Add attention layer to Seq2Seq model - python-3.x

Question

I have build a Seq2Seq model of encoder-decoder. I want to add an attention layer to it. I tried adding attention layer through this but it didn't help.
Here is my initial code without attention
# Encoder
encoder_inputs = Input(shape=(None,))
enc_emb = Embedding(num_encoder_tokens, latent_dim, mask_zero = True)(encoder_inputs)
encoder_lstm = LSTM(latent_dim, return_state=True)
encoder_outputs, state_h, state_c = encoder_lstm(enc_emb)
# We discard `encoder_outputs` and only keep the states.
encoder_states = [state_h, state_c]
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,))
dec_emb_layer = Embedding(num_decoder_tokens, latent_dim, mask_zero = True)
dec_emb = dec_emb_layer(decoder_inputs)
# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_outputs, _, _ = decoder_lstm(dec_emb,
initial_state=encoder_states)
decoder_dense = Dense(num_decoder_tokens, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)
# Define the model that will turn
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
model.summary()
And this is the code after I added attention layer in decoder (the encoder layer is same as in initial code)
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,))
dec_emb_layer = Embedding(num_decoder_tokens, latent_dim, mask_zero = True)
dec_emb = dec_emb_layer(decoder_inputs)
# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
attention = dot([decoder_lstm, encoder_lstm], axes=[2, 2])
attention = Activation('softmax')(attention)
context = dot([attention, encoder_lstm], axes=[2,1])
decoder_combined_context = concatenate([context, decoder_lstm])
decoder_outputs, _, _ = decoder_combined_context(dec_emb,
initial_state=encoder_states)
decoder_dense = Dense(num_decoder_tokens, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)
# Define the model that will turn
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
model.summary()
While doing this, I got an error
Layer dot_1 was called with an input that isn't a symbolic tensor. Received type: <class 'keras.layers.recurrent.LSTM'>. Full input: [<keras.layers.recurrent.LSTM object at 0x7f8f77e2f3c8>, <keras.layers.recurrent.LSTM object at 0x7f8f770beb70>]. All inputs to the layer should be tensors.
Can someone please help in fitting an attention layer in this architecture?

Answer 1

the dot products need to be computed on tensor outputs... in encoder you correctly define the encoder_output, in decoder you have to add decoder_outputs, state_h, state_c = decoder_lstm(enc_emb, initial_state=encoder_states)
the dot products now are
attention = dot([decoder_outputs, encoder_outputs], axes=[2, 2])
attention = Activation('softmax')(attention)
context = dot([attention, encoder_outputs], axes=[2,1])
the concatenation doesn't need initial_states. you have to define it in your rnn layer: decoder_outputs, state_h, state_c = decoder_lstm(enc_emb, initial_state=encoder_states)
here the full example
ENCODER + DECODER
# dummy variables
num_encoder_tokens = 30
num_decoder_tokens = 10
latent_dim = 100
encoder_inputs = Input(shape=(None,))
enc_emb = Embedding(num_encoder_tokens, latent_dim, mask_zero = True)(encoder_inputs)
encoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
encoder_outputs, state_h, state_c = encoder_lstm(enc_emb)
# We discard `encoder_outputs` and only keep the states.
encoder_states = [state_h, state_c]
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,))
dec_emb_layer = Embedding(num_decoder_tokens, latent_dim, mask_zero = True)
dec_emb = dec_emb_layer(decoder_inputs)
# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_outputs, _, _ = decoder_lstm(dec_emb,
initial_state=encoder_states)
decoder_dense = Dense(num_decoder_tokens, activation='softmax')
decoder_outputs = decoder_dense(decoder_outputs)
# Define the model that will turn
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
model = Model([encoder_inputs, decoder_inputs], decoder_outputs)
model.summary()
DECODER w\ ATTENTION
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,))
dec_emb_layer = Embedding(num_decoder_tokens, latent_dim, mask_zero = True)
dec_emb = dec_emb_layer(decoder_inputs)
# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_outputs, state_h, state_c = decoder_lstm(dec_emb, initial_state=encoder_states)
attention = dot([decoder_outputs, encoder_outputs], axes=[2, 2])
attention = Activation('softmax')(attention)
context = dot([attention, encoder_outputs], axes=[2,1])
decoder_outputs = concatenate([context, decoder_outputs])
decoder_dense = Dense(num_decoder_tokens, activation='softmax')(decoder_outputs)
# Define the model that will turn
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
model = Model([encoder_inputs, decoder_inputs], decoder_dense)
model.summary()

Answer 2

Marco's answer from above works, but one has to change the lines that involve the dot function in the second chunk. It takes one positional argument as in tensorflow's example here.
Finally, the chunk bellow includes the correction and will work:
# Set up the decoder, using `encoder_states` as initial state.
decoder_inputs = Input(shape=(None,))
dec_emb_layer = Embedding(num_decoder_tokens, latent_dim, mask_zero = True)
dec_emb = dec_emb_layer(decoder_inputs)
# We set up our decoder to return full output sequences,
# and to return internal states as well. We don't use the
# return states in the training model, but we will use them in inference.
decoder_lstm = LSTM(latent_dim, return_sequences=True, return_state=True)
decoder_outputs, state_h, state_c = decoder_lstm(dec_emb, initial_state=encoder_states)
attention = Dot(axes=[2, 2])([decoder_outputs, encoder_outputs])
attention = Activation('softmax')(attention)
context = Dot(axes=[2,1])([attention, encoder_outputs])
decoder_outputs = concatenate([context, decoder_outputs])
decoder_dense = Dense(num_decoder_tokens, activation='softmax')(decoder_outputs)
# Define the model that will turn
# `encoder_input_data` & `decoder_input_data` into `decoder_target_data`
model = Model([encoder_inputs, decoder_inputs], decoder_dense)
model.summary()