CNNSupervisedTrainer_CifarClassifierNetwork.py 20.9 KB
Newer Older
Nicola Gatto's avatar
Nicola Gatto committed
1 2 3 4 5 6
import mxnet as mx
import logging
import numpy as np
import time
import os
import shutil
Christian Fuß's avatar
Christian Fuß committed
7
import pickle
Sebastian N.'s avatar
Sebastian N. committed
8 9
import math
import sys
Nicola Gatto's avatar
Nicola Gatto committed
10 11
from mxnet import gluon, autograd, nd

Eyüp Harputlu's avatar
Eyüp Harputlu committed
12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
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)

Eyüp Harputlu's avatar
Eyüp Harputlu committed
28 29 30 31 32 33 34 35 36
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)

Sebastian N.'s avatar
Sebastian N. committed
37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56
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:
57
            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))
Sebastian N.'s avatar
Sebastian N. committed
58 59
        return loss.mean(axis=self._batch_axis, exclude=True)

Sebastian N.'s avatar
Sebastian N. committed
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
@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

117 118
                if (match_counts / counts) > 0:
                    precisions[n] = match_counts / counts
Sebastian N.'s avatar
Sebastian N. committed
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

        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
Christian Fuß's avatar
Christian Fuß committed
170

Sebastian N.'s avatar
Sebastian N. committed
171 172


173
class CNNSupervisedTrainer_CifarClassifierNetwork:
174
    def __init__(self, data_loader, net_constructor):
Nicola Gatto's avatar
Nicola Gatto committed
175 176
        self._data_loader = data_loader
        self._net_creator = net_constructor
177
        self._networks = {}
Nicola Gatto's avatar
Nicola Gatto committed
178 179 180 181

    def train(self, batch_size=64,
              num_epoch=10,
              eval_metric='acc',
Sebastian N.'s avatar
Sebastian N. committed
182
              eval_metric_params={},
183
              eval_train=False,
Eyüp Harputlu's avatar
Eyüp Harputlu committed
184 185
              loss ='softmax_cross_entropy',
              loss_params={},
Nicola Gatto's avatar
Nicola Gatto committed
186 187 188 189
              optimizer='adam',
              optimizer_params=(('learning_rate', 0.001),),
              load_checkpoint=True,
              checkpoint_period=5,
190 191
              log_period=50,
              context='gpu',
192
              save_attention_image=False,
193
              use_teacher_forcing=False,
194
              normalize=True,
195 196
              shuffle_data=False,
              clip_global_grad_norm=None,
197
              preprocessing = False):
Nicola Gatto's avatar
Nicola Gatto committed
198 199 200 201 202 203 204
        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'.")

205 206
        if preprocessing:
            preproc_lib = "CNNPreprocessor_CifarClassifierNetwork_executor"
207
            train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_preprocessed_data(batch_size, preproc_lib, shuffle_data)
208
        else:
209
            train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_data(batch_size, shuffle_data)
210

Nicola Gatto's avatar
Nicola Gatto committed
211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
        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']

226 227 228 229
        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)
Nicola Gatto's avatar
Nicola Gatto committed
230 231 232 233 234 235 236 237

        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_)

238
        self._networks = self._net_creator.networks
Nicola Gatto's avatar
Nicola Gatto committed
239 240 241 242 243 244 245

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

Sebastian N.'s avatar
Sebastian N. committed
246
        trainers = [mx.gluon.Trainer(network.collect_params(), optimizer, optimizer_params) for network in self._networks.values() if len(network.collect_params().values()) != 0]
Nicola Gatto's avatar
Nicola Gatto committed
247

Eyüp Harputlu's avatar
Eyüp Harputlu committed
248 249
        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
250
        ignore_indices = [loss_params['ignore_indices']] if 'ignore_indices' in loss_params else []
Eyüp Harputlu's avatar
Eyüp Harputlu committed
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(from_logits=fromLogits, sparse_label=sparseLabel)
254
        elif loss == 'softmax_cross_entropy_ignore_indices':
255
            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)
Eyüp Harputlu's avatar
Eyüp Harputlu committed
257
        elif loss == 'sigmoid_binary_cross_entropy':
Nicola Gatto's avatar
Nicola Gatto committed
258
            loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
259 260 261
        elif loss == 'cross_entropy':
            loss_function = CrossEntropyLoss(sparse_label=sparseLabel)
        elif loss == 'l2':
Nicola Gatto's avatar
Nicola Gatto committed
262
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
263
        elif loss == 'l1':
Nicola Gatto's avatar
Nicola Gatto committed
264
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
265 266 267 268 269 270 271 272 273 274 275 276 277
        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)
Eyüp Harputlu's avatar
Eyüp Harputlu committed
278 279
        elif loss == 'log_cosh':
            loss_function = LogCoshLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
280 281
        else:
            logging.error("Invalid loss parameter.")
Nicola Gatto's avatar
Nicola Gatto committed
282 283 284 285

