use attention_3d_block in many to many mapping

[Opdoop]

Hi, I'm beginner of Keras and tring to use attention_3d_block in translation module.
I have input of 5 sentences, each sentences has padding to 6 words, each word is presented in 620 dim(as embedding dim).
And the output is 5 sentences, sentences padding to 9 words, and word is presented in 1-of-k in 30 dim(as vocabulary size)
How to use attention_3d_block in this scenario as the LSTM is many to many?

[Opdoop]

I add a TimeDistribute(Dense()) to fit input_dim, but how to fit the time_step.
As the module structure is below:

Get Error when run
ValueError: Error when checking target: expected activation_1 to have shape (None, 6, 30) but got array with shape (5, 9, 30)

[philipperemy]

@Opdoop can you please share the minimal code to reproduce your error?

[Opdoop]

@philipperemy
Data prepare is :

    input_text = ['中国 的 首都 是 北京'
                  , '日本 的 首都 是 东京'
                  , '美国 的 首都 是 华盛顿'
                  , '英国 的 首都 是 伦敦'
                  , '德国 的 首都 是 柏林']
    
    tar_text = ['Beijing is the capital of China'
                , 'Tokyo is the capital of Japan'
                , 'Washington is the capital of the United States'
                , 'London is the capital of England'
                , 'Berlin is the capital of Germany']
                
    input_list = []
    tar_list = []
    END = ' EOS'
    for tmp_input in input_text:
        tmp_input = tmp_input+END
        input_list.append(tokenize(tmp_input))
    for tmp_tar in tar_text:
        tmp_tar = tmp_tar+END
        tar_list.append(tokenize(tmp_tar))
    vocab = sorted(reduce(lambda x, y: x | y, (set(tmp_list) for tmp_list in input_list + tar_list)))

    vocab_size = len(vocab) + 1  # keras embedding need len(vocab)+1
    input_maxlen = max(map(len, (x for x in input_list)))
    tar_maxlen = max(map(len, (x for x in tar_list)))
    output_dim = vocab_size
    hidden_dim = 1000
    INPUT_DIM = hidden_dim #senten length
    TIME_STEPS = input_maxlen #

    word_to_idx = dict((c, i + 1) for i, c in enumerate(vocab))  
    idx_to_word = dict((i + 1, c) for i, c in enumerate(vocab))  

    inputs_train, tars_train = vectorize_stories(input_list, tar_list, word_to_idx, input_maxlen, tar_maxlen, vocab_size)
    #inputs_trains shape is (5,6) and tars_tratin shape is (5,9,30)

Module is :

def model_attention_applied_before_lstm():
    inputs = Input(shape=(input_maxlen,))
    embed = Embedding(input_dim=vocab_size,
                              output_dim=620,
                              input_length=input_maxlen)(inputs)
    attention_mul = LSTM(hidden_dim, return_sequences=True)(embed)    
    attention_mul = attention_3d_block(attention_mul)
    attention_mul = LSTM(500, return_sequences=True)(attention_mul)
    output = TimeDistributed(Dense(output_dim, activation='sigmoid'),input_shape=(tar_maxlen, output_dim))(attention_mul)
    #output = TimeDistributed(Dense(output_dim))(output)
    #output = Dense(output_dim, activation='sigmoid')(attention_mul)
    #output = RepeatVector(tar_maxlen)(output)
    #output = Permute((tar_maxlen,output_dim),name='reshapeLayer')(output)
    output = Activation('softmax')(output)
    model = Model(input=[inputs], output=output)
    return model

And run the model as:

    m = model_attention_applied_before_lstm()
    m.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy'])
    print(m.summary())

    m.fit(inputs_train, tars_train, epochs=10, batch_size=3, validation_split=0.1)

[philipperemy]

In that case, what you need to have is a sequence to sequence attention. This project does not support this.

• http://distill.pub/2016/augmented-rnns/
• https://www.tensorflow.org/tutorials/seq2seq
• https://github.com/harvardnlp/seq2seq-attn

[Opdoop]

Ooooooooh. Thanks a lot.
Do you know any sequence to sequence attention implement using keras?

[philipperemy]

I'm not sure about this. This project:

• https://github.com/farizrahman4u/seq2seq

is the most famous seq2seq in Keras. Maybe there's an attention mechanism there. To be checked

[philipperemy]

The Seq2seq framework includes a ready made attention model which does the same.

Yes it does have one!