-
Notifications
You must be signed in to change notification settings - Fork 196
/
train_IAN_simple.py
864 lines (661 loc) · 45.2 KB
/
train_IAN_simple.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
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
###
# Hierarchical Adversarial Introspective Autoencoder Main training Function
#
# A Brock 2016
# TO DO:
# 1. Cleanup
# 2. Split validation/test sets appropriately, do early stopping on validation set, do reconstructions on test set
# 13. Learn to render to HTML with VTK, and do the 2d-manifold thing. Consider rendering directly to a webpage?
# 16. Read and UNDERSTAND Batch Norm and Improved Gan Training papers. Really no excuse for not absolutely understanding these; consider checking WeightNorm (by salimans?) too.
# 19. Add code to automatically test if a saved npz of the weights already exists
# 27. Figure out how to incorporate log likelihood into introspective reconstruction error, along with new log_sigma_thetas
# 32. Do we want the classifier to take in the LATENTS or the layer just before the latents?
# 33. Do we want to propagate the log_sigma through the decoder network somehow to make log-likelihood work better? I'd prefer to use some entropy measure, honestly
# 36. Add support for class-conditional-ness
# 38. Test regime: number of latent variables vs. accuracy
# 40. Multi-GPU? Split a batch across two GPUs? Send the autoencoding batch to one gpu, send the adversarial batch to another
# 42. Get a validation set together
#
# Consider hacking the adversarial BCE by changing from zeros/ones to -1's and 2's.
# consider only using adversarial gen loss on random samples
# Consider the effects of batch normalization on an X_hat batch split into half-recon and half-generated,
# specifically: 1. Does batch-norm depend on the 0-axis of the tensor at all, such that we need to shuffle that tensor? I don't think it doesn
# 2. Should we be splitting and separately batch norming the recon slice and the generated slice? It wouldn't increase
# computational complexity too much, I don't think, though it depends on what layer we split the bnorm in--does this split
# persist at all intermediate layers, or just on the output? How exactly does this split work anyhow?
# Consider replacing inception modules with inception-style context aggregation modules using dilated convolutions, or concatenating them
# to get multiple feature maps? i.e. including dilated convs as one of the possible pieces of an inception module
# Consider the information path from the intermediate layers to the hierarchical latents--can we reduce the number of parameters
# And improve expressiveness by replacing FC layers with convolutional/inception-style/context aggregation layers, and if so
# what exactly are we doing by doing this? How does replacing the FC layer with a set of conv layers differ from just FC'ing to a
# deeper layer? Can we do this in the output of the hierarchical layer as well, replacing it with an inception upscale layer?
# Consider replacing nesterov Momentum with Adam or adamax, using our own implementation that doesn't toss preexisting grads
# Consider doing everything as SGD in the middle, then applying ADAM or Nesterov Momentum at the very end to the entire Updates dict
# (i.e. producing an "apply_adam" method a la lasagne's implementation of Nesterov Momentum)
# Consider adding in provisions to completely abandon the pixel-wise reconstruction, and only use introspection+adversarial loss
# Consider returning more meaningful adversarial loss metrics so that we can observe adversarial performance more accurately.
# Consider weighting adversarial loss so that the discriminator doesn't just learn to be dumb and let the generator walk all over it,
# i.e. running into a super-local optimum such that it always outputs "True," thereby always being right when examining a real image and
# letting the generator always get its full win when examining a generated image.
# learn the difference between "i.e." and "e.g."
# Some sort of attention-esque mechanism...maybe give each latent a particular receptive field? Or weight pixel-wise reconstruction
# by each pixel's distance from center, forcing it to produce outer details better
# Is the data shuffled? How is it ordered? Do we need to shuffle it ourselves?
# Is a 500-dim hierarchical latent representation realy equivalent to a 500-dim non-hierarchical rep if
# it necessarily increases the number of layer parameters?
# Add in text to our generated images, indicating configuration parameters, epoch number, and accuracy, for easy reference
# Scale the adversarial gradients by the reconstruction loss so that one objective does not overtake the other
# Develop gui that allows you to explore 3D latent space
# Maybe even do that for hierarchical latent space where the lower-level inputs are set to zero. We can run inference on 32*32*32 in real time,
# we can damn well do it in 3*(2*32)*(2*32) = 12*32*32
# Version Notes:
#
#
# introspection notes:
# Get more improvement/size of latent space; seems to also generalize better (the avg training error is closer to the validation error)
# Figure out a way to shuffle back and forth such that we reconstruct/adversarialize on alternating halves of the dataset and
