7da9bf6e3f5ed3dcd6f3adc6cf674e11c307741e,baseline/pytorch/tagger/model.py,TaggerModelBase,create,#Any#Any#Any#,51

Before Change

        model = cls()
        model.lengths_key = kwargs.get("lengths_key")
        model.activation_type = kwargs.get("activation", "tanh")
        model.embeddings = embeddings
        model.pdrop = float(kwargs.get("dropout", 0.5))
        model.dropin_values = kwargs.get("dropin", {})
        model.labels = labels

        embed_model = model.init_embed(**kwargs)
        transducer_model = model.init_encoder(embed_model.output_dim, **kwargs)

        use_crf = bool(kwargs.get("crf", False))

        constraint_mask = kwargs.get("constraint_mask")
        if constraint_mask is not None:
            constraint_mask = constraint_mask.unsqueeze(0)

        if use_crf:
            decoder_model = CRF(len(labels), constraint_mask=constraint_mask, batch_first=True)
        else:
            decoder_model = TaggerGreedyDecoder(
                len(labels),
                constraint_mask=constraint_mask,
                batch_first=True,
                reduction=kwargs.get("reduction", "batch")
            )

        model.layers = TagSequenceModel(len(labels), embed_model, transducer_model, decoder_model)
        logger.info(model.layers)
        return model

    def drop_inputs(self, key, x):

After Change

        model.pdrop = float(kwargs.get("dropout", 0.5))
        model.dropin_values = kwargs.get("dropin", {})
        model.labels = labels
        model.create_layers(embeddings, **kwargs)
        return model

    def create_layers(self, embeddings: Dict[str, TensorDef], **kwargs):
        This method defines the model itself, and must be overloaded by derived classes

Instances