保存微调的 Tensorflow 模型时列表索引超出范围

Question

我正在尝试使用 Tensorflow 从 Huggingface 微调预训练的 BERT 模型。一切运行顺利，模型构建和训练没有错误。但是当我尝试保存模型时，它会因错误“IndexError：list index out of range”而停止。我正在使用带有 TPU 的 Google Colab。

任何帮助将不胜感激！

代码：

import tensorflow as tf
from tensorflow.keras import activations, optimizers, losses
from transformers import TFBertModel

def create_model(max_sequence, model_name, num_labels):
    bert_model = TFBertModel.from_pretrained(model_name)
    input_ids = tf.keras.layers.Input(shape=(max_sequence,), dtype=tf.int32, name='input_ids')
    attention_mask = tf.keras.layers.Input((max_sequence,), dtype=tf.int32, name='attention_mask')
    output = bert_model([input_ids, attention_mask])[0]
    output = output[:, 0, :]
    output = tf.keras.layers.Dense(num_labels, activation='sigmoid')(output)
    model = tf.keras.models.Model(inputs=[input_ids, attention_mask], outputs=output)
    return model

with strategy.scope():
  model = create_model(20, 'bert-base-uncased', 1)
  opt = optimizers.Adam(learning_rate=3e-5)
  loss = 'binary_crossentropy'
  model.compile(optimizer=opt, loss=loss, metrics=['accuracy'])

model.fit(tfdataset_train, batch_size=32, epochs=2)
SAVE_PATH = 'path/to/save/location'
model.save(SAVE_PATH)

错误：

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-22-255116b49022> in <module>()
      1 SAVE_PATH = 'path/to/save/location'
----> 2 model.save(SAVE_PATH,save_format='tf')

50 frames
/usr/local/lib/python3.7/dist-packages/transformers/modeling_tf_utils.py in input_processing(func, config, input_ids, **kwargs)
    372                     output[tensor_name] = input
    373                 else:
--> 374                     output[parameter_names[i]] = input
    375             elif isinstance(input, allowed_types) or input is None:
    376                 output[parameter_names[i]] = input

IndexError: list index out of range

用形状绘制的模型： Tensorflow Model

Answer 1

解决办法是改变：

output = bert_model([input_ids, attention_mask])[0]

到

output = bert_model.bert([input_ids, attention_mask])[0]

参考：https://github.com/huggingface/transformers/issues/3627

我赞成您发布的解决方案，但后来我在训练时发现模型存在问题。它不能很好地收敛。

Answer 2

我在尝试保存使用所有三个输入张量（input_ids、token_type_ids 和 attention_mask）的微调 BERT 模型（使用 huggingface 和 tensorflow）时遇到了同样的错误。

对我有用的解决方案是对user_007的答案稍作修改，如下：

output = bert_model.bert(
    input_ids=input_ids, token_type_ids=token_type_ids, attention_mask=attention_mask
)[0]

Answer 3

问题解决了！

移除两个输入层之一（即 attention_mask）解决了这个问题。

工作代码->

import tensorflow as tf
from tensorflow.keras import activations, optimizers, losses
from transformers import TFBertModel

def create_model(max_sequence, model_name, num_labels):
    bert_model = TFBertModel.from_pretrained(model_name)
    input_ids = tf.keras.layers.Input(shape=(max_sequence,), dtype=tf.int32, name='input_ids')
    #attention_mask = tf.keras.layers.Input((max_sequence,), dtype=tf.int32, name='attention_mask')
    #output = bert_model([input_ids, attention_mask])[0]
    output = bert_model([input_ids])[0]
    output = output[:, 0, :]
    output = tf.keras.layers.Dense(num_labels, activation='sigmoid')(output)
    #model = tf.keras.models.Model(inputs=[input_ids, attention_mask], outputs=output)
    model = tf.keras.models.Model(inputs=[input_ids], outputs=output)
    return model

with strategy.scope():
  model = create_model(20, 'bert-base-uncased', 1)
  opt = optimizers.Adam(learning_rate=3e-5)
  loss = 'binary_crossentropy'
  model.compile(optimizer=opt, loss=loss, metrics=['accuracy'])

model.fit(tfdataset_train, batch_size=32, epochs=2)
SAVE_PATH = 'path/to/save/location'
model.save(SAVE_PATH)