# Add option to choose between nesterov momentum and adam
# Consider adding in log_sigma parameters for the introspective losses as well
#
## Hierarchical Adversarial Autoencoder
#
# Consider throwing in a "output pictures and save only if validation accuracy improves"
# Version 3+: Changing from (0,1) to (-1,1) and using Tanh in place of sigmoids
## Notes: orthogonal regularization
# -Only applied to weight vectors! make sure this only happens on convolution filters or square matrices!
# instead of Eye, do we need to scale by eye/num_filts?
import argparse
import imp
import time
import logging
import itertools
import numpy as np
from path import Path
import theano
import theano.tensor as T
from theano.tensor.opt import register_canonicalize
from theano.sandbox.rng_mrg import MRG_RandomStreams as RandomStreams
import lasagne
from lasagne.layers import SliceLayer as SL
import voxnet
import CAcheckpoints
import GANcheckpoints
from collections import OrderedDict
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from fuel.datasets import CelebA
from discgen_utils import plot_image_grid
## Utilities:
# to_tanh: transforms an array in the range [0,255] to the range [-1,1]
# from_tanh: transforms an array in the range [-1,1] to the range[0,255]
def to_tanh(input):
return 2.0*(input/255.0)-1.0
# return input/255.0
def from_tanh(input):
return 255.0*(input+1)/2.0
# return 255.0*input
### Make Training Functions Method
# This function defines and compiles the computational graphs that define the training, validation, and test functions.
def make_training_functions(cfg,model):
# Define input tensors
# Tensor axes are batch-channel-dim1-dim2
# Image Input
X = T.TensorType('float32', [False]*4)('X')
# Latent Input, for providing latent values from the main function
Z = T.TensorType('float32', [False]*2)('Z') # Latents
# Classification Tensor, only used if including a supervised or class-conditional task
y = T.TensorType('float32', [False]*2)('y')
# Shared and Utility Variables
X_shared = lasagne.utils.shared_empty(4, dtype='float32')
y_shared = lasagne.utils.shared_empty(2, dtype='float32')
Z_shared = lasagne.utils.shared_empty(2, dtype='float32')
pi = np.cast[theano.config.floatX](np.pi)
# Input layer
l_in = model['l_in']
# Output layer
l_out = model['l_out']
# Latent Layer
l_latents = model['l_latents']
# Means
l_mu = model['l_mu']
# Log-sigmas
l_ls = model['l_ls']
# Introspective loss layers
l_introspect = model['l_introspect']
# Adversarial Discriminator
l_discrim = model['l_discrim']
# Classifier
l_classifier = model['l_classifier']
# Class-conditional latents
l_cc = model['l_cc']
# Decoder Layers, including final output layer. Consider calling this something difference to indicate that it is
# actually a list of layers, not just a single layer.
l_decoder = lasagne.layers.get_all_layers(l_out)[len(lasagne.layers.get_all_layers(l_latents)):]
# Batch Indexing Parameters
batch_index = T.iscalar('batch_index')
batch_slice = slice(batch_index*cfg['batch_size'], (batch_index+1)*cfg['batch_size'])
# Define RNG
rng = RandomStreams(lasagne.random.get_rng().randint(1,69))
###############################################################################
# Step 1: Compute full forward pass, save the outputs of all relevant layers? #
###############################################################################
# Build the main computational graph
# if cfg['reconstruct'] or cfg['introspect']:
# if cfg['adversarial']:
# outputs = lasagne.layers.get_output([l_out]+[l_mu]+[l_ls]+\
# [l_classifier]+[l_discrim]+[SL(l,indices=slice(0,cfg['batch_size']//2),axis=0)\
# for l in l_introspect],\
# {l_in:X,
# model['l_cc']:y,
# model['l_Z_rand']:rng.normal(cfg['batch_size']//2,cfg['num_latents'])}) # Consider swapping l_classifier in for l_latents. may need to tab this properly
# else:
# outputs = lasagne.layers.get_output([l_out]+[lasagne.layers.ConcatLayer([lasagne.layers.flatten(mu) for mu in l_mu])]+[lasagne.layers.ConcatLayer([lasagne.layers.flatten(ls) for ls in l_ls])]+\
# [l_classifier]+[l_discrim]+l_introspect,\
# {l_in:X,
# model['l_cc']:y}) # Consider swapping l_classifier in for l_latents. may need to tab this properly
# elif cfg['adversarial']:
outputs = lasagne.layers.get_output([l_out]+[l_mu]+[l_ls]+\
[l_classifier]+[l_discrim]+l_introspect,\
{l_in:X,
model['l_cc']:y,
model['l_Z_rand']:Z})
# Reconstruction
X_hat = outputs[0]
# Latent means
Z_mu = outputs[1]
# Latent log-sigmas
Z_ls = outputs[2]
# Classification Predictions
y_hat = outputs[3]
# Discriminator Output
p_X = outputs[4]
# Output of the encoder layers (selected for introspection) as a function of the input image
g_X = outputs[5:]
# Build the second half of the computational graph
# if cfg['adversarial']:
# out_hat = lasagne.layers.get_output([l_discrim]+[SL(l,indices=slice(0,cfg['batch_size']//2),axis=0) for l in l_introspect],{l_in:X_hat})
# else:
# out_hat = lasagne.layers.get_output([l_discrim]+l_introspect,{l_in:X_hat})
out_hat = lasagne.layers.get_output([l_discrim]+l_introspect,{l_in:X_hat})
# Discriminator Output given Reconstruction/samples
p_X_hat = out_hat[0]
# Output of the encoder layers (selected for introspection) as a function of the reconstruction
g_X_hat = out_hat[1:]
p_X_gen = lasagne.layers.get_output(l_discrim,{l_in:lasagne.layers.get_output(l_out,{model['l_latents']:Z})})
# Build the testing computational graph
out_d = lasagne.layers.get_output([l_out,l_classifier]+[SL(lat,indices=slice(0,cfg['batch_size']//2),axis=0) for lat in l_latents],
