CNNSupervisedTrainer_Alexnet.py 15.9 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 57 58 59 60
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)
        #loss = _apply_weighting(F, loss, self._weight, sample_weight)
        # 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))
        return loss.mean(axis=self._batch_axis, exclude=True)

Sebastian N.'s avatar
Sebastian N. committed
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
@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

                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
Christian Fuß's avatar
Christian Fuß committed
170

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


173
class CNNSupervisedTrainer_Alexnet:
Christian Fuß's avatar
Christian Fuß committed
174
    def applyBeamSearch(input, length, width, maxLength, currProb, netIndex, bestOutput):
175
        bestProb = 0.0
Christian Fuß's avatar
Christian Fuß committed
176 177
        while length < maxLength:
            length += 1
178 179 180 181 182 183 184 185
            batchIndex = 0
            for batchEntry in input:
                top_k_indices = mx.nd.topk(batchEntry, axis=0, k=width)
                top_k_values = mx.nd.topk(batchEntry, ret_typ='value', axis=0, k=width)
                for index in range(top_k_indices.size):

                    #print mx.nd.array(top_k_indices[index])
                    #print top_k_values[index]
Christian Fuß's avatar
Christian Fuß committed
186
                    if length == 1:
187
                        #print mx.nd.array(top_k_indices[index])
Christian Fuß's avatar
Christian Fuß committed
188
                        result = applyBeamSearch(self._networks[netIndex](mx.nd.array(top_k_indices[index])), length, width, maxLength,
189 190
                            currProb * top_k_values[index], netIndex, self._networks[netIndex](mx.nd.array(top_k_indices[index])))
                    else:
Christian Fuß's avatar
Christian Fuß committed
191
                        result = applyBeamSearch(self._networks[netIndex](mx.nd.array(top_k_indices[index])), length, width, maxLength,
192 193
                            currProb * top_k_values[index], netIndex, bestOutput)

Christian Fuß's avatar
Christian Fuß committed
194
                    if length == maxLength:
195 196 197 198 199 200 201 202 203 204 205 206
                        #print currProb
                        if currProb > bestProb:
                            bestProb = currProb
                            bestOutput[batchIndex] = result[batchIndex]
                            #print "new bestOutput: ", bestOutput

                batchIndex += 1
        #print bestOutput
        #print bestProb
        return bestOutput


207
    def __init__(self, data_loader, net_constructor):
Nicola Gatto's avatar
Nicola Gatto committed
208 209
        self._data_loader = data_loader
        self._net_creator = net_constructor
210
        self._networks = {}
Nicola Gatto's avatar
Nicola Gatto committed
211 212 213 214

    def train(self, batch_size=64,
              num_epoch=10,
              eval_metric='acc',
Sebastian N.'s avatar
Sebastian N. committed
215
              eval_metric_params={},
Eyüp Harputlu's avatar
Eyüp Harputlu committed
216 217
              loss ='softmax_cross_entropy',
              loss_params={},
Nicola Gatto's avatar
Nicola Gatto committed
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
              optimizer='adam',
              optimizer_params=(('learning_rate', 0.001),),
              load_checkpoint=True,
              context='gpu',
              checkpoint_period=5,
              normalize=True):
        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 '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']


        train_iter, test_iter, data_mean, data_std = self._data_loader.load_data(batch_size)
248 249 250 251 252

        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
253 254 255 256 257 258 259 260

        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_)

261
        self._networks = self._net_creator.networks
Nicola Gatto's avatar
Nicola Gatto committed
262 263 264 265 266 267 268

        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
269
        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
270

