Aufgrund einer Störung des s3 Storage, könnten in nächster Zeit folgende GitLab Funktionen nicht zur Verfügung stehen: LFS, Container Registry, Job Artifacs, Uploads (Wiki, Bilder, Projekt-Exporte). Wir bitten um Verständnis. Es wird mit Hochdruck an der Behebung des Problems gearbeitet. Weitere Informationen zur Störung des Object Storage finden Sie hier: https://maintenance.itc.rwth-aachen.de/ticket/status/messages/59-object-storage-pilot

CNNSupervisedTrainer_RNNtest.py 27.2 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
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 []
249
        loss_axis = loss_params['loss_axis'] if 'loss_axis' in loss_params else -1
250
        batch_axis = loss_params['batch_axis'] if 'batch_axis' in loss_params else 0
251 252
        if loss == 'softmax_cross_entropy':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
253
            loss_function = mx.gluon.loss.SoftmaxCrossEntropyLoss(axis=loss_axis, from_logits=fromLogits, sparse_label=sparseLabel, batch_axis=batch_axis)
254 255
        elif loss == 'softmax_cross_entropy_ignore_indices':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
256
            loss_function = SoftmaxCrossEntropyLossIgnoreIndices(ignore_indices=ignore_indices, from_logits=fromLogits, sparse_label=sparseLabel, batch_axis=batch_axis)
257 258 259
        elif loss == 'sigmoid_binary_cross_entropy':
            loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
        elif loss == 'cross_entropy':
260
            loss_function = CrossEntropyLoss(axis=loss_axis, sparse_label=sparseLabel, batch_axis=batch_axis)
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
        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):
Julian Treiber's avatar
Julian Treiber committed
463
                if True:
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
                    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)