-
Notifications
You must be signed in to change notification settings - Fork 2
/
test_metrics_MNIST_natDigit.py
316 lines (261 loc) · 16.6 KB
/
test_metrics_MNIST_natDigit.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
# -*- coding: utf-8 -*-
from __future__ import division
import time
import numpy as np
from six.moves import xrange
import scipy.io as sio
import torch
from utils import *
from trans_opt_objectives import *
def log_likelihood(encoder,decoder,transNet,sampler_c,Psi,x,sample_labels,anchors,to_noise_std,num_anchor,M,numRestart,scale,opt,save_folder,k =10):
'''
Compute log_likelihood on rotated MNIST digits
Inputs:
- encoder: Encoder network
- decoder: Decoder network
- transNet: Transport operator layer
- sampler_c: Layer that samples the transport operator coefficients
- Psi: Current transport operator dictionary elements
- x: Batch of data [batch_size,H,W,C]
- labels: Batch of labels [batch_size,y_dim]
- to_noise_std: Sampling noise standard deviation for Gaussian prior distribution
- num_anchor: Number of anchors per class
- M: Number of transport operator dictionary elements
- numRestart: Number of restarts for coefficient inference
- scale: Value to scale the latent vectors by to get them in a range that is suitable for coefficient inference
- save_folder: Directory for saving data
- k: Number of samples of each latent vector for computing the LL
Outputs:
- LL_total: Average log-likelihood with all constants added
- LL_inner: Array of log-likelihood with all constants added for each latent vector
- LL_total_no_add: Average log-likelihood with no constants added
- LL_inner_no_add: Array of log-likelihood with no constants added
'''
# Compute the test log likelihood
batch_size_use = x.size(0)
input_h = opt.img_size
D = np.prod(x.size())/batch_size_use
#num_anchor = anchors.size(0)
a_mu= encoder(anchors)
a_mu_scale = torch.div(a_mu,scale)
a_mu_scale_np = a_mu_scale.detach().numpy()
d = a_mu.size(1)
sigma_recon = np.sqrt(1.0/(opt.recon_weight))
p_x_add = -D/2*np.log(2*np.pi)-D*np.log(sigma_recon)
#gamma_post = 1/np.sqrt(opt.post_TO_weight)
gamma_post = to_noise_std
post_TO_weight = 1.0/(gamma_post**2)
b_post = 1.0/opt.post_l1_weight
q_z_x_add = -d/2.0*np.log(2.0*np.pi)-d*np.log(gamma_post) - M*np.log(2.0*b_post)
gamma_prior = to_noise_std
#gamma_prior = 1/np.sqrt(opt.prior_weight)
prior_weight =1.0/(gamma_prior**2)
#b_prior = 1/opt.prior_l1_weight
b_prior = 1.0/opt.post_l1_weight
prior_l1_weight = opt.post_l1_weight
p_z_add = -d/2.0*np.log(2*np.pi)-d*np.log(gamma_prior) - M*np.log(2.0*b_prior)
LL_inner = np.zeros((batch_size_use,1))
LL_inner_no_add = np.zeros((batch_size_use,1))
time_save = np.zeros((batch_size_use))
#log_q_z_x_no_add_store = np.zeros((batch_size_use))
#log_p_x_z_no_add_store = np.zeros((batch_size_use))
for n in range(0,batch_size_use):
batch_time_start = time.time()
x_ind_torch = torch.unsqueeze(x[n,:,:,:],0)
z_mu = encoder(x_ind_torch)
z_mu_scale = torch.div(z_mu,scale)
z_mu_scale_np = z_mu_scale.detach().numpy()
x_repeat = x_ind_torch.repeat(k,1,1,1) # This may need to be expanded for images
z_mu_scale_repeat = z_mu_scale.repeat(k,1)
z_coeff = sampler_c(k,M,opt.post_l1_weight)
z_scale = transNet(z_mu_scale_repeat.double(),z_coeff.double(),Psi,to_noise_std)
z = torch.mul(z_scale,scale)
z_scale_np = z_scale.detach().numpy()
log_p_x_z_no_add = -0.5*opt.recon_weight*torch.sum((decoder(z.float()).double().reshape(k,-1)-x_repeat.double().reshape(k,-1))**2,1)
