CNNSupervisedTrainer_RNNtest.py 26.9 KB
Newer Older
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
import mxnet as mx
import logging
import numpy as np
import time
import os
import shutil
import pickle
import math
import sys
from mxnet import gluon, autograd, nd

class CrossEntropyLoss(gluon.loss.Loss):
    def __init__(self, axis=-1, sparse_label=True, weight=None, batch_axis=0, **kwargs):
        super(CrossEntropyLoss, self).__init__(weight, batch_axis, **kwargs)
        self._axis = axis
        self._sparse_label = sparse_label

    def hybrid_forward(self, F, pred, label, sample_weight=None):
        pred = F.log(pred)
        if self._sparse_label:
            loss = -F.pick(pred, label, axis=self._axis, keepdims=True)
        else:
            label = gluon.loss._reshape_like(F, label, pred)
            loss = -F.sum(pred * label, axis=self._axis, keepdims=True)
        loss = gluon.loss._apply_weighting(F, loss, self._weight, sample_weight)
        return F.mean(loss, axis=self._batch_axis, exclude=True)

class LogCoshLoss(gluon.loss.Loss):
    def __init__(self, weight=None, batch_axis=0, **kwargs):
        super(LogCoshLoss, self).__init__(weight, batch_axis, **kwargs)

    def hybrid_forward(self, F, pred, label, sample_weight=None):
        loss = F.log(F.cosh(pred - label))
        loss = gluon.loss._apply_weighting(F, loss, self._weight, sample_weight)
        return F.mean(loss, axis=self._batch_axis, exclude=True)

class SoftmaxCrossEntropyLossIgnoreIndices(gluon.loss.Loss):
    def __init__(self, axis=-1, ignore_indices=[], sparse_label=True, from_logits=False, weight=None, batch_axis=0, **kwargs):
        super(SoftmaxCrossEntropyLossIgnoreIndices, self).__init__(weight, batch_axis, **kwargs)
        self._axis = axis
        self._ignore_indices = ignore_indices
        self._sparse_label = sparse_label
        self._from_logits = from_logits

    def hybrid_forward(self, F, pred, label, sample_weight=None):
        log_softmax = F.log_softmax
        pick = F.pick
        if not self._from_logits:
            pred = log_softmax(pred, self._axis)
        if self._sparse_label:
            loss = -pick(pred, label, axis=self._axis, keepdims=True)
        else:
            label = _reshape_like(F, label, pred)
            loss = -(pred * label).sum(axis=self._axis, keepdims=True)
        # ignore some indices for loss, e.g. <pad> tokens in NLP applications
        for i in self._ignore_indices:
            loss = loss * mx.nd.logical_not(mx.nd.equal(mx.nd.argmax(pred, axis=1), mx.nd.ones_like(mx.nd.argmax(pred, axis=1))*i) * mx.nd.equal(mx.nd.argmax(pred, axis=1), label))
        return loss.mean(axis=self._batch_axis, exclude=True)

@mx.metric.register
class BLEU(mx.metric.EvalMetric):
    N = 4

    def __init__(self, exclude=None, name='bleu', output_names=None, label_names=None):
        super(BLEU, self).__init__(name=name, output_names=output_names, label_names=label_names)

        self._exclude = exclude or []

        self._match_counts = [0 for _ in range(self.N)]
        self._counts = [0 for _ in range(self.N)]

        self._size_ref = 0
        self._size_hyp = 0

    def update(self, labels, preds):
        labels, preds = mx.metric.check_label_shapes(labels, preds, True)

        new_labels = self._convert(labels)
        new_preds = self._convert(preds)

        for label, pred in zip(new_labels, new_preds):
            reference = [word for word in label if word not in self._exclude]
            hypothesis = [word for word in pred if word not in self._exclude]

            self._size_ref += len(reference)
            self._size_hyp += len(hypothesis)

            for n in range(self.N):
                reference_ngrams = self._get_ngrams(reference, n + 1)
                hypothesis_ngrams = self._get_ngrams(hypothesis, n + 1)

                match_count = 0

                for ngram in hypothesis_ngrams:
                    if ngram in reference_ngrams:
                        reference_ngrams.remove(ngram)

                        match_count += 1

                self._match_counts[n] += match_count
                self._counts[n] += len(hypothesis_ngrams)

    def get(self):
        precisions = [sys.float_info.min for n in range(self.N)]

        i = 1

        for n in range(self.N):
            match_counts = self._match_counts[n]
            counts = self._counts[n]

            if counts != 0:
                if match_counts == 0:
                    i *= 2
                    match_counts = 1 / i

                if (match_counts / counts) > 0:
                    precisions[n] = match_counts / counts

        bleu = self._get_brevity_penalty() * math.exp(sum(map(math.log, precisions)) / self.N)

        return (self.name, bleu)

    def calculate(self):
        precisions = [sys.float_info.min for n in range(self.N)]

        i = 1

        for n in range(self.N):
            match_counts = self._match_counts[n]
            counts = self._counts[n]

            if counts != 0:
                if match_counts == 0:
                    i *= 2
                    match_counts = 1 / i

                precisions[n] = match_counts / counts

        return self._get_brevity_penalty() * math.exp(sum(map(math.log, precisions)) / self.N)

    def _get_brevity_penalty(self):
        if self._size_hyp >= self._size_ref:
            return 1
        else:
            return math.exp(1 - (self._size_ref / self._size_hyp))

    @staticmethod
    def _get_ngrams(sentence, n):
        ngrams = []

        if len(sentence) >= n:
            for i in range(len(sentence) - n + 1):
                ngrams.append(sentence[i:i+n])

        return ngrams

    @staticmethod
    def _convert(nd_list):
        if len(nd_list) == 0:
            return []

        new_list = [[] for _ in range(nd_list[0].shape[0])]

        for element in nd_list:
            for i in range(element.shape[0]):
                new_list[i].append(element[i].asscalar())

        return new_list



class CNNSupervisedTrainer_RNNtest:
    def __init__(self, data_loader, net_constructor):
        self._data_loader = data_loader
        self._net_creator = net_constructor
        self._networks = {}

    def train(self, batch_size=64,
              num_epoch=10,
              eval_metric='acc',
              eval_metric_params={},
              eval_train=False,
              loss ='softmax_cross_entropy',
              loss_params={},
              optimizer='adam',
              optimizer_params=(('learning_rate', 0.001),),
              load_checkpoint=True,
              checkpoint_period=5,
              log_period=50,
              context='gpu',
              save_attention_image=False,
              use_teacher_forcing=False,
              normalize=True,
              preprocessing = False):
        if context == 'gpu':
            mx_context = mx.gpu()
        elif context == 'cpu':
            mx_context = mx.cpu()
        else:
            logging.error("Context argument is '" + context + "'. Only 'cpu' and 'gpu are valid arguments'.")

        if preprocessing:
            preproc_lib = "CNNPreprocessor_RNNtest_executor"
            train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_preprocessed_data(batch_size, preproc_lib)
        else:
            train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_data(batch_size)

        if 'weight_decay' in optimizer_params:
            optimizer_params['wd'] = optimizer_params['weight_decay']
            del optimizer_params['weight_decay']
        if 'learning_rate_decay' in optimizer_params:
            min_learning_rate = 1e-08
            if 'learning_rate_minimum' in optimizer_params:
                min_learning_rate = optimizer_params['learning_rate_minimum']
                del optimizer_params['learning_rate_minimum']
            optimizer_params['lr_scheduler'] = mx.lr_scheduler.FactorScheduler(
                                                   optimizer_params['step_size'],
                                                   factor=optimizer_params['learning_rate_decay'],
                                                   stop_factor_lr=min_learning_rate)
            del optimizer_params['step_size']
            del optimizer_params['learning_rate_decay']

        if normalize:
            self._net_creator.construct(context=mx_context, data_mean=data_mean, data_std=data_std)
        else:
            self._net_creator.construct(context=mx_context)

