pytorch_translate/char_source_hybrid.py [195:260]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class CharCNNEncoder(FairseqEncoder):
    """
    Character-level CNN encoder to generate word representations, as input to
    transformer encoder.
    """

    def __init__(
        self,
        args,
        dictionary,
        embed_tokens,
        num_chars=50,
        embed_dim=32,
        char_cnn_params="[(128, 3), (128, 5)]",
        char_cnn_nonlinear_fn="tanh",
        char_cnn_num_highway_layers=0,
        char_cnn_output_dim=-1,
        use_pretrained_weights=False,
        finetune_pretrained_weights=False,
        weights_file=None,
    ):
        super().__init__(dictionary)

        convolutions_params = literal_eval(char_cnn_params)
        self.char_cnn_encoder = char_encoder.CharCNNModel(
            dictionary,
            num_chars,
            embed_dim,
            convolutions_params,
            char_cnn_nonlinear_fn,
            char_cnn_num_highway_layers,
            char_cnn_output_dim,
            use_pretrained_weights,
            finetune_pretrained_weights,
            weights_file,
        )

        self.embed_tokens = embed_tokens
        token_embed_dim = embed_tokens.embedding_dim
        self.word_layer_norm = nn.LayerNorm(token_embed_dim)

        char_embed_dim = (
            char_cnn_output_dim
            if char_cnn_output_dim != -1
            else sum(out_dim for (out_dim, _) in convolutions_params)
        )
        self.char_layer_norm = nn.LayerNorm(char_embed_dim)
        self.word_dim = char_embed_dim + token_embed_dim
        self.char_scale = math.sqrt(char_embed_dim / self.word_dim)
        self.word_scale = math.sqrt(token_embed_dim / self.word_dim)
        if self.word_dim != args.encoder_embed_dim:
            self.word_to_transformer_embed = fairseq_transformer.Linear(
                self.word_dim, args.encoder_embed_dim
            )

        self.dropout = args.dropout

        self.padding_idx = dictionary.pad()
        self.embed_positions = fairseq_transformer.PositionalEmbedding(
            1024,
            args.encoder_embed_dim,
            self.padding_idx,
            learned=args.encoder_learned_pos,
        )

        self.transformer_encoder_given_embeddings = TransformerEncoderGivenEmbeddings(
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



pytorch_translate/char_source_transformer_model.py [197:262]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
class CharCNNEncoder(FairseqEncoder):
    """
    Character-level CNN encoder to generate word representations, as input to
    transformer encoder.
    """

    def __init__(
        self,
        args,
        dictionary,
        embed_tokens,
        num_chars=50,
        embed_dim=32,
        char_cnn_params="[(128, 3), (128, 5)]",
        char_cnn_nonlinear_fn="tanh",
        char_cnn_num_highway_layers=0,
        char_cnn_output_dim=-1,
        use_pretrained_weights=False,
        finetune_pretrained_weights=False,
        weights_file=None,
    ):
        super().__init__(dictionary)

        convolutions_params = literal_eval(char_cnn_params)
        self.char_cnn_encoder = char_encoder.CharCNNModel(
            dictionary,
            num_chars,
            embed_dim,
            convolutions_params,
            char_cnn_nonlinear_fn,
            char_cnn_num_highway_layers,
            char_cnn_output_dim,
            use_pretrained_weights,
            finetune_pretrained_weights,
            weights_file,
        )

        self.embed_tokens = embed_tokens
        token_embed_dim = embed_tokens.embedding_dim
        self.word_layer_norm = nn.LayerNorm(token_embed_dim)

        char_embed_dim = (
            char_cnn_output_dim
            if char_cnn_output_dim != -1
            else sum(out_dim for (out_dim, _) in convolutions_params)
        )
        self.char_layer_norm = nn.LayerNorm(char_embed_dim)
        self.word_dim = char_embed_dim + token_embed_dim
        self.char_scale = math.sqrt(char_embed_dim / self.word_dim)
        self.word_scale = math.sqrt(token_embed_dim / self.word_dim)
        if self.word_dim != args.encoder_embed_dim:
            self.word_to_transformer_embed = fairseq_transformer.Linear(
                self.word_dim, args.encoder_embed_dim
            )

        self.dropout = args.dropout

        self.padding_idx = dictionary.pad()
        self.embed_positions = fairseq_transformer.PositionalEmbedding(
            1024,
            args.encoder_embed_dim,
            self.padding_idx,
            learned=args.encoder_learned_pos,
        )

        self.transformer_encoder_given_embeddings = TransformerEncoderGivenEmbeddings(
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -