fcea9fee573e854177b4b9af1cfd1b20029ed21e,models/common/char_model.py,CharacterModel,forward,#CharacterModel#Any#Any#Any#Any#Any#,25

Before Change

    def forward(self, chars, chars_mask, word_orig_idx, sentlens, wordlens):
        embs = self.dropout(self.char_emb(chars))
        char_reps = self.charlstm(embs, wordlens, hx=(self.charlstm_h_init.expand(self.args["char_num_layers"], embs.size(0), self.args["char_hidden_dim"]).contiguous(), self.charlstm_c_init.expand(self.args["char_num_layers"], embs.size(0), self.args["char_hidden_dim"]).contiguous()))[0]

        // attention
        weights = torch.sigmoid(self.char_attn(self.dropout(char_reps))).masked_fill(chars_mask.unsqueeze(2), 0)
        weights = weights.transpose(1, 2)

        res = weights.bmm(char_reps).squeeze(1)
        res = tensor_unsort(res, word_orig_idx)

After Change

        char_reps = self.charlstm(embs, wordlens, hx=(self.charlstm_h_init.expand(self.args["char_num_layers"], batch_size, self.args["char_hidden_dim"]).contiguous(), self.charlstm_c_init.expand(self.args["char_num_layers"], batch_size, self.args["char_hidden_dim"]).contiguous()))[0]

        // attention
        weights = torch.sigmoid(self.char_attn(self.dropout(char_reps.data)))

        char_reps = PackedSequence(char_reps.data * weights, char_reps.batch_sizes)
        char_reps, _ = pad_packed_sequence(char_reps, batch_first=True)
        res = char_reps.sum(1)
        res = tensor_unsort(res, word_orig_idx)

Instances