def __init__()

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

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

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

        Args:
            bart_layer (`torch.nn.Module`):
                The original `BartEncoderLayer` 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(
                [
                    bart_layer.self_attn.q_proj.weight,
                    bart_layer.self_attn.k_proj.weight,
                    bart_layer.self_attn.v_proj.weight,
                ]
            )
        )

        self.in_proj_bias = nn.Parameter(
            torch.cat(
                [
                    bart_layer.self_attn.q_proj.bias,
                    bart_layer.self_attn.k_proj.bias,
                    bart_layer.self_attn.v_proj.bias,
                ]
            )
        )

        # Out proj layer
        self.out_proj_weight = bart_layer.self_attn.out_proj.weight
        self.out_proj_bias = bart_layer.self_attn.out_proj.bias

        # Linear layer 1
        self.linear1_weight = bart_layer.fc1.weight
        self.linear1_bias = bart_layer.fc1.bias

        # Linear layer 2
        self.linear2_weight = bart_layer.fc2.weight
        self.linear2_bias = bart_layer.fc2.bias

        # Layer norm 1
        self.norm1_eps = bart_layer.self_attn_layer_norm.eps
        self.norm1_weight = bart_layer.self_attn_layer_norm.weight
        self.norm1_bias = bart_layer.self_attn_layer_norm.bias

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

        # Model hyper parameters
        self.num_heads = bart_layer.self_attn.num_heads
        self.embed_dim = bart_layer.self_attn.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": ["self_attn.q_proj.weight", "self_attn.k_proj.weight", "self_attn.v_proj.weight"],
            "in_proj_bias": ["self_attn.q_proj.bias", "self_attn.k_proj.bias", "self_attn.v_proj.bias"],
            "out_proj_weight": "self_attn.out_proj.weight",
            "out_proj_bias": "self_attn.out_proj.bias",
            "linear1_weight": "fc1.weight",
            "linear1_bias": "fc1.bias",
            "linear2_weight": "fc2.weight",
            "linear2_bias": "fc2.bias",
            "norm1_eps": "self_attn_layer_norm.eps",
            "norm1_weight": "self_attn_layer_norm.weight",
            "norm1_bias": "self_attn_layer_norm.bias",
            "norm2_eps": "final_layer_norm.eps",
            "norm2_weight": "final_layer_norm.weight",
            "norm2_bias": "final_layer_norm.bias",
        }
        self.dropout = config.attention_dropout
        self.activation_dropout = config.activation_dropout
        self.attention_head_size = config.d_model // config.encoder_attention_heads
        self.act_fn_callable = ACT2FN[self.act_fn]

        self.validate_bettertransformer()