-
Notifications
You must be signed in to change notification settings - Fork 17
/
module_NN_ens.py
519 lines (414 loc) · 19.2 KB
/
module_NN_ens.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
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import norm
import tensorflow as tf
import datetime
from scipy.special import erf
import importlib
import utils
importlib.reload(utils)
from utils import *
class NN():
def __init__(self,
activation_fn, x_dim, y_dim, hidden_size,
init_stddev_1_w, init_stddev_1_b, init_stddev_2_w, optimiser_in, n, learning_rate, decay_rate=1.0,
drop_out=False, deep_NN=False):
'''set up one single-layer NN'''
''' unless deep_NN = True, in which case add two layers'''
self.activation_fn = activation_fn
self.x_dim = x_dim
self.y_dim = y_dim
self.hidden_size = hidden_size
self.optimiser_in = optimiser_in
self.n = n
self.learning_rate = learning_rate
self.decay_rate = decay_rate
self.drop_out = drop_out
self.deep_NN = deep_NN
if activation_fn == 'tanh':
fn_use = tf.nn.tanh
elif activation_fn == 'relu':
fn_use = tf.nn.relu
elif activation_fn == 'Lrelu':
def tp_Lrelu(x):
a=0.2
return tf.maximum(a*x,x)
# not sure why but this didn't work
# fn_use = tf.nn.leaky_relu # alpha=0.2 by default
fn_use = tp_Lrelu
elif activation_fn == 'erf':
fn_use = tf.erf
elif activation_fn == 'softplus':
def tp_softplus(x):
# manually adjust so it is more similar to relu
c=3. # if this is bigger -> relu behaviour, but less 'soft'
return tf.divide(tf.log(tf.exp(tf.multiply(x,c)) + 1),c)
# https://stackoverflow.com/questions/44230635/avoid-overflow-with-softplus-function-in-python
# to avoid overflow we could do something like if x>30/c, return x
# return tf.cond(x>30/c, lambda: tf.divide(tf.log(tf.exp(tf.multiply(x,c)) + 1),c), lambda: x)
# def f1(): return tf.divide(tf.log(tf.exp(tf.multiply(x,c)) + 1),c)
# def f2(): return x
# return tf.cond(tf.less(x,30/c), f1, f2)
fn_use = tp_softplus
# used to have these as float32
self.inputs = tf.placeholder(tf.float64, [None, x_dim], name='inputs')
self.y_target = tf.placeholder(tf.float64, [None, y_dim], name='target')
anchor_factor = 1 # could inflate the init. dist. by this
if activation_fn != 'rbf':
# we use 'Dense' instead of 'dense' - so can access weights more easily
self.layer_1_w = tf.layers.Dense(hidden_size,
activation=fn_use, #trainable=False,
kernel_initializer=tf.random_normal_initializer(mean=0.,stddev=anchor_factor*init_stddev_1_w),
bias_initializer=tf.random_normal_initializer(mean=0.,stddev=anchor_factor*init_stddev_1_b))
self.layer_1 = self.layer_1_w.apply(self.inputs)
if self.drop_out:
self.layer_1 = tf.layers.dropout(self.layer_1,rate=0.4,training=True)
self.output_w = tf.layers.Dense(y_dim,
activation=None, use_bias=False,
kernel_initializer=tf.random_normal_initializer(mean=0.,stddev=anchor_factor*init_stddev_2_w))
if self.deep_NN:
print('going deep...')
