CNNSupervisedTrainer.ftl 16.1 KB
Newer Older
Bernhard Rumpe's avatar
BR-sy  
Bernhard Rumpe committed
1
<#-- (c) https://github.com/MontiCore/monticore -->
Nicola Gatto's avatar
Nicola Gatto committed
2 3 4 5 6 7
import mxnet as mx
import logging
import numpy as np
import time
import os
import shutil
8
import pickle
Sebastian N.'s avatar
Sebastian N. committed
9 10
import math
import sys
Nicola Gatto's avatar
Nicola Gatto committed
11 12
from mxnet import gluon, autograd, nd

Eyüp Harputlu's avatar
Eyüp Harputlu committed
13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28
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
29 30 31 32 33 34 35 36 37
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)

38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
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:
58
            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))
59 60
        return loss.mean(axis=self._batch_axis, exclude=True)

Sebastian N.'s avatar
Sebastian N. committed
61 62 63 64 65 66 67
@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)

Sebastian N.'s avatar
Sebastian N. committed
68
        self._exclude = exclude or []
Sebastian N.'s avatar
Sebastian N. committed
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

        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

118 119
                if (match_counts / counts) > 0:
                    precisions[n] = match_counts / counts
Sebastian N.'s avatar
Sebastian N. committed
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

        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
171

172 173


174
class ${tc.fileNameWithoutEnding}:
175
    def __init__(self, data_loader, net_constructor):
Nicola Gatto's avatar
Nicola Gatto committed
176 177
        self._data_loader = data_loader
        self._net_creator = net_constructor
178
        self._networks = {}
Nicola Gatto's avatar
Nicola Gatto committed
179 180 181 182

    def train(self, batch_size=64,
              num_epoch=10,
              eval_metric='acc',
Sebastian N.'s avatar
Sebastian N. committed
183
              eval_metric_params={},
184
              eval_train=False,
Eyüp Harputlu's avatar
Eyüp Harputlu committed
185 186
              loss ='softmax_cross_entropy',
              loss_params={},
Nicola Gatto's avatar
Nicola Gatto committed
187 188 189 190
              optimizer='adam',
              optimizer_params=(('learning_rate', 0.001),),
              load_checkpoint=True,
              checkpoint_period=5,
191 192
              log_period=50,
              context='gpu',
193
              save_attention_image=False,
194
              use_teacher_forcing=False,
195
              normalize=True,
196 197
              shuffle_data=False,
              clip_global_grad_norm=None,
198
              preprocessing = False):
Nicola Gatto's avatar
Nicola Gatto committed
199 200 201 202 203 204 205
        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'.")

206 207
        if preprocessing:
            preproc_lib = "CNNPreprocessor_${tc.fileNameWithoutEnding?keep_after("CNNSupervisedTrainer_")}_executor"
208
            train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_preprocessed_data(batch_size, preproc_lib, shuffle_data)
209
        else:
210
            train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_data(batch_size, shuffle_data)
211

Nicola Gatto's avatar
Nicola Gatto committed
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226
        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']

227 228 229 230
        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
231 232 233 234 235 236 237 238

        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_)

239
        self._networks = self._net_creator.networks
Nicola Gatto's avatar
Nicola Gatto committed
240 241 242 243 244 245 246

        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
247
        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
248

Eyüp Harputlu's avatar
Eyüp Harputlu committed
249 250
        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
251
        ignore_indices = [loss_params['ignore_indices']] if 'ignore_indices' in loss_params else []
Eyüp Harputlu's avatar
Eyüp Harputlu committed
252 253
        if loss == 'softmax_cross_entropy':
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
254
            loss_function = mx.gluon.loss.SoftmaxCrossEntropyLoss(from_logits=fromLogits, sparse_label=sparseLabel)
255
        elif loss == 'softmax_cross_entropy_ignore_indices':
256
            fromLogits = loss_params['from_logits'] if 'from_logits' in loss_params else False
257
            loss_function = SoftmaxCrossEntropyLossIgnoreIndices(ignore_indices=ignore_indices, from_logits=fromLogits, sparse_label=sparseLabel)
Eyüp Harputlu's avatar
Eyüp Harputlu committed
258
        elif loss == 'sigmoid_binary_cross_entropy':
Nicola Gatto's avatar
Nicola Gatto committed
259
            loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
260 261 262
        elif loss == 'cross_entropy':
            loss_function = CrossEntropyLoss(sparse_label=sparseLabel)
        elif loss == 'l2':
Nicola Gatto's avatar
Nicola Gatto committed
263
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
264
        elif loss == 'l1':
Nicola Gatto's avatar
Nicola Gatto committed
265
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
266 267 268 269 270 271 272 273 274 275 276 277 278
        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
279 280
        elif loss == 'log_cosh':
            loss_function = LogCoshLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
281 282
        else:
            logging.error("Invalid loss parameter.")
Nicola Gatto's avatar
Nicola Gatto committed
283 284 285 286

