Has anyone implement this networks proposed in paper Hierarchical Attention Networks for Document Classification?

[460130107]

The paper URL is https://www.researchgate.net/publication/305334401_Hierarchical_Attention_Networks_for_Document_Classification

[124399839]

You can contact the author to request the code.

[124399839]

欢迎加群367355275讨论。。。。

[bkj]

I think that the HN-MAX is (roughly)

max_sents = # maximum number of sentences per document
max_words = # maximum number of words per sentence

x = Input(shape=(max_sents, max_words,))

emb_words = TimeDistributed(Embedding(input_dim=max_features, output_dim=200, mask_zero=True))(x)

emb_sents = TimeDistributed(Bidirectional(GRU(50, consume_less='gpu', return_sequences=True)))(emb_words)
emb_sents = TimeDistributed(GlobalMaxPooling1D())(emb_sents)

emb_docs = Bidirectional(GRU(50, consume_less='gpu', return_sequences=True))(emb_sents)
emb_docs = GlobalMaxPooling1D()(emb_docs)

prediction = Dense(y_train.shape[1], activation='softmax')(emb_docs)
model = Model(input=x, output=prediction)
model.compile(loss='categorical_crossentropy', optimizer=SGD(momentum=0.9), metrics=['accuracy'])

Turning this into the HN-AVG variant is fairly straightforward, and to the HN-ATT you'd have to write a little attention unit, but that shouldn't be particularly difficult I don't think.

[richliao]

I have the paper implemented. Here is the blog https://richliao.github.io/ and the code is at https://github.com/richliao/textClassifier.

[RahulKulhari]

pytorch implementation of paper in pytorch https://github.com/EdGENetworks/attention-networks-for-classification by @EdGENetworks and @Sandeep42.

@zcyang please provide your feedback on implementation.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs, but feel free to re-open a closed issue if needed.