        tic = None

        for epoch in range(begin_epoch, begin_epoch + num_epoch):
286 287

            loss_total = 0
Nicola Gatto's avatar
Nicola Gatto committed
288 289
            train_iter.reset()
            for batch_i, batch in enumerate(train_iter):
290 291
                with autograd.record():
                    labels = [batch.label[i].as_in_context(mx_context) for i in range(1)]
292

293 294
                    data_ = batch.data[0].as_in_context(mx_context)

295
                    softmax_ = mx.nd.zeros((batch_size, 10,), ctx=mx_context)
Christian Fuß's avatar
Christian Fuß committed
296

297

298 299
                    nd.waitall()

300
                    lossList = []
301

302
                    softmax_ = self._networks[0](data_)
303 304

                    lossList.append(loss_function(softmax_, labels[0]))
305

306 307 308
                    loss = 0
                    for element in lossList:
                        loss = loss + element
Nicola Gatto's avatar
Nicola Gatto committed
309 310

                loss.backward()
311

312 313
                loss_total += loss.sum().asscalar()

314 315 316 317 318 319 320 321
                if clip_global_grad_norm:
                    grads = []

                    for network in self._networks.values():
                        grads.extend([param.grad(mx_context) for param in network.collect_params().values()])

                    gluon.utils.clip_global_norm(grads, clip_global_grad_norm)

322 323
                for trainer in trainers:
                    trainer.step(batch_size)
Nicola Gatto's avatar
Nicola Gatto committed
324 325 326 327

                if tic is None:
                    tic = time.time()
                else:
328
                    if batch_i % log_period == 0:
Nicola Gatto's avatar
Nicola Gatto committed
329
                        try:
330
                            speed = log_period * batch_size / (time.time() - tic)
Nicola Gatto's avatar
Nicola Gatto committed
331 332 333
                        except ZeroDivisionError:
                            speed = float("inf")

334 335 336 337
                        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))
Nicola Gatto's avatar
Nicola Gatto committed
338 339 340 341 342

                        tic = time.time()

            tic = None

343 344 345 346 347

            if eval_train:
                train_iter.reset()
                metric = mx.metric.create(eval_metric, **eval_metric_params)
                for batch_i, batch in enumerate(train_iter):
348
                    labels = [batch.label[i].as_in_context(mx_context) for i in range(1)]
349

350
                    data_ = batch.data[0].as_in_context(mx_context)
351

352
                    softmax_ = mx.nd.zeros((batch_size, 10,), ctx=mx_context)
353

354

355 356
                    nd.waitall()

357
                    outputs = []
358
                    attentionList=[]
359
                    softmax_ = self._networks[0](data_)
360

361
                    outputs.append(softmax_)
362

363 364

                    if save_attention_image == "True":
365 366
                        import matplotlib
                        matplotlib.use('Agg')
367 368 369 370 371 372 373
                        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)

374 375 376 377
                        plt.clf()
                        fig = plt.figure(figsize=(15,15))
                        max_length = len(labels)-1

378
                        ax = fig.add_subplot(max_length//3, max_length//4, 1)
379
                        ax.imshow(train_images[0+batch_size*(batch_i)].transpose(1,2,0))
380

381 382
                        for l in range(max_length):
                            attention = attentionList[l]
383
                            attention = mx.nd.slice_axis(attention, axis=0, begin=0, end=1).squeeze()
384
                            attention_resized = np.resize(attention.asnumpy(), (8, 8))
385
                            ax = fig.add_subplot(max_length//3, max_length//4, l+2)
386 387 388
                            if int(labels[l+1][0].asscalar()) > len(dict):
                                ax.set_title("<unk>")
                            elif dict[int(labels[l+1][0].asscalar())] == "<end>":
389
                                ax.set_title(".")
390
                                img = ax.imshow(train_images[0+batch_size*(batch_i)].transpose(1,2,0))
391 392 393 394
                                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())])
395
                            img = ax.imshow(train_images[0+batch_size*(batch_i)].transpose(1,2,0))
396
                            ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())
397 398 399 400

