def __init__()

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

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

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

        Args:
            vit_layer (`torch.nn.Module`):
                The original `ViTLayer` 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(
                [
                    vit_layer.attention.attention.query.weight,
                    vit_layer.attention.attention.key.weight,
                    vit_layer.attention.attention.value.weight,
                ]
            )
        )
        self.in_proj_bias = nn.Parameter(
            torch.cat(
                [
                    vit_layer.attention.attention.query.bias,
                    vit_layer.attention.attention.key.bias,
                    vit_layer.attention.attention.value.bias,
                ]
            )
        )

        # Out proj layer
        self.out_proj_weight = vit_layer.attention.output.dense.weight
        self.out_proj_bias = vit_layer.attention.output.dense.bias

        # Linear layer 1
        self.linear1_weight = vit_layer.intermediate.dense.weight
        self.linear1_bias = vit_layer.intermediate.dense.bias

        # Linear layer 2
        self.linear2_weight = vit_layer.output.dense.weight
        self.linear2_bias = vit_layer.output.dense.bias

        # Layer norm 1
        self.norm1_eps = vit_layer.layernorm_before.eps
        self.norm1_weight = vit_layer.layernorm_before.weight
        self.norm1_bias = vit_layer.layernorm_before.bias

        # Layer norm 2
        self.norm2_eps = vit_layer.layernorm_after.eps
        self.norm2_weight = vit_layer.layernorm_after.weight
        self.norm2_bias = vit_layer.layernorm_after.bias

        # Model hyper parameters
        self.num_heads = vit_layer.attention.attention.num_attention_heads
        self.embed_dim = int(vit_layer.attention.attention.attention_head_size * self.num_heads)

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

        self.original_layers_mapping = {
            "in_proj_weight": [
                "attention.attention.query.weight",
                "attention.attention.key.weight",
                "attention.attention.value.weight",
            ],
            "in_proj_bias": [
                "attention.attention.query.bias",
                "attention.attention.key.bias",
                "attention.attention.value.bias",
            ],
            "out_proj_weight": "attention.output.dense.weight",
            "out_proj_bias": "attention.output.dense.bias",
            "linear1_weight": "intermediate.dense.weight",
            "linear1_bias": "intermediate.dense.bias",
            "linear2_weight": "output.dense.weight",
            "linear2_bias": "output.dense.bias",
            "norm1_weight": "layernorm_before.weight",
            "norm1_bias": "layernorm_before.bias",
            "norm2_weight": "layernorm_after.weight",
            "norm2_bias": "layernorm_after.bias",
        }

        self.validate_bettertransformer()