Attention cell input shapes and Seq2Seq

[lambdaofgod]

How do Attention Cell's call method parameters map to parameters in Seq2Seq?

From what I understand in Seq2Seq decoder

• query corresponds to decoder states: this makes (batch_size, query_length, query_dim) equal to (batch_size, output_sequence_length, num_hidden_decoder)
• key corresponds to encoder outputs: (batch_size, memory_length, key_dim) is (batch_size, input_sequence_length, num_hidden_encoder).
• value is just key, at least for dot and MLP attention cells.

Is my understanding right?

[szhengac]

Your understanding is correct.