{l_in:X,
model['l_cc']:y,
model['l_Z_rand']:rng.normal((cfg['batch_size']//2,cfg['num_latents']))},
deterministic=True) if cfg['adversarial'] and cfg['reconstruct'] else\
lasagne.layers.get_output([l_out,
l_classifier]+l_latents,
{l_in:X,
model['l_cc']:y},
deterministic=True) if cfg['hierarchical'] else\
lasagne.layers.get_output([l_out,
l_classifier,l_latents],
{model['l_Z_rand']:rng.normal((cfg['batch_size'],cfg['num_latents']))},
deterministic=True)
X_hat_deterministic = out_d[0]
y_hat_deterministic = out_d[1]
latent_values = out_d[2:] if cfg['hierarchical'] else None
#################################
# Step 2: Define loss functions #
#################################
# Orthogonal normalization for all parameters
# Define orthonormal residual
def ortho_res(z):
s = 0
for x in z:
if x.name[4:8] is 'conv':
y = T.batched_tensordot(x,x.dimshuffle(0,1,3,2),[[1,3],[1,2]])
y-=T.eye(x.shape[2],x.shape[3]).dimshuffle('x',0,1).repeat(x.shape[0],0)
s+=T.sum(T.abs(y))
return(s)
# return sum([T.sum( T.abs_(T.batched_tensordot(x,x.dimshuffle(0,1,3,2),[[1,3],[1,2]])-T.eye(x.shape[2],x.shape[3]).dimshuffle('x',0,1).repeat(x.shape[0],0))) for x in z])
# return T.sum(T.batched_dot(x,x.dimshuffle(1,0,2))-T.identity_like(x))
# y = T.batched_dot(x,x.T)
# return T.sum(y-T.identity_like(x))
# y = [T.batched_dot(x,x.T) for x in z]
# return T.sum([T.sum(x-T.identity_like(x)) for x in y])
l2_all = lasagne.regularization.regularize_network_params(l_out,
lasagne.regularization.l2,tags={'regularizable':True,'trainable':True})
# Create log_sigma parameter
log_sigma_theta = lasagne.utils.create_param(spec=np.zeros((3, 64, 64)),shape=(3,64,64), name='log_sigma_theta')
log_sigma = log_sigma_theta.dimshuffle('x', 0, 1, 2)
# Define Pixel-wise reconstruction loss
# if cfg['reconstruct'] or cfg['introspect']:
# pixel_loss = 0.5 * T.mean(T.log(2 * pi) + 2 * log_sigma + T.sqr(lasagne.nonlinearities.sigmoid(X_hat[:cfg['batch_size']//2,:,:,:])\
# - X[:cfg['batch_size']//2,:,:,:]) / T.exp(2 * log_sigma)) if cfg['adversarial']\
# else 0.5 * T.mean(T.log(2 * pi) + 2 * log_sigma + T.sqr(lasagne.nonlinearities.sigmoid(X_hat) - X) / (T.exp(2 * log_sigma)*(T.abs_(X-0.5)+0.5)))
# else:
# pixel_loss = None
# pixel_loss = 0.5 * T.mean(T.log(2 * pi) + 2 * log_sigma + T.sqr(X_hat - X) / (T.exp(2 * log_sigma)))
# pixel_loss = T.mean(2.0*T.log(0.5*( T.exp(100*(X_hat-X)) + T.exp(-100*(X_hat-X)) ) ) )
pixel_loss = T.mean(2*T.abs_(X_hat-X+1e-8))
# KL Divergence between latents and Standard Normal prior
kl_div = -0.5 * T.mean(1 + 2*Z_ls - T.sqr(Z_mu) - T.exp(2 * Z_ls))
# kl_div = -0.5 * T.mean(1 + 2*Z_ls[:cfg['batch_size']//2,:] - T.sqr(Z_mu[:cfg['batch_size']//2,:]) - T.exp(2 * Z_ls[:cfg['batch_size']//2,:])) if cfg['adversarial'] and cfg['reconstruct']\
# else -0.5 * T.mean(1 + 2*Z_ls - T.sqr(Z_mu) - T.exp(2 * Z_ls)) if cfg['reconstruct'] or cfg['introspect']\
# else None
# Classification objective losses
if cfg['discriminative']:
print('Calculating Classification Loss and Grads...')
# Calculate Classifier Loss
classifier_loss = T.cast(T.mean(T.nnet.categorical_crossentropy(T.nnet.softmax(y_hat), y)), 'float32')
# Classifier Training Accuracy
classifier_error_rate = T.cast( T.mean( T.neq(T.argmax(y_hat,axis=1), T.argmax(y,axis=1)) ), 'float32' )
# Classifier Validation/Test Accuracy
classifier_test_error_rate = T.cast( T.mean( T.neq(T.argmax(y_hat_deterministic,axis=1), T.argmax(y,axis=1))), 'float32' )
# Combined losses
reg_pixel_loss = pixel_loss + cfg['reg']*l2_all +classifier_loss+kl_div if cfg['kl_div'] else pixel_loss + cfg['reg']*l2_all +classifier_loss
else:
classifier_loss = None
classifier_error_rate = None
classifier_test_error_rate = None
reg_pixel_loss = pixel_loss + cfg['reg']*l2_all+kl_div if cfg['kl_div'] and cfg['reconstruct'] else pixel_loss + cfg['reg']*l2_all if cfg['reconstruct'] else None
##########################
# Step 3: Define Updates #
##########################
# Get Parameters
# All network parameters, including log_sigma
params = lasagne.layers.get_all_params(l_out,trainable=True)+[log_sigma_theta]
# Encoder Parameters
encoder_params = lasagne.layers.get_all_params(l_latents,trainable=True)+l_discrim.get_params(trainable=True)
# Decoder Params--consider including log_sigma_theta in this, too?
decoder_params = [p for p in lasagne.layers.get_all_params(l_out,trainable=True) if p not in encoder_params]
# decoder_params = lasagne.layers.get_all_params(l_out,trainable=True)
Z_params = [p for p in lasagne.layers.get_all_params(l_latents,trainable=True) if p not in lasagne.layers.get_all_params(l_discrim,trainable=True)]
# Define learning rate, with provisions made for annealing schedule
if isinstance(cfg['learning_rate'], dict):
learning_rate = theano.shared(np.float32(cfg['learning_rate'][0]))
else:
learning_rate = theano.shared(np.float32(cfg['learning_rate']))
# Prepare the pixel-wise reconstruction updates
# if cfg['reconstruct']:
# print('Calculating Pixel-wise Loss and Grads...')
# updates = lasagne.updates.adam(reg_pixel_loss,params,learning_rate)
# elif cfg['kl_div']:
# updates = lasagne.updates.adam(kl_div,encoder_params,learning_rate)
# else:
# updates = OrderedDict()
# updates=lasagne.updates.adam(
# grads = T.grad(cost = reg_pixel_loss, wrt = params) # Optionally calculate gradients directly
# Adversarial Stuff
if cfg['adversarial']:
print('Calculating Adversarial Loss and Grads...')
