CTCModel.py 49.1 KB
Newer Older
Yann SOULLARD's avatar
Yann SOULLARD committed
1 2 3 4
import keras.backend as K
import tensorflow as tf
import numpy as np

Yann SOULLARD's avatar
Yann SOULLARD committed
5
import os
Yann SOULLARD's avatar
Yann SOULLARD committed
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
from keras import Input
from keras.engine import Model
from keras.layers import Lambda
from keras.models import model_from_json
import pickle
from tensorflow.python.ops import ctc_ops as ctc

from keras.utils import Sequence, GeneratorEnqueuer, OrderedEnqueuer
import warnings
from keras.utils.generic_utils import Progbar

#from ocr_ctc.utils.utils_analysis import tf_edit_distance
#from ocr_ctc.utils.utils_keras import Kreshape_To1D
from keras.preprocessing import sequence

"""
authors: Yann Soullard, Cyprien Ruffino (2017)
LITIS lab, university of Rouen (France)

The CTCModel class extends the Keras Model for the use of the Connectionist Temporal Classification (CTC) 
One makes use of the CTC proposed in tensorflow. Thus CTCModel can only be used with the backend tensorflow.

The CTCModel structure is composed of 3 branches. Each branch is a Keras Model:
    - One for computing the CTC loss (model_train)
    - One for predicting using the ctc_decode method (model_pred)
    - One for analyzing (model_eval) that computes the Label Error Rate (LER) and Sequence Error Rate (SER).

In a Keras Model, x is the input features and y the labels. 
Here, x data are of the form [input_sequences, label_sequences, inputs_lengths, labels_length] 
and y are not used as in a Keras Model (this is an array which is not considered,
the labeling is given in the x data structure). 
"""


class CTCModel:

    def __init__(self, inputs, outputs, greedy=True, beam_width=100, top_paths=1, charset=None):
        """
        Initialization of a CTC Model. 
        :param inputs: Input layer of the neural network
            outputs: Last layer of the neural network before CTC (e.g. a TimeDistributed Dense)
            greedy, beam_width, top_paths: Parameters of the CTC decoding (see ctc decoding tensorflow for more details)
            charset: labels related to the input of the CTC approach
        """
        self.model_train = None
        self.model_pred = None
        self.model_eval = None
        self.inputs = inputs
        self.outputs = outputs

        self.greedy = greedy
        self.beam_width = beam_width
        self.top_paths = top_paths
        self.charset = charset


    def compile(self, optimizer):
        """
        Configures the CTC Model for training.
        
        There is 3 Keras models:
            - one for training
            - one for predicting
            - one for analyzing
        
        Lambda layers are used to compute:
            - the CTC loss function 
            - the CTC decoding
            - the CTC evaluation
        
        :param optimizer: The optimizer used during training
        """


        # Others inputs for the CTC approach
        labels = Input(name='labels', shape=[None])
        input_length = Input(name='input_length', shape=[1])
        label_length = Input(name='label_length', shape=[1])

        # Lambda layer for computing the loss function
        loss_out = Lambda(self.ctc_loss_lambda_func, output_shape=(1,), name='CTCloss')(
            self.outputs + [labels, input_length, label_length])

        # Lambda layer for the decoding function
        out_decoded_dense = Lambda(self.ctc_complete_decoding_lambda_func, output_shape=(None, None), name='CTCdecode', arguments={'greedy': self.greedy,
                                     'beam_width': self.beam_width, 'top_paths': self.top_paths},dtype="float32")(
            self.outputs + [input_length])

        # Lambda layer to perform an analysis (CER and SER)
        out_analysis = Lambda(self.ctc_complete_analysis_lambda_func, output_shape=(None,), name='CTCanalysis',
                                   arguments={'greedy': self.greedy,
                                              'beam_width': self.beam_width, 'top_paths': self.top_paths},dtype="float32")(
                    self.outputs + [labels, input_length, label_length])


        # create Keras models
        self.model_init = Model(inputs=self.inputs, outputs=self.outputs)
        self.model_train = Model(inputs=self.inputs + [labels, input_length, label_length], outputs=loss_out)
        self.model_pred = Model(inputs=self.inputs + [input_length], outputs=out_decoded_dense)
        self.model_eval = Model(inputs=self.inputs + [labels, input_length, label_length], outputs=out_analysis)

        # Compile models
        self.model_train.compile(loss={'CTCloss': lambda yt, yp: yp}, optimizer=optimizer)
        self.model_pred.compile(loss={'CTCdecode': lambda yt, yp: yp}, optimizer=optimizer)
        self.model_eval.compile(loss={'CTCanalysis': lambda yt, yp: yp}, optimizer=optimizer)


    def get_model_train(self):
        """
        :return: Model used for training using the CTC approach
        """
        return self.model_train

    def get_model_pred(self):
        """
        :return: Model used for testing using the CTC approach
        """
        return self.model_pred


    def get_model_eval(self):
        """
        :return: Model used for evaluating using the CTC approach
        """
        return self.model_eval