# add an extra hidden layer
self.layer_2_w = tf.layers.Dense(hidden_size,
activation=fn_use, #trainable=False,
kernel_initializer=tf.random_normal_initializer(mean=0.,stddev=anchor_factor*init_stddev_2_w),
bias_initializer=tf.random_normal_initializer(mean=0.,stddev=anchor_factor*init_stddev_2_w))
self.layer_2 = self.layer_2_w.apply(self.layer_1)
if self.drop_out:
self.layer_2 = tf.layers.dropout(self.layer_2,rate=0.4,training=True)
self.output = self.output_w.apply(self.layer_2)
else:
self.output = self.output_w.apply(self.layer_1)
else:
self.c = tf.Variable(tf.random_normal([x_dim,hidden_size],mean=0.,stddev=anchor_factor*init_stddev_1_w, dtype=tf.float64),trainable=True, dtype=tf.float64)
self.beta = tf.Variable(initial_value=[init_stddev_1_b],trainable=False, dtype=tf.float64)
self.beta_2 = tf.pow(2*tf.square(self.beta),-1)
self.w2_rbf = tf.Variable(tf.random_normal([hidden_size,y_dim],mean=0.,stddev=anchor_factor*init_stddev_2_w, dtype=tf.float64), dtype=tf.float64)
self.layer_1_rbf = tf.exp(-self.beta_2*tf.square(self.inputs - self.c))
if self.drop_out:
self.layer_1_rbf = tf.layers.dropout(self.layer_1_rbf,rate=0.4,training=True)
self.output = tf.matmul(self.layer_1_rbf,self.w2_rbf)
# self.output = tf.matmul(tf.exp(-tf.multiply(self.beta_2,tf.square(self.inputs - self.c))),self.w2_rbf)
self.batch = tf.Variable(0, trainable=False)
# decayed_learning_rate = learning_rate *
# decay_rate ^ (global_step / decay_steps)
self.l_rate_decay = tf.train.exponential_decay(
self.learning_rate, # Base learning rate.
global_step=self.batch, # Current index into the dataset.
decay_steps=20, # Decay step.
decay_rate=self.decay_rate, # Decay rate.
staircase=True)
if optimiser_in == 'adam':
# self.opt_method = tf.train.AdamOptimizer(learning_rate)
self.opt_method = tf.train.AdamOptimizer(self.l_rate_decay)
elif optimiser_in == 'SGD':
self.opt_method = tf.train.GradientDescentOptimizer(learning_rate)
elif optimiser_in == 'AdaDel':
self.opt_method = tf.train.AdadeltaOptimizer(learning_rate)
elif optimiser_in == 'RMS':
self.opt_method = tf.train.RMSPropOptimizer(learning_rate)
elif optimiser_in == 'AdaGrad':
self.opt_method = tf.train.AdagradOptimizer(learning_rate)
# self.loss_ = tf.reduce_mean(tf.square(self.y_target - self.output))
# self.loss_ = 1/self.n * tf.reduce_sum(tf.square(self.y_target - self.output))
self.loss_ = 1/tf.shape(self.inputs, out_type=tf.int64)[0] * tf.reduce_sum(tf.square(self.y_target - self.output))
# self.mse_loss = 1/self.n * tf.reduce_sum(tf.square(self.y_target - self.output)) # useful for val
self.mse_loss = 1/tf.shape(self.inputs, out_type=tf.int64)[0] * tf.reduce_sum(tf.square(self.y_target - self.output)) # useful for val
# self.loss_ = 1/tf.cast(tf.size(self.inputs),tf.float32) * tf.reduce_sum(tf.square(self.y_target - self.output))
self.optimizer = self.opt_method.minimize(self.loss_, global_step=self.batch)
return
def get_weights(self, sess):
'''method to return current params - yes it rly does seem this hard..'''
if self.activation_fn != 'rbf':
ops = [self.layer_1_w.kernel, self.layer_1_w.bias, self.output_w.kernel]
else:
ops = [self.c, self.beta, self.w2_rbf]
w1, b1, w2 = sess.run(ops)
# b2 = sess.run(self.output_w.bias)
return w1, b1, w2
def anchor(self, sess, lambda_anchor, regularise=False, unconstrain=False):
'''method to set loss to account for anchoring'''
# lambda_anchor=[0,0,0] # hardcode for testing effect of anchoring
# regularise = True ### hardcode for testing effect of anchoring
# lambda_anchor = lambda_anchor*0.01
# print('\nlambda_anchor',lambda_anchor)
if unconstrain:
# turn off effect of prior
lambda_anchor=[0,0,0]
print('unconstraining!!!')
