-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcond_indep_test.py
290 lines (239 loc) · 13.8 KB
/
cond_indep_test.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
import numpy as np
from flows import Basic_Flow, cop_flow_constructor, marg_flow_constructor
from exp_runner import ExperimentBuilder
from utils import set_optimizer_scheduler
import torch
from utils import nll_error, create_folders, set_seeds
from options import TrainOptions
from data_provider import split_data_marginal, split_data_copula
import os
import json
import scipy.stats
from eval.plots import visualize_joint, visualize1d
from utils.load_and_save import save_statistics
eps = 1e-7
def marginal_estimator(loader_train: torch.utils.data.DataLoader, loader_val: torch.utils.data.DataLoader,
loader_test: torch.utils.data.DataLoader, exp_name: str, device: str,
epochs: int =50, variable_num: int =0, disable_tqdm: bool =False, **kwargs) -> tuple:
# Initialize marginal transform
marg_flow = marg_flow_constructor(**kwargs)
optimizer, scheduler = set_optimizer_scheduler(marg_flow,
lr=kwargs['lr'],
weight_decay=kwargs['weight_decay'],
amsgrad=kwargs['amsgrad'],
epochs=epochs)
experiment = ExperimentBuilder(network_model=marg_flow,
optimizer=optimizer,
scheduler=scheduler,
error=nll_error,
exp_name=exp_name,
flow_name= 'mf_' + str(variable_num),
epochs=epochs,
train_data=loader_train,
val_data=loader_val,
test_data=loader_test,
device=device,
disable_tqdm=disable_tqdm) # build an experiment object
# Train marginal flow
experiment_metrics, test_metrics = experiment.run_experiment() # run experiment and return experiment metrics
return experiment, experiment_metrics, test_metrics
def marginal_transform_1d(inputs: np.ndarray, exp_name: str, device: str, epochs: int =100, batch_size: int =128,
num_workers: int =0, variable_num: int =0, disable_tqdm=False, **kwargs) -> np.ndarray:
# Transform into data object
data_train_scaled, data_val_scaled, data_test_scaled, loader_train, loader_val, loader_test = split_data_marginal(inputs,
batch_size,
num_workers=num_workers,
return_datasets=True)
experiment, __, __ = marginal_estimator(loader_train=loader_train,
loader_val=loader_val,
loader_test=loader_test,
exp_name=exp_name,
device=device,
epochs=epochs,
batch_size=batch_size,
num_workers=num_workers,
variable_num=variable_num,
disable_tqdm=disable_tqdm,
**kwargs)
#
inputs_scaled = np.concatenate([data_train_scaled, data_val_scaled, data_test_scaled], axis=0)
# Transform
inputs_scaled = torch.from_numpy(inputs_scaled).float().to(device)
outputs = experiment.model.transform_to_noise(inputs_scaled)
# Plot results
visualize1d(experiment.model,
device=device,
path=experiment.figures_path,
true_samples=inputs_scaled.detach().cpu().numpy(),
obs=1000,
name='marg_flow')
norm_distr = scipy.stats.norm()
visualize_joint(norm_distr.cdf(torch.cat([outputs, outputs], axis=1).detach().cpu().numpy()), figures_path=experiment.figures_path, name='marg_flow_output')
return outputs
def marginal_transform(inputs: np.ndarray, exp_name: str, device: str, disable_tqdm=False, **kwargs) -> torch.Tensor:
if inputs.ndim > 1:
outputs = torch.empty_like(torch.from_numpy(inputs)).to(device).detach()
for dim in range(inputs.shape[1]):
outputs[:, dim: dim + 1] = marginal_transform_1d(inputs=inputs[:, dim: dim+1],
exp_name=exp_name,
device=device,
variable_num=dim,
disable_tqdm=disable_tqdm,
**kwargs).reshape(-1, 1).detach()
elif inputs.ndim == 1:
outputs = marginal_transform_1d(inputs=inputs.reshape(-1,1), exp_name=exp_name,
device=device, disable_tqdm=disable_tqdm, **kwargs).reshape(-1, 1).detach()
else:
raise ValueError('Invalid input shape.')
return outputs
def copula_estimator(loader_train: torch.utils.data.DataLoader, loader_val: torch.utils.data.DataLoader,
loader_test: torch.utils.data.DataLoader, cond_set_dim: int, exp_name: str, device: str, epochs: int =100,
disable_tqdm: bool =False, **kwargs) -> tuple:
# Initialize Copula Transform
cop_flow = cop_flow_constructor(context_dim=cond_set_dim, **kwargs)
optimizer, scheduler = set_optimizer_scheduler(cop_flow,
kwargs['lr'],
kwargs['weight_decay'],
kwargs['amsgrad'],
epochs)
experiment = ExperimentBuilder(network_model=cop_flow,
optimizer=optimizer,
scheduler=scheduler,
error=nll_error,
exp_name=exp_name,
flow_name= 'cf',
epochs=epochs,
train_data=loader_train,
val_data=loader_val,
test_data=loader_test,
device=device,
disable_tqdm=disable_tqdm) # build an experiment object
# Train marginal flow
experiment_metrics, test_metrics = experiment.run_experiment() # run experiment and return experiment metrics
return experiment, experiment_metrics, test_metrics
def mi_estimator(cop_flow: Basic_Flow, device: str, obs_n: int =1000, obs_m: int =1000, cond_set: int =None) -> float:
log_density = torch.empty((cond_set.shape[0] if cond_set is not None else obs_n, obs_m))
norm_distr = torch.distributions.normal.Normal(0, 1)
for mm in range(obs_m):
cop_samples = cop_flow.sample_copula(num_samples=cond_set.shape[0] if cond_set is not None else obs_n,
context=cond_set.to(device) if cond_set is not None else None)
log_density[:, mm] = cop_flow.log_pdf_uniform(cop_samples,
context=norm_distr.cdf(cond_set).to(device) if cond_set is not None else None)
mi = torch.mean(log_density)
return mi.cpu().numpy()
def independence_test(mi: float, threshold: float =0.05):
if mi < threshold:
return True
else:
return False
def copula_indep_test(x_input: np.ndarray, y_input: np.ndarray,
cond_set: np.ndarray, exp_name: str, device: str, kwargs_m, kwargs_c,
num_runs: int=30, batch_size_m: int =128, batch_size_c: int =128, num_workers: int =0,
visualize=False, transform_marginals=True, disable_tqdm=False) -> bool:
if transform_marginals:
print('Estimating x marginal...')