Yann SOULLARD's avatar
Yann SOULLARD committed
132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153
    def get_loss_on_batch(self, inputs, verbose=False):
        """
        Computation the loss
        inputs is a list of 4 elements:
            x_features, y_label, x_len, y_len (similarly to the CTC in tensorflow)
        :return: Probabilities (output of the TimeDistributedDense layer)
        """

        x = inputs[0]
        x_len = inputs[2]
        y = inputs[1]
        y_len = inputs[3]

        no_lab = True if 0 in y_len else False

        if no_lab is False:
            loss_data = self.model_train.predict_on_batch([x, y, x_len, y_len], verbose=verbose)

        loss = np.sum(loss_data)

        return loss, loss_data

Yann SOULLARD's avatar
Yann SOULLARD committed
154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372

    def get_loss(self, inputs, verbose=False):
        """
        Computation the loss
        inputs is a list of 4 elements:
            x_features, y_label, x_len, y_len (similarly to the CTC in tensorflow)
        :return: Probabilities (output of the TimeDistributedDense layer)
        """

        x = inputs[0]
        x_len = inputs[2]
        y = inputs[1]
        y_len = inputs[3]
        batch_size = x.shape[0]

        no_lab = True if 0 in y_len else False

        if no_lab is False:
            loss_data = self.model_train.predict([x, y, x_len, y_len], batch_size=batch_size, verbose=verbose)

        loss = np.sum(loss_data)

        return loss, loss_data


    def get_loss_generator(self, generator, nb_batchs, verbose=False):
        """
        The generator must provide x as [input_sequences, label_sequences, inputs_lengths, labels_length]
        :return: loss on the entire dataset_manager and the loss per data
        """

        loss_per_data = []

        for k in range(nb_batchs):

            data = next(generator)

            x = data[0][0]
            x_len = data[0][2]
            y = data[0][1]
            y_len = data[0][3]
            batch_size = x.shape[0]

            no_lab = True if 0 in y_len else False

            if no_lab is False:
                loss_data = self.model_train.predict([x, y, x_len, y_len], batch_size=batch_size, verbose=verbose)
                loss_per_data += [elmt[0] for elmt in loss_data]

        loss = np.sum(loss_per_data)
        return loss, loss_per_data


    def get_probas_generator(self, generator, nb_batchs, verbose=False):
        """
        Get the probabilities of each label at each time of an observation sequence (matrix T x D)
        This is the output of the softmax function after the recurrent layers (the input of the CTC computations)
        
        Computation is done in batches using a generator. This function does not exist in a Keras Model.
        
        :return: A set of probabilities for each sequence and each time frame, one probability per label + the blank
            (this is the output of the TimeDistributed Dense layer, the blank label is the last probability)
        """

        probs_epoch = []

        for k in range(nb_batchs):

            data = next(generator)

            x = data[0][0]
            x_len = data[0][2]
            batch_size = x.shape[0]

            # Find the output of the softmax function
            probs = self.model_init.predict(x, batch_size=batch_size, verbose=verbose)

            # Select the outputs that do not refer to padding
            probs_epoch += [np.asarray(probs[data_idx, :x_len[data_idx][0], :]) for data_idx in range(batch_size)]

        return probs_epoch

    def get_probas_on_batch(self, inputs, verbose=False):
        """
        Get the probabilities of each label at each time of an observation sequence (matrix T x D)
        This is the output of the softmax function after the recurrent layers (the input of the CTC computations)

        Computation is done for a batch. This function does not exist in a Keras Model.

        :return: A set of probabilities for each sequence and each time frame, one probability per label + the blank
            (this is the output of the TimeDistributed Dense layer, the blank label is the last probability)
        """

        x = inputs[0]
        x_len = inputs[2]
        batch_size = x.shape[0]

        #  Find the output of the softmax function
        probs = self.model_init.predict(x, batch_size=batch_size, verbose=verbose)

        # Select the outputs that do not refer to padding
        probs_epoch = [np.asarray(probs[data_idx, :x_len[data_idx][0], :]) for data_idx in range(batch_size)]

        return probs_epoch


    def get_probas(self, inputs, batch_size, verbose=False):
        """
        Get the probabilities of each label at each time of an observation sequence (matrix T x D)
        This is the output of the softmax function after the recurrent layers (the input of the CTC computations)

        Computation is done for a batch. This function does not exist in a Keras Model.

        :return: A set of probabilities for each sequence and each time frame, one probability per label + the blank
            (this is the output of the TimeDistributed Dense layer, the blank label is the last probability)
        """


        x = inputs[0]
        x_len = inputs[2]

        #  Find the output of the softmax function
        probs = self.model_init.predict(x, batch_size=batch_size, verbose=verbose)

        # Select the outputs that do not refer to padding
        probs_epoch = [np.asarray(probs[data_idx, :x_len[data_idx][0], :]) for data_idx in range(batch_size)]

        return probs_epoch



    def fit_generator(self, generator,
                      steps_per_epoch,
                      epochs=1,
                      verbose=1,
                      callbacks=None,
                      validation_data=None,
                      validation_steps=None,
                      class_weight=None,
                      max_q_size=10,
                      workers=1,
                      pickle_safe=False,
                      initial_epoch=0):
        """
        Model training on data yielded batch-by-batch by a Python generator.
        
        The generator is run in parallel to the model, for efficiency. 
        For instance, this allows you to do real-time data augmentation on images on CPU in parallel to training your model on GPU.
        
        A major modification concerns the generator that must provide x data of the form:
          [input_sequences, label_sequences, inputs_lengths, labels_length]
        (in a similar way than for using CTC in tensorflow)
        
        :param: See keras.engine.Model.fit_generator()
        :return: A History object
        """
        out = self.model_train.fit_generator(generator, steps_per_epoch, epochs=epochs, verbose=verbose,
                                             callbacks=callbacks, validation_data=validation_data,
                                             validation_steps=validation_steps, class_weight=class_weight,
                                             max_q_size=max_q_size, workers=workers, pickle_safe=pickle_safe,
                                             initial_epoch=initial_epoch)

        self.model_pred.set_weights(self.model_train.get_weights())  # required??
        self.model_eval.set_weights(self.model_train.get_weights())
        return out


    def fit(self, x=None,
            y=None,
            batch_size=None,
            epochs=1,
            verbose=1,
            callbacks=None,
            validation_split=0.0,
            validation_data=None,
            shuffle=True,
            class_weight=None,
            sample_weight=None,
            initial_epoch=0,
            steps_per_epoch=None,
            validation_steps=None):
        """
        Model training on data.