log_p_x_z_no_add_store = log_p_x_z_no_add.detach().numpy()
log_p_x_z = log_p_x_z_no_add +p_x_add
x0 = z_mu_scale_np[0,:].astype('double')
c_est_mu = np.zeros((k,M))
for b in range(0,k):
x1 = z_scale_np[b,:].astype('double')
c_est_mu[b,:],E,nit = infer_transOpt_coeff(x0,x1,Psi.detach().numpy().astype('double'),opt.post_cInfer_weight,0.0,1.0)
z_est_mu_scale = transNet(z_mu_scale_repeat.double(),torch.from_numpy(c_est_mu),Psi,0.0)
log_q_z_x_no_add = -post_TO_weight*0.5*torch.sum((scale*(z_scale.double()-z_est_mu_scale))**2,1) -opt.post_l1_weight*torch.sum(torch.abs(torch.from_numpy(c_est_mu)),1)
log_q_z_x_no_add_store = log_q_z_x_no_add.detach().numpy()
log_q_z_x = log_q_z_x_no_add + q_z_x_add
# Compute the prior loss function
if sample_labels.ndim == 1:
label_use = sample_labels[n] # Change this depending on the application
else:
label_use = np.where(sample_labels[n,:]==1)[0]
anchors_use_np = a_mu_scale_np[int(num_anchor*label_use):int(num_anchor*(label_use+1)),:]
a_mu_use = a_mu_scale[int(num_anchor*label_use):int(num_anchor*(label_use+1)),:]
log_p_z_no_add = torch.zeros(k)
log_p_z = torch.zeros(k)
anchor_idx_use = np.zeros((opt.batch_size))
for b in range(0,k):
x1 = z_scale_np[b,:].astype('double')
prior_TO_anchor_sum = 0.0
c_est_a = np.zeros((num_anchor,M))
E_anchor= np.zeros((num_anchor,numRestart))
arc_len_min = 1000000.0
for a_idx in range(0,num_anchor):
# Infer the coefficients between anchors and z
x0 = anchors_use_np[a_idx,:].astype('double')
E_single = np.zeros((numRestart))
c_est_a_store = np.zeros((numRestart,M))
for r_idx in range(0,numRestart):
#rangeMin = 0.0
#rangeMax = 1.0
#rangeMin = -600+r_idx*300
#rangeMax = rangeMin + 300
rangeMin = -2.5 + r_idx*5
rangeMax = rangeMin + 2.5
c_est_a_store[r_idx,:],E_anchor[a_idx,r_idx],nit_anchor = infer_transOpt_coeff(x0,x1,Psi.detach().numpy().astype('double'),opt.prior_cInfer_weight,rangeMin,rangeMax)
E_single[r_idx] = E_anchor[a_idx,r_idx]
minIdx = np.argmin(E_single)
c_est_a[a_idx,:] = c_est_a_store[minIdx,:]
#c_est_a[a_idx,:],E_anchor[b,a_idx],nit_anchor[b,a_idx] = infer_transOpt_coeff(x0,x1,Psi_use.astype('double'),opt.prior_cInfer_weight,-100.0,100.0)
c_est_a_ind = c_est_a[a_idx,:]
test = 0
arc_len = E_anchor[a_idx,minIdx]
if opt.closest_anchor_flag == 0:
z_est_a_scale_ind = transNet(torch.unsqueeze(a_mu_use[a_idx,:].double(),0),torch.from_numpy(np.expand_dims(c_est_a_ind,axis =0)),Psi,0.0)
z_est_a_ind = torch.mul(z_est_a_scale_ind,scale)
prior_TO_temp = torch.exp(-0.5*prior_weight*torch.sum(torch.pow(scale*(z_scale[b,:].double()-z_est_a_scale_ind),2))-prior_l1_weight*torch.sum(torch.abs(torch.from_numpy(c_est_a_ind))))
prior_TO_anchor_sum = prior_TO_anchor_sum+prior_TO_temp
elif opt.closest_anchor_flag == 1:
if arc_len < arc_len_min:
z_est_a_scale_ind = transNet(torch.unsqueeze(a_mu_use[a_idx,:].double(),0),torch.from_numpy(np.expand_dims(c_est_a_ind,axis =0)),Psi,0.0)
z_est_a_ind = torch.mul(z_est_a_scale_ind,scale)
prior_TO_anchor_sum = torch.exp(-0.5*prior_weight*torch.sum(torch.pow(scale*(z_scale[b,:].double()-z_est_a_scale_ind),2))-prior_l1_weight*torch.sum(torch.abs(torch.from_numpy(c_est_a_ind))))
#print('Change: arc orig: ' + str(arc_len_min) + ' arc new: ' + str(arc_len) + ' prior: ' + str(prior_TO_anchor_sum.detach().numpy()))
test = 1
arc_len_min = arc_len
anchor_idx_use[b] = a_idx