if regularise:
# to do normal regularisation
print('regularising!!!')
w1, b1, w2 = self.get_weights(sess)
self.w1_init, self.b1_init, self.w2_init = np.zeros_like(w1),np.zeros_like(b1),np.zeros_like(w2) # overwrite for normal regulariser
else:
# get weights
w1, b1, w2 = self.get_weights(sess)
# set around initial params
self.w1_init, self.b1_init, self.w2_init = w1, b1, w2
# print('w1_init',self.w1_init)
# print('b1_init',self.b1_init)
# print('w2_init',self.w2_init)
if self.activation_fn != 'rbf':
# set squared loss around it
loss_anchor = lambda_anchor[0]*tf.reduce_sum(tf.square(self.w1_init - self.layer_1_w.kernel))
loss_anchor += lambda_anchor[1]*tf.reduce_sum(tf.square(self.b1_init - self.layer_1_w.bias))
loss_anchor += lambda_anchor[2]*tf.reduce_sum(tf.square(self.w2_init - self.output_w.kernel))
else:
loss_anchor = lambda_anchor[0]*tf.reduce_sum(tf.square(self.w1_init - self.c))
loss_anchor += lambda_anchor[1]*tf.reduce_sum(tf.square(self.b1_init - self.beta))
loss_anchor += lambda_anchor[2]*tf.reduce_sum(tf.square(self.w2_init - self.w2_rbf))
# combine with original loss
self.loss_ = self.loss_ + 1/tf.shape(self.inputs, out_type=tf.int64)[0] * loss_anchor
# I spent a long time analysing if we need to divide this by n
# although we should in the eqn, actually tf doesn't repeat loss_anchor
# n times, so no need!
# 25 aug, actually I got this wrong - do need to. cost me a lot of time...
# reset optimiser
self.optimizer = self.opt_method.minimize(self.loss_, global_step=self.batch)
return
def get_weights_deep(self, sess):
'''method to return current params - yes it rly does seem this hard..'''
ops = [self.layer_1_w.kernel, self.layer_1_w.bias, self.layer_2_w.kernel, self.layer_2_w.bias, self.output_w.kernel]
w1, b1, w2, b2, w3 = sess.run(ops)
# b2 = sess.run(self.output_w.bias)
return w1, b1, w2, b2, w3
def anchor_deep(self, sess, lambda_anchor, regularise=False, unconstrain=False):
'''method to set loss to account for anchoring for a deep NN'''
if unconstrain:
# turn off effect of prior
lambda_anchor=[0,0,0,0,0]
print('unconstraining!!!')
if regularise:
# to do normal regularisation
print('regularising!!!')
w1, b1, w2, b2, w3 = self.get_weights_deep(sess)
self.w1_init, self.b1_init, self.w2_init, self.b2_init, self.w3_init = np.zeros_like(w1),np.zeros_like(b1),np.zeros_like(w2),np.zeros_like(b2),np.zeros_like(w3)
else:
# get weights
w1, b1, w2, b2, w3 = self.get_weights_deep(sess)
# set around initial params
self.w1_init, self.b1_init, self.w2_init, self.b2_init, self.w3_init = w1, b1, w2, b2, w3
# print('w1_init',self.w1_init)
if self.activation_fn == 'rbf':
raise Exception('tp: deep NN not set up for rbf activations')
# set squared loss around it
loss_anchor = lambda_anchor[0]*tf.reduce_sum(tf.square(self.w1_init - self.layer_1_w.kernel))
loss_anchor += lambda_anchor[1]*tf.reduce_sum(tf.square(self.b1_init - self.layer_1_w.bias))
loss_anchor += lambda_anchor[2]*tf.reduce_sum(tf.square(self.w2_init - self.layer_2_w.kernel))
loss_anchor += lambda_anchor[2]*tf.reduce_sum(tf.square(self.b2_init - self.layer_2_w.bias))
loss_anchor += lambda_anchor[2]*tf.reduce_sum(tf.square(self.w3_init - self.output_w.kernel))
# combine with original loss
self.loss_ = self.loss_ + 1/tf.shape(self.inputs, out_type=tf.int64)[0] * loss_anchor
