Question about Build_conv_block

[sonnguyen129]

When I read your code, I see this function def build_conv_block(self, dim, padding_type, norm_layer, use_dropout, use_bias, cated_stream2=False, cal_att=False)

Can you explain two params cated_stream2 and cal_att? Thank you so much.

[sonnguyen129]

Hi @Amazingren, please explain to me. Thank you so much.

[Amazingren]

Hi @sonnguyen129 ,

Thanks for the interest in our work, also sorry for the late reply since we are hurried on another project. The function build_conv_block is just the code used in our previous projects, the two params cated_stream2 and cal_att in the CrossMLP projects are just redundant and won't affect the performance of the experiments, please ignore them. Sorry for bringing you the confusion.

Best regards Bin Ren

[sonnguyen129]

Hi @Amazingren Thanks for your reply. Can you explain why use Bacthnorm in build_conv_block and use_bias = InstanceNorm? Hope to hear from you soon.

[sonnguyen129]

Hi @Amazingren I have another question. What if I use input F_s is a normal feature map instead of a semantic map in Cross MLP-Mixer block?

[Amazingren]

Hi @Amazingren Thanks for your reply. Can you explain why use Bacthnorm in build_conv_block and use_bias = InstanceNorm? Hope to hear from you soon. Hi @sonnguyen129.

That's used to figure out which kind of normalization method will boot the performance of our method. And we found InstansNorm here is better than. But it depends, you can also modify them to do further exploration.

[Amazingren]

Hi @Amazingren I have another question. What if I use input F_s is a normal feature map instead of a semantic map in Cross MLP-Mixer block?

I think it is possible to take F_s as a normal feature based on your task. The direct thing you need to do maybe is just to adjust give some small modifications to the first block, for example, change the channel numbers of the first convolutional layers.

[sonnguyen129]

Hi @Amazingren Thank you for your reply. Is the first block you mentioned Conv blocks circled in red?