Question

经过Phillippe Remy's PhasedLSTMCell example的努力，我想知道是否可以使用更高级别的Tensorflow Keras API在各层中使用PhasedLSTM单元。

对于tf.keras.layers.RNN docs，RNN层应接受 “一个RNN单元实例或一个RNN单元实例列表[，]”，所以我希望这样的方法可以起作用：

plstm = tf.keras.layers.RNN(tf.contrib.rnn.PhasedLSTMCell(32))([time, reshaped])

但是它抛出：

ValueError: An `initial_state` was passed that is not compatible with `cell.state_size`. Received `state_spec`=[InputSpec(shape=(None, 784, 1), ndim=3)]; however `cell.state_size` is [32, 32]

（下面是完整的MWE。）

是否可以在图层中使用PhasedLSTMCell？否则我会搞砸了吗？

import numpy as np
import tensorflow as tf

def build_evenly_space_t(mnist_img_size, batch_size):
        return np.reshape(np.tile(np.array(range(mnist_img_size)), (batch_size, 1)), (batch_size, mnist_img_size, 1))

mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

# This works
imgs = tf.keras.layers.Input(shape=(28, 28))
reshaped = tf.keras.layers.Reshape((28 * 28, 1), input_shape=(28*28,))(imgs)
lstm = tf.keras.layers.RNN(tf.nn.rnn_cell.LSTMCell(32))(reshaped)
out = tf.keras.layers.Dense(10, activation=tf.nn.softmax)(lstm)
this_model_will = tf.keras.models.Model(inputs=imgs, outputs=out)
this_model_will.compile(optimizer='adam',
                            loss='sparse_categorical_crossentropy',
                            metrics=['accuracy'])

# this does not
time = tf.keras.layers.Input(shape=(1,))
imgs = tf.keras.layers.Input(shape=(28, 28))
reshaped = tf.keras.layers.Reshape((28 * 28, 1), input_shape=(28*28,))(imgs)
plstm = tf.keras.layers.RNN(tf.contrib.rnn.PhasedLSTMCell(32))([time, reshaped])
out = tf.keras.layers.Dense(10, activation=tf.nn.softmax)(plstm)
this_model_will_not = tf.keras.models.Model(inputs=[time, imgs], outputs=out)

this_model_will_not.compile(optimizer='adam',
                            loss='sparse_categorical_crossentropy',
                            metrics=['accuracy'])

Tensorflow的PhasedLSTMCell可以在层中使用吗？

0 个答案: