7db6246fae613a5abf1fbea0b2b987b615ae5a13,visdialch/encoders/lf.py,LateFusionEncoder,forward,#LateFusionEncoder#,42

Before Change

        ques_embed = self.ques_rnn(ques_embed, batch["ques_len"])

        // embed history
        hist = hist.view(-1, hist.size(2))
        hist_embed = self.word_embed(hist)
        hist_embed = self.hist_rnn(hist_embed, batch["hist_len"])

        fused_vector = torch.cat((img, ques_embed, hist_embed), 1)

After Change

        batch_size, num_rounds, max_sequence_length = ques.size()

        // average bottom-up features of all proposals, to form fc7-like features
        if img.dim() == 3:
            img = torch.mean(img, dim=1)  // shape: (batch, img_feature_size)

        // repeat image feature vectors to be provided for every round
        img = img.view(batch_size, 1, self.config["img_feature_size"])
        img = img.repeat(1, num_rounds, 1)
        img = img.view(batch_size * num_rounds, self.config["img_feature_size"])

Instances