def __init__()

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

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

    def __init__(self, bert_layer, config):
        r"""
        A simple conversion of the Distill-BERTLayer to its `BetterTransformer` implementation.

        Args:
            bert_layer (`torch.nn.Module`):
                The original Distill-BERT Layer 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(
                [
                    bert_layer.attention.q_lin.weight,
                    bert_layer.attention.k_lin.weight,
                    bert_layer.attention.v_lin.weight,
                ]
            )
        )
        self.in_proj_bias = nn.Parameter(
            torch.cat(
                [
                    bert_layer.attention.q_lin.bias,
                    bert_layer.attention.k_lin.bias,
                    bert_layer.attention.v_lin.bias,
                ]
            )
        )

        # Out proj layer
        self.out_proj_weight = bert_layer.attention.out_lin.weight
        self.out_proj_bias = bert_layer.attention.out_lin.bias

        # Linear layer 1
        self.linear1_weight = bert_layer.ffn.lin1.weight
        self.linear1_bias = bert_layer.ffn.lin1.bias

        # Linear layer 2
        self.linear2_weight = bert_layer.ffn.lin2.weight
        self.linear2_bias = bert_layer.ffn.lin2.bias

        # Layer norm 1
        self.norm1_eps = bert_layer.sa_layer_norm.eps
        self.norm1_weight = bert_layer.sa_layer_norm.weight
        self.norm1_bias = bert_layer.sa_layer_norm.bias

        # Layer norm 2
        self.norm2_eps = bert_layer.output_layer_norm.eps
        self.norm2_weight = bert_layer.output_layer_norm.weight
        self.norm2_bias = bert_layer.output_layer_norm.bias

        # Model hyper parameters
        self.num_heads = bert_layer.attention.n_heads
        self.embed_dim = bert_layer.attention.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_lin.weight", "attention.k_lin.weight", "attention.v_lin.weight"],
            "in_proj_bias": ["attention.q_lin.bias", "attention.k_lin.bias", "attention.v_lin.bias"],
            "out_proj_weight": "attention.out_lin.weight",
            "out_proj_bias": "attention.out_lin.bias",
            "linear1_weight": "ffn.lin1.weight",
            "linear1_bias": "ffn.lin1.bias",
            "linear2_weight": "ffn.lin2.weight",
            "linear2_bias": "ffn.lin2.bias",
            "norm1_weight": "sa_layer_norm.weight",
            "norm1_bias": "sa_layer_norm.bias",
            "norm2_weight": "output_layer_norm.weight",
            "norm2_bias": "output_layer_norm.bias",
        }
        self.attention_dropout = config.attention_dropout
        self.dropout = config.dropout
        self.attention_head_size = config.dim // config.n_heads
        self.act_fn_callable = ACT2FN[self.act_fn]

        self.validate_bettertransformer()