无法为KerasTensor构建TypeSpec

Question

运行以下命令：

import tensorflow as tf

from tensorflow.keras import layers
from tensorflow import keras


feat_shape = (50, 66, 3)

inputs = layers.Input(shape=(None,) + feat_shape[1:], dtype=tf.float32)
x = inputs

shape = tf.shape(x)
b, t, f, c = x.get_shape().as_list()
x = layers.Lambda(tf.reshape, arguments=dict(shape=(shape[0], shape[1], shape[2] * shape[3])))(x)
x.set_shape((b, t, f * c))

x = layers.Dense(filters)(x)

lstm_out = layers.LSTM(lstm_units, return_sequences=True, return_state=True)(x)
x = lstm_out[0]

model = keras.Model(inputs=inputs, outputs=x)

抛给我这个错误：

TypeError: Could not build a TypeSpec for <KerasTensor: shape=(None, None, None) dtype=float32 (created by layer 'tf.reshape')> with type KerasTensor

我使用的Tensorflow版本是2.4.0，我非常确定这样的东西应该可以工作。

这里的问题是什么?我如何解决这个问题？

Answer 1

def reshape_(t):
    shape = tf.shape(t)
    B, T, F, C = t.get_shape().as_list()
    t = tf.reshape(t, shape=(shape[0], shape[1], shape[2] * shape[3]))
    t.set_shape((B, T, F * C))
    return t

inputs = layers.Input(shape=(None,) + feat_shape[1:], dtype=tf.float32)
x = inputs
x = layers.Lambda(reshape_)(x)
x = layers.Dense(filters)(x)

lstm_out = layers.LSTM(lstm_units, return_sequences=True, return_state=True)(x)
x = lstm_out[0]

model = keras.Model(inputs=inputs, outputs=x)

​

​