        tic = None

        for epoch in range(begin_epoch, begin_epoch + num_epoch):
287 288 289 290 291 292
            if shuffle_data:
                if preprocessing:
                    preproc_lib = "CNNPreprocessor_${tc.fileNameWithoutEnding?keep_after("CNNSupervisedTrainer_")}_executor"
                    train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_preprocessed_data(batch_size, preproc_lib, shuffle_data)
                else:
                    train_iter, test_iter, data_mean, data_std, train_images, test_images = self._data_loader.load_data(batch_size, shuffle_data)
293

294 295 296
            global_loss_train = 0.0
            train_batches = 0

297
            loss_total = 0
Nicola Gatto's avatar
Nicola Gatto committed
298 299 300
            train_iter.reset()
            for batch_i, batch in enumerate(train_iter):
                with autograd.record():
301
<#include "pythonExecuteTrain.ftl">
302

303 304 305
                    loss = 0
                    for element in lossList:
                        loss = loss + element
Nicola Gatto's avatar
Nicola Gatto committed
306 307

                loss.backward()
308

309 310
                loss_total += loss.sum().asscalar()

311 312 313
                global_loss_train += float(loss.mean().asscalar())
                train_batches += 1

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

                        tic = time.time()

341 342 343
            if train_batches > 0:
                global_loss_train /= train_batches

Nicola Gatto's avatar
Nicola Gatto committed
344 345
            tic = None

346

347 348 349 350
            if eval_train:
                train_iter.reset()
                metric = mx.metric.create(eval_metric, **eval_metric_params)
                for batch_i, batch in enumerate(train_iter):
351
<#include "pythonExecuteTest.ftl">
352

353

354 355 356
<#include "saveAttentionImageTrain.ftl">


357 358 359 360 361 362
                    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)
363

364 365 366 367
                    metric.update(preds=predictions, labels=labels)
                train_metric_score = metric.get()[1]
            else:
                train_metric_score = 0
Nicola Gatto's avatar
Nicola Gatto committed
368

369 370 371
            global_loss_test = 0.0
            test_batches = 0

Nicola Gatto's avatar
Nicola Gatto committed
372
            test_iter.reset()
Sebastian N.'s avatar
Sebastian N. committed
373
            metric = mx.metric.create(eval_metric, **eval_metric_params)
Nicola Gatto's avatar
Nicola Gatto committed
374
            for batch_i, batch in enumerate(test_iter):
375
                if True: <#-- Fix indentation -->
376
<#include "pythonExecuteTest.ftl">
377 378


379 380
<#include "saveAttentionImageTest.ftl">

381 382 383 384 385 386
                loss = 0
                for element in lossList:
                    loss = loss + element

                global_loss_test += float(loss.mean().asscalar())
                test_batches += 1
387

388
                predictions = []
389
                for output_name in outputs:
390
                    if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
391 392 393 394
                        predictions.append(mx.nd.argmax(output_name, axis=1))
                    #ArgMax already applied
                    else:
                        predictions.append(output_name)
395 396

                metric.update(preds=predictions, labels=labels)
Nicola Gatto's avatar
Nicola Gatto committed
397 398
            test_metric_score = metric.get()[1]

399 400
            if test_batches > 0:
                global_loss_test /= test_batches
Nicola Gatto's avatar
Nicola Gatto committed
401

402
            logging.info("Epoch[%d] Train: %f, Test: %f, Train Loss: %f, Test Loss: %f" % (epoch, train_metric_score, test_metric_score, global_loss_train, global_loss_test))
403

Nicola Gatto's avatar
Nicola Gatto committed
404
            if (epoch - begin_epoch) % checkpoint_period == 0:
405 406
                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
407

408 409 410
        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
411

412
    def parameter_path(self, index):
Bernhard Rumpe's avatar
BR-sy  
Bernhard Rumpe committed
413
        return self._net_creator._model_dir_ + self._net_creator._model_prefix_ + '_' + str(index)