Python recurrent.GRU属性代码示例

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def test_temporal_clf(self):
        print('temporal classification data:')
        (X_train, y_train), (X_test, y_test) = get_test_data(nb_train=1000, nb_test=200, input_shape=(5,10), 
                                                             classification=True, nb_class=2)
        print('X_train:', X_train.shape)
        print('X_test:', X_test.shape)
        print('y_train:', y_train.shape)
        print('y_test:', y_test.shape)

        y_train = to_categorical(y_train)
        y_test = to_categorical(y_test)

        model = Sequential()
        model.add(GRU(X_train.shape[-1], y_train.shape[-1]))
        model.add(Activation('softmax'))
        model.compile(loss='categorical_crossentropy', optimizer='adadelta')
        history = model.fit(X_train, y_train, nb_epoch=12, batch_size=16, validation_data=(X_test, y_test), show_accuracy=True, verbose=2)
        self.assertTrue(history.history['val_acc'][-1] > 0.9) 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def understand_variable_length_handle():
    """用来帮助理解如何用 recurrent layer 处理变长序列"""
    model = Sequential()
    model.add(GRU(input_dim=256, output_dim=256, return_sequences=True))
    model.compile(loss='mean_squared_error', optimizer='sgd')
    train_x = np.random.randn(100, 78, 256)
    train_y = np.random.randn(100, 78, 256)
    model.fit(train_x, train_y, verbose=0)

    inz_1 = np.random.randn(1, 78, 256)
    rez_1 = model.predict_proba(inz_1, verbose=0)

    inz_2 = np.random.randn(1, 87, 256)
    rez_2 = model.predict_proba(inz_2, verbose=0)

    print()
    print('=========== understand variable length =================')
    print('With `return_sequence=True`')
    print('Input shape is: {}, output shae is {}'.format(inz_1.shape, rez_1.shape))
    print('Input shape is: {}, output shae is {}'.format(inz_2.shape, rez_2.shape))
    print('====================== end =============================') 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def try_variable_length_train():
    """变长序列训练实验

    实验失败，这样得到的 train_x 和 train_y 的 dtype 是 object 类型，
    取其 shape 得到的是 (100,) ，这将导致训练出错
    """
    model = Sequential()
    model.add(GRU(input_dim=256, output_dim=256, return_sequences=True))
    model.compile(loss='mean_squared_error', optimizer='sgd')

    train_x = []
    train_y = []
    for i in range(100):
        seq_length = np.random.randint(78, 87 + 1)
        sequence = []
        for _ in range(seq_length):
            sequence.append([np.random.randn() for _ in range(256)])

        train_x.append(np.array(sequence))
        train_y.append(np.array(sequence))

    train_x = np.array(train_x)
    train_y = np.array(train_y)

    model.fit(np.array(train_x), np.array(train_y)) 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def gen_model(vocab_size=100, embedding_size=128, maxlen=100, output_size=6, hidden_layer_size=100, num_hidden_layers = 1, RNN_LAYER_TYPE="LSTM"):
    RNN_CLASS = LSTM
    if RNN_LAYER_TYPE == "GRU":
        RNN_CLASS = GRU
    logger.info("Parameters: vocab_size = %s, embedding_size = %s, maxlen = %s, output_size = %s, hidden_layer_size = %s, " %\
            (vocab_size, embedding_size, maxlen, output_size, hidden_layer_size))
    logger.info("Building Model")
    model = Sequential()
    logger.info("Init Model with vocab_size = %s, embedding_size = %s, maxlen = %s" % (vocab_size, embedding_size, maxlen))
    model.add(Embedding(vocab_size, embedding_size, input_length=maxlen))
    logger.info("Added Embedding Layer")
    model.add(Dropout(0.5))
    logger.info("Added Dropout Layer")
    for i in xrange(num_hidden_layers):
        model.add(RNN_CLASS(output_dim=hidden_layer_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True))
        logger.info("Added %s Layer" % RNN_LAYER_TYPE)
        model.add(Dropout(0.5))
        logger.info("Added Dropout Layer")
    model.add(RNN_CLASS(output_dim=output_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True))
    logger.info("Added %s Layer" % RNN_LAYER_TYPE)
    model.add(Dropout(0.5))
    logger.info("Added Dropout Layer")
    model.add(TimeDistributedDense(output_size, activation="softmax"))
    logger.info("Added Dropout Layer")
    logger.info("Created model with following config:\n%s" % json.dumps(model.get_config(), indent=4))
    logger.info("Compiling model with optimizer %s" % optimizer)
    start_time = time.time()
    model.compile(loss='categorical_crossentropy', optimizer=optimizer)
    total_time = time.time() - start_time
    logger.info("Model compiled in %.4f seconds." % total_time)
    return model 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def get_gru(units):
    """GRU(Gated Recurrent Unit)
    Build GRU Model.

