Attention

Complexity

Total Complexity

10

Size/Duplication

Total Lines	53
Duplicated Lines	0 %

Importance

Changes	1
Bugs	0	Features	0

5 Methods

Rating	Name	Duplication	Size	Complexity
B	compute_alignments()	0	22	4
A	prepare()	0	5	1
A	compute_context_vector()	0	8	2
A	precompute()	0	5	1
A	__init__()	0	5	2

#!/usr/bin/env python
# -*- coding: utf-8 -*-

from layer import NeuralLayer
import deepy.tensor as T

class Attention(NeuralLayer):


    def __init__(self, hidden_size, input_dim):
        super(Attention, self).__init__("attention")
        self.input_dim = input_dim if input_dim else hidden_size
        self.hidden_size = hidden_size
        self.init(input_dim)

    def prepare(self):
        self.Ua = self.create_weight(self.input_dim, self.hidden_size, "ua")
        self.Wa = self.create_weight(self.hidden_size, self.hidden_size, "wa")
        self.Va = self.create_weight(label="va", shape=(self.hidden_size,))
        self.register_parameters(self.Va, self.Wa, self.Ua)


    def precompute(self, inputs):
        """
        Precompute partial values in the score function.
        """
        return T.dot(inputs, self.Ua)

    def compute_alignments(self, prev_state, precomputed_values, mask=None):
        """
        Compute the alignment weights based on the previous state.
        """

        WaSp = T.dot(prev_state, self.Wa)
        UaH = precomputed_values
        # For test time the UaH will be (time, output_dim)
        if UaH.ndim == 2:
            preact = WaSp[:, None, :] + UaH[None, :, :]
        else:
            preact = WaSp[:, None, :] + UaH
        act = T.activate(preact, 'tanh')
        align_scores = T.dot(act, self.Va)  # ~ (batch, time)
        if mask:
            mask = (1 - mask) * -99.00
            if align_scores.ndim == 3:
                align_scores += mask[None, :]
            else:
                align_scores += mask
        align_weights = T.nnet.softmax(align_scores)
        return align_weights

    def compute_context_vector(self, prev_state, inputs, precomputed_values=None, mask=None):
        """
        Compute the context vector with soft attention.
        """
        precomputed_values = precomputed_values if precomputed_values else self.precompute(inputs)
        align_weights = self.compute_alignments(prev_state, precomputed_values, mask)
        context_vector = T.sum(align_weights[:, :, None] * inputs, axis=1)
        return context_vector