首页 > 解决方案 > 使用预训练的 BERT 嵌入作为带有 tensorflow.keras 的 CNN 的输入会导致 ValueError

问题描述

我是 NLP 和深度学习的新手,所以我遇到了(可能)一个非常基本的问题。

我正在尝试创建一个基于预先训练的 BERT 嵌入作为特征的二进制分类器。到目前为止,我已经成功创建了嵌入,并使用 tensorflow.keras 构建了一个简单的 Sequential() 模型。下面的代码有效:

model = tf.keras.Sequential([
    Dense(4, activation = 'relu', input_shape = (768,)),
    Dense(4, activation = 'relu'),
    Dense(1, activation = 'sigmoid')])

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

我想做的是将这段代码改编为现在的 CNN。但是,当我添加卷积层时,出现错误:

model = tf.keras.Sequential([
    Conv1D(filters = 250, kernel_size = 3, padding='valid', activation='relu', strides=1, input_shape = (768,)),
    GlobalMaxPooling1D(),
    Dense(4, activation = 'relu'),
    Dense(1, activation = 'sigmoid')])

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

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-23-59695050a94e> in <module>()
      3     GlobalMaxPooling1D(),
      4     Dense(4, activation = 'relu'),
----> 5     Dense(1, activation = 'sigmoid')])
      6 
      7 model.compile(optimizer = 'adam',

5 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/input_spec.py in assert_input_compatibility(input_spec, inputs, layer_name)
    178                          'expected ndim=' + str(spec.ndim) + ', found ndim=' +
    179                          str(ndim) + '. Full shape received: ' +
--> 180                          str(x.shape.as_list()))
    181     if spec.max_ndim is not None:
    182       ndim = x.shape.ndims

ValueError: Input 0 of layer conv1d_2 is incompatible with the layer: expected ndim=3, found ndim=2. Full shape received: [None, 768]

这是我使用的数据的样子。

特征:

train_features[0]

array([-4.97862399e-01,  1.49541467e-01,  5.81708886e-02,  1.63668215e-01,
       -2.77605206e-01,  3.57868642e-01,  1.70950562e-01,  2.69330859e-01,
       -3.29369396e-01,  2.12891083e-02, -4.02462274e-01, -1.98120754e-02,
       -2.18944401e-01,  4.34780568e-01, -2.75409579e-01,  2.03015730e-01,...

train_features[0].shape
(768,)

标签:

train_labels.iloc[0:3]
turnout       
0        73446    0
1        53640    1
         16895    1
Name: turnout, dtype: int64

非常感谢任何建议。太感谢了!

标签: pythontensorflowkerasdeep-learningconv-neural-network

解决方案


2D 卷积需要 4D 输入:(batch_size, width1, width2, channels).

您的数据是具有 shape 的单个数组(batch_size, 768)。如果您真的想使用卷积(如果您认为数据中可能存在空间关系),则需要在将其输入模型之前对其进行适当的整形。

1D 卷积需要 3D 输入:(batch_size, length, channels).


推荐阅读