Eyüp Harputlu's avatar
Eyüp Harputlu committed
271 272 273 274 275 276
        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
        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 == 'sigmoid_binary_cross_entropy':
Nicola Gatto's avatar
Nicola Gatto committed
277
            loss_function = mx.gluon.loss.SigmoidBinaryCrossEntropyLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
278 279 280
        elif loss == 'cross_entropy':
            loss_function = CrossEntropyLoss(sparse_label=sparseLabel)
        elif loss == 'l2':
Nicola Gatto's avatar
Nicola Gatto committed
281
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
282
        elif loss == 'l1':
Nicola Gatto's avatar
Nicola Gatto committed
283
            loss_function = mx.gluon.loss.L2Loss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
284 285 286 287 288 289 290 291 292 293 294 295 296
        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
297 298
        elif loss == 'log_cosh':
            loss_function = LogCoshLoss()
Eyüp Harputlu's avatar
Eyüp Harputlu committed
299 300
        else:
            logging.error("Invalid loss parameter.")
Nicola Gatto's avatar
Nicola Gatto committed
301 302 303 304 305 306 307

        speed_period = 50
        tic = None

        for epoch in range(begin_epoch, begin_epoch + num_epoch):
            train_iter.reset()
            for batch_i, batch in enumerate(train_iter):
308
                data_ = batch.data[0].as_in_context(mx_context)
309
                predictions_label = batch.label[0].as_in_context(mx_context)
310

Christian Fuß's avatar
Christian Fuß committed
311 312
                outputs=[]

Nicola Gatto's avatar
Nicola Gatto committed
313
                with autograd.record():
314
                    predictions_ = mx.nd.zeros((batch_size, 10,), ctx=mx_context)
315

316
                    lossList = []
317
                    predictions_ = self._networks[0](data_)
318 319 320 321 322
                    lossList.append(loss_function(predictions_, predictions_label))

                    loss = 0
                    for element in lossList:
                        loss = loss + element
323

Nicola Gatto's avatar
Nicola Gatto committed
324
                loss.backward()
325 326 327

                for trainer in trainers:
                    trainer.step(batch_size)
Nicola Gatto's avatar
Nicola Gatto committed
328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344

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

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

                        tic = time.time()

            tic = None

            train_iter.reset()
Sebastian N.'s avatar
Sebastian N. committed
345
            metric = mx.metric.create(eval_metric, **eval_metric_params)
Nicola Gatto's avatar
Nicola Gatto committed
346
            for batch_i, batch in enumerate(train_iter):
347
                data_ = batch.data[0].as_in_context(mx_context)
348 349 350 351 352

                labels = [
                    batch.label[0].as_in_context(mx_context)
                ]

353
                outputs=[]
354

Sebastian N.'s avatar
Sebastian N. committed
355
                if True:
356
                    predictions_ = mx.nd.zeros((batch_size, 10,), ctx=mx_context)
357 358

                    predictions_ = self._networks[0](data_)
359
                    outputs.append(predictions_)
360

361
                predictions = []
362
                for output_name in outputs:
Sebastian N.'s avatar
Sebastian N. committed
363
                    if mx.nd.shape_array(mx.nd.squeeze(output_name)).size > 1:
364 365 366 367
                        predictions.append(mx.nd.argmax(output_name, axis=1))
                    #ArgMax already applied
                    else:
                        predictions.append(output_name)
368

369
                metric.update(preds=predictions, labels=labels)
Nicola Gatto's avatar
Nicola Gatto committed
370 371 372
            train_metric_score = metric.get()[1]

            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
                data_ = batch.data[0].as_in_context(mx_context)
376 377 378 379 380

                labels = [
                    batch.label[0].as_in_context(mx_context)
                ]

381 382
                outputs=[]

Sebastian N.'s avatar
Sebastian N. committed
383
                if True:
384
                    predictions_ = mx.nd.zeros((batch_size, 10,), ctx=mx_context)
385 386

                    predictions_ = self._networks[0](data_)
387
                    outputs.append(predictions_)
388

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

397
                metric.update(preds=predictions, labels=labels)
Nicola Gatto's avatar
Nicola Gatto committed
398 399 400 401
            test_metric_score = metric.get()[1]

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

402

Nicola Gatto's avatar
Nicola Gatto committed
403
            if (epoch - begin_epoch) % checkpoint_period == 0:
404 405
                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
406

407 408 409
        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
410

411 412
    def parameter_path(self, index):
        return self._net_creator._model_dir_ + self._net_creator._model_prefix_ + '_' + str(index)