# Regularizations
# l2_discrim = lasagne.regularization.regularize_network_params(l_discrim,
# lasagne.regularization.l2,tags={'regularizable':True,'trainable':True})
l2_discrim = lasagne.regularization.apply_penalty(encoder_params,lasagne.regularization.l2)
l2_gen = lasagne.regularization.apply_penalty(decoder_params,lasagne.regularization.l2)
# l2_gen = lasagne.regularization.regularize_network_params(l_out,
# lasagne.regularization.l2,tags={'regularizable':True,'trainable':True})
# Adversarial Loss for Discriminator
# adversarial_discrim_loss = T.mean(T.nnet.binary_crossentropy(T.clip( p_X_hat , 1e-7, 1.0 - 1e-7), T.zeros(p_X_hat.shape)))\
# + T.mean(T.nnet.binary_crossentropy(T.clip( p_X , 1e-7, 1.0 - 1e-7), T.ones(p_X.shape)))+cfg['reg']*l2_discrim
feature_loss = T.cast(T.mean([T.mean(lasagne.objectives.squared_error(g_X[i],g_X_hat[i])) for i in xrange(len(g_X_hat))]),'float32')
discrim_g_loss = T.mean(T.nnet.binary_crossentropy(T.clip( p_X_hat , 1e-7, 1.0 - 1e-7), T.zeros(p_X_hat.shape)))+\
T.mean(T.nnet.binary_crossentropy(T.clip( p_X_gen , 1e-7, 1.0 - 1e-7), T.zeros(p_X_gen.shape)))
discrim_d_loss = T.mean(T.nnet.binary_crossentropy(T.clip( p_X , 1e-7, 1.0 - 1e-7), T.ones(p_X.shape)))
adversarial_discrim_loss = discrim_g_loss+discrim_d_loss+cfg['reg']*l2_discrim#+kl_div+cfg['recon_weight']*pixel_loss
# Discriminator Accuracy
discrim_accuracy = (T.mean(T.ge(p_X,0.5))+T.mean(T.lt(p_X_hat,0.5)))/2
# Adversarial Loss for Generator
adversarial_gen_loss = T.mean(T.nnet.binary_crossentropy(T.clip( p_X_hat , 1e-7, 1.0 - 1e-7), T.ones(p_X_hat.shape)))+\
T.mean(T.nnet.binary_crossentropy(T.clip( p_X_gen , 1e-7, 1.0 - 1e-7), T.ones(p_X_gen.shape)))+\
cfg['reg']*l2_gen
# adversarial_gen_loss = T.mean(-T.log(T.clip( p_X_hat , 1e-7, 1.0 - 1e-7)/(1-T.clip( p_X_hat , 1e-7, 1.0 - 1e-7)))) + cfg['reg']*l2_gen
# Total Adversarial Loss
# adversarial_loss = adversarial_discrim_loss+adversarial_gen_loss
# Optional: Expressions not to backpropagate through, for use with "consider_constant" in T.grad
# block = list(itertools.chain.from_iterable([i.get_params() for i in lasagne.layers.get_all_layers(model['l_dec_fc1'])[len(lasagne.layers.get_all_layers(model['l_latents']))+1:-1]]))
# Adversarial Gradients for Discriminator
# adversarial_discrim_grads = T.grad(cost = adversarial_discrim_loss, wrt = lasagne.layers.get_all_params(l_discrim,trainable=True))#, consider_constant = [X_hat])
# Adversarial Gradients for Generator
# adversarial_gen_grads = T.grad(cost = adversarial_gen_loss, wrt = lasagne.layers.get_all_params(l_out,trainable=True))
# Prepare Adversarial Updates with Adam
# updates = lasagne.updates.adam(adversarial_discrim_grads+adversarial_gen_grads,
# lasagne.layers.get_all_params(l_discrim,trainable=True)+decoder_params,learning_rate,beta1=cfg['beta1']) if cfg['optimizer']=='Adam'\
# else lasagne.updates.nesterov_momentum(adversarial_discrim_grads+adversarial_gen_grads,
# lasagne.layers.get_all_params(l_discrim,trainable=True)+decoder_params,learning_rate,cfg['momentum'])
discrim_updates = lasagne.updates.adam(adversarial_discrim_loss,encoder_params,learning_rate,beta1=cfg['beta1'])
discrim_to_latent_updates = lasagne.updates.adam(cfg['feature_weight']*feature_loss+cfg['recon_weight']*pixel_loss+kl_div,
Z_params,
learning_rate=learning_rate,beta1=cfg['beta1'])
for ud in discrim_to_latent_updates:
discrim_updates[ud] = discrim_to_latent_updates[ud]
gen_updates = lasagne.updates.adam(adversarial_gen_loss+cfg['recon_weight']*pixel_loss+cfg['feature_weight']*feature_loss,decoder_params,learning_rate,beta1=cfg['beta1'])
# for param in lasagne.layers.get_all_params(model['l_mu'],trainable=True)[-3:]+lasagne.layers.get_all_params(model['l_ls'],trainable=True)[-3:]:
# gen_updates[param] = lasagne.updates.adam(cfg['feature_weight']*feature_loss+cfg['recon_weight']*pixel_loss+kl_div,[param],learning_rate,beta1=cfg['beta1'])
for ud in discrim_to_latent_updates:
gen_updates[ud] = discrim_to_latent_updates[ud]
# if cfg['reconstruct'] or cfg[:
# for param in adversarial_updates:
# updates[param] = updates[param] + adversarial_updates[param] - param if param in updates else adversarial_updates[param]
# Prepare Adversarial Updates with Nesterov Momentum
# for param,grad in zip(lasagne.layers.get_all_params(l_discrim,trainable=True)+decoder_params,adversarial_discrim_grads+adversarial_gen_grads):
# value = param.get_value(borrow=True)
# velocity = theano.shared(np.zeros(value.shape, dtype=value.dtype),
# broadcastable=param.broadcastable)
# x = cfg['momentum'] * velocity - learning_rate*grad
# updates[velocity] = x
# updates[param] = updates[param] + x*cfg['momentum'] - learning_rate*grad if param in updates else param + x*cfg['momentum'] - learning_rate*grad
else:
adversarial_gen_loss = None
adversarial_discrim_loss = None
if cfg['introspect']:
print('Calculating Introspective Loss and Grads...')