        A major modification concerns the x input of the form:
          [input_sequences, label_sequences, inputs_lengths, labels_length]
        (in a similar way than for using CTC in tensorflow)

        :param: See keras.engine.Model.fit()
        :return: A History object
        """

        out = self.model_train.fit(x=x, y=y, batch_size=batch_size, epochs=epochs, verbose=verbose,
            callbacks=callbacks, validation_split=validation_split, validation_data=validation_data,
            shuffle=shuffle, class_weight=class_weight, sample_weight=sample_weight, initial_epoch=initial_epoch,
            steps_per_epoch=steps_per_epoch, validation_steps=validation_steps)

        self.model_pred.set_weights(self.model_train.get_weights())
        self.model_eval.set_weights(self.model_train.get_weights())

        return out


    def train_on_batch(self, x, y, sample_weight=None, class_weight=None):
        """ Runs a single gradient update on a single batch of data.
        See Keras.Model for more details.
        
        
        """

        out = self.model_train.train_on_batch(x, y, sample_weight=sample_weight,
                       class_weight=class_weight)

        self.model_pred.set_weights(self.model_train.get_weights())
        self.model_eval.set_weights(self.model_train.get_weights())

        return out


Yann SOULLARD's avatar
Yann SOULLARD committed
373
    def evaluate(self, x=None, batch_size=None, verbose=1, steps=None, metrics=['loss', 'ler', 'ser']):
Yann SOULLARD's avatar
Yann SOULLARD committed
374 375 376 377 378 379 380 381 382 383 384 385
        """ Evaluates the model on a dataset_manager.

                :param: See keras.engine.Model.predict()
                :return: A History object

                CTC evaluation on data yielded batch-by-batch by a Python generator.

                Inputs x:
                        x_input = Input data as a 3D Tensor (batch_size, max_input_len, dim_features)
                        y = Input data as a 2D Tensor (batch_size, max_label_len)
                        x_len = 1D array with the length of each data in batch_size
                        y_len = 1D array with the length of each labeling
Yann SOULLARD's avatar
Yann SOULLARD committed
386 387 388 389 390
                        
                metrics = list of metrics that are computed. This is elements among the 3 following metrics:
                    'loss' : compute the loss function on x
                    'ler' : compute the label error rate
                    'ser' : compute the sequence error rate
Yann SOULLARD's avatar
Yann SOULLARD committed
391 392 393 394 395 396 397 398 399 400 401 402 403

                Outputs: a list containing:
                    ler_dataset = label error rate for each data (a list)
                    seq_error = sequence error rate on the dataset_manager
        """
        seq_error = 0

        x_input = x[0]
        x_len = x[2]
        y = x[1]
        y_len = x[3]
        nb_data = x_input.shape[0]

Yann SOULLARD's avatar
Yann SOULLARD committed
404 405
        if 'ler' in metrics or 'ser' in metrics:
            eval_batch = self.model_eval.predict([x_input, y, x_len, y_len], batch_size=batch_size, verbose=verbose, steps=steps)
Yann SOULLARD's avatar
Yann SOULLARD committed
406

Yann SOULLARD's avatar
Yann SOULLARD committed
407 408 409
        if 'ser' in metrics:
            seq_error += np.sum([1 for ler_data in eval_batch if ler_data != 0])
            seq_error = seq_error / nb_data if nb_data > 0 else -1.
Yann SOULLARD's avatar
Yann SOULLARD committed
410

Yann SOULLARD's avatar
Yann SOULLARD committed
411 412 413 414 415 416 417
        outmetrics = []
        if 'loss' in metrics:
            outmetrics.append(self.get_loss(x))
        if 'ler' in metrics:
            outmetrics.append(eval_batch)
        if 'ser' in metrics:
            outmetrics.append(seq_error)
Yann SOULLARD's avatar
Yann SOULLARD committed
418

Yann SOULLARD's avatar
Yann SOULLARD committed
419 420 421
        return outmetrics

    def test_on_batch(self, x=None, metrics=['loss', 'ler', 'ser']):
Yann SOULLARD's avatar
Yann SOULLARD committed
422 423 424 425
        """ Name of a Keras Model function: this relates to evaluate on batch """
        return self.evaluate_on_batch(x)


Yann SOULLARD's avatar
Yann SOULLARD committed
426
    def evaluate_on_batch(self, x=None, metrics=['loss', 'ler', 'ser']):
Yann SOULLARD's avatar
Yann SOULLARD committed
427 428 429 430 431 432 433 434 435 436 437 438
        """ Evaluates the model on a dataset_manager.

                :param: See keras.engine.Model.predict_on_batch()
                :return: A History object

                CTC evaluation on data yielded batch-by-batch by a Python generator.

                Inputs x:
                        x_input = Input data as a 3D Tensor (batch_size, max_input_len, dim_features)
                        y = Input data as a 2D Tensor (batch_size, max_label_len)
                        x_len = 1D array with the length of each data in batch_size
                        y_len = 1D array with the length of each labeling
Yann SOULLARD's avatar
Yann SOULLARD committed
439 440 441 442 443
                        
                metrics = list of metrics that are computed. This is elements among the 3 following metrics:
                    'loss' : compute the loss function on x
                    'ler' : compute the label error rate
                    'ser' : compute the sequence error rate
Yann SOULLARD's avatar
Yann SOULLARD committed
444 445 446 447 448 449 450 451 452 453 454 455 456

                Outputs: a list containing:
                    ler_dataset = label error rate for each data (a list)
                    seq_error = sequence error rate on the dataset_manager
        """
        seq_error = 0

        x_input = x[0]
        x_len = x[2]
        y = x[1]
        y_len = x[3]
        nb_data = x_input.shape[0]

