Question

我正在用tf.keras建立一个非常小的顺序模型来进行文本分类，如下所示：

model = tf.keras.Sequential([
            embedding_layer,
            tf.keras.layers.Conv1D(64, 5, activation='relu', name="conv1d"),
            tf.keras.layers.MaxPool1D(pool_size=4, name="maxpool1d"),
            tf.keras.layers.Flatten(name="flatten"),
            tf.keras.layers.Dense(1, activation='sigmoid', name="dense")
        ])

Layer (type)                 Output Shape              Param #   
=================================================================
inputs (InputLayer)          [(None, 2762)]            0         
_________________________________________________________________
embedding (Embedding)        (None, 2762, 50)          1893250   
_________________________________________________________________
conv1d (Conv1D)              (None, 2758, 64)          16064     
_________________________________________________________________
maxpool1d (MaxPooling1D)     (None, 689, 64)           0         
_________________________________________________________________
flatten (Flatten)            (None, 44096)             0         
_________________________________________________________________
dense (Dense)                (None, 1)                 44097     
=================================================================

在测试时，对于输入中给定的句子，我想找到对预测最有用的单词窗口。我想我必须看一下conv1d层或最大池化层的输出，但是我不确定要寻找什么。我认为这与某些层的最高激活有关。

我已经知道我可以使用以下代码行访问给定层的输出值：

intermediate_layer_model = tf.keras.Model(inputs=model.input, outputs=model.get_layer(name="flatten").output)
    intermediate_output = intermediate_layer_model.predict(x=test_input, workers=1, use_multiprocessing=False)

在CNNN预测中找到最有用的词

0 个答案: