def _init_combine_layer(self, attr1, attr2):
config_attr = attr1 + "_" + attr2 + "_modal_combine"
multi_modal_combine_layer = ModalCombineLayer(
self.config[config_attr].type,
getattr(self, self._get_embeddings_attr(attr1)),
getattr(self, self._get_embeddings_attr(attr2)),
**self.config[config_attr].params)
setattr(
self,
attr1 + "_" + attr2 + "_multi_modal_combine_layer",
multi_modal_combine_layer,
)
def __init__(self, image_dim, question_dim, **kwargs):
super().__init__()
combine_type = kwargs["modal_combine"]["type"]
combine_params = kwargs["modal_combine"]["params"]
modal_combine_layer = ModalCombineLayer(combine_type, image_dim,
question_dim, **combine_params)
transform_type = kwargs["transform"]["type"]
transform_params = kwargs["transform"]["params"]
transform_layer = TransformLayer(transform_type,
modal_combine_layer.out_dim,
**transform_params)
normalization = kwargs["normalization"]
self.module = TopDownAttention(modal_combine_layer, transform_layer,
normalization)
if hasattr(self.module, "out_dim"):
self.out_dim = self.module.out_dim