# Introspective Loss Term
# Optionally include term to scale losses such that deeper layers are considered more important than more shallow layers
feature_loss = T.cast(T.mean([T.mean(lasagne.objectives.squared_error(g_X[i],g_X_hat[i])) for i in xrange(len(g_X_hat))]),'float32')
# Introspective Gradients
feature_grads = lasagne.updates.get_or_compute_grads(feature_loss,decoder_params)
# Prepare Introspective Updates with Adam
feature_updates = lasagne.updates.adam(feature_grads,decoder_params,learning_rate) if cfg['optimizer']=='Adam'\
else lasagne.updates.nesterov_momentum(feature_grads,decoder_params,learning_rate,cfg['momentum'])
for param in feature_updates:
updates[param] = updates[param] + feature_updates[param] - param if param in updates else feature_updates[param]
# Prepare Introspective Updates with Nesterov Momentum
# for param,grad in zip(decoder_params,feature_grads):
# value = param.get_value(borrow=True)
# velocity = theano.shared(np.zeros(value.shape, dtype=value.dtype),
# broadcastable=param.broadcastable)
# x = cfg['momentum'] * velocity - learning_rate*grad
# updates[velocity] = x
# updates[param] = updates[param] + x*cfg['momentum'] - learning_rate*grad if param in updates else param + x*cfg['momentum'] - learning_rate * grad
else:
feature_loss = None
# Pixel-wise Training MSE
# error_rate = T.cast( T.mean( T.sqr(lasagne.nonlinearities.sigmoid(X_hat[:cfg['batch_size']//2,:,:,:])-X[:cfg['batch_size']//2,:,:,:])), 'float32' ) if cfg['adversarial'] and cfg['reconstruct']\
# else T.cast( T.mean( T.sqr(lasagne.nonlinearities.sigmoid(X_hat)-X)), 'float32' ) if cfg['reconstruct'] or cfg['introspect']\
# else None
error_rate = T.cast( T.mean( T.sqr(X_hat-X)), 'float32' )
# Pixel-wise Test MSE
test_error_rate = T.cast( T.mean( T.sqr(lasagne.nonlinearities.sigmoid(X_hat_deterministic[:cfg['batch_size']//2,:,:,:])-X[:cfg['batch_size']//2,:,:,:])), 'float32' ) if cfg['adversarial'] and cfg['reconstruct']\
else T.cast( T.mean( T.sqr(lasagne.nonlinearities.sigmoid(X_hat_deterministic)-X)), 'float32' ) if cfg['reconstruct'] or cfg['introspect']\
else None
# Sample function
if cfg['hierarchical']:
sample = theano.function([Z],lasagne.nonlinearities.sigmoid(lasagne.layers.get_output(model['l_out'],\
{l_Z:T.reshape(Z[:,a[0]:a[1]],(T.shape(Z)[0],)+dim) for l_Z,a,dim in zip(model['l_latents'], zip(np.append(0,cfg['latent_indices'][:-1]),cfg['latent_indices']),cfg['latent_dims'])},deterministic=True)),on_unused_input='warn')
else:
sample = theano.function([Z],lasagne.layers.get_output(model['l_out'],{model['l_latents']:Z},deterministic=True))
# Inference Function--Infer latents given an image
# Zfn = theano.function([X],T.concatenate([T.flatten(l,2) for l in latent_values],axis=1),on_unused_input='warn') if cfg['reconstruct'] or cfg['introspect'] else None
Zfn = theano.function([X],lasagne.layers.get_output(model['l_latents'],{model['l_in']:X},deterministic=True),on_unused_input='warn')
# Outputs for Update Function
update_outs = [x for x in [pixel_loss,
feature_loss,
classifier_loss,
adversarial_gen_loss,
adversarial_discrim_loss,
kl_div,
classifier_error_rate,
error_rate,
] if x is not None]
# Define Update Function
# update_iter = theano.function([batch_index],update_outs,
# updates=updates, givens={
# X: X_shared[batch_slice],
# y: y_shared[batch_slice],
# Z: Z_shared[batch_slice],
# },on_unused_input='warn' )
update_gen = theano.function([batch_index],[adversarial_gen_loss,pixel_loss,1-error_rate],
updates=gen_updates,
givens = {X: X_shared[batch_slice], y: y_shared[batch_slice],Z: Z_shared[batch_slice]},
on_unused_input = 'warn')
update_discrim = theano.function([batch_index],[discrim_g_loss,discrim_d_loss,discrim_accuracy,pixel_loss,1-error_rate],
updates=discrim_updates,
givens = {X: X_shared[batch_slice], y: y_shared[batch_slice],Z: Z_shared[batch_slice]},
on_unused_input = 'warn')
# outputs for test/validation function
test_outs = [x for x in [test_error_rate,
classifier_test_error_rate] if x is not None]
# Define test/validation function
test_error_fn = theano.function([batch_index],
test_outs, givens={
X: X_shared[batch_slice],
y: y_shared[batch_slice]
},on_unused_input='warn' )
# Dictionary of Theano Functions
# tfuncs = {'update_iter':update_iter,
tfuncs = {'update_gen': update_gen,
'update_discrim': update_discrim,
'test_function':test_error_fn,
'sample': sample,
'Zfn' : Zfn
}
# Dictionary of Theano Variables
tvars = {'X' : X,
'y' : y,
'Z' : Z,
'X_shared' : X_shared,
'y_shared' : y_shared,
'Z_shared' : Z_shared,
'batch_slice' : batch_slice,
'batch_index' : batch_index,
'learning_rate' : learning_rate,
'log_sigma': log_sigma_theta
}
return tfuncs, tvars, model
# Data Loading Function
#
# This function interfaces with a Fuel dataset and returns numpy arrays containing the requested data
def data_loader(cfg,set,offset=0,shuffle=False,seed=42):
# Define chunk size
chunk_size = cfg['batch_size']*cfg['batches_per_chunk']
np.random.seed(seed)
index = np.random.permutation(set.num_examples-offset) if shuffle else np.asarray(range(set.num_examples-offset))
# Open Dataset
set.open()
# Loop across all data
for i in xrange(set.num_examples//chunk_size):
yield to_tanh(np.float32(set.get_data(request = list(index[range(offset+chunk_size*i,offset+chunk_size*(i+1))]))[0]))
# Close dataset
set.close(state=None)
# Main Function
def main(args):
# Load Config Module from source file
config_module = imp.load_source('config', args.config_path)
# Get configuration parameters
cfg = config_module.cfg
# Define name of npz file to which the model parameters will be saved
weights_fname = str(args.config_path)[:-3]+'.npz'
# Define the name of the jsonl file to which the training log will be saved
metrics_fname = weights_fname[:-4]+'METRICS.jsonl'
# Prepare logs
logging.basicConfig(level=logging.INFO, format='%(asctime)s %(levelname)s| %(message)s')
logging.info('Metrics will be saved to {}'.format(metrics_fname))
mlog = voxnet.metrics_logging.MetricsLogger(metrics_fname, reinitialize=True)
model = config_module.get_model(interp=False)
logging.info('Compiling theano functions...')
# Compile functions
tfuncs, tvars,model = make_training_functions(cfg,model)
logging.info('Training...')
# Iteration Counter, indicates total number of minibatches processed
itr = 0
# Best validation accuracy variable
best_acc = 0
# Test set for interpolations
test_set = CelebA('64',('test',),sources=('features',))
# Loop across epochs
offset = True
params = list(set(lasagne.layers.get_all_params(model['l_out'],trainable=True)+[tvars['log_sigma']]+lasagne.layers.get_all_params(model['l_discrim'],trainable=True)+[x for x in lasagne.layers.get_all_params(model['l_out']) if x.name[-4:]=='mean' or x.name[-7:]=='inv_std']))
# Ratio of gen updates to discrim updates
update_ratio = cfg['update_ratio']
for epoch in xrange(cfg['max_epochs']):
offset = not offset
# Get generator for data
loader = data_loader(cfg,
CelebA('64',('train',),sources=('features',)),
offset=offset*cfg['batch_size']//2,shuffle=cfg['shuffle'],
seed=epoch) # Does this need to happen every epoch?
# Update Learning Rate, either with annealing schedule or decay rate
if isinstance(cfg['learning_rate'], dict) and epoch > 0:
if any(x==epoch for x in cfg['learning_rate'].keys()):
lr = np.float32(tvars['learning_rate'].get_value())
new_lr = cfg['learning_rate'][epoch]
logging.info('Changing learning rate from {} to {}'.format(lr, new_lr))
tvars['learning_rate'].set_value(np.float32(new_lr))
if cfg['decay_rate'] and epoch > 0:
lr = np.float32(tvars['learning_rate'].get_value())
new_lr = lr*(1-cfg['decay_rate'])
logging.info('Changing learning rate from {} to {}'.format(lr, new_lr))
tvars['learning_rate'].set_value(np.float32(new_lr))
# Number of Chunks
iter_counter = 0
# Epoch-Wise Metrics
# vloss_e, floss_e, closs_e, a_g_loss_e, a_d_loss_e, d_kl_e, c_acc_e, acc_e = 0, 0, 0, 0, 0, 0, 0, 0
# Loop across all chunks
for x_shared in loader:
# Increment Chunk Counter
iter_counter+=1
# Figure out number of batches
num_batches = len(x_shared)//cfg['batch_size']
# Shuffle chunk
# np.random.seed(42*epoch)
index = np.random.permutation(len(x_shared))
# Load data onto GPU
tvars['X_shared'].set_value(x_shared[index], borrow=True)
tvars['Z_shared'].set_value(np.float32(np.random.randn(len(x_shared),cfg['num_latents'])),borrow=True)
# Chunk Metrics
# voxel_lvs,feature_lvs,class_lvs,a_g_lvs,a_d_lvs, kl_divs,class_accs,accs = [],[],[],[],[],[],[],[]
a_g_lvs,a_dg_lvs,a_dd_lvs,discrim_acc,pixel_lvs,pixel_accs = [],[],[],[],[],[]
# Loop across all batches in chunk
for bi in xrange(num_batches):
if itr % (update_ratio+1)==0:
[a_gen_loss,pixel_loss,pixel_acc] = tfuncs['update_gen'](bi)
a_g_lvs.append(a_gen_loss)
pixel_lvs.append(pixel_loss)
pixel_accs.append(pixel_acc)
else:
[a_discrim_gloss,a_discrim_dloss,discrim_accuracy,pixel_loss,pixel_acc] = tfuncs['update_discrim'](bi)
a_dg_lvs.append(a_discrim_gloss)
a_dd_lvs.append(a_discrim_dloss)
discrim_acc.append(discrim_accuracy)
pixel_lvs.append(pixel_loss)
pixel_accs.append(pixel_acc)