x_input = marginal_transform(x_input, exp_name, device=device, disable_tqdm=disable_tqdm, **kwargs_m)
print('Estimating y marginal...')
y_input = marginal_transform(y_input, exp_name, device=device, disable_tqdm=disable_tqdm, **kwargs_m)
if cond_set is not None:
print('Estimating cond set marginals...')
cond_set = marginal_transform(cond_set, exp_name, device=device, **kwargs_m).float()
cond_set_dim = cond_set.shape[1]
else:
cond_set_dim = None
else:
x_input = torch.from_numpy(x_input).float().to(device)
y_input = torch.from_numpy(y_input).float().to(device)
cond_set = torch.from_numpy(cond_set).float().to(device) if cond_set is not None else None
cond_set_dim = cond_set.shape[1] if cond_set is not None else None
if visualize:
# Generate the directory names
exp_path = os.path.join('results', exp_name, 'cf')
figures_path = os.path.join(exp_path, 'figures')
if not os.path.exists(figures_path):
os.makedirs(figures_path)
inputs = torch.cat([x_input[:1000, :], y_input[:1000, :]], axis=1).cpu().numpy()
visualize_joint(inputs, figures_path, name='input_dataset')
norm_distr = scipy.stats.norm()
visualize_joint(norm_distr.cdf(inputs), figures_path, name='input_uni_dataset')
if cond_set is not None:
visualize_joint(cond_set[:1000, :].detach().cpu().numpy(), figures_path, name='cond_input_dataset')
visualize_joint(norm_distr.cdf(cond_set[:1000, :].detach().cpu().numpy()), figures_path, name='cond_input_uni_dataset')
# Transform into data object
print('Creating copula dataset..')
__, __, __, loader_train, loader_val, loader_test = split_data_copula(x_input,
y_input,
cond_set,
batch_size=128,
num_workers=0,
return_datasets=True)
print('Estimating copula..')
experiment, __, test_metrics = copula_estimator(loader_train, loader_val, loader_test, cond_set_dim=cond_set_dim,
exp_name=exp_name, device=device, batch_size=batch_size_c,
num_workers=num_workers, disable_tqdm=disable_tqdm, **kwargs_c)
cop_flow = experiment.model
if visualize:
# Plot copula samples
samples = cop_flow.sample_copula(np.min([1000, cond_set.shape[0]]) if cond_set is not None else 1000,
context=cond_set[:1000, :] if cond_set is not None else None).detach().cpu().numpy()
visualize_joint(samples, figures_path, name='cop_samples')
print('Estimating mutual information..')
with torch.no_grad():
cop_flow.eval()
mi = mi_estimator(cop_flow, device=device, cond_set=cond_set)
test_metrics['mi'] = [mi]
experiment_logs = os.path.join('results', experiment.exp_name, 'cf', 'stats')
save_statistics(experiment_logs, 'test_summary.csv', test_metrics, current_epoch=0, continue_from_mode=False)
return mi
if __name__ == '__main__':
# Training settings
args = TrainOptions().parse() # get training options
args.flow_name = ''
# Create Folders
create_folders(args)
with open(os.path.join(args.experiment_logs, 'args'), 'w') as f:
json.dump(args.__dict__, f, indent=2)
# Cuda settings
use_cuda = torch.cuda.is_available()
args.device = torch.cuda.current_device()
# Set Seed
set_seeds(seed=args.seed, use_cuda=use_cuda)
# Get inputs
obs = 50
x = np.random.uniform(size=(obs, 1))
y = np.random.uniform(size=(obs, 1))
z = np.random.uniform(size=(obs, 5))
# kwargs marginal
kwargs_m = {'n_layers': args.n_layers_m,
'lr': args.lr_m,
'weight_decay': args.weight_decay_m,
'amsgrad': args.amsgrad_m,
'n_bins': args.n_bins_m,
'tail_bound': args.tail_bound_m,
'hidden_units': args.hidden_units_m,
'tails': args.tails_m,
'identity_init': args.identity_init_m,
'epochs': args.epochs_m,
'batch_size': args.batch_size_m}
# kwargs copula
kwargs_c = {'n_layers': args.n_layers_c,
'lr': args.lr_c,
'weight_decay': args.weight_decay_c,
'amsgrad': args.amsgrad_c,
'n_bins': args.n_bins_c,
'tail_bound': args.tail_bound_c,
'hidden_units': args.hidden_units_m,
'tails': args.tails_m,
'identity_init': args.identity_init_m,
'epochs': args.epochs_c,
'batch_size': args.batch_size_c}
#
print(copula_indep_test(x, y, z, exp_name=args.exp_name,
device=args.device,
kwargs_m=kwargs_m,
kwargs_c=kwargs_c,
epochs_m=args.epochs_m,
epochs_c=args.epochs_c,
batch_size_m=args.batch_size_m,
batch_size_c=args.batch_size_c,
num_workers=args.num_workers))