Keras IMDB文本二分类训练报错：无法将NumPy数组转为Tensor

Hey there! Let's fix this error you're hitting. The problem here is straightforward: the raw IMDB sequences you're loading are variable-length integer lists, and Keras can't directly convert those into a Tensor for training—your Embedding layer expects uniformly shaped input data.

解决方案：统一序列长度 #

The fix is to pad (or truncate) all your sequences to the same length using keras.preprocessing.sequence.pad_sequences. Here's your adjusted code with this critical step added:

import tensorflow as tf
from tensorflow import keras
import numpy as np
data = keras.datasets.imdb
(x_train,y_train),(x_test,y_test) = data.load_data()
dictionary = data.get_word_index()
dictionary = {k:(v+3) for k,v in dictionary.items()}
dictionary['<PAD>'] = 0
dictionary['<START>'] = 1
dictionary['<UNKNOWN>'] = 2
dictionary['<UNUSED>'] = 3
dictionary = dict([(v,k) for (k,v) in dictionary.items()])

# ---------------------- 新增的序列统一处理步骤 ----------------------
# 将所有序列调整为256长度，超长截断，不足则用<PAD>（对应值0）补在末尾
max_sequence_length = 256
x_train = keras.preprocessing.sequence.pad_sequences(
    x_train,
    value=dictionary['<PAD>'],
    padding='post',
    maxlen=max_sequence_length
)
x_test = keras.preprocessing.sequence.pad_sequences(
    x_test,
    value=dictionary['<PAD>'],
    padding='post',
    maxlen=max_sequence_length
)
# -------------------------------------------------------------------

model = keras.Sequential([
    keras.layers.Embedding(10000,16),
    keras.layers.GlobalAveragePooling1D(),
    keras.layers.Dense(16,activation='relu'),
    keras.layers.Dense(1,activation='sigmoid')
])
model.compile(
    optimizer='adam',
    loss='binary_crossentropy',
    metrics=['accuracy']
)
print(model.summary())
history = model.fit(x_train,y_train,epochs=50,batch_size=32,verbose=1)
prediction = model.predict(x_test)
print(prediction)

关键步骤解释 #

• pad_sequences 把原本长度不一的列表转换成了二维NumPy数组（形状为 (样本数量, 统一序列长度)），这样TensorFlow就能正常将其转换为模型可接受的Tensor。
• padding='post' 把补全的<PAD> token放在序列末尾，这种方式更适配你用的GlobalAveragePooling1D层，不会影响有效文本的特征提取。
• 你可以根据需求调整max_sequence_length，常用值有100、256或512；如果设置的长度短于部分序列，默认会截断序列末尾的多余token。

内容的提问来源于stack exchange，提问作者Philip Purwoko