Keras：功能API輸入層應該用於嵌入層？

Question

我正在使用Keras功能API來創建一個神經網絡，它將單詞嵌入層作為句子分類任務的輸入。 但是我的代碼在連接輸入和嵌入層的開始就斷開了。 按照https://medium.com/tensorflow/predicting-the-price-of-wine-with-the-keras-functional-api-and-tensorflow-a95d1c2c1b03上的教程，我有如下代碼：

max_seq_length=100 #i.e., sentence has a max of 100 words 
word_weight_matrix = ... #this has a shape of 9825, 300, i.e., the vocabulary has 9825 words and each is a 300 dimension vector 
deep_inputs = Input(shape=(max_seq_length,))
embedding = Embedding(9825, 300, input_length=max_seq_length,
                          weights=word_weight_matrix, trainable=False)(deep_inputs) # line A
hidden = Dense(targets, activation="softmax")(embedding)
model = Model(inputs=deep_inputs, outputs=hidden)

然后，A行會導致以下錯誤：

ValueError: You called `set_weights(weights)` on layer "embedding_1" with a  weight list of length 9825, but the layer was expecting 1 weights. Provided weights: [[-0.04057981  0.05743935  0.0109863  ...,  0.0072...

我真的不明白這個錯誤意味着什么......

似乎沒有正確定義輸入層...以前當我使用順序模型並且嵌入層定義完全​​相同時，一切正常。 但是當我切換到功能API時，我有這個錯誤。

任何幫助非常感謝，提前感謝

Answer 1

試試這個更新的代碼：你必須在嵌入層使用len(vocabulary) + 1 ！ 和weights=[word_weight_matrix]

max_seq_length=100 #i.e., sentence has a max of 100 words 
word_weight_matrix = ... #this has a shape of 9825, 300, i.e., the vocabulary has 9825 words and each is a 300 dimension vector 
deep_inputs = Input(shape=(max_seq_length,))
embedding = Embedding(9826, 300, input_length=max_seq_length,
                      weights=[word_weight_matrix], trainable=False)(deep_inputs) # line A
hidden = Dense(targets, activation="softmax")(embedding)
model = Model(inputs=deep_inputs, outputs=hidden)