-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathvalidation.py
executable file
·261 lines (226 loc) · 14.3 KB
/
validation.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
import os
from torch.utils.data import DataLoader
import torch
import pandas as pd
import numpy as np
from HL_VAE import read_functions
from elbo_functions import deviance_upper_bound, elbo
from utils import batch_predict_varying_T, HensmanDataLoader
import HL_VAE.read_functions as rd
num_workers = 4
def validation_dubo(latent_dim, covar_module0, covar_module1, likelihood, train_xt, m, log_v, z, P, T, eps):
"""
Efficient KL divergence using the variational mean and variance instead of a sample from the latent space (DUBO).
See L-VAE supplementary material.
:param covar_module0: additive kernel (sum of cross-covariances) without id covariate
:param covar_module1: additive kernel (sum of cross-covariances) with id covariate
:param likelihood: GPyTorch likelihood model
:param train_xt: auxiliary covariate information
:param m: variational mean
:param log_v: (log) variational variance
:param z: inducing points
:param P: number of unique instances
:param T: number of longitudinal samples per individual
:param eps: jitter
:return: KL divergence between variational distribution and additive GP prior (DUBO)
"""
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
v = torch.exp(log_v)
torch_dtype = torch.double
x_st = torch.reshape(train_xt, [P, T, train_xt.shape[1]]).to(device)
stacked_x_st = torch.stack([x_st for i in range(latent_dim)], dim=1)
K0xz = covar_module0(train_xt, z).evaluate().to(device)
K0zz = (covar_module0(z, z).evaluate() + eps * torch.eye(z.shape[1], dtype=torch_dtype).to(device)).to(device)
LK0zz = torch.cholesky(K0zz).to(device)
iK0zz = torch.cholesky_solve(torch.eye(z.shape[1], dtype=torch_dtype).to(device), LK0zz).to(device)
K0_st = covar_module0(stacked_x_st, stacked_x_st).evaluate().transpose(0,1)
B_st = (covar_module1(stacked_x_st, stacked_x_st).evaluate() + torch.eye(T, dtype=torch.double).to(device) * likelihood.noise_covar.noise.unsqueeze(dim=2)).transpose(0,1)
LB_st = torch.cholesky(B_st).to(device)
iB_st = torch.cholesky_solve(torch.eye(T, dtype=torch_dtype).to(device), LB_st)
dubo_sum = torch.tensor([0.0]).double().to(device)
for i in range(latent_dim):
m_st = torch.reshape(m[:, i], [P, T, 1]).to(device)
v_st = torch.reshape(v[:, i], [P, T]).to(device)
K0xz_st = torch.reshape(K0xz[i], [P, T, K0xz.shape[2]]).to(device)
iB_K0xz = torch.matmul(iB_st[i], K0xz_st).to(device)
K0zx_iB_K0xz = torch.matmul(torch.transpose(K0xz[i], 0, 1), torch.reshape(iB_K0xz, [P*T, K0xz.shape[2]])).to(device)
W = K0zz[i] + K0zx_iB_K0xz
W = (W + W.T) / 2
LW = torch.cholesky(W).to(device)
logDetK0zz = 2 * torch.sum(torch.log(torch.diagonal(LK0zz[i]))).to(device)
logDetB = 2 * torch.sum(torch.log(torch.diagonal(LB_st[i], dim1=-2, dim2=-1))).to(device)
logDetW = 2 * torch.sum(torch.log(torch.diagonal(LW))).to(device)
logDetSigma = -logDetK0zz + logDetB + logDetW
iB_m_st = torch.solve(m_st, B_st[i])[0].to(device)
qF1 = torch.sum(m_st*iB_m_st).to(device)
p = torch.matmul(K0xz[i].T, torch.reshape(iB_m_st, [P * T])).to(device)
qF2 = torch.sum(torch.triangular_solve(p[:,None], LW, upper=False)[0] ** 2).to(device)
qF = qF1 - qF2
tr = torch.sum(iB_st[i] * K0_st[i]) - torch.sum(K0zx_iB_K0xz * iK0zz[i])
logDetD = torch.sum(torch.log(v[:, i])).to(device)
tr_iB_D = torch.sum(torch.diagonal(iB_st[i], dim1=-2, dim2=-1)*v_st).to(device)
D05_iB_K0xz = torch.reshape(iB_K0xz*torch.sqrt(v_st)[:,:,None], [P*T, K0xz.shape[2]])
K0zx_iB_D_iB_K0zx = torch.matmul(torch.transpose(D05_iB_K0xz,0,1), D05_iB_K0xz).to(device)
tr_iB_K0xz_iW_K0zx_iB_D = torch.sum(torch.diagonal(torch.cholesky_solve(K0zx_iB_D_iB_K0zx, LW))).to(device)
tr_iSigma_D = tr_iB_D - tr_iB_K0xz_iW_K0zx_iB_D
dubo = 0.5*(tr_iSigma_D + qF - P*T + logDetSigma - logDetD + tr)
dubo_sum = dubo_sum + dubo
return dubo_sum
def validate(nnet_model, dataset, type_KL, num_samples, latent_dim, covar_module0, covar_module1, likelihoods,
zt_list, T, train_mu, train_x, id_covariate, results_path, eps=1e-6):
print("Testing the model with a validation set")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
validation_results = []
batch_size =len(dataset)
assert (type_KL == 'GPapprox_closed' or type_KL == 'GPapprox')
# set up Data Loader for training
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
Q = len(dataset[0]['label'])
P = pd.unique(dataset.label_source.iloc[:, id_covariate]).size
full_mu = torch.zeros(len(dataset), latent_dim, dtype=torch.double, requires_grad=True).to(device)
full_log_var = torch.zeros(len(dataset), latent_dim, dtype=torch.double, requires_grad=True).to(device)
full_labels = torch.zeros(len(dataset), Q, dtype=torch.double, requires_grad=False).to(device)
recon_loss_sum = 0
miss_recon_loss_sum = 0
recon_loss_sum_mse = 0
nll_loss_sum = 0
for batch_idx, sample_batched in enumerate(dataloader):
indices = sample_batched['idx']
data = sample_batched['digit'].double().to(device)
mask = sample_batched['mask'].double().to(device)
full_labels[indices] = sample_batched['label'].double().to(device)
param_mask = sample_batched['param_mask'].view(batch_size, -1)
param_mask = param_mask.to(device)
data = torch.squeeze(data)
mask = torch.squeeze(mask)
p_samples, mu, log_var, log_p_x, log_p_x_missing, p_params, q_samples, q_params = nnet_model(data, mask, param_mask, nnet_model.types_info)
nll = nnet_model.loss_function(log_p_x)
p_params_complete = read_functions.p_params_concatenation_by_key([p_params], nnet_model.types_info,
data.shape[0], data.device, 'x')
data_transformed = rd.discrete_variables_transformation(data, nnet_model.types_info)
recon_x_transformed, _ = read_functions.statistics(p_params_complete,
nnet_model.types_info, data.device, nnet_model.conv,
[nnet_model._log_vy_real,
nnet_model._log_vy_pos])
recon_loss_het, miss_recon_loss_het, _ = read_functions.error_computation(data_transformed,
recon_x_transformed,
nnet_model.types_info,
mask, dim=0)
recon_loss = recon_loss_het
full_mu[indices] = mu
full_log_var[indices] = log_var
recon_loss = torch.sum(recon_loss)
recon_loss_sum = recon_loss_sum + recon_loss.item()
nll = torch.sum(nll)
nll_loss_sum = nll_loss_sum + nll.item()
gp_loss_sum = 0
if isinstance(covar_module0, list):
if type_KL == 'GPapprox':
for sample in range(0, num_samples):
Z = nnet_model.sample_latent(full_mu, full_log_var)
for i in range(0, latent_dim):
Z_dim = Z[:, i]
gp_loss = -elbo(covar_module0[i], covar_module1[i], likelihoods[i], full_labels, Z_dim,
zt_list[i].to(device), P, T, eps)
gp_loss_sum = gp_loss.item() + gp_loss_sum
gp_loss_sum /= num_samples
elif type_KL == 'GPapprox_closed':
for i in range(0, latent_dim):
mu_sliced = full_mu[:, i]
log_var_sliced = full_log_var[:, i]
gp_loss = deviance_upper_bound(covar_module0[i], covar_module1[i],
likelihoods[i], full_labels,
mu_sliced, log_var_sliced,
zt_list[i].to(device), P,
T, eps)
gp_loss_sum = gp_loss.item() + gp_loss_sum
else:
if type_KL == 'GPapprox_closed':
df_lbl = pd.DataFrame(np.array(full_labels.cpu()))
for sz in df_lbl.groupby(0).size().unique():
ids = (df_lbl.groupby(0).size() == sz)[(df_lbl.groupby(0).size() == sz)].index
cur_df = df_lbl.loc[df_lbl[0].isin(ids)]
par_labels = full_labels[cur_df.index]
par_mu = full_mu[cur_df.index]
par_log_var = full_log_var[cur_df.index]
par_P = len(ids)
gp_loss = validation_dubo(latent_dim, covar_module0, covar_module1,
likelihoods, par_labels,
par_mu, par_log_var,
zt_list, par_P, sz, eps)
gp_loss_sum += gp_loss.item()
net_loss_sum = gp_loss_sum + nll_loss_sum
#Do logging
print('Validation set - Loss: %.3f - GP loss: %.3f - NLL loss: %.3f - Recon Loss: %.3f' % (
net_loss_sum, gp_loss_sum, nll_loss_sum, recon_loss_sum))
if nnet_model.conv:
halfway = P//2
subjects = halfway
l1 = [i*T + k for i in range(0,subjects) for k in range(0,5)]
l2 = [i*T + k for i in range(halfway,halfway+subjects) for k in range(0,5)]
prediction_mu = torch.cat((train_mu,
full_mu[l1],
full_mu[l2]))
prediction_x = torch.cat((train_x,
full_labels[l1],
full_labels[l2]))
test_x = torch.cat((full_labels[0:subjects*T], full_labels[halfway*T:(halfway+subjects)*T]))
prediction_dataloader_p1 = DataLoader(dataset[0:subjects*T], batch_size=subjects*T, shuffle=False)
for batch_idx, sample_batched in enumerate(prediction_dataloader_p1):
data_p1 = sample_batched['digit'].double().to(device)
mask_p1 = sample_batched['mask'].double().to(device)
param_mask1 = sample_batched['param_mask'].view(sample_batched['param_mask'].shape[0], -1)
prediction_dataloader_p2 = DataLoader(dataset[halfway*T:(halfway+subjects)*T], batch_size=subjects*T, shuffle=False)
for batch_idx, sample_batched in enumerate(prediction_dataloader_p2):
data_p2 = sample_batched['digit'].double().to(device)
mask_p2 = sample_batched['mask'].double().to(device)
param_mask2 = sample_batched['param_mask'].view(sample_batched['param_mask'].shape[0], -1)
prediction_data = torch.cat((data_p1, data_p2))
prediction_mask = torch.cat((mask_p1, mask_p2))
prediction_param_mask = torch.cat((param_mask1, param_mask2))
else:
indexes_for_prediction = []
for no in df_lbl[0].unique():
indexes_for_prediction = indexes_for_prediction + list(df_lbl[df_lbl[0] == no].index[:2].astype(int))
prediction_mu = torch.cat((train_mu,
full_mu[indexes_for_prediction]))
prediction_x = torch.cat((train_x,
full_labels[indexes_for_prediction]))
test_x = full_labels
prediction_data = sample_batched['digit'].double().to(device)
prediction_mask = sample_batched['mask'].double().to(device)
prediction_true_mask = sample_batched['true_mask'].double().to(device)
prediction_param_mask = sample_batched['param_mask'].view(sample_batched['param_mask'].shape[0], -1)
Z_pred = batch_predict_varying_T(latent_dim, covar_module0, covar_module1, likelihoods, prediction_x, test_x, prediction_mu, zt_list, id_covariate, eps=1e-6)
prediction_param_mask = prediction_param_mask.to(device).to(torch.float64)
prediction_data = torch.squeeze(prediction_data).to(torch.float64)
prediction_mask = torch.squeeze(prediction_mask).to(torch.float64)
log_p_x, _, recon_batch, params_x = nnet_model.decode(Z_pred, prediction_data, prediction_mask, prediction_param_mask)
p_params_complete = read_functions.p_params_concatenation_by_key([params_x], nnet_model.types_info,
prediction_data.shape[0], data.device, 'x')
data_transformed = rd.discrete_variables_transformation(prediction_data, nnet_model.types_info)
recon_x_transformed, _ = read_functions.statistics(p_params_complete,
nnet_model.types_info, prediction_data.device, nnet_model.conv,
[nnet_model._log_vy_real,
nnet_model._log_vy_pos])
recon_loss_GP, miss_recon_loss_GP, _ = read_functions.error_computation(data_transformed,
recon_x_transformed,
nnet_model.types_info,
prediction_mask)
print('Validation set - GP prediction Error: %.4f, Mean of Error %.4f' % (torch.sum(recon_loss_GP).item(),
(torch.sum(recon_loss_GP).item())/recon_loss_GP.shape[0]))
validation_results.append(recon_loss_sum/len(dataset))
validation_results.append(torch.sum(recon_loss_GP).item()/recon_loss_GP.shape[0])
validation_results.append(recon_loss_sum_mse/len(dataset))
validation_results.append(miss_recon_loss_sum/len(dataset))
validation_results.append(torch.sum(miss_recon_loss_GP).item()/miss_recon_loss_GP.shape[0])
validation_results.append(net_loss_sum)
validation_results.append(gp_loss_sum)
validation_results.append(nll_loss_sum)
validation_results.append(recon_loss_sum)
validation_results.append(torch.sum(recon_loss_GP).item())
df_res = pd.DataFrame(validation_results, index=['vae_error','GP_error', 'vae_mse',#'GP_mse',
'miss_vae_error','miss_GP_error', 'net_loss', 'GP_loss', 'nll_loss',
'recon_loss_sum', 'GP_recon_loss_sum'])
df_res.to_csv(os.path.join(results_path, 'validation_results.csv'),header=False)
return df_res