Yann SOULLARD's avatar
Yann SOULLARD committed
457 458 459 460 461 462
        if 'ler' in metrics or 'ser' in metrics:
            eval_batch = self.model_eval.predict_on_batch([x_input, y, x_len, y_len])

        if 'ser' in metrics:
            seq_error += np.sum([1 for ler_data in eval_batch if ler_data != 0])
            seq_error = seq_error / nb_data if nb_data > 0 else -1.
Yann SOULLARD's avatar
Yann SOULLARD committed
463

Yann SOULLARD's avatar
Yann SOULLARD committed
464 465 466 467 468 469 470
        outmetrics = []
        if 'loss' in metrics:
            outmetrics.append(self.get_loss(x))
        if 'ler' in metrics:
            outmetrics.append(eval_batch)
        if 'ser' in metrics:
            outmetrics.append(seq_error)
Yann SOULLARD's avatar
Yann SOULLARD committed
471

Yann SOULLARD's avatar
Yann SOULLARD committed
472
        return outmetrics
Yann SOULLARD's avatar
Yann SOULLARD committed
473 474


Yann SOULLARD's avatar
Yann SOULLARD committed
475
    def evaluate_generator(self, generator, steps=None, max_queue_size=10, workers=1, use_multiprocessing=False, verbose=0, metrics=['ler', 'ser']):
Yann SOULLARD's avatar
Yann SOULLARD committed
476 477 478 479 480 481 482 483 484 485 486 487 488 489
        """ Evaluates the model on a data generator.
        
        :param: See keras.engine.Model.fit()
        :return: A History object
        
        CTC evaluation on data yielded batch-by-batch by a Python generator.

        Inputs:
            generator = DataGenerator class that returns:
                    x = Input data as a 3D Tensor (batch_size, max_input_len, dim_features)
                    y = Input data as a 2D Tensor (batch_size, max_label_len)
                    x_len = 1D array with the length of each data in batch_size
                    y_len = 1D array with the length of each labeling
            nb_batchs = number of batchs that are evaluated
Yann SOULLARD's avatar
Yann SOULLARD committed
490 491 492 493 494 495 496 497 498 499
            
            metrics = list of metrics that are computed. This is elements among the 3 following metrics:
                    'loss' : compute the loss function on x
                    'ler' : compute the label error rate
                    'ser' : compute the sequence error rate
            Warning: if the 'loss' and another metric are requested, make sure that the number of steps allows to evaluate the entire dataset,
                   even if the data given by the generator will be not the same for all metrics. To make sure, you can only compute 'ler' and 'ser' here
                   then initialize again the generator and call get_loss_generator. 
            
        
Yann SOULLARD's avatar
Yann SOULLARD committed
500

Yann SOULLARD's avatar
Yann SOULLARD committed
501 502 503 504
        Outputs: a list containing the metrics given in argument:
            loss : the loss on the set
            ler : the label error rate for each data (a list)
            seq_error : the sequence error rate on the dataset
Yann SOULLARD's avatar
Yann SOULLARD committed
505 506
                 """

Yann SOULLARD's avatar
Yann SOULLARD committed
507 508
        if 'ler' in metrics or 'ser' in metrics:
            ler_dataset = self.model_eval.predict_generator(generator, steps,
Yann SOULLARD's avatar
Yann SOULLARD committed
509 510 511 512
                          max_queue_size=max_queue_size,
                          workers=workers,
                          use_multiprocessing=use_multiprocessing,
                          verbose=verbose)
Yann SOULLARD's avatar
Yann SOULLARD committed
513
        if 'ser' in metrics:
Yann SOULLARD's avatar
Yann SOULLARD committed
514
            seq_error = float(np.sum([1 for ler_data in ler_dataset if ler_data != 0])) / len(ler_dataset) if len(ler_dataset)>0 else 1.
Yann SOULLARD's avatar
Yann SOULLARD committed
515 516 517 518 519

        outmetrics = []
        if 'loss' in metrics:
            outmetrics.append(self.get_loss_generator(generator, steps))
        if 'ler' in metrics:
Yann SOULLARD's avatar
Yann SOULLARD committed
520
            outmetrics.append(ler_dataset)
Yann SOULLARD's avatar
Yann SOULLARD committed
521 522 523 524
        if 'ser' in metrics:
            outmetrics.append(seq_error)

        return outmetrics
Yann SOULLARD's avatar
Yann SOULLARD committed
525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740


    def predict_on_batch(self, x):
        """Returns predictions for a single batch of samples.

                # Arguments
                    x: [Input samples as a Numpy array, Input length as a numpy array]

                # Returns
                    Numpy array(s) of predictions.
        """
        batch_size = x[0].shape[0]

        return self.predict(x, batch_size=batch_size)


    def predict_generator(self, generator, steps,
                          max_queue_size=10,
                          workers=1,
                          use_multiprocessing=False,
                          verbose=0,
                          decode_func=None):
        """Generates predictions for the input samples from a data generator.

        The generator should return the same kind of data as accepted by
        `predict_on_batch`.
        
        generator = DataGenerator class that returns:
                        x = Input data as a 3D Tensor (batch_size, max_input_len, dim_features)
                        y = Input data as a 2D Tensor (batch_size, max_label_len)
                        x_len = 1D array with the length of each data in batch_size
                        y_len = 1D array with the length of each labeling

        # Arguments
            generator: Generator yielding batches of input samples
                    or an instance of Sequence (keras.utils.Sequence)
                    object in order to avoid duplicate data
                    when using multiprocessing.
            steps: Total number of steps (batches of samples)
                to yield from `generator` before stopping.
            max_queue_size: Maximum size for the generator queue.
            workers: Maximum number of processes to spin up
                when using process based threading
            use_multiprocessing: If `True`, use process based threading.
                Note that because
                this implementation relies on multiprocessing,
                you should not pass
                non picklable arguments to the generator
                as they can't be passed
                easily to children processes.
            verbose: verbosity mode, 0 or 1.
            decode_func: a function for decoding a list of predicted sequences (using self.charset)

        # Returns
            A tuple containing:
                A numpy array(s) of predictions.
                A numpy array(s) of ground truth.

