seujung

for i, (rel_img, rel_qst, rel_label, norel_img, norel_qst, norel_label) in enumerate(dataloader_train):
    # Relational data        
    rel_imgs = gluon.utils.split_and_load(rel_img, ctx)
    rel_qsts = gluon.utils.split_and_load(rel_qst, ctx)
    rel_labels = gluon.utils.split_and_load(rel_label, ctx)
        
    # No-relationtional data       
    norel_imgs = gluon.utils.split_and_load(norel_img, ctx)
    norel_qsts = gluon.utils.split_and_load(norel_qst, ctx)
    norel_labels = gluon.utils.split_and_load(norel_label, ctx)        

seujung

class RN_Model(nn.HybridBlock):
    def __init__(self, **kwargs):
        super(RN_Model, self).__init__(**kwargs)
    
        with self.name_scope():
            self.conv = ConvInputModel()
            
            self.g_fc1 = nn.Dense(256, activation='relu', flatten=False)
            self.g_fc2 = nn.Dense(256, activation='relu', flatten=False)
            self.g_fc3 = nn.Dense(256, activation='relu', flatten=False)

seujung

class FCOutputModel(nn.HybridSequential):
    def __init__(self,**kwargs):
        super(FCOutputModel,self).__init__(**kwargs)
        
        with self.name_scope():
            self.add(
                nn.Dense(256, activation='relu'),
                nn.Dropout(0.5),
                nn.Dense(10)
            )

seujung

class ConvInputModel(nn.HybridSequential):
    def __init__(self,**kwargs):
        super(ConvInputModel,self).__init__(**kwargs)                
        with self.name_scope():
            self.add(
                nn.Conv2D(channels=24, kernel_size=3, strides=2, padding=1, activation='relu'),
                nn.BatchNorm(),
                nn.Conv2D(channels=24, kernel_size=3, strides=2, padding=1, activation='relu'),
                nn.BatchNorm(),
                nn.Conv2D(channels=24, kernel_size=3, strides=2, padding=1, activation='relu'),