def __init__(self,
seq_dim,
img_dim,
hidden_dim,
output_dim,
bili_norm_type,
bili_affine,
bili_non_linear,
bili_dropout=None,
**kwargs):
super(HdmProdBilinearFusion, self).__init__()
self.SeqTrans = nn.Linear(seq_dim, hidden_dim)
self.ImgTrans = nn.Linear(img_dim, hidden_dim)
self.OutTrans = nn.Linear(hidden_dim, output_dim)
if bili_dropout is None:
self.Dropout = nn.Identity()
else:
self.Dropout = nn.Dropout(bili_dropout)
self.Norm = buildNormalization(norm_name=bili_norm_type,
feature_shape=output_dim,
affine=bili_affine,
norm_name_map={'bn': 'bn_1d'})
self.NonLinear = buildActivation(bili_non_linear)
def __init__(self,
input_dim,
dnn_hidden_dims,
dnn_activations,
dnn_dropouts,
dnn_norm_type=None,
feature_dims={},
**kwargs):
super().__init__(input_dim, dnn_hidden_dims, dnn_activations,
dnn_dropouts, **kwargs)
try:
seq_dim = feature_dims['seq']
except KeyError as e:
raise ValueError(
f"[DNNCatRetCatFusion] Unable to fetch sequence dim from param: {str(e)}"
)
cat_dim = seq_dim + self._DnnOutputDim
self.Norm = buildNormalization(dnn_norm_type,
cat_dim,
norm_name_map={
'bn': 'bn_1d',
'in': 'in_1d'
})
def __init__(self,
seq_dim,
img_dim,
output_dim,
bili_norm_type,
bili_affine,
bili_non_linear,
bili_dropout=None,
**kwargs):
super(BilinearFusion, self).__init__()
self.Trans = nn.Bilinear(seq_dim, img_dim, output_dim, bias=False)
if bili_dropout is None:
self.Dropout = nn.Identity()
else:
self.Dropout = nn.Dropout(bili_dropout)
self.Norm = buildNormalization(norm_name=bili_norm_type,
feature_shape=output_dim,
affine=bili_affine,
norm_name_map={
'bn': 'bn_1d',
'ln': "ln_1d"
})
self.NonLinear = buildActivation(bili_non_linear)
def __init__(self,
input_dim,
dnn_hidden_dims,
dnn_activations,
dnn_dropouts,
dnn_norm_type=None,
feature_dims={},
**kwargs):
super().__init__(input_dim, dnn_hidden_dims, dnn_activations,
dnn_dropouts, **kwargs)
cat_dim = self._DnnOutputDim
self.Norm = buildNormalization(dnn_norm_type,
cat_dim,
norm_name_map={
'bn': 'bn_1d',
'in': 'in_1d'
})