        # Raises
            ValueError: In case the generator yields
                data in an invalid format.
        """
        self.model_pred._make_predict_function()

        steps_done = 0
        wait_time = 0.01
        all_outs = []
        all_lab = []
        is_sequence = isinstance(generator, Sequence)
        if not is_sequence and use_multiprocessing and workers > 1:
            warnings.warn(
                UserWarning('Using a generator with `use_multiprocessing=True`'
                            ' and multiple workers may duplicate your data.'
                            ' Please consider using the`keras.utils.Sequence'
                            ' class.'))
        enqueuer = None

        try:
            if is_sequence:
                enqueuer = OrderedEnqueuer(generator,
                                           use_multiprocessing=use_multiprocessing)
            else:
                enqueuer = GeneratorEnqueuer(generator,
                                             use_multiprocessing=use_multiprocessing,
                                             wait_time=wait_time)
            enqueuer.start(workers=workers, max_queue_size=max_queue_size)
            output_generator = enqueuer.get()

            if verbose == 1:
                progbar = Progbar(target=steps)

            while steps_done < steps:
                generator_output = next(output_generator)
                if isinstance(generator_output, tuple):
                    # Compatibility with the generators
                    # used for training.
                    if len(generator_output) == 2:
                        x, _ = generator_output
                    elif len(generator_output) == 3:
                        x, _, _ = generator_output
                    else:
                        raise ValueError('Output of generator should be '
                                         'a tuple `(x, y, sample_weight)` '
                                         'or `(x, y)`. Found: ' +
                                         str(generator_output))
                else:
                    # Assumes a generator that only
                    # yields inputs (not targets and sample weights).
                    x = generator_output

                [x_input, y, x_length, y_length] = x
                outs = self.predict_on_batch([x_input, x_length])
                if not isinstance(outs, list):
                    outs = [outs]

                if not all_outs:
                    for out in outs:
                        all_outs.append([])
                        all_lab.append([])

                for i, out in enumerate(outs):
                    all_outs[i].append([val_out for val_out in out if val_out!=-1])
                    if isinstance(y_length[i], list):
                        all_lab[i].append(y[i][:y_length[i][0]])
                    elif isinstance(y_length[i], int):
                        all_lab[i].append(y[i][:y_length[i]])
                    elif isinstance(y_length[i], float):
                        all_lab[i].append(y[i][:int(y_length[i])])
                    else:
                        all_lab[i].append(y[i])

                steps_done += 1
                if verbose == 1:
                    progbar.update(steps_done)

        finally:
            if enqueuer is not None:
                enqueuer.stop()

        batch_size = len(all_outs)
        nb_data = len(all_outs[0])
        pred_out = []
        lab_out = []
        for i in range(nb_data):
            pred_out += [all_outs[b][i] for b in range(batch_size)]
            lab_out +=  [all_lab[b][i] for b in range(batch_size)]

        if decode_func is not None:  # convert model prediction (a label between 0 to nb_labels to an original label sequence)
            pred_out = decode_func(pred_out, self.charset)
            lab_out = decode_func(lab_out, self.charset)

        return pred_out, lab_out


    def predict(self, x, batch_size=None, verbose=0, steps=None, max_len=None, max_value=999):

        """
        The same function as in the Keras Model but with a different function predict_loop for dealing with variable length predictions
        Except that x = [x_features, x_len]
        
        Generates output predictions for the input samples.

                Computation is done in batches.

                # Arguments
                    x: The input data, as a Numpy array
                        (or list of Numpy arrays if the model has multiple outputs).
                    batch_size: Integer. If unspecified, it will default to 32.
                    verbose: Verbosity mode, 0 or 1.
                    steps: Total number of steps (batches of samples)
                        before declaring the prediction round finished.
                        Ignored with the default value of `None`.

                # Returns
                    Numpy array(s) of predictions.

                # Raises
                    ValueError: In case of mismatch between the provided
                        input data and the model's expectations,
                        or in case a stateful model receives a number of samples
                        that is not a multiple of the batch size.
                """
        [x_inputs, x_len] = x
        if max_len is None:
            max_len = np.max(x_len)

        # Backwards compatibility.
        if batch_size is None and steps is None:
            batch_size = 32
        if x is None and steps is None:
            raise ValueError('If predicting from data tensors, '
                             'you should specify the `steps` '
                             'argument.')
        # Validate user data.
        x = _standardize_input_data(x, self.model_pred._feed_input_names,
                                    self.model_pred._feed_input_shapes,
                                    check_batch_axis=False)
        if self.model_pred.stateful:
            if x[0].shape[0] > batch_size and x[0].shape[0] % batch_size != 0:
                raise ValueError('In a stateful network, '
                                 'you should only pass inputs with '
                                 'a number of samples that can be '
                                 'divided by the batch size. Found: ' +
                                 str(x[0].shape[0]) + ' samples. '
                                                      'Batch size: ' + str(batch_size) + '.')

        # Prepare inputs, delegate logic to `_predict_loop`.
        if self.model_pred.uses_learning_phase and not isinstance(K.learning_phase(), int):
            ins = x + [0.]
        else:
            ins = x
        self.model_pred._make_predict_function()
        f = self.model_pred.predict_function
        out = self._predict_loop(f, ins, batch_size=batch_size, max_value=max_value,
                                  verbose=verbose, steps=steps, max_len=max_len)

Yann SOULLARD's avatar
Yann SOULLARD committed
741
        out_decode = [dec_data[:list(dec_data).index(max_value)] if max_value in dec_data else dec_data for i,dec_data in enumerate(out)]
Yann SOULLARD's avatar
Yann SOULLARD committed
742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763
        return out_decode

    def _predict_loop(self, f, ins, max_len=100, max_value=999, batch_size=32, verbose=0, steps=None):
        """Abstract method to loop over some data in batches.

