Beispiel #1
0
def new_rnn(cfg, layer_num):
    use_cudnnlstm = K.backend() == 'tensorflow' and len(config.get_visible_devices('GPU')) > 0
    if use_cudnnlstm:
        if cfg['rnn_bidirectional']:
            return Bidirectional(LSTM(cfg['rnn_size'],
                                           return_sequences=True),
                                 name='rnn_{}'.format(layer_num))

        return LSTM(cfg['rnn_size'],
                         return_sequences=True,
                         name='rnn_{}'.format(layer_num))
    else:
        if cfg['rnn_bidirectional']:
            return Bidirectional(LSTM(cfg['rnn_size'],
                                      return_sequences=True,
                                      recurrent_activation='sigmoid'),
                                 name='rnn_{}'.format(layer_num))

        return LSTM(cfg['rnn_size'],
                    return_sequences=True,
                    recurrent_activation='sigmoid',
                    name='rnn_{}'.format(layer_num))
Beispiel #2
0
    def train_on_texts(self,
                       texts,
                       context_labels=None,
                       batch_size=128,
                       num_epochs=50,
                       verbose=1,
                       new_model=False,
                       gen_epochs=1,
                       train_size=1.0,
                       max_gen_length=300,
                       validation=True,
                       dropout=0.0,
                       via_new_model=False,
                       save_epochs=0,
                       multi_gpu=False,
                       **kwargs):

        if new_model and not via_new_model:
            self.train_new_model(texts,
                                 context_labels=context_labels,
                                 num_epochs=num_epochs,
                                 gen_epochs=gen_epochs,
                                 train_size=train_size,
                                 batch_size=batch_size,
                                 dropout=dropout,
                                 validation=validation,
                                 save_epochs=save_epochs,
                                 multi_gpu=multi_gpu,
                                 **kwargs)
            return

        if context_labels:
            context_labels = LabelBinarizer().fit_transform(context_labels)

        if 'prop_keep' in kwargs:
            train_size = prop_keep

        if self.config['word_level']:
            # If training word level, must add spaces around each
            # punctuation. https://stackoverflow.com/a/3645946/9314418
            punct = '!"#$%&()*+,-./:;<=>?@[\]^_`{|}~\\n\\t\'‘’“”’–—…'
            for i in range(len(texts)):
                texts[i] = re.sub('([{}])'.format(punct), r' \1 ', texts[i])
                texts[i] = re.sub(' {2,}', ' ', texts[i])
            texts = [text_to_word_sequence(text, filters='') for text in texts]

        # calculate all combinations of text indices + token indices
        indices_list = [
            np.meshgrid(np.array(i), np.arange(len(text) + 1))
            for i, text in enumerate(texts)
        ]
        # indices_list = np.block(indices_list) # this hangs when indices_list is large enough
        # FIX BEGIN ------
        indices_list_o = np.block(indices_list[0])
        for i in range(len(indices_list) - 1):
            tmp = np.block(indices_list[i + 1])
            indices_list_o = np.concatenate([indices_list_o, tmp])
        indices_list = indices_list_o
        # FIX END ------

        # If a single text, there will be 2 extra indices, so remove them
        # Also remove first sequences which use padding
        if self.config['single_text']:
            indices_list = indices_list[self.config['max_length']:-2, :]

        indices_mask = np.random.rand(indices_list.shape[0]) < train_size

        if multi_gpu:
            num_gpus = len(config.get_visible_devices('GPU'))
            batch_size = batch_size * num_gpus

        gen_val = None
        val_steps = None
        if train_size < 1.0 and validation:
            indices_list_val = indices_list[~indices_mask, :]
            gen_val = generate_sequences_from_texts(texts, indices_list_val,
                                                    self, context_labels,
                                                    batch_size)
            val_steps = max(
                int(np.floor(indices_list_val.shape[0] / batch_size)), 1)

        indices_list = indices_list[indices_mask, :]

        num_tokens = indices_list.shape[0]
        assert num_tokens >= batch_size, "Fewer tokens than batch_size."

        level = 'word' if self.config['word_level'] else 'character'
        print("Training on {:,} {} sequences.".format(num_tokens, level))

        steps_per_epoch = max(int(np.floor(num_tokens / batch_size)), 1)

        gen = generate_sequences_from_texts(texts, indices_list, self,
                                            context_labels, batch_size)

        base_lr = 4e-3

        # scheduler function must be defined inline.
        def lr_linear_decay(epoch):
            return (base_lr * (1 - (epoch / num_epochs)))

        '''
        FIXME
        This part is a bit messy as we need to initialize the model within
        strategy.scope() when using multi-GPU. Can probably be cleaned up a bit.
        '''

        if context_labels is not None:
            if new_model:
                weights_path = None
            else:
                weights_path = "{}_weights.hdf5".format(self.config['name'])
                self.save(weights_path)

            if multi_gpu:
                from tensorflow import distribute as distribute
                strategy = distribute.MirroredStrategy()
                with strategy.scope():
                    parallel_model = textgenrnn_model(
                        self.num_classes,
                        dropout=dropout,
                        cfg=self.config,
                        context_size=context_labels.shape[1],
                        weights_path=weights_path)
                    parallel_model.compile(loss='categorical_crossentropy',
                                           optimizer=Adam(lr=4e-3))
                model_t = parallel_model
                print("Training on {} GPUs.".format(num_gpus))
            else:
                model_t = self.model
        else:
            if multi_gpu:
                from tensorflow import distribute as distribute
                if new_model:
                    weights_path = None
                else:
                    weights_path = "{}_weights.hdf5".format(
                        self.config['name'])

                strategy = distribute.MirroredStrategy()
                with strategy.scope():
                    # Do not locate model/merge on CPU since sample sizes are small.
                    parallel_model = textgenrnn_model(
                        self.num_classes,
                        cfg=self.config,
                        weights_path=weights_path)
                    parallel_model.compile(loss='categorical_crossentropy',
                                           optimizer=Adam(lr=4e-3))

                model_t = parallel_model
                print("Training on {} GPUs.".format(num_gpus))
            else:
                model_t = self.model

        model_t.fit(gen,
                    steps_per_epoch=steps_per_epoch,
                    epochs=num_epochs,
                    callbacks=[
                        LearningRateScheduler(lr_linear_decay),
                        generate_after_epoch(self, gen_epochs, max_gen_length),
                        save_model_weights(self, num_epochs, save_epochs)
                    ],
                    verbose=verbose,
                    max_queue_size=10,
                    validation_data=gen_val,
                    validation_steps=val_steps)

        # Keep the text-only version of the model if using context labels
        if context_labels is not None:
            self.model = Model(inputs=self.model.input[0],
                               outputs=self.model.output[1])