itr += 1
# if not a_dg_lvs:
# a_dg_lvs,a_dd_lvs,discrim_acc = 0,0,0
# if not a_g_lvs:
# a_g_lvs = 0
# Train!
# results = tfuncs['update_iter'](bi)
# Assign results
# TODO: Clean up the assignment so that the variable things are just on the end of the assignment and
# this can be done in one or two lines
# voxel_loss = results[0] if cfg['reconstruct'] or cfg['introspect'] else 0
# feature_loss = results[(cfg['reconstruct'] or cfg['introspect'])] if cfg['introspect'] else 0
# classifier_loss = results[cfg['introspect']+(cfg['reconstruct'] or cfg['introspect'])] if cfg['discriminative'] else 0
# a_gen_loss = results[cfg['introspect']+cfg['discriminative']+(cfg['reconstruct'] or cfg['introspect'])] if cfg['adversarial'] else 0
# a_discrim_loss = results[1+cfg['introspect']+cfg['discriminative']+(cfg['reconstruct'] or cfg['introspect'])] if cfg['adversarial'] else 0
# kl_div = results[cfg['introspect']+cfg['discriminative']+2*cfg['adversarial']+(cfg['reconstruct'] or cfg['introspect'])] if cfg['kl_div'] else 0
# class_acc = results[cfg['introspect']+cfg['discriminative']+2*cfg['adversarial']+cfg['kl_div']+(cfg['reconstruct'] or cfg['introspect'])] if cfg['discriminative'] else 0
# acc = results[cfg['introspect']+2*cfg['discriminative']+2*cfg['adversarial']+cfg['kl_div']+(cfg['reconstruct'] or cfg['introspect'])] if cfg['reconstruct'] or cfg['introspect'] else 0
# voxel_lvs.append(voxel_loss)
# feature_lvs.append(feature_loss)
# class_lvs.append(classifier_loss)
# kl_divs.append(kl_div)
# class_accs.append(class_acc)
# accs.append(acc)
[agloss,adgloss,addloss,accuracy] = [float(np.mean(a_g_lvs)),float(np.mean(a_dg_lvs)),float(np.mean(a_dd_lvs)),float(np.mean(discrim_acc))]
[ploss,pixel_accuracy] = [float(np.mean(pixel_lvs)),float(np.mean(pixel_accs))]
# Chunk-wise metrics
# [vloss, floss,closs, agloss, adloss, d_kl,c_acc,acc] = [float(np.mean(voxel_lvs)), float(np.mean(feature_lvs)),
# float(np.mean(class_lvs)), float(np.mean(a_g_lvs)),float(np.mean(a_d_lvs)),float(np.mean(kl_divs)),
# 1.0-float(np.mean(class_accs)), 1.0-float(np.mean(accs))]
# Epoch-wise metrics
# vloss_e, floss_e, closs_e, a_g_loss_e, a_d_loss_e, d_kl_e, c_acc_e, acc_e = [vloss_e+vloss, floss_e+floss, closs_e+closs, a_g_loss_e+agloss, a_d_loss_e+adloss, d_kl_e+d_kl, c_acc_e+c_acc, acc_e+acc]
# Report Chunk Metrics
# logging.info('epoch: {:4d}, itr: {:8d}, p_loss: {:8.5f}, f_loss: {:8.5f}, a_g_loss: {:8.5f}, a_d_loss: {:8.5f}, D_kl: {:8.5f}, acc: {:6.5f}'.format(epoch, itr, vloss, floss,
# agloss, adloss, d_kl, acc))
logging.info('epoch: {:4d}, itr: {:8d}, ag_loss: {:7.4f}, adg_loss: {:7.4f}, add_loss: {:7.4f}, acc: {:5.3f}, ploss: {:7.4f}, pacc: {:5.3f}'.format(epoch,itr,agloss,adgloss,addloss,accuracy,ploss,pixel_accuracy))
mlog.log(epoch=epoch, itr=itr, agloss=agloss,adgloss = adgloss,addloss=addloss,discrim_accuracy=accuracy,ploss=ploss,pixel_accuracy=pixel_accuracy)
# Log Chunk Metrics
# mlog.log(epoch=epoch, itr=itr, vloss=vloss,floss=floss, agloss=agloss,adloss = adloss, acc=acc,d_kl=d_kl,c_acc=c_acc)
# Average Epoch-wise Metrics
# vloss_e, floss_e, closs_e, a_g_loss_e, a_d_loss_e, d_kl_e, c_acc_e, acc_e = [vloss_e/iter_counter, floss_e/iter_counter,
# closs_e/iter_counter, a_g_loss_e/iter_counter, a_d_loss_e/iter_counter, d_kl_e/iter_counter,
# c_acc_e/iter_counter, acc_e/iter_counter]
# Report Epoch-wise metrics
# logging.info('Training metrics, Epoch {}, p_loss: {}, f_loss: {}, a_g_loss: {}, a_d_loss: {}, c_loss: {}, D_kl: {}, class_acc: {}, acc: {}'.format(epoch, vloss_e, floss_e,a_g_loss_e, a_d_loss_e, closs_e,d_kl_e,c_acc_e,acc_e))
# Log Epoch-wise metrics
# mlog.log(epoch=epoch, vloss_e=vloss_e, floss_e=floss_e, a_g_loss_e = a_g_loss_e,a_d_loss_e=a_d_loss_e,closs_e=closs_e, d_kl_e=d_kl_e, c_acc_e=c_acc_e, acc_e=acc_e)
if cfg['reconstruct'] or cfg['introspect']:
logging.info('Examining performance on validation set')
# Validation Metrics
test_error,test_class_error = [],[],
# Prepare Test Loader
for o in xrange(2):
test_loader = data_loader(cfg,CelebA('64',('valid',),sources=('features',)),offset=o*cfg['batch_size']//2)
# Loop Across Chunks
for x_shared in test_loader:
# Figure Out Number of Batches
num_batches = len(x_shared)//cfg['batch_size']
# Load chunk onto GPU
tvars['X_shared'].set_value(x_shared, borrow=True)
# Loop Across Batches
for bi in xrange(num_batches):