    # Arguments
        units: List(int), number of input, output and hidden units.
    # Returns
        model: Model, nn model.
    """

    model = Sequential()
    model.add(GRU(units[1], input_shape=(units[0], 1), return_sequences=True))
    model.add(GRU(units[2]))
    model.add(Dropout(0.2))
    model.add(Dense(units[3], activation='sigmoid'))

    return model 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def __init__(self, input_shape, gru_size=10, tdd_size=4):
    self.model = Sequential()
    self.model.add(
        GRU(gru_size, return_sequences=True, input_shape=input_shape))
    if tdd_size is not None:
      self.model.add(TimeDistributedDense(tdd_size))
    self.model.add(Flatten())
    self.model.add(Dense(1))
    self.model.add(Activation('sigmoid'))
    print('Compiling model...')
    self.model.compile(optimizer='adam', loss='binary_crossentropy') 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def _buildDecoder(self, z, latent_rep_size, max_length, charset_length):
    h = Dense(latent_rep_size, name='latent_input', activation='relu')(z)
    h = RepeatVector(max_length, name='repeat_vector')(h)
    h = GRU(501, return_sequences=True, name='gru_1')(h)
    h = GRU(501, return_sequences=True, name='gru_2')(h)
    h = GRU(501, return_sequences=True, name='gru_3')(h)
    return TimeDistributed(
        Dense(charset_length, activation='softmax'), name='decoded_mean')(h) 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def test_gru(self):
        _runner(recurrent.GRU) 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def test_temporal_reg(self):
        print('temporal regression data:')
        (X_train, y_train), (X_test, y_test) = get_test_data(nb_train=1000, nb_test=200, input_shape=(5, 10), output_shape=(2,),
                                                             classification=False)
        print('X_train:', X_train.shape)
        print('X_test:', X_test.shape)
        print('y_train:', y_train.shape)
        print('y_test:', y_test.shape)

        model = Sequential()
        model.add(GRU(X_train.shape[-1], y_train.shape[-1]))
        model.compile(loss='hinge', optimizer='adam')
        history = model.fit(X_train, y_train, nb_epoch=12, batch_size=16, validation_data=(X_test, y_test), verbose=2)
        self.assertTrue(history.history['val_loss'][-1] < 0.8) 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def understand_return_sequence():
    """用来帮助理解 recurrent layer 中的 return_sequences 参数"""
    model_1 = Sequential()
    model_1.add(GRU(input_dim=256, output_dim=256, return_sequences=True))
    model_1.compile(loss='mean_squared_error', optimizer='sgd')
    train_x = np.random.randn(100, 78, 256)
    train_y = np.random.randn(100, 78, 256)
    model_1.fit(train_x, train_y, verbose=0)

    model_2 = Sequential()
    model_2.add(GRU(input_dim=256, output_dim=256, return_sequences=False))
    model_2.compile(loss='mean_squared_error', optimizer='sgd')
    train_x = np.random.randn(100, 78, 256)
    train_y = np.random.randn(100, 256)
    model_2.fit(train_x, train_y, verbose=0)

    inz = np.random.randn(100, 78, 256)
    rez_1 = model_1.predict_proba(inz, verbose=0)
    rez_2 = model_2.predict_proba(inz, verbose=0)

    print()
    print('=========== understand return_sequence =================')
    print('Input shape is: {}'.format(inz.shape))
    print('Output shape of model with `return_sequences=True`: {}'.format(rez_1.shape))
    print('Output shape of model with `return_sequences=False`: {}'.format(rez_2.shape))
    print('====================== end =============================') 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def try_variable_length_train_in_batch():
    """变长序列训练实验(2)"""
    model = Sequential()
    model.add(GRU(input_dim=256, output_dim=256, return_sequences=True))
    model.compile(loss='mean_squared_error', optimizer='sgd')