                        plt.tight_layout()
                        target_dir = 'target/attention_images'
                        if not os.path.exists(target_dir):
401
                            os.makedirs(target_dir)
402 403 404
                        plt.savefig(target_dir + '/attention_train.png')
                        plt.close()

405 406 407 408 409 410
                    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)
411

412 413 414 415
                    metric.update(preds=predictions, labels=labels)
                train_metric_score = metric.get()[1]
            else:
                train_metric_score = 0
Nicola Gatto's avatar
Nicola Gatto committed
416 417

            test_iter.reset()
Sebastian N.'s avatar
Sebastian N. committed
418
            metric = mx.metric.create(eval_metric, **eval_metric_params)
Nicola Gatto's avatar
Nicola Gatto committed
419
            for batch_i, batch in enumerate(test_iter):
420
                if True: 
421
                    labels = [batch.label[i].as_in_context(mx_context) for i in range(1)]
422

423
                    data_ = batch.data[0].as_in_context(mx_context)
424

425
                    softmax_ = mx.nd.zeros((batch_size, 10,), ctx=mx_context)
426

427

428 429
                    nd.waitall()

430
                    outputs = []
431
                    attentionList=[]
432
                    softmax_ = self._networks[0](data_)
433

434
                    outputs.append(softmax_)
435

436 437

                    if save_attention_image == "True":
438 439 440 441 442 443 444 445 446 447
                        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)

448
                        plt.clf()
449
                        fig = plt.figure(figsize=(15,15))
450 451
                        max_length = len(labels)-1

452
                        ax = fig.add_subplot(max_length//3, max_length//4, 1)
453
                        ax.imshow(test_images[0+batch_size*(batch_i)].transpose(1,2,0))
454

455 456
                        for l in range(max_length):
                            attention = attentionList[l]
457
                            attention = mx.nd.slice_axis(attention, axis=0, begin=0, end=1).squeeze()
458
                            attention_resized = np.resize(attention.asnumpy(), (8, 8))
459
                            ax = fig.add_subplot(max_length//3, max_length//4, l+2)
460 461 462
                            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>":
463
                                ax.set_title(".")
464
                                img = ax.imshow(test_images[0+batch_size*(batch_i)].transpose(1,2,0))
465 466 467
                                ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())
                                break
                            else:
468
                                ax.set_title(dict[int(mx.nd.slice_axis(outputs[l+1], axis=0, begin=0, end=1).squeeze().asscalar())])
469
                            img = ax.imshow(test_images[0+batch_size*(batch_i)].transpose(1,2,0))
470
                            ax.imshow(attention_resized, cmap='gray', alpha=0.6, extent=img.get_extent())
471 472

                        plt.tight_layout()
473 474 475
                        target_dir = 'target/attention_images'
                        if not os.path.exists(target_dir):
                            os.makedirs(target_dir)
476 477 478
                        plt.savefig(target_dir + '/attention_test.png')
                        plt.close()

479
                predictions = []
480
                for output_name in outputs:
Sebastian N.'s avatar
Sebastian N. committed
481
                    if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
482 483 484 485
                        predictions.append(mx.nd.argmax(output_name, axis=1))
                    #ArgMax already applied
                    else:
                        predictions.append(output_name)
486

487
                metric.update(preds=predictions, labels=labels)
Nicola Gatto's avatar
Nicola Gatto committed
488 489 490 491
            test_metric_score = metric.get()[1]

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

492

Nicola Gatto's avatar
Nicola Gatto committed
493
            if (epoch - begin_epoch) % checkpoint_period == 0:
494 495
                for i, network in self._networks.items():
                    network.save_parameters(self.parameter_path(i) + '-' + str(epoch).zfill(4) + '.params')
Nicola Gatto's avatar
Nicola Gatto committed
496

497 498 499
        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)
Nicola Gatto's avatar
Nicola Gatto committed
500

501
    def parameter_path(self, index):
Bernhard Rumpe's avatar
BR-sy  
Bernhard Rumpe committed
502
        return self._net_creator._model_dir_ + self._net_creator._model_prefix_ + '_' + str(index)