跟随Chollet机器学习书籍时，KerasTensor传入tf函数报错求助

问题背景 #

运行以下代码时第3行报错：A KerasTensor cannot be used as input to a TensorFlow function，报错原因是直接将Keras张量传入了原生TensorFlow函数，尝试创建新类但不知具体操作。

原代码：

import tensorflow as tf
inputs = keras.Input(shape=(None,), dtype="int64")
embedded = tf.one_hot(inputs, depth=max_tokens)
x = layers.Bidirectional(layers.LSTM(32))(embedded)
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(1, activation="sigmoid")(x)
model = keras.Model(inputs, outputs)
model.compile(optimizer="rmsprop",
              loss="binary_crossentropy",
              metrics=["accuracy"])
model.summary()

解决方案 #

方法一：用Lambda层封装TensorFlow操作 #

将tf.one_hot包装进Keras的Lambda层，让操作兼容KerasTensor：

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

max_tokens = 1000  # 替换为你的实际词汇量
inputs = keras.Input(shape=(None,), dtype="int64")
# 用Lambda层包装tf.one_hot
embedded = layers.Lambda(lambda x: tf.one_hot(x, depth=max_tokens))(inputs)
x = layers.Bidirectional(layers.LSTM(32))(embedded)
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(1, activation="sigmoid")(x)
model = keras.Model(inputs, outputs)
model.compile(optimizer="rmsprop",
              loss="binary_crossentropy",
              metrics=["accuracy"])
model.summary()

方法二：改用Keras Embedding层（更推荐） #

对于文本嵌入场景，Keras的Embedding层是更高效的原生解决方案，替代手动one-hot：

import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

max_tokens = 1000  # 替换为你的实际词汇量
inputs = keras.Input(shape=(None,), dtype="int64")
# 使用Embedding层，input_dim对应词汇量，output_dim设为与one-hot维度一致
embedded = layers.Embedding(input_dim=max_tokens, output_dim=max_tokens)(inputs)
x = layers.Bidirectional(layers.LSTM(32))(embedded)
x = layers.Dropout(0.5)(x)
outputs = layers.Dense(1, activation="sigmoid")(x)
model = keras.Model(inputs, outputs)
model.compile(optimizer="rmsprop",
              loss="binary_crossentropy",
              metrics=["accuracy"])
model.summary()

注意事项 #

• 必须定义max_tokens的具体数值，否则代码会因未定义变量报错
• 补全keras和layers的导入语句，原代码缺少这部分导入

内容的提问来源于stack exchange，提问作者kelvin