# I spent a long time analysing if we need to divide this by n
# although we should in the eqn, actually tf doesn't repeat loss_anchor
# n times, so no need!
# 25 aug, actually I got this wrong - do need to. cost me a lot of time...
# reset optimiser
self.optimizer = self.opt_method.minimize(self.loss_, global_step=self.batch)
return
def predict(self, x, sess):
feed = {self.inputs: x}
y_pred = sess.run(self.output, feed_dict=feed)
return y_pred
class NN_ens:
def __init__(self,
activation_fn,
data_noise,
b_0_var=1.0, w_0_var=1.0, u_var=1.0, g_var=1.0,
optimiser_in = 'adam',
learning_rate = 0.001,
hidden_size = 100,
n_epochs = 100,
cycle_print = 10,
n_ensembles = 3,
regularise = False,
unconstrain = False,
drop_out = False,
deep_NN = False,
batch_size = 32,
total_trained=0,
decay_rate=1.0
):
''' create object that will be an ensemble of NNs '''
self.activation_fn = activation_fn
self.data_noise = data_noise
self.optimiser_in = optimiser_in
self.learning_rate = learning_rate
self.decay_rate = decay_rate
self.hidden_size = hidden_size
self.n_epochs = n_epochs
self.cycle_print = cycle_print
self.n_ensembles = n_ensembles
self.regularise = regularise # regularise around zero, not anchor
self.unconstrain = unconstrain # set regularisation lambdas to zero
self.drop_out = drop_out # use dropout for training and test time
self.deep_NN = deep_NN # use more than one hidden layer
self.total_trained = total_trained
self.batch_size = batch_size
self.drop_out = drop_out
if self.drop_out:
self.name_ = 'NN_drop_h' + str(hidden_size) + '_ens' + str(n_ensembles+total_trained)
elif self.regularise:
self.name_ = 'NN_regular_h' + str(hidden_size) + '_ens' + str(n_ensembles+total_trained)
elif self.unconstrain:
self.name_ = 'NN_uncons_h' + str(hidden_size) + '_ens' + str(n_ensembles+total_trained)
elif self.deep_NN:
self.name_ = 'NN_deepanch_h' + str(hidden_size) + '_ens' + str(n_ensembles+total_trained)
else:
self.name_ = 'NN_anch_h' + str(hidden_size) + '_ens' + str(n_ensembles+total_trained)
# variance for step fn, relu, erf
self.b_0_var = b_0_var # first layer bias variance
self.w_0_var = w_0_var # first layer weight variance
# variance for rbf - we use williams 1996 notation
# i.e. node = exp(-(x-u)^2 / 2*var_g)
self.g_var = g_var # param of rbf fn (fixed)
self.u_var = u_var # var of centers, as -> inf, goes to stationary cov dist
# place holders
self.mse_unnorm = 0.
self.rmse = 0.
self.nll = 0.
return
def train(self, X_train, y_train, X_val=None, y_val=None, is_print=True):
''' train an ensemble of NNs '''
# note we use different notation in this file,
# so b_1 is first bias - elsewhere we call this b_0
if self.activation_fn == 'relu' or self.activation_fn == 'softplus' or self.activation_fn == 'Lrelu':
init_stddev_1_w = np.sqrt(self.w_0_var) # /np.sqrt(self.hidden_size)
init_stddev_1_b = np.sqrt(self.b_0_var) # /np.sqrt(self.hidden_size)
init_stddev_2_w = 1.0/np.sqrt(self.hidden_size)#*np.sqrt(10) # 2nd layer init. dist
# init_stddev_2_w = np.sqrt(10.0)/np.sqrt(self.hidden_size)#*np.sqrt(10) # 2nd layer init. dist
lambda_anchor = self.data_noise/(np.array([init_stddev_1_w,init_stddev_1_b,init_stddev_2_w*1])**2)#/X_train.shape[0]
# lambda_anchor = [0.,0.,0.]
elif self.activation_fn == 'tanh' or self.activation_fn == 'erf':
init_stddev_1_w = np.sqrt(self.w_0_var) # 1st layer init. dist for weights
init_stddev_1_b = np.sqrt(self.b_0_var) # for bias
init_stddev_2_w = 1.0/np.sqrt(self.hidden_size) # 2nd layer init. dist
# lambda_anchor = [0.,0.,0.] # lambda for weight layer 1, bias layer 1, weight layer 2
lambda_anchor = self.data_noise/(np.array([init_stddev_1_w,init_stddev_1_b,init_stddev_2_w])**2)
elif self.activation_fn == 'rbf':
init_stddev_1_w = np.sqrt(self.u_var) # centres = sig_u
init_stddev_1_b = np.sqrt(self.g_var) # fixed /beta
init_stddev_2_w = 1.0/np.sqrt(self.hidden_size) # 2nd layer init. dist
lambda_anchor = self.data_noise/(np.array([init_stddev_1_w,init_stddev_1_b,init_stddev_2_w])**2)
n = X_train.shape[0]
X_dim = X_train.shape[1]
y_dim = 1 #y_train.shape[1]
# batch_size = n
# --- ensembles w proper anchoring! ---
NNs=[]
y_pred=[]
y_prior=[]
tf.reset_default_graph()
sess = tf.Session()
for ens in range(0,self.n_ensembles):
if is_print:
print('\n\n-- working on ensemble number '+ str(self.total_trained + ens) + ' ---')
else:
print('-- working on ensemble number '+ str(self.total_trained + ens) + ' ---', end='\r')
# create a NN
NNs.append(NN(self.activation_fn, X_dim, y_dim, self.hidden_size,
init_stddev_1_w, init_stddev_1_b, init_stddev_2_w,
self.optimiser_in, n, self.learning_rate, decay_rate=self.decay_rate, drop_out=self.drop_out, deep_NN=self.deep_NN))
# sess.run(tf.global_variables_initializer()) # must do this after NN created
# sess.run(tf.initialize_variables([NNs[ens].layer_1_w.kernel, NNs[ens].layer_1_w.bias, NNs[ens].output_w.kernel]))
# initialise only unitialized variables
global_vars = tf.global_variables()
is_not_initialized = sess.run([tf.is_variable_initialized(var) for var in global_vars])
not_initialized_vars = [v for (v, f) in zip(global_vars, is_not_initialized) if not f]
if len(not_initialized_vars):
sess.run(tf.variables_initializer(not_initialized_vars))
# set anchoring
if self.deep_NN == False:
NNs[ens].anchor(sess, lambda_anchor, regularise=self.regularise, unconstrain=self.unconstrain)
else:
NNs[ens].anchor_deep(sess, lambda_anchor, regularise=self.regularise, unconstrain=self.unconstrain)
# prior
# if X_val != None:
y_prior.append(NNs[ens].predict(X_val, sess))
# do training
feed = {}
feed[NNs[ens].inputs] = X_train