    # 分作两个 batch, 不同 batch 中的 sequence 长度不一样
    seq_lens = [78, 87]
    for i in range(2):
        train_x = np.random.randn(20, seq_lens[i], 256)
        train_y = np.random.randn(20, seq_lens[i], 256)
        model.train_on_batch(train_x, train_y) 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def build_model(input_size, seq_len, hidden_size):
    """建立一个 sequence to sequence 模型"""
    model = Sequential()
    model.add(GRU(input_dim=input_size, output_dim=hidden_size, return_sequences=False))
    model.add(Dense(hidden_size, activation="relu"))
    model.add(RepeatVector(seq_len))
    model.add(GRU(hidden_size, return_sequences=True))
    model.add(TimeDistributed(Dense(output_dim=input_size, activation="linear")))
    model.compile(loss="mse", optimizer='adam')

    return model 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def gen_model_brnn(vocab_size=100, embedding_size=128, maxlen=100, output_size=6, hidden_layer_size=100, num_hidden_layers = 1, RNN_LAYER_TYPE="LSTM"):
    RNN_CLASS = LSTM
    if RNN_LAYER_TYPE == "GRU":
        RNN_CLASS = GRU
    logger.info("Parameters: vocab_size = %s, embedding_size = %s, maxlen = %s, output_size = %s, hidden_layer_size = %s, " %\
            (vocab_size, embedding_size, maxlen, output_size, hidden_layer_size))
    logger.info("Building Graph model for Bidirectional RNN")
    model = Graph()
    model.add_input(name='input', input_shape=(maxlen,), dtype=int)
    logger.info("Added Input node")
    logger.info("Init Model with vocab_size = %s, embedding_size = %s, maxlen = %s" % (vocab_size, embedding_size, maxlen))
    model.add_node(Embedding(vocab_size, embedding_size, input_length=maxlen), name='embedding', input='input')
    logger.info("Added Embedding node")
    model.add_node(Dropout(0.5), name="dropout_0", input="embedding")
    logger.info("Added Dropout Node")
    for i in xrange(num_hidden_layers):
        last_dropout_name = "dropout_%s" % i
        forward_name, backward_name, dropout_name = ["%s_%s" % (k, i + 1) for k in ["forward", "backward", "dropout"]]
        model.add_node(RNN_CLASS(output_dim=hidden_layer_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True), name=forward_name, input=last_dropout_name)
        logger.info("Added %s forward node[%s]" % (RNN_LAYER_TYPE, i+1))
        model.add_node(RNN_CLASS(output_dim=hidden_layer_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True, go_backwards=True), name=backward_name, input=last_dropout_name)
        logger.info("Added %s backward node[%s]" % (RNN_LAYER_TYPE, i+1))
        model.add_node(Dropout(0.5), name=dropout_name, inputs=[forward_name, backward_name])
        logger.info("Added Dropout node[%s]" % (i+1))
    model.add_node(TimeDistributedDense(output_size, activation="softmax"), name="tdd", input=dropout_name)
    logger.info("Added TimeDistributedDense node")
    model.add_output(name="output", input="tdd")
    logger.info("Added Output node")
    logger.info("Created model with following config:\n%s" % model.get_config())
    logger.info("Compiling model with optimizer %s" % optimizer)
    start_time = time.time()
    model.compile(optimizer, {"output": 'categorical_crossentropy'})
    total_time = time.time() - start_time
    logger.info("Model compiled in %.4f seconds." % total_time)
    return model 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def gen_model_brnn_multitask(vocab_size=100, embedding_size=128, maxlen=100, output_size=[6, 96], hidden_layer_size=100, num_hidden_layers = 1, RNN_LAYER_TYPE="LSTM"):
    RNN_CLASS = LSTM
    if RNN_LAYER_TYPE == "GRU":
        RNN_CLASS = GRU
    logger.info("Parameters: vocab_size = %s, embedding_size = %s, maxlen = %s, output_size = %s, hidden_layer_size = %s, " %\
            (vocab_size, embedding_size, maxlen, output_size, hidden_layer_size))
    logger.info("Building Graph model for Bidirectional RNN")
    model = Graph()
    model.add_input(name='input', input_shape=(maxlen,), dtype=int)
    logger.info("Added Input node")
    logger.info("Init Model with vocab_size = %s, embedding_size = %s, maxlen = %s" % (vocab_size, embedding_size, maxlen))
    model.add_node(Embedding(vocab_size, embedding_size, input_length=maxlen, mask_zero=True), name='embedding', input='input')
    logger.info("Added Embedding node")
    model.add_node(Dropout(0.5), name="dropout_0", input="embedding")
    logger.info("Added Dropout Node")
    for i in xrange(num_hidden_layers):
        last_dropout_name = "dropout_%s" % i
        forward_name, backward_name, dropout_name = ["%s_%s" % (k, i + 1) for k in ["forward", "backward", "dropout"]]
        model.add_node(RNN_CLASS(output_dim=hidden_layer_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True), name=forward_name, input=last_dropout_name)
        logger.info("Added %s forward node[%s]" % (RNN_LAYER_TYPE, i+1))
        model.add_node(RNN_CLASS(output_dim=hidden_layer_size, activation='sigmoid', inner_activation='hard_sigmoid', return_sequences=True, go_backwards=True), name=backward_name, input=last_dropout_name)
        logger.info("Added %s backward node[%s]" % (RNN_LAYER_TYPE, i+1))
        model.add_node(Dropout(0.5), name=dropout_name, inputs=[forward_name, backward_name])
        logger.info("Added Dropout node[%s]" % (i+1))
    output_names = []
    for i, output_task_size in enumerate(output_size):
        tdd_name, output_name = "tdd_%s" % i, "output_%s" % i
        model.add_node(TimeDistributedDense(output_task_size, activation="softmax"), name=tdd_name, input=dropout_name)
        logger.info("Added TimeDistributedDense node %s with output_size %s" % (i, output_task_size))
        model.add_output(name=output_name, input=tdd_name)
        output_names.append(output_name)
    logger.info("Added Output node")
    logger.info("Created model with following config:\n%s" % model.get_config())
    logger.info("Compiling model with optimizer %s" % optimizer)
    start_time = time.time()
    model.compile(optimizer, {k: 'categorical_crossentropy' for k in output_names})
    total_time = time.time() - start_time
    logger.info("Model compiled in %.4f seconds." % total_time)
    return model, output_names 

# 需要导入模块: from keras.layers import recurrent [as 别名]
# 或者: from keras.layers.recurrent import GRU [as 别名]
def build(self, input_shape):
        bs, input_length, input_dim = input_shape

        self.controller_input_dim, self.controller_output_dim = controller_input_output_shape(
                input_dim, self.units, self.m_depth, self.n_slots, self.shift_range, self.read_heads,
                self.write_heads)
            
        # Now that we've calculated the shape of the controller, we have add it to the layer/model.
        if self.controller is None:
            self.controller = Dense(
                name = "controller",
                activation = 'linear',
                bias_initializer = 'zeros',
                units = self.controller_output_dim,
                input_shape = (bs, input_length, self.controller_input_dim))
            self.controller.build(input_shape=(self.batch_size, input_length, self.controller_input_dim))
            self.controller_with_state = False


        # This is a fixed shift matrix
        self.C = _circulant(self.n_slots, self.shift_range)

        self.trainable_weights = self.controller.trainable_weights 

        # We need to declare the number of states we want to carry around.
        # In our case the dimension seems to be 6 (LSTM) or 5 (GRU) or 4 (FF),
        # see self.get_initial_states, those respond to:
        # [old_ntm_output] + [init_M, init_wr, init_ww] +  [init_h] (LSMT and GRU) + [(init_c] (LSTM only))
        # old_ntm_output does not make sense in our world, but is required by the definition of the step function we
        # intend to use.
        # WARNING: What self.state_spec does is only poorly understood,
        # I only copied it from keras/recurrent.py.
        self.states = [None, None, None, None]
        self.state_spec = [InputSpec(shape=(None, self.output_dim)),                            # old_ntm_output
                            InputSpec(shape=(None, self.n_slots, self.m_depth)),                # Memory
                            InputSpec(shape=(None, self.read_heads, self.n_slots)),   # weights_read
                            InputSpec(shape=(None, self.write_heads, self.n_slots))]  # weights_write

        super(NeuralTuringMachine, self).build(input_shape)