def __init__()

in pytext/models/representations/lightconv.py [0:0]

• 62 lines of code
• 10 McCabe index (conditional complexity)

    def __init__(self, config: Config, embed_dim: int, padding_idx: Tensor) -> None:
        super().__init__(config)
        self.padding_idx = padding_idx
        self.pooling_type = config.pooling_type
        self.dropout = nn.Dropout(config.encoder_config.dropout)
        input_embed_dim = embed_dim
        self.embed_scale = math.sqrt(input_embed_dim)  # todo: try with input_embed_dim
        self.max_source_positions = config.encoder_config.max_source_positions
        self.no_token_positional_embeddings = (
            config.encoder_config.no_token_positional_embeddings
        )

        # creating this is also conditional
        self.project_in_dim = (
            Linear(input_embed_dim, config.encoder_config.encoder_embed_dim)
            if config.encoder_config.encoder_embed_dim != input_embed_dim
            else PlaceholderIdentity()
        )

        layers = []
        # Overwrite the config.layer_config.encoder_embed_dim so that it will always match with config.encoder_config.encoder_embed_dim
        config.layer_config.encoder_embed_dim = config.encoder_config.encoder_embed_dim
        for size in config.encoder_kernel_size_list:
            layers.append(create_module(config.layer_config, kernel_size=size))

        self.layers = nn.ModuleList(layers)
        self.embed_layer_norm = LayerNorm(config.encoder_config.encoder_embed_dim)
        self.combine_pos_embed = config.encoder_config.combine_pos_embed.value

        if config.encoder_config.combine_pos_embed == PostionalEmbedCombine.SUM:
            pos_embed_dim = config.encoder_config.encoder_embed_dim
        elif config.encoder_config.combine_pos_embed == PostionalEmbedCombine.CONCAT:
            pos_embed_dim = config.encoder_config.encoder_embed_dim - input_embed_dim
        else:
            raise NotImplementedError

        if not config.encoder_config.no_token_positional_embeddings:
            if (
                config.encoder_config.positional_embedding_type
                == PostionalEmbedType.LEARNED
            ):
                self.embed_positions = PositionalEmbedding(
                    config.encoder_config.max_source_positions,
                    pos_embed_dim,
                    self.padding_idx,
                )
            elif (
                config.encoder_config.positional_embedding_type
                == PostionalEmbedType.SINUSOIDAL
                or config.encoder_config.positional_embedding_type
                == PostionalEmbedType.HYBRID
            ):
                self.embed_positions = SinusoidalPositionalEmbedding(
                    pos_embed_dim,
                    self.padding_idx,
                    init_size=config.encoder_config.max_source_positions,
                    learned_embed=config.encoder_config.positional_embedding_type
                    == PostionalEmbedType.HYBRID,
                )
            else:
                raise NotImplementedError("Positional embedding type not supported")
        else:
            self.embed_positions = PlaceholderIdentity()

        self.normalize = config.encoder_config.encoder_normalize_before
        if self.normalize:
            self.layer_norm = LayerNorm(config.encoder_config.encoder_embed_dim)
        else:
            self.layer_norm = PlaceholderIdentity()

        log_class_usage(__class__)