feed[NNs[ens].y_target] = y_train
# feed[NNs[ens].l_rate_in] = 0.1
# print('\n\nhhhhh\n\n\n')
# print(self.learning_rate)
# if (X_val!=None)[0,0]:
feed_val = {}
feed_val[NNs[ens].inputs] = X_val
feed_val[NNs[ens].y_target] = y_val
# feed_val[NNs[ens].l_rate_in] = self.learning_rate
ep_ = 0; train_complete=False
while not train_complete:
# for ep_ in range(0,self.n_epochs):
if False and ep_==0:
# view training as it progresses
y_pred_temp = NNs[ens].predict(X_val,sess)
plot_1d_grid(X_val, y_pred_temp, 0.01, X_train, y_train,title='ep '+str(ep_))
# blank = sess.run(NNs[ens].optimizer, feed_dict=feed)
# train in batches
perm = np.random.permutation(X_train.shape[0])
X_train_shuff = X_train[perm]
y_train_shuff = y_train[perm]
n_batches = int(np.ceil(X_train.shape[0]/self.batch_size))
for b in range(0,n_batches):
# if not final batch
if b != n_batches:
X_train_b = X_train_shuff[b*self.batch_size:(b+1)*self.batch_size]
y_train_b = y_train_shuff[b*self.batch_size:(b+1)*self.batch_size]
else: # use whatever is left
X_train_b = X_train_shuff[b*self.batch_size:]
y_train_b = y_train_shuff[b*self.batch_size:]
feed_b = {}
feed_b[NNs[ens].inputs] = X_train_b
feed_b[NNs[ens].y_target] = y_train_b
blank = sess.run(NNs[ens].optimizer, feed_dict=feed_b)
if ep_ % self.cycle_print == 0 or ep_ == self.n_epochs-1:
if False:
# view training as it progresses
y_pred_temp = NNs[ens].predict(X_val,sess)
plot_1d_grid(X_val, y_pred_temp, 0.01, X_train, y_train,title='ep '+str(ep_))
loss = sess.run(NNs[ens].loss_, feed_dict=feed)
# if (X_val!=None)[0,0]:
loss_val = sess.run(NNs[ens].mse_loss, feed_dict=feed_val)
l_rate_curr = sess.run(NNs[ens].l_rate_decay, feed_dict=feed_val)
if is_print:
print('ep:',ep_,'/', self.n_epochs-1, 'train:',np.round(loss,5), 'val mse:',np.round(loss_val,5), 'lr',np.round(l_rate_curr,5))#, end='\r')
# useful to do a stability check here
# if one NN doesnt train perfectly it can mess up whole ensemble
if ep_ == self.n_epochs-1: # if last run
# train further if increased since last check
if (loss - loss_old)/loss > 0.02:
ep_ = np.max(ep_-int(self.n_epochs/10),0)
print(' !!! one was unstable !!!, continuing training')
continue
loss_old = loss.copy()
ep_+=1
if ep_ == self.n_epochs: train_complete=True
# make prediction - used to do here so don't worry about reinit other NNs
# but we found a way to only reinit new variables now
# y_pred.append(NNs[ens].predict(x_s))
self.NNs = NNs
self.sess = sess
# priors
# if X_val != None:
y_priors = np.array(y_prior)
y_priors = y_priors[:,:,0]
y_prior_mu = np.mean(y_prior,axis=0)
y_prior_std = np.std(y_prior,axis=0, ddof=1)
return y_priors, y_prior_mu, y_prior_std
def predict(self, X_val):
# predict for whole ensemble
y_pred=[]
for ens in range(0,self.n_ensembles):
y_pred.append(self.NNs[ens].predict(X_val, self.sess))
y_preds = np.array(y_pred)
y_preds = y_preds[:,:,0]
y_pred_mu = np.mean(y_preds,axis=0)
y_pred_std = np.std(y_preds,axis=0, ddof=1)
# add on data noise
y_pred_std = np.sqrt(np.square(y_pred_std) + self.data_noise)
self.y_pred_mu = y_pred_mu
self.y_pred_std = y_pred_std
self.y_preds = y_preds
return y_preds, y_pred_mu, y_pred_std