        Keras function that has been modified. 
        
        # Arguments
            f: Keras function returning a list of tensors.
            ins: list of tensors to be fed to `f`.
            batch_size: integer batch size.
            verbose: verbosity mode.
            steps: Total number of steps (batches of samples)
                before declaring `_predict_loop` finished.
                Ignored with the default value of `None`.

        # Returns
            Array of predictions (if the model has a single output)
            or list of arrays of predictions
            (if the model has multiple outputs).
        """

764 765 766 767 768
        num_samples = check_num_samples(ins,
                                        batch_size=batch_size,
                                        steps=steps,
                                        steps_name='steps')
        
Yann SOULLARD's avatar
Yann SOULLARD committed
769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933
        if steps is not None:
            # Step-based predictions.
            # Since we do not know how many samples
            # we will see, we cannot pre-allocate
            # the returned Numpy arrays.
            # Instead, we store one array per batch seen
            # and concatenate them upon returning.
            unconcatenated_outs = []
            for step in range(steps):
                batch_outs = f(ins)
                if not isinstance(batch_outs, list):
                    batch_outs = [batch_outs]
                if step == 0:
                    for batch_out in batch_outs:
                        unconcatenated_outs.append([])
                for i, batch_out in enumerate(batch_outs):
                    unconcatenated_outs[i].append(batch_out)

            if len(unconcatenated_outs) == 1:
                return np.concatenate(unconcatenated_outs[0], axis=0)
            return [np.concatenate(unconcatenated_outs[i], axis=0)
                    for i in range(len(unconcatenated_outs))]
        else:
            # Sample-based predictions.
            outs = []
            batches = _make_batches(num_samples, batch_size)
            index_array = np.arange(num_samples)
            for batch_index, (batch_start, batch_end) in enumerate(batches):
                batch_ids = index_array[batch_start:batch_end]
                if ins and isinstance(ins[-1], float):
                    # Do not slice the training phase flag.
                    ins_batch = _slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
                else:
                    ins_batch = _slice_arrays(ins, batch_ids)
                batch_outs = f(ins_batch)
                if not isinstance(batch_outs, list):
                    batch_outs = [batch_outs]
                if batch_index == 0:
                    # Pre-allocate the results arrays.
                    for batch_out in batch_outs:
                        shape = (num_samples,max_len)
                        outs.append(np.zeros(shape, dtype=batch_out.dtype))
                for i, batch_out in enumerate(batch_outs):
                    outs[i][batch_start:batch_end] = sequence.pad_sequences(batch_out, value=float(max_value), maxlen=max_len,
                                     dtype=batch_out.dtype, padding="post")

            if len(outs) == 1:
                return outs[0]
            return outs


    @staticmethod
    def ctc_loss_lambda_func(args):
        """
        Function for computing the ctc loss (can be put in a Lambda layer)
        :param args: 
            y_pred, labels, input_length, label_length
        :return: CTC loss 
        """

        y_pred, labels, input_length, label_length = args
        return K.ctc_batch_cost(labels, y_pred, input_length, label_length)#, ignore_longer_outputs_than_inputs=True)


    @staticmethod
    def ctc_complete_decoding_lambda_func(args, **arguments):
        """
        Complete CTC decoding using Keras (function K.ctc_decode)
        :param args: 
            y_pred, input_length
        :param arguments:
            greedy, beam_width, top_paths
        :return: 
            K.ctc_decode with dtype='float32'
        """

        #import tensorflow as tf # Require for loading a model saved

        y_pred, input_length = args
        my_params = arguments

        assert (K.backend() == 'tensorflow')

        return K.cast(K.ctc_decode(y_pred, tf.squeeze(input_length), greedy=my_params['greedy'], beam_width=my_params['beam_width'], top_paths=my_params['top_paths'])[0][0], dtype='float32')

    @staticmethod
    def ctc_complete_analysis_lambda_func(args, **arguments):
        """
        Complete CTC analysis using Keras and tensorflow
        WARNING : tf is required 
        :param args: 
            y_pred, labels, input_length, label_len
        :param arguments:
            greedy, beam_width, top_paths
        :return: 
            ler = label error rate
        """

        #import tensorflow as tf # Require for loading a model saved

        y_pred, labels, input_length, label_len = args
        my_params = arguments

        assert (K.backend() == 'tensorflow')

        batch = tf.log(tf.transpose(y_pred, perm=[1, 0, 2]) + 1e-8)
        input_length = tf.to_int32(tf.squeeze(input_length))

        greedy = my_params['greedy']
        beam_width = my_params['beam_width']
        top_paths = my_params['top_paths']

        if greedy:
            (decoded, log_prob) = ctc.ctc_greedy_decoder(
                inputs=batch,
                sequence_length=input_length)
        else:
            (decoded, log_prob) = ctc.ctc_beam_search_decoder(
                inputs=batch, sequence_length=input_length,
                beam_width=beam_width, top_paths=top_paths)

        cast_decoded = tf.cast(decoded[0], tf.float32)

        sparse_y = K.ctc_label_dense_to_sparse(labels, tf.cast(tf.squeeze(label_len), tf.int32))
        ed_tensor = tf_edit_distance(cast_decoded, sparse_y, norm=True)
        ler_per_seq = Kreshape_To1D(ed_tensor)

        return K.cast(ler_per_seq, dtype='float32')



    def save_model(self, path_dir, charset=None):
        """ Save a model in path_dir 
        save model_train, model_pred and model_eval in json 
        save inputs and outputs in json
        save model CTC parameters in a pickle 
        
