多尺度融合的实现问题

[czisok]

您好，看了您的论文和代码，您在论文中提到multi-scale fusion mechanism at the end of DeepSTN+ to improve both the accuracy of prediction and the stability of the model.
但是在代码中，您是将ResPlus最后输出的结果进行了激活、正则化、卷积、正则化的操作，
` if is_plus:
for i in range(R_N):
cpt=Res_plus('Res_plus_'+str(i+1),conv_F,plus,rate,drop,H,W)(cpt)

else:  
    for i in range(R_N):
        cpt=Res_normal('Res_normal_'+str(i+1),conv_F,drop,H,W)(cpt)

cpt_conv2=Activation('relu')(cpt)
cpt_out2=BatchNormalization()(cpt_conv2)
cpt_conv1=Dropout(drop)(cpt_out2)
cpt_conv1=Conv2D(filters=channel,kernel_size=(1, 1),padding="same")(cpt_conv1)
cpt_out1=Activation('tanh')(cpt_conv1)`

并不是像论文架构图中将多个ResPlus Unit 的输出进行融合在输送到最后的卷积模块，而是只将ResPlus最后的输出送到了最后的卷积模块，是为什么呢？

[vonfeng]

Sorry, this may be a negligence in the released codes. You can easily add end-fusion mechanism in the end of network by yourself. It will help improve the performance in some cases.

[czisok]

Sorry, this may be a negligence in the released codes. You can easily add end-fusion mechanism in the end of network by yourself. It will help improve the performance in some cases.

Thanks, got it.