        begin_epoch = 0
        if load_checkpoint:
            begin_epoch = self._net_creator.load(mx_context)
        else:
            if os.path.isdir(self._net_creator._model_dir_):
                shutil.rmtree(self._net_creator._model_dir_)

        self._networks = self._net_creator.networks

        try:
            os.makedirs(self._net_creator._model_dir_)
        except OSError:
            if not os.path.isdir(self._net_creator._model_dir_):
                raise

        trainers = [mx.gluon.Trainer(network.collect_params(), optimizer, optimizer_params) for network in self._networks.values() if len(network.collect_params().values()) != 0]

        margin = loss_params['margin'] if 'margin' in loss_params else 1.0
        sparseLabel = loss_params['sparse_label'] if 'sparse_label' in loss_params else True
        ignore_indices = [loss_params['ignore_indices']] if 'ignore_indices' in loss_params else []
        if loss == 'softmax_cross_entropy':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
            loss_function = mx.gluon.loss.SoftmaxCrossEntropyLoss(from_logits=fromLogits, sparse_label=sparseLabel)
        elif loss == 'softmax_cross_entropy_ignore_indices':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
            loss_function = SoftmaxCrossEntropyLossIgnoreIndices(ignore_indices=ignore_indices, from_logits=fromLogits, sparse_label=sparseLabel)
        elif loss == 'sigmoid_binary_cross_entropy':
            loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
        elif loss == 'cross_entropy':
            loss_function = CrossEntropyLoss(sparse_label=sparseLabel)
        elif loss == 'l2':
            loss_function = mx.gluon.loss.L2Loss()
        elif loss == 'l1':
            loss_function = mx.gluon.loss.L2Loss()
        elif loss == 'huber':
            rho = loss_params['rho'] if 'rho' in loss_params else 1
            loss_function = mx.gluon.loss.HuberLoss(rho=rho)
        elif loss == 'hinge':
            loss_function = mx.gluon.loss.HingeLoss(margin=margin)
        elif loss == 'squared_hinge':
            loss_function = mx.gluon.loss.SquaredHingeLoss(margin=margin)
        elif loss == 'logistic':
            labelFormat = loss_params['label_format'] if 'label_format' in loss_params else 'signed'
            loss_function = mx.gluon.loss.LogisticLoss(label_format=labelFormat)
        elif loss == 'kullback_leibler':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else True
            loss_function = mx.gluon.loss.KLDivLoss(from_logits=fromLogits)
        elif loss == 'log_cosh':
            loss_function = LogCoshLoss()
        else:
            logging.error("Invalid loss parameter.")

        tic = None

        for epoch in range(begin_epoch, begin_epoch + num_epoch):

            loss_total = 0
            train_iter.reset()
            for batch_i, batch in enumerate(train_iter):
                with autograd.record():
                    labels = [batch.label[i].as_in_context(mx_context) for i in range(5)]

                    source_ = batch.data[0].as_in_context(mx_context)

                    target_ = [mx.nd.zeros((batch_size, 30000,), ctx=mx_context) for i in range(5)]



                    nd.waitall()

                    lossList = []

                    target_[0] = self._networks[0](source_)

                    lossList.append(loss_function(target_[0], labels[0]))
                    for i in range(1, 5):
                        target_[i-1+1] = self._networks[1](target_[i-1+0])

                        lossList.append(loss_function(target_[i-1+1], labels[i-1+1]))
                        if use_teacher_forcing == "True":
                            target_[i-1+1] = mx.nd.expand_dims(labels[i-1+1], axis=1)

                    loss = 0
                    for element in lossList:
                        loss = loss + element

                loss.backward()

                loss_total += loss.sum().asscalar()

                for trainer in trainers:
                    trainer.step(batch_size)

                if tic is None:
                    tic = time.time()
                else:
                    if batch_i % log_period == 0:
                        try:
                            speed = log_period * batch_size / (time.time() - tic)
                        except ZeroDivisionError:
                            speed = float("inf")

                        loss_avg = loss_total / (batch_size * log_period)
                        loss_total = 0

                        logging.info("Epoch[%d] Batch[%d] Speed: %.2f samples/sec Loss: %.5f" % (epoch, batch_i, speed, loss_avg))

                        tic = time.time()

            tic = None


            if eval_train:
                train_iter.reset()
                metric = mx.metric.create(eval_metric, **eval_metric_params)
                for batch_i, batch in enumerate(train_iter):
                    labels = [batch.label[i].as_in_context(mx_context) for i in range(5)]

                    source_ = batch.data[0].as_in_context(mx_context)

                    target_ = [mx.nd.zeros((batch_size, 30000,), ctx=mx_context) for i in range(5)]



                    nd.waitall()

                    outputs = []
                    attentionList=[]
                    target_[0] = self._networks[0](source_)

                    outputs.append(target_[0])
                    k = 1
                    sequences = [([target_[1-1+0]], mx.nd.full((batch_size, 1,), 1.0, ctx=mx_context), [mx.nd.full((batch_size, 64,), 0.0, ctx=mx_context)])]

                    for i in range(1, 5):
                        all_candidates = []

                        for seq, score, attention in sequences:
                            target_[i-1+0] = seq[-1]
                            target_[i-1+1] = self._networks[1](target_[i-1+0])
                            out = target_[i-1+1]

                            topk = out.topk(k=k)

                            for top_index in range(len(topk[0])):
                                j = mx.nd.slice_axis(topk, axis=1, begin=top_index, end=top_index+1)
                                currentScore = mx.nd.slice_axis(out, axis=1, begin=top_index, end=top_index+1)
                                newScore = mx.nd.expand_dims(score.squeeze() * currentScore.squeeze(), axis=1)
                                candidate = (seq + [j],  newScore, attention + [])
                                all_candidates.append(candidate)

                        ordered = []
                        newSequences = []
                        for batch_entry in range(batch_size):
                            ordered.append([])
                            batchCandidate = [([seq[batch_entry] for seq in candidate[0]], candidate[1][batch_entry], [attention[batch_entry].expand_dims(axis=0) for attention in candidate[2]]) for candidate in all_candidates]
                            ordered[batch_entry] = sorted(batchCandidate, key=lambda tup: tup[1].asscalar())
                            if batch_entry == 0:
                                newSequences = ordered[batch_entry]
                            else:
                                newSequences = [([mx.nd.concat(newSequences[sequenceIndex][0][seqIndex], ordered[batch_entry][sequenceIndex][0][seqIndex], dim=0) for seqIndex in range(len(newSequences[sequenceIndex][0]))],
                                    mx.nd.concat(newSequences[sequenceIndex][1], ordered[batch_entry][sequenceIndex][1], dim=0),
                                    [mx.nd.concat(newSequences[sequenceIndex][2][attentionIndex], ordered[batch_entry][sequenceIndex][2][attentionIndex], dim=0) for attentionIndex in range(len(newSequences[sequenceIndex][2]))])
                                    for sequenceIndex in range(len(newSequences))]

                        newSequences = [([newSequences[sequenceIndex][0][seqIndex].expand_dims(axis=1) for seqIndex in range(len(newSequences[sequenceIndex][0]))],
                            newSequences[sequenceIndex][1].expand_dims(axis=1), [newSequences[sequenceIndex][2][attentionIndex] for attentionIndex in range(len(newSequences[sequenceIndex][2]))])
                            for sequenceIndex in range(len(newSequences))]

                        sequences = newSequences[:][:k]

                    for i in range(1, 5):
                        target_[i-1+1] = sequences[0][0][i]
                        outputs.append(target_[i-1+1])


                    if save_attention_image == "True":
                        import matplotlib
                        matplotlib.use('Agg')
                        import matplotlib.pyplot as plt
                        logging.getLogger('matplotlib').setLevel(logging.ERROR)

                        if(os.path.isfile('src/test/resources/training_data/Show_attend_tell/dict.pkl')):
                            with open('src/test/resources/training_data/Show_attend_tell/dict.pkl', 'rb') as f:
                                dict = pickle.load(f)