#z_est_a_ind = transNet(torch.unsqueeze(a_mu_scale[a_idx,:].double(),0),torch.from_numpy(np.expand_dims(c_est_a_ind,axis =0)),Psi,0.0)
#prior_TO_temp = torch.exp(-0.5*opt.prior_weight*torch.sum(torch.pow(z[b,:].double()-z_est_a_ind,2))-opt.prior_l1_weight*torch.sum(torch.abs(torch.from_numpy(c_est_a_ind))))
#prior_TO_anchor_sum = prior_TO_anchor_sum+prior_TO_temp
if opt.closest_anchor_flag == 1:
num_anchor_use = 1
else:
num_anchor_use = num_anchor
log_p_z_no_add[b] = torch.log(prior_TO_anchor_sum/num_anchor_use)
log_p_z[b] = log_p_z_no_add[b] + p_z_add
LL_inner_no_add[n] = ((log_p_x_z_no_add + log_p_z_no_add.double() - log_q_z_x_no_add).logsumexp(-1) - np.log(k)).detach().numpy()
LL_inner[n] = ((log_p_x_z + log_p_z.double() - log_q_z_x).logsumexp(-1) - np.log(k)).detach().numpy()
time_save[n] = time.time()-batch_time_start
print ("LL: [Test Sample %d/%d] time: %4.4f" % (n, batch_size_use,time.time()-batch_time_start))
sio.savemat(save_folder + 'testMetrics_batch' + str(opt.batch_size) + '_' + str(k) + 'samp_startPt' + str(opt.startPt) + '_progress.mat',{'LL_inner':LL_inner,'LL_inner_no_add':LL_inner_no_add,'step':n,'time_save':time_save});
LL_total = np.mean(LL_inner)
LL_total_no_add = np.mean(LL_inner_no_add)
return LL_total, LL_inner,LL_total_no_add, LL_inner_no_add
def test_metrics(encoder,decoder,transNet,sampler_c,Psi,x,sample_labels,anchors,to_noise_std,num_anchor,M,numRestart,opt,mse_loss_sum,latent_mse_loss,scale):
'''
Compute log_likelihood
Inputs:
- encoder: Encoder network
- decoder: Decoder network
- transNet: Transport operator layer
- sampler_c: Layer that samples the transport operator coefficients
- Psi: Current transport operator dictionary elements
- x: Batch of data [batch_size,H,W,C]
- labels: Batch of labels [batch_size,y_dim]
- to_noise_std: Sampling noise standard deviation for Gaussian prior distribution
- num_anchor: Number of anchors per class
- M: Number of transport operator dictionary elements
- numRestart: Number of restarts for coefficient inference
- opt: Set of parameters
- mse_loss_sum: MSE loss function definition for output data
- latent_mse_loss: MSE loss function used for comparing the latent vectors transformed by transport operators
- scale: Value to scale the latent vectors by to get them in a range that is suitable for coefficient inference
Outputs:
- ELBO: ELBO Computed without added constants
- MSE: Mean squared error between the input data and reconstructed data outputs
'''
test_size = 50
a_mu= encoder(anchors)
a_mu_scale = torch.div(a_mu,scale)
a_mu_scale_np = a_mu_scale.detach().numpy()
batch_size = x.shape[0]
input_h = opt.img_size
batch_idxs = batch_size // test_size
MSE_total = np.zeros((batch_idxs))
ELBO_total = np.zeros((batch_idxs))
for idx in xrange(0, batch_idxs):
batch_time_start = time.time()
X_use = x[idx*test_size:(idx+1)*test_size]
X_use = X_use.float()
z_mu = encoder(X_use)
z_mu_scale = torch.div(z_mu,scale)
z_mu_scale_np = z_mu_scale.detach().numpy()
# Sample the coefficients
z_coeff = sampler_c(test_size,M,opt.post_l1_weight)
z_scale = transNet(z_mu_scale.double(),z_coeff.double(),Psi,to_noise_std)
z = torch.mul(z_scale,scale)
z_scale_np = z_scale.detach().numpy()
# Compute the reconstruction loss
MSE = 0.5*mse_loss_sum(decoder(z.float()).double(),X_use.double())/test_size
MSE_new = torch.sum(torch.mean((decoder(z.float()).double().reshape(test_size,-1)-X_use.double().reshape(test_size,-1))**2,1),0)/test_size
MSE_total[idx] = MSE_new.detach().numpy()