# Test!
test_results = tfuncs['test_function'](bi) # Get the test
# Assign results
batch_test_error=test_results[0]
batch_test_class_error = test_results[1] if cfg['discriminative'] else 0
test_error.append(batch_test_error)
test_class_error.append(batch_test_class_error)
# Average Results
t_error = 1-float(np.mean(test_error))
t_class_error = 1-float(np.mean(test_class_error))
# Report Validation Results
logging.info('Epoch {} Test Accuracy: {}, Classification Test Accuracy: {}, Best_acc: {} '.format(epoch, t_error,t_class_error,best_acc))
# Log Validation Results
mlog.log(test_error=t_error,t_class_error = t_class_error)
# If we see improvement, save weights and produce output images
# if cfg['reconstruct'] or cfg['introspect']:
if not (epoch%cfg['checkpoint_every_nth']):
# Update Best-yet accuracy
# if t_error > best_acc:
# best_acc = t_error
# Save Weights
# Open Test Set
test_set.open()
np.random.seed(epoch*42+5)
# Generate Random Samples
samples = np.uint8(from_tanh(tfuncs['sample'](np.random.randn(27,cfg['num_latents']).astype(np.float32))))
np.random.seed(epoch*42+5)
# Get Reconstruction/Interpolation Endpoints
endpoints = np.uint8(test_set.get_data(request = list(np.random.choice(test_set.num_examples,6,replace=False)))[0])
# Get reconstruction latents
Ze = np.asarray(tfuncs['Zfn'](to_tanh(np.float32(endpoints))))
print(np.shape(Ze))
# Get Interpolant Latents
Z = np.asarray([Ze[2 * i, :] * (1 - j) + Ze[2 * i + 1, :] * j for i in range(3) for j in [x/6.0 for x in range(7)]],dtype=np.float32)
# Get all images
images = np.append(samples,np.concatenate([np.insert(endpoints[2*i:2*(i+1),:,:,:],1,np.uint8(from_tanh(tfuncs['sample'](Z[7*i:7*(i+1),:]))),axis=0) for i in range(3)],axis=0),axis=0)
# Plot images
plot_image_grid(images,6,9,'pics/'+str(args.config_path)[:-3]+'_'+str(epoch)+'.png')
# Close test set
test_set.close(state=None)
params = list(set(lasagne.layers.get_all_params(model['l_out'],trainable=True)+[tvars['log_sigma']]+lasagne.layers.get_all_params(model['l_discrim'],trainable=True)+[x for x in lasagne.layers.get_all_params(model['l_out'])+lasagne.layers.get_all_params(model['l_discrim']) if x.name[-4:]=='mean' or x.name[-7:]=='inv_std']))
GANcheckpoints.save_weights(weights_fname, params,{'itr': itr, 'ts': time.time(),'learning_rate':np.float32(tvars['learning_rate'].get_value())})
# Save weights
# CAcheckpoints.save_weights(weights_fname, model['l_out'],tvars['log_sigma'],
# {'itr': itr, 'ts': time.time(),'best_acc':best_acc,'learning_rate':np.float32(tvars['learning_rate'].get_value())})
# elif not (epoch%cfg['checkpoint_every_nth']):
# logging.info('Checkpoint: Saving weights and generating samples')
# np.random.seed(epoch*42+5)
# samples = np.uint8(from_tanh(tfuncs['sample'](np.random.randn(54,cfg['num_latents']).astype(np.float32))))
# Plot images
# plot_image_grid(samples,6,9,'pics/'+str(args.config_path)[:-3]+'_'+str(epoch)+'.png')
# params = list(set(lasagne.layers.get_all_params(model['l_out'],trainable=True)+[tvars['log_sigma']]+lasagne.layers.get_all_params(model['l_discrim'],trainable=True)+[x for x in lasagne.layers.get_all_params(model['l_out'])+lasagne.layers.get_all_params(model['l_discrim']) if x.name[-4:]=='mean' or x.name[-7:]=='inv_std']))
# GANcheckpoints.save_weights(weights_fname, params,{'itr': itr, 'ts': time.time(),'learning_rate':np.float32(tvars['learning_rate'].get_value())})
logging.info('training done')
if __name__=='__main__':
parser = argparse.ArgumentParser()
parser.add_argument('config_path', type=Path, help='config .py file')
args = parser.parse_args()
main(args)