                        plt.clf()
                        fig = plt.figure(figsize=(15,15))
                        max_length = len(labels)-1

                        ax = fig.add_subplot(max_length//3, max_length//4, 1)
                        ax.imshow(train_images[0+batch_size*(batch_i)].transpose(1,2,0))

                        for l in range(max_length):
                            attention = attentionList[l]
                            attention = mx.nd.slice_axis(attention, axis=0, begin=0, end=1).squeeze()
                            attention_resized = np.resize(attention.asnumpy(), (8, 8))
                            ax = fig.add_subplot(max_length//3, max_length//4, l+2)
                            if int(labels[l+1][0].asscalar()) > len(dict):
                                ax.set_title("<unk>")
                            elif dict[int(labels[l+1][0].asscalar())] == "<end>":
                                ax.set_title(".")
                                img = ax.imshow(train_images[0+batch_size*(batch_i)].transpose(1,2,0))
                                ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())
                                break
                            else:
                                ax.set_title(dict[int(labels[l+1][0].asscalar())])
                            img = ax.imshow(train_images[0+batch_size*(batch_i)].transpose(1,2,0))
                            ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())

                        plt.tight_layout()
                        target_dir = 'target/attention_images'
                        if not os.path.exists(target_dir):
                            os.makedirs(target_dir)
                        plt.savefig(target_dir + '/attention_train.png')
                        plt.close()

                    predictions = []
                    for output_name in outputs:
                        if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
                            predictions.append(mx.nd.argmax(output_name, axis=1))
                        else:
                            predictions.append(output_name)

                    metric.update(preds=predictions, labels=labels)
                train_metric_score = metric.get()[1]
            else:
                train_metric_score = 0

            test_iter.reset()
            metric = mx.metric.create(eval_metric, **eval_metric_params)
            for batch_i, batch in enumerate(test_iter):
                if True: 
                    labels = [batch.label[i].as_in_context(mx_context) for i in range(5)]

                    source_ = batch.data[0].as_in_context(mx_context)

                    target_ = [mx.nd.zeros((batch_size, 30000,), ctx=mx_context) for i in range(5)]



                    nd.waitall()

                    outputs = []
                    attentionList=[]
                    target_[0] = self._networks[0](source_)

                    outputs.append(target_[0])
                    k = 1
                    sequences = [([target_[1-1+0]], mx.nd.full((batch_size, 1,), 1.0, ctx=mx_context), [mx.nd.full((batch_size, 64,), 0.0, ctx=mx_context)])]

                    for i in range(1, 5):
                        all_candidates = []

                        for seq, score, attention in sequences:
                            target_[i-1+0] = seq[-1]
                            target_[i-1+1] = self._networks[1](target_[i-1+0])
                            out = target_[i-1+1]

                            topk = out.topk(k=k)

                            for top_index in range(len(topk[0])):
                                j = mx.nd.slice_axis(topk, axis=1, begin=top_index, end=top_index+1)
                                currentScore = mx.nd.slice_axis(out, axis=1, begin=top_index, end=top_index+1)
                                newScore = mx.nd.expand_dims(score.squeeze() * currentScore.squeeze(), axis=1)
                                candidate = (seq + [j],  newScore, attention + [])
                                all_candidates.append(candidate)

                        ordered = []
                        newSequences = []
                        for batch_entry in range(batch_size):
                            ordered.append([])
                            batchCandidate = [([seq[batch_entry] for seq in candidate[0]], candidate[1][batch_entry], [attention[batch_entry].expand_dims(axis=0) for attention in candidate[2]]) for candidate in all_candidates]
                            ordered[batch_entry] = sorted(batchCandidate, key=lambda tup: tup[1].asscalar())
                            if batch_entry == 0:
                                newSequences = ordered[batch_entry]
                            else:
                                newSequences = [([mx.nd.concat(newSequences[sequenceIndex][0][seqIndex], ordered[batch_entry][sequenceIndex][0][seqIndex], dim=0) for seqIndex in range(len(newSequences[sequenceIndex][0]))],
                                    mx.nd.concat(newSequences[sequenceIndex][1], ordered[batch_entry][sequenceIndex][1], dim=0),
                                    [mx.nd.concat(newSequences[sequenceIndex][2][attentionIndex], ordered[batch_entry][sequenceIndex][2][attentionIndex], dim=0) for attentionIndex in range(len(newSequences[sequenceIndex][2]))])
                                    for sequenceIndex in range(len(newSequences))]

