def __init__()

in optimum/bettertransformer/models/encoder_models.py [0:0]

• 56 lines of code
• 2 McCabe index (conditional complexity)

    def __init__(self, wav2vec2_layer, config):
        r"""
        A simple conversion of the Wav2Vec2EncoderLayer to its `BetterTransformer` implementation.

        Args:
            wav2vec2_layer (`torch.nn.Module`):
                The original `Wav2Vec2EncoderLayer` where the weights needs to be retrieved.
        """
        super().__init__(config)
        super(BetterTransformerBaseLayer, self).__init__()
        # In_proj layer
        self.in_proj_weight = nn.Parameter(
            torch.cat(
                [
                    wav2vec2_layer.attention.q_proj.weight,
                    wav2vec2_layer.attention.k_proj.weight,
                    wav2vec2_layer.attention.v_proj.weight,
                ]
            )
        )
        self.in_proj_bias = nn.Parameter(
            torch.cat(
                [
                    wav2vec2_layer.attention.q_proj.bias,
                    wav2vec2_layer.attention.k_proj.bias,
                    wav2vec2_layer.attention.v_proj.bias,
                ]
            )
        )

        # Out proj layer
        self.out_proj_weight = wav2vec2_layer.attention.out_proj.weight
        self.out_proj_bias = wav2vec2_layer.attention.out_proj.bias

        # Linear layer 1
        self.linear1_weight = wav2vec2_layer.feed_forward.intermediate_dense.weight
        self.linear1_bias = wav2vec2_layer.feed_forward.intermediate_dense.bias

        # Linear layer 2
        self.linear2_weight = wav2vec2_layer.feed_forward.output_dense.weight
        self.linear2_bias = wav2vec2_layer.feed_forward.output_dense.bias

        # Layer norm 1
        self.norm1_eps = wav2vec2_layer.layer_norm.eps
        self.norm1_weight = wav2vec2_layer.layer_norm.weight
        self.norm1_bias = wav2vec2_layer.layer_norm.bias

        # Layer norm 2
        self.norm2_eps = wav2vec2_layer.final_layer_norm.eps
        self.norm2_weight = wav2vec2_layer.final_layer_norm.weight
        self.norm2_bias = wav2vec2_layer.final_layer_norm.bias

        # Model hyper parameters
        self.num_heads = wav2vec2_layer.attention.num_heads
        self.embed_dim = wav2vec2_layer.attention.embed_dim

        # Last step: set the last layer to `False` -> this will be set to `True` when converting the model
        self.is_last_layer = False

        self.original_layers_mapping = {
            "in_proj_weight": ["attention.q_proj.weight", "attention.k_proj.weight", "attention.v_proj.weight"],
            "in_proj_bias": ["attention.q_proj.bias", "attention.k_proj.bias", "attention.v_proj.bias"],
            "out_proj_weight": "attention.out_proj.weight",
            "out_proj_bias": "attention.out_proj.bias",
            "linear1_weight": "feed_forward.intermediate_dense.weight",
            "linear1_bias": "feed_forward.intermediate_dense.bias",
            "linear2_weight": "feed_forward.output_dense.weight",
            "linear2_bias": "feed_forward.output_dense.bias",
            "norm1_weight": "layer_norm.weight",
            "norm1_bias": "layer_norm.bias",
            "norm1_eps": "layer_norm.eps",
            "norm2_weight": "final_layer_norm.weight",
            "norm2_bias": "final_layer_norm.bias",
            "norm2_eps": "final_layer_norm.eps",
        }

        if config.do_stable_layer_norm:
            self.norm_first = True

        self.validate_bettertransformer()