d1f352a71ba7056a4dc3adb6698e0e792d78b0b2,models/AttModel.py,AttModel,_sample_beam,#AttModel#Any#Any#Any#Any#,116

Before Change

        // embed fc and att feats
        fc_feats = self.fc_embed(fc_feats)
        _att_feats = self.att_embed(att_feats.view(-1, self.att_feat_size))
        att_feats = _att_feats.view(*(att_feats.size()[:-1] + (self.rnn_size,)))

        // Project the attention feats first to reduce memory and computation comsumptions.
        p_att_feats = self.ctx2att(att_feats.view(-1, self.rnn_size))
        p_att_feats = p_att_feats.view(*(att_feats.size()[:-1] + (self.att_hid_size,)))

        assert beam_size <= self.vocab_size + 1, "lets assume this for now, otherwise this corner case causes a few headaches down the road. can be dealt with in future if needed"
        seq = torch.LongTensor(self.seq_length, batch_size).zero_()
        seqLogprobs = torch.FloatTensor(self.seq_length, batch_size)

After Change

        return logprobs, state

    def _sample_beam(self, fc_feats, att_feats, att_masks=None, opt={}):
        beam_size = opt.get("beam_size", 10)
        batch_size = fc_feats.size(0)

        // embed fc and att feats
        fc_feats = self.fc_embed(fc_feats)
        att_feats = pack_wrapper(self.att_embed, att_feats, att_masks)

        // Project the attention feats first to reduce memory and computation comsumptions.
        p_att_feats = self.ctx2att(att_feats)

Instances