# Compute the posterior loss function
# Infer the coefficients between z_mu and z
c_est_mu = np.zeros((test_size,M))
for b in range(0,test_size):
x0 = z_mu_scale_np[b,:].astype('double')
x1 = z_scale_np[b,:].astype('double')
c_est_mu[b,:],E,nit = infer_transOpt_coeff(x0,x1,Psi.detach().numpy().astype('double'),opt.post_cInfer_weight,0.0,1.0)
# Transform mu with no noise
z_est_mu_scale = transNet(z_mu_scale.double(),torch.from_numpy(c_est_mu),Psi,0.0)
post_TO_loss = (-0.5*opt.post_TO_weight*latent_mse_loss(scale*z_scale.double(),scale*z_est_mu_scale))/test_size
post_l1_loss = -torch.sum(torch.abs(torch.from_numpy(c_est_mu)))/test_size
# Compute the prior loss function
prior_TO_sum = 0.0
for b in range(0,test_size):
x1 = z_scale_np[b,:].astype('double')
prior_TO_anchor_sum = 0.0
c_est_a = np.zeros((num_anchor,M))
if sample_labels.ndim == 1:
label_use = sample_labels[b] # Change this depending on the application
else:
label_use = np.where(sample_labels[b,:]==1)[0]
anchors_use_np = a_mu_scale_np[int(num_anchor*label_use):int(num_anchor*(label_use+1)),:]
a_mu_use = a_mu_scale[int(num_anchor*label_use):int(num_anchor*(label_use+1)),:]
E_anchor= np.zeros((num_anchor,numRestart))
arc_len_min = 1000000.0
for a_idx in range(0,num_anchor):
# Infer the coefficients between anchors and z
x0 = anchors_use_np[a_idx,:].astype('double')
E_single = np.zeros((numRestart))
c_est_a_store = np.zeros((numRestart,M))
for r_idx in range(0,numRestart):
rangeMin = -2.5 + r_idx*5
rangeMax = rangeMin + 2.5
c_est_a_store[r_idx,:],E_anchor[a_idx,r_idx],nit_anchor = infer_transOpt_coeff(x0,x1,Psi.detach().numpy().astype('double'),opt.prior_cInfer_weight,rangeMin,rangeMax)
E_single[r_idx] = E_anchor[a_idx,r_idx]
minIdx = np.argmin(E_single)
c_est_a[a_idx,:] = c_est_a_store[minIdx,:]
c_est_a_ind = c_est_a[a_idx,:]
arc_len = E_anchor[a_idx,minIdx]
if opt.closest_anchor_flag == 0:
z_est_a_scale_ind = transNet(torch.unsqueeze(a_mu_use[a_idx,:].double(),0),torch.from_numpy(np.expand_dims(c_est_a_ind,axis =0)),Psi,0.0)
z_est_a_ind = torch.mul(z_est_a_scale_ind,scale)
prior_TO_temp = torch.exp(-0.5*opt.prior_weight*torch.sum(torch.pow(scale*(z_scale[b,:].double()-z_est_a_scale_ind),2))-opt.prior_l1_weight*torch.sum(torch.abs(torch.from_numpy(c_est_a_ind))))
prior_TO_anchor_sum = prior_TO_anchor_sum+prior_TO_temp
elif opt.closest_anchor_flag == 1:
if arc_len < arc_len_min:
z_est_a_scale_ind = transNet(torch.unsqueeze(a_mu_use[a_idx,:].double(),0),torch.from_numpy(np.expand_dims(c_est_a_ind,axis =0)),Psi,0.0)
prior_TO_anchor_sum = torch.exp(-0.5*opt.prior_weight*torch.sum(torch.pow(scale*(z_scale[b,:].double()-z_est_a_scale_ind),2))-opt.prior_l1_weight*torch.sum(torch.abs(torch.from_numpy(c_est_a_ind))))
#print('Change: arc orig: ' + str(arc_len_min) + ' arc new: ' + str(arc_len) + ' prior: ' + str(prior_TO_anchor_sum.detach().numpy()))
arc_len_min = arc_len
if opt.closest_anchor_flag == 1:
num_anchor_use = 1
else:
num_anchor_use = num_anchor
#print(torch.log(prior_TO_anchor_sum/num_anchor).detach().numpy())
prior_TO_sum = prior_TO_sum - torch.log(prior_TO_anchor_sum/num_anchor_use)
print ("Metrics: [Test Sample %d/%d] time: %4.4f" % (idx, batch_idxs,time.time()-batch_time_start))
prior_TO_sum = prior_TO_sum/batch_size
ELBO = -1*(opt.recon_weight*MSE + post_TO_loss + opt.post_l1_weight*post_l1_loss + prior_TO_sum)
ELBO_total[idx] = ELBO.detach().numpy()
return MSE_total, ELBO_total