调用model.fit()抛出ValueError张量转换错误，求排查原因

问题背景与代码重现 #

先通过BERT生成嵌入矩阵，再以此为初始权重构建LSTM模型，调用model.fit()时触发类型转换错误。

生成BERT嵌入矩阵的代码 #

# Load pre-trained model tokenizer and model
tokenizer = BertTokenizer.from_pretrained('bert-base-multilingual-cased')
model = BertModel.from_pretrained('bert-base-multilingual-cased')

# Define batch size
batch_size = 1

# Tokenize and encode input data in batches
encoded_inputs = []
for i in range(0, len(labeled_data), batch_size):
    inputs = labeled_data[i:i+batch_size]
    encoded_inputs.append(tokenizer.batch_encode_plus(inputs, padding=True, truncation=True, return_tensors="pt"))

# Generate embeddings for each batch
embeddings_new = []
for encoded_input in tqdm(encoded_inputs):
    with torch.no_grad():
        model_output = model(**encoded_input)
    batch_embeddings = model_output.last_hidden_state.mean(dim=1)
    embeddings_new.append(batch_embeddings)

embeddings_new = tf.concat(embeddings_new, axis=0) 
embedding_matrix = model.embeddings.word_embeddings.weight
embedding_matrix = embedding_matrix.cpu().detach().numpy()
embed_tensor = tf.convert_to_tensor(embedding_matrix, dtype=tf.float32)

构建LSTM模型的代码 #

lstm_out1 = 150
embed_dim = 768
   
model = Sequential()
model.add(Embedding(embedding_matrix.shape[0], embed_dim, weights=[embed_tensor], input_length=50, trainable=False))
model.add(LSTM(lstm_out1, dropout=0.2, recurrent_dropout=0.2))

model.add(Dense(64, activation='relu'))

model.add(Dense(1, activation='sigmoid'))

adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)

model.compile(loss='binary_crossentropy',
                  optimizer=adam,
                  metrics=['accuracy'])

model.summary()

触发错误的调用与错误信息 #

调用代码：

model.fit(tokenized_sentences, labels, batch_size=5, epochs=1, shuffle=True)

错误信息：

ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type list).

输入数据说明 #

• tokenized_sentences是二维整数列表，由以下代码生成：

tokenized_sentences = []
for sentence in labeled_data:
    # Apply the tokenizer to each sentence to obtain its tokens
    tokens = tokenizer.encode(sentence, add_special_tokens=True)
    # Append the tokenized sentence to the list
    tokenized_sentences.append(tokens)

示例内容：[[101, 10406, 10161, ..., 102], ...]

• labels是一维整数列表，示例内容：[1, 0, 1, 1, 0, ..., 0]

错误原因 #

1. 序列长度不统一：tokenized_sentences中每个子列表（单句token序列）的长度不一致，但LSTM模型的Embedding层设置了input_length=50，要求输入必须是固定长度的规则张量。TensorFlow无法将长度参差不齐的嵌套列表直接转换为符合要求的张量，因此抛出类型转换错误。
2. 数据格式不符合要求：嵌套列表不属于TensorFlow默认支持的输入格式，需要先转换为固定长度的NumPy数组或TensorFlow张量。

解决方法 #

方法一：tokenize阶段直接生成固定长度序列 #

修改tokenize代码，利用BERT tokenizer的参数直接生成符合input_length要求的序列：

tokenized_sentences = []
max_len = 50  # 与Embedding层的input_length保持一致
for sentence in labeled_data:
    tokens = tokenizer.encode(
        sentence, 
        add_special_tokens=True, 
        max_length=max_len, 
        padding='max_length', 
        truncation=True
    )
    tokenized_sentences.append(tokens)
# 转换为NumPy数组
import numpy as np
tokenized_sentences = np.array(tokenized_sentences)
labels = np.array(labels)

方法二：手动补全/截断已有序列 #

如果已经生成了tokenized_sentences，可以手动调整为固定长度：

max_len = 50
processed_sentences = []
for tokens in tokenized_sentences:
    if len(tokens) < max_len:
        # 补零至指定长度
        padded_tokens = tokens + [0]*(max_len - len(tokens))
    else:
        # 截断至指定长度
        padded_tokens = tokens[:max_len]
    processed_sentences.append(padded_tokens)
# 转换为数组
tokenized_sentences = np.array(processed_sentences)
labels = np.array(labels)

方法三：使用Keras内置工具处理序列 #

利用pad_sequences快速统一序列长度：

from tensorflow.keras.preprocessing.sequence import pad_sequences

max_len = 50
tokenized_sentences = pad_sequences(
    tokenized_sentences, 
    maxlen=max_len, 
    padding='post', 
    truncating='post'
)
labels = np.array(labels)

完成上述处理后，再调用model.fit()即可正常运行。

内容的提问来源于stack exchange，提问作者Debbie