LMF.py

"""
paper: Efficient Low-rank Multimodal Fusion with Modality-Specific Factors
From: https://github.com/Justin1904/Low-rank-Multimodal-Fusion
"""
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.nn.parameter import Parameter
from torch.nn.init import xavier_normal_

from models.subNets.FeatureNets import SubNet, TextSubNet

__all__ = ['LMF']

class LMF(nn.Module):
    '''
    Low-rank Multimodal Fusion
    '''

    def __init__(self, args):
        '''
        Args:
            input_dims - a length-3 tuple, contains (audio_dim, video_dim, text_dim)
            hidden_dims - another length-3 tuple, hidden dims of the sub-networks
            text_out - int, specifying the resulting dimensions of the text subnetwork
            dropouts - a length-4 tuple, contains (audio_dropout, video_dropout, text_dropout, post_fusion_dropout)
            output_dim - int, specifying the size of output
            rank - int, specifying the size of rank in LMF
        Output:
            (return value in forward) a scalar value between -3 and 3
        '''
        super(LMF, self).__init__()

        # dimensions are specified in the order of audio, video and text
        self.text_in, self.audio_in, self.video_in = args.feature_dims
        self.text_hidden, self.audio_hidden, self.video_hidden = args.hidden_dims

        self.text_out= self.text_hidden // 2
        self.output_dim = args.num_classes if args.train_mode == "classification" else 1
        self.rank = args.rank

        self.audio_prob, self.video_prob, self.text_prob, self.post_fusion_prob = args.dropouts

        # define the pre-fusion subnetworks
        self.audio_subnet = SubNet(self.audio_in, self.audio_hidden, self.audio_prob)
        self.video_subnet = SubNet(self.video_in, self.video_hidden, self.video_prob)
        self.text_subnet = TextSubNet(self.text_in, self.text_hidden, self.text_out, dropout=self.text_prob)

        # define the post_fusion layers
        self.post_fusion_dropout = nn.Dropout(p=self.post_fusion_prob)
        # self.post_fusion_layer_1 = nn.Linear((self.text_out + 1) * (self.video_hidden + 1) * (self.audio_hidden + 1), self.post_fusion_dim)
        self.audio_factor = Parameter(torch.Tensor(self.rank, self.audio_hidden + 1, self.output_dim))
        self.video_factor = Parameter(torch.Tensor(self.rank, self.video_hidden + 1, self.output_dim))
        self.text_factor = Parameter(torch.Tensor(self.rank, self.text_out + 1, self.output_dim))
        self.fusion_weights = Parameter(torch.Tensor(1, self.rank))
        self.fusion_bias = Parameter(torch.Tensor(1, self.output_dim))

        # init teh factors
        xavier_normal_(self.audio_factor)
        xavier_normal_(self.video_factor)
        xavier_normal_(self.text_factor)
        xavier_normal_(self.fusion_weights)
        self.fusion_bias.data.fill_(0)

    def forward(self, text_x, audio_x, video_x):
        '''
        Args:
            audio_x: tensor of shape (batch_size, audio_in)
            video_x: tensor of shape (batch_size, video_in)
            text_x: tensor of shape (batch_size, sequence_len, text_in)
        '''
        audio_x = audio_x.squeeze(1)
        video_x = video_x.squeeze(1)

        audio_h = self.audio_subnet(audio_x)
        video_h = self.video_subnet(video_x)
        text_h = self.text_subnet(text_x)
        batch_size = audio_h.data.shape[0]

        # next we perform low-rank multimodal fusion
        # here is a more efficient implementation than the one the paper describes
        # basically swapping the order of summation and elementwise product
        # next we perform "tensor fusion", which is essentially appending 1s to the tensors and take Kronecker product
        add_one = torch.ones(size=[batch_size, 1], requires_grad=False).type_as(audio_h).to(text_x.device)
        _audio_h = torch.cat((add_one, audio_h), dim=1)
        _video_h = torch.cat((add_one, video_h), dim=1)
        _text_h = torch.cat((add_one, text_h), dim=1)
        
        fusion_audio = torch.matmul(_audio_h, self.audio_factor)
        fusion_video = torch.matmul(_video_h, self.video_factor)
        fusion_text = torch.matmul(_text_h, self.text_factor)
        fusion_zy = fusion_audio * fusion_video * fusion_text

        # output = torch.sum(fusion_zy, dim=0).squeeze()
        # use linear transformation instead of simple summation, more flexibility
        output = torch.matmul(self.fusion_weights, fusion_zy.permute(1, 0, 2)).squeeze() + self.fusion_bias
        output = output.view(-1, self.output_dim)

        res = {
            'Feature_t': text_h,
            'Feature_a': audio_h,
            'Feature_v': video_h,
            'Feature_f': fusion_zy.permute(1, 0, 2).squeeze(),
            'M': output
        }
        return res