tgt_embed_layer.initialize(1)
second_input = tgt_embed_layer.belongs_to(decoder).compute(
feedback_var, mask=tgt_mask_var)
second_input = DimShuffle(1, 0, 2).compute(second_input)
recurrent_unit = GRU(args.hidden_size,
input_type="sequence",
output_type="sequence",
additional_input_dims=[args.word_embed])
expander_chain = Chain(Dense(600), Dense(args.tgt_vocab_size))
expander_chain.initialize(args.hidden_size)
attention_layer = SoftAttentionalLayer(recurrent_unit)
attention_var = attention_layer.belongs_to(decoder).compute(
hidden_layer,
mask=src_mask_var,
feedback=second_input,
steps=tgt_var.shape(1))
# expander
output_var = expander_chain.belongs_to(expander).compute(attention_var)
cost = TMCostLayer(tgt_var, tgt_mask_var,
args.tgt_vocab_size).compute(output_var)
model = ComputationalGraph(input_dim=[src_var, src_mask_var])
model.training_callbacks.append(training_monitor)
seq_input_var = create_var(T.matrix('seq'),
dim=args.hidden_size * 2,
test_shape=[64, args.hidden_size * 2])
state_var = create_var(T.matrix("s"),
dim=args.hidden_size,
test_shape=[64, args.hidden_size])
input_embed = WordEmbedding(
args.word_embed,
args.tgt_vocab_size).belongs_to(decoder).compute(last_token_var)
recurrent_unit = GRU(args.hidden_size,
input_type="sequence",
output_type="sequence",
additional_input_dims=[input_embed.dim()])
attention_layer = SoftAttentionalLayer(recurrent_unit, test=True)
attention_layer.belongs_to(decoder).initialize(args.hidden_size * 2)
new_state = attention_layer.step({
"UaH":
T.dot(seq_input_var.tensor, attention_layer.Ua),
"feedback":
input_embed.tensor,
"inputs":
seq_input_var.tensor,
"state":
state_var.tensor
})["state"]
decoder_output_var = create_var(new_state, dim=args.hidden_size)
# expander