                        newSequences = [([newSequences[sequenceIndex][0][seqIndex].expand_dims(axis=1) for seqIndex in range(len(newSequences[sequenceIndex][0]))],
                            newSequences[sequenceIndex][1].expand_dims(axis=1), [newSequences[sequenceIndex][2][attentionIndex] for attentionIndex in range(len(newSequences[sequenceIndex][2]))])
                            for sequenceIndex in range(len(newSequences))]

                        sequences = newSequences[:][:k]

                    for i in range(1, 5):
                        target_[i-1+1] = sequences[0][0][i]
                        outputs.append(target_[i-1+1])


                    if save_attention_image == "True":
                        if not eval_train:
                            import matplotlib
                            matplotlib.use('Agg')
                            import matplotlib.pyplot as plt
                            logging.getLogger('matplotlib').setLevel(logging.ERROR)

                            if(os.path.isfile('src/test/resources/training_data/Show_attend_tell/dict.pkl')):
                                with open('src/test/resources/training_data/Show_attend_tell/dict.pkl', 'rb') as f:
                                    dict = pickle.load(f)

                        plt.clf()
                        fig = plt.figure(figsize=(15,15))
                        max_length = len(labels)-1

                        ax = fig.add_subplot(max_length//3, max_length//4, 1)
                        ax.imshow(test_images[0+batch_size*(batch_i)].transpose(1,2,0))

                        for l in range(max_length):
                            attention = attentionList[l]
                            attention = mx.nd.slice_axis(attention, axis=0, begin=0, end=1).squeeze()
                            attention_resized = np.resize(attention.asnumpy(), (8, 8))
                            ax = fig.add_subplot(max_length//3, max_length//4, l+2)
                            if int(mx.nd.slice_axis(outputs[l+1], axis=0, begin=0, end=1).squeeze().asscalar()) > len(dict):
                                ax.set_title("<unk>")
                            elif dict[int(mx.nd.slice_axis(outputs[l+1], axis=0, begin=0, end=1).squeeze().asscalar())] == "<end>":
                                ax.set_title(".")
                                img = ax.imshow(test_images[0+batch_size*(batch_i)].transpose(1,2,0))
                                ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())
                                break
                            else:
                                ax.set_title(dict[int(mx.nd.slice_axis(outputs[l+1], axis=0, begin=0, end=1).squeeze().asscalar())])
                            img = ax.imshow(test_images[0+batch_size*(batch_i)].transpose(1,2,0))
                            ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())

                        plt.tight_layout()
                        target_dir = 'target/attention_images'
                        if not os.path.exists(target_dir):
                            os.makedirs(target_dir)
                        plt.savefig(target_dir + '/attention_test.png')
                        plt.close()

                predictions = []
                for output_name in outputs:
                    if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
                        predictions.append(mx.nd.argmax(output_name, axis=1))
                    #ArgMax already applied
                    else:
                        predictions.append(output_name)

                metric.update(preds=predictions, labels=labels)
            test_metric_score = metric.get()[1]

            logging.info("Epoch[%d] Train: %f, Test: %f" % (epoch, train_metric_score, test_metric_score))


            if (epoch - begin_epoch) % checkpoint_period == 0:
                for i, network in self._networks.items():
                    network.save_parameters(self.parameter_path(i) + '-' + str(epoch).zfill(4) + '.params')

        for i, network in self._networks.items():
            network.save_parameters(self.parameter_path(i) + '-' + str(num_epoch + begin_epoch).zfill(4) + '.params')
            network.export(self.parameter_path(i) + '_newest', epoch=0)

    def parameter_path(self, index):
        return self._net_creator._model_dir_ + self._net_creator._model_prefix_ + '_' + str(index)