        :param path_dir: directory where the model architecture will be saved
        :param charset: set of labels (useful to keep the label order)
        """

        model_json = self.model_train.to_json()
        with open(path_dir + "/model_train.json", "w") as json_file:
            json_file.write(model_json)

        model_json = self.model_pred.to_json()
        with open(path_dir + "/model_pred.json", "w") as json_file:
            json_file.write(model_json)

        model_json = self.model_eval.to_json()
        with open(path_dir + "/model_eval.json", "w") as json_file:
            json_file.write(model_json)

        model_json = self.model_init.to_json()
        with open(path_dir + "/model_init.json", "w") as json_file:
            json_file.write(model_json)

        param = {'greedy': self.greedy, 'beam_width': self.beam_width, 'top_paths': self.top_paths, 'charset': self.charset}

        output = open(path_dir + "/model_param.pkl", 'wb')
        p = pickle.Pickler(output)
        p.dump(param)
        output.close()


Yann SOULLARD's avatar
Yann SOULLARD committed
934
    def load_model(self, path_dir, optimizer, file_weights=None):
Yann SOULLARD's avatar
Yann SOULLARD committed
935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981
        """ Load a model in path_dir 
        load model_train, model_pred and model_eval from json 
        load inputs and outputs from json
        load model CTC parameters from a pickle 
        
        :param path_dir: directory where the model is saved
        :param optimizer: The optimizer used during training
        """


        json_file = open(path_dir + '/model_train.json', 'r')
        loaded_model_json = json_file.read()
        json_file.close()
        self.model_train = model_from_json(loaded_model_json)

        json_file = open(path_dir + '/model_pred.json', 'r')
        loaded_model_json = json_file.read()
        json_file.close()
        self.model_pred = model_from_json(loaded_model_json, custom_objects={"tf": tf})

        json_file = open(path_dir + '/model_eval.json', 'r')
        loaded_model_json = json_file.read()
        json_file.close()
        self.model_eval = model_from_json(loaded_model_json, custom_objects={"tf": tf, "ctc": ctc,
                                                                             "tf_edit_distance": tf_edit_distance,
                                                                             "Kreshape_To1D": Kreshape_To1D})

        json_file = open(path_dir + '/model_init.json', 'r')
        loaded_model_json = json_file.read()
        json_file.close()
        self.model_init = model_from_json(loaded_model_json, custom_objects={"tf": tf})

        self.inputs = self.model_init.inputs
        self.outputs = self.model_init.outputs

        input = open(path_dir + "/model_param.pkl", 'rb')
        p = pickle.Unpickler(input)
        param = p.load()
        input.close()

        self.greedy = param['greedy'] if 'greedy' in param.keys() else self.greedy
        self.beam_width = param['beam_width'] if 'beam_width' in param.keys() else self.beam_width
        self.top_paths = param['top_paths'] if 'top_paths' in param.keys() else self.top_paths
        self.charset = param['charset'] if 'charset' in param.keys() else self.charset

        self.compile(optimizer)

Yann SOULLARD's avatar
Yann SOULLARD committed
982 983 984 985 986 987 988 989 990
        if file_weights is not None:
            if os.path.exists(file_weights):
                self.model_train.load_weights(file_weights)
                self.model_pred.set_weights(self.model_train.get_weights())
                self.model_eval.set_weights(self.model_train.get_weights())
            elif os.path.exists(path_dir + file_weights):
                self.model_train.load_weights(path_dir + file_weights)
                self.model_pred.set_weights(self.model_train.get_weights())
                self.model_eval.set_weights(self.model_train.get_weights())
Yann SOULLARD's avatar
Yann SOULLARD committed
991

Yann SOULLARD's avatar
Yann SOULLARD committed
992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186


def _standardize_input_data(data, names, shapes=None,
                            check_batch_axis=True,
                            exception_prefix=''):
    """Normalizes inputs and targets provided by users.

    Users may pass data as a list of arrays, dictionary of arrays,
    or as a single array. We normalize this to an ordered list of
    arrays (same order as `names`), while checking that the provided
    arrays have shapes that match the network's expectations.

    # Arguments
        data: User-provided input data (polymorphic).
        names: List of expected array names.
        shapes: Optional list of expected array shapes.
        check_batch_axis: Boolean; whether to check that
            the batch axis of the arrays matches the expected
            value found in `shapes`.
        exception_prefix: String prefix used for exception formatting.

    Keras function that has been modified. 
    
    # Returns
        List of standardized input arrays (one array per model input).

    # Raises
        ValueError: in case of improperly formatted user-provided data.
    """
    if not names:
        if data is not None and hasattr(data, '__len__') and len(data):
            raise ValueError('Error when checking model ' +
                             exception_prefix + ': '
                             'expected no data, but got:', data)
        return []
    if data is None:
        return [None for _ in range(len(names))]
    if isinstance(data, dict):
        arrays = []
        for name in names:
            if name not in data:
                raise ValueError('No data provided for "' +
                                 name + '". Need data for each key in: ' +
                                 str(names))
            arrays.append(data[name])
    elif isinstance(data, list):
        if len(data) != len(names):
            if data and hasattr(data[0], 'shape'):
                raise ValueError('Error when checking model ' +
                                 exception_prefix +
                                 ': the list of Numpy arrays '
                                 'that you are passing to your model '
                                 'is not the size the model expected. '
                                 'Expected to see ' + str(len(names)) +
                                 ' array(s), but instead got '
                                 'the following list of ' + str(len(data)) +
                                 ' arrays: ' + str(data)[:200] +
                                 '...')
            else:
                if len(names) == 1:
                    data = [np.asarray(data)]
                else:
                    raise ValueError(
                        'Error when checking model ' +
                        exception_prefix +
                        ': you are passing a list as '
                        'input to your model, '
                        'but the model expects '
                        'a list of ' + str(len(names)) +
                        ' Numpy arrays instead. '
                        'The list you passed was: ' +
                        str(data)[:200])
        arrays = data
    else:
        if not hasattr(data, 'shape'):
            raise TypeError('Error when checking model ' +
                            exception_prefix +
                            ': data should be a Numpy array, '
                            'or list/dict of Numpy arrays. '
                            'Found: ' + str(data)[:200] + '...')
        if len(names) > 1:
            # Case: model expects multiple inputs but only received
            # a single Numpy array.
            raise ValueError('The model expects ' + str(len(names)) + ' ' +
                             exception_prefix +
                             ' arrays, but only received one array. '
                             'Found: array with shape ' + str(data.shape))
        arrays = [data]

