_steps'] = args.nsteps
az_svi.posterior.attrs['svi_loss_trajectory'] = losses
az_svi.posterior.attrs['machine_info'] = psutil.subprocess.run(['lscpu'],
capture_output=True,
text=True).stdout.split('\n')
# Make folder for arviz data if it doesn't already exist
print('saving files...')
save_traces(file_path = arviz_path,
arviz_trace = az_svi,
dict_trace = svi_samples_new,
model = args.model,
idx = data_idx,
backend = 'pyro',
infer_type = 'variational',
machine = args.machine,
nparticles = args.nparticles)
elif args.modeltype == 'hierarchical':
model = args.model
model_config = ssms.config.model_config[model]
central_data_hierarchical = pickle.load(open('data/hierarchical/' + model + \
'_nsamples_1000_nsubjects_20_nparams_200_stdfracdenom_6.pickle',
'rb'))
for data_idx in range(args.idmin, args.idmax, 1):
data = central_data_hierarchical['data'][data_idx]['pyro']
gt_params = central_data_hierarchical['data'][data_idx]['gt_params']
n_samples = central_data_hierarchical['data'][data_idx]['pyro'].shape[0]
n_subjects = central_data_hierarchical['data'][data_idx]['pyro'].shape[1]
network = load_network(model = model,
machine = args.machine)
ssm_model_hierarchical = model_maker_hierarchical(model = model)
pyro.clear_param_store()
# These should be reset each training loop.
if args.guide == 'normal':
auto_guide_hierarchical = pyro.infer.autoguide.AutoNormal(ssm_model_hierarchical,
init_scale = 0.02,
init_loc_fn = init_to_uniform_custom)
elif args.guide == 'mvnormal':
auto_guide_hierarchical = pyro.infer.autoguide.AutoMultivariateNormal(ssm_model_hierarchical,
init_scale = 0.02,
init_loc_fn = init_to_uniform_custom)
adam = pyro.optim.Adam({"lr": 0.02})
elbo = pyro.infer.Trace_ELBO(num_particles = args.nparticles, vectorize_particles = True)
svi = pyro.infer.SVI(ssm_model_hierarchical, auto_guide_hierarchical, adam, elbo)
losses = []
start_t = time()
# Main loop
for step in range(args.nsteps):
loss = svi.step(*(n_subjects, n_samples, data, network, model))
losses.append(loss)
if step % 100 == 0:
logging.info("Step {} out of {}, Elbo loss: {}".format(step, args.nsteps, loss))
end_t = time()
print('Inference took: ', end_t - start_t, ' seconds...')
# plt.figure(figsize=(5, 2))
# plt.plot(losses)
# plt.xlabel("SVI step");
# plt.ylabel("ELBO loss");
# Predictive
predictive = pyro.infer.Predictive(ssm_model_hierarchical, guide = auto_guide_hierarchical, num_samples = 2000)
svi_samples = predictive(n_subjects, n_samples, data, network, model) #torch.tensor(data))
# Get ready for arviz
svi_samples_new = {key_: svi_samples[key_].numpy().swapaxes(0, 1) for key_ in svi_samples.keys() if key_ != 'obs'}
# Make into inference data
az_svi = az.from_dict(posterior = svi_samples_new,
posterior_predictive = {'obs': np.expand_dims(svi_samples['obs'].numpy(), 0)})
az_svi.posterior.attrs['runtime'] = end_t - start_t
az_svi.posterior.attrs['svi_steps'] = args.nsteps
az_svi.posterior.attrs['svi_loss_trajectory'] = losses
az_svi.posterior.attrs['machine_info'] = psutil.subprocess.run(['lscpu'],
capture_output=True,
text=True).stdout.split('\n')
# Make folder for arviz data if it doesn't already exist
arviz_path = 'data/hierarchical/' + \
model + '_nsamples_1000_nsubjects_20_nparams_200_stdfracdenom_6'
pathlib.Path(arviz_path).mkdir(parents = True,
exist_ok = True)
print('saving files...')
save_traces(file_path = arviz_path,
arviz_trace = az_svi,
dict_trace = svi_samples_new,
model = args.model,
idx = data_idx,
backend = 'pyro',
infer_type = 'variational',
machine = args.machine,
nparticles = args.nparticles)