attention.py

import torch
import torch.nn as nn
from torch.nn.utils.weight_norm import weight_norm
from fc import FCNet
from bc import BCNet

class Attention(nn.Module):
    def __init__(self, v_dim, q_dim, num_hid):
        super(Attention, self).__init__()
        self.nonlinear = FCNet([v_dim + q_dim, num_hid])
        self.linear = weight_norm(nn.Linear(num_hid, 1), dim=None)

    def forward(self, v, q):
        """
        v: [batch, k, vdim]
        q: [batch, qdim]
        """
        logits = self.logits(v, q)
        w = nn.functional.softmax(logits, 1)
        return w

    def logits(self, v, q):
        num_objs = v.size(1)
        q = q.unsqueeze(1).repeat(1, num_objs, 1)
        vq = torch.cat((v, q), 2)
        joint_repr = self.nonlinear(vq)
        logits = self.linear(joint_repr)
        return logits


class NewAttention(nn.Module):
    def __init__(self, v_dim, q_dim, num_hid, dropout=0.2):
        super(NewAttention, self).__init__()

        self.v_proj = FCNet([v_dim, num_hid])
        self.q_proj = FCNet([q_dim, num_hid])
        self.dropout = nn.Dropout(dropout)
        self.linear = weight_norm(nn.Linear(q_dim, 1), dim=None)

    def forward(self, v, q):
        """
        v: [batch, k, vdim]
        q: [batch, qdim]
        """
        logits = self.logits(v, q)
        return logits

    def logits(self, v, q):
        batch, k, _ = v.size()
        v_proj = self.v_proj(v) # [batch, k, qdim]
        q_proj = self.q_proj(q).unsqueeze(1).repeat(1, k, 1)
        joint_repr = v_proj * q_proj
        joint_repr = self.dropout(joint_repr)
        logits = self.linear(joint_repr)
        return logits
class BiAttention(nn.Module):
    def __init__(self, x_dim, y_dim, z_dim, glimpse, dropout=[.2,.5]):
        super(BiAttention, self).__init__()

        self.glimpse = glimpse
        self.logits = weight_norm(BCNet(x_dim, y_dim, z_dim, glimpse, dropout=dropout, k=3), \
            name='h_mat', dim=None)

    def forward(self, v, q, v_mask=True):
        """
        v: [batch, k, vdim]
        q: [batch, qdim]
        """
        p, logits = self.forward_all(v, q, v_mask)
        return p, logits

    def forward_all(self, v, q, v_mask=True, logit=False, mask_with=-float('inf')):
        v_num = v.size(1)
        q_num = q.size(1)
        logits = self.logits(v,q) # b x g x v x q

        if v_mask:
            mask = (0 == v.abs().sum(2)).unsqueeze(1).unsqueeze(3).expand(logits.size())
            logits.data.masked_fill_(mask.data, mask_with)

        if not logit:
            p = nn.functional.softmax(logits.view(-1, self.glimpse, v_num * q_num), 2)
            return p.view(-1, self.glimpse, v_num, q_num), logits

        return logits