    # Make arrays at least 2D.
    for i in range(len(names)):
        array = arrays[i]
        if len(array.shape) == 1:
            array = np.expand_dims(array, 1)
            arrays[i] = array

    # Check shapes compatibility.
    if shapes:
        for i in range(len(names)):
            if shapes[i] is None:
                continue
            array = arrays[i]
            if len(array.shape) != len(shapes[i]):
                raise ValueError('Error when checking ' + exception_prefix +
                                 ': expected ' + names[i] +
                                 ' to have ' + str(len(shapes[i])) +
                                 ' dimensions, but got array with shape ' +
                                 str(array.shape))
            for j, (dim, ref_dim) in enumerate(zip(array.shape, shapes[i])):
                if not j and not check_batch_axis:
                    # skip the first axis
                    continue
                if ref_dim:
                    if ref_dim != dim:
                        raise ValueError(
                            'Error when checking ' + exception_prefix +
                            ': expected ' + names[i] +
                            ' to have shape ' + str(shapes[i]) +
                            ' but got array with shape ' +
                            str(array.shape))
    return arrays


def _slice_arrays(arrays, start=None, stop=None):
    """Slice an array or list of arrays.

    This takes an array-like, or a list of
    array-likes, and outputs:
        - arrays[start:stop] if `arrays` is an array-like
        - [x[start:stop] for x in arrays] if `arrays` is a list

    Can also work on list/array of indices: `_slice_arrays(x, indices)`

    Keras function that has been modified. 
    
    # Arguments
        arrays: Single array or list of arrays.
        start: can be an integer index (start index)
            or a list/array of indices
        stop: integer (stop index); should be None if
            `start` was a list.

    # Returns
        A slice of the array(s).
    """
    if arrays is None:
        return [None]
    elif isinstance(arrays, list):
        if hasattr(start, '__len__'):
            # hdf5 datasets only support list objects as indices
            if hasattr(start, 'shape'):
                start = start.tolist()
            return [None if x is None else x[start] for x in arrays]
        else:
            return [None if x is None else x[start:stop] for x in arrays]
    else:
        if hasattr(start, '__len__'):
            if hasattr(start, 'shape'):
                start = start.tolist()
            return arrays[start]
        elif hasattr(start, '__getitem__'):
            return arrays[start:stop]
        else:
            return [None]


def _make_batches(size, batch_size):
    """Returns a list of batch indices (tuples of indices).

    Keras function that has been modified. 
    
    # Arguments
        size: Integer, total size of the data to slice into batches.
        batch_size: Integer, batch size.

    # Returns
        A list of tuples of array indices.
    """
    num_batches = int(np.ceil(size / float(batch_size)))
    return [(i * batch_size, min(size, (i + 1) * batch_size))
            for i in range(0, num_batches)]


def Kreshape_To1D(my_tensor):
    """ Reshape to a 1D Tensor using K.reshape"""

    sum_shape = K.sum(K.shape(my_tensor))
    return K.reshape(my_tensor, (sum_shape,))


def tf_edit_distance(hypothesis, truth, norm=False):
    """ Edit distance using tensorflow 

    inputs are tf.Sparse_tensors """

1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233
    return tf.edit_distance(hypothesis, truth, normalize=norm, name='edit_distance')


def check_num_samples(ins,
                      batch_size=None,
                      steps=None,
                      steps_name='steps'):
    """Checks the number of samples provided for training and evaluation.
    The number of samples is not defined when running with `steps`,
    in which case the number of samples is set to `None`.
    # Arguments
        ins: List of tensors to be fed to the Keras function.
        batch_size: Integer batch size or `None` if not defined.
        steps: Total number of steps (batches of samples)
            before declaring `predict_loop` finished.
            Ignored with the default value of `None`.
        steps_name: The public API's parameter name for `steps`.
    # Raises
        ValueError: when `steps` is `None` and the attribute `ins.shape`
        does not exist. Also raises ValueError when `steps` is not `None`
        and `batch_size` is not `None` because they are mutually
        exclusive.
    # Returns
        When `steps` is `None`, returns the number of samples to be
        processed based on the size of the first dimension of the
        first input Numpy array. When `steps` is not `None` and
        `batch_size` is `None`, returns `None`.
    # Raises
        ValueError: In case of invalid arguments.
    """
    if steps is not None and batch_size is not None:
        raise ValueError(
            'If ' + steps_name + ' is set, the `batch_size` must be None.')

    if not ins or any(K.is_tensor(x) for x in ins):
        if steps is None:
            raise ValueError(
                'If your data is in the form of symbolic tensors, '
                'you should specify the `' + steps_name + '` argument '
                '(instead of the `batch_size` argument, '
                'because symbolic tensors are expected to produce '
                'batches of input data).')
        return None

    if hasattr(ins[0], 'shape'):
        return int(ins[0].shape[0])
    return None # Edge case where ins == [static_learning_phase]