enc_blk_nums middle_blk_num dec_blk_nums

[Ken1256]

Different tasks need required different settings?

[mayorx]

Hi, Ken1256,

For different tasks, the differentiated design of the network structure is not our original creation, we set this up empirically. In essence, we believe that although denoise/deblur are both called "image restoration tasks", they require different capabilities of the network, such as: Work related to deblur usually emphasizes the "receptive field", while work related to denoise does not seem to emphasize this. I hope that alleviates your concerns.

Thanks.

[Ken1256]

Why enc_blk_nums[3] "receptive field" better than middle_blk_num? How does the deblur + denoise task allocate 36 blocks?

[mayorx]

These are the result of trade-offs between different quantities. E.g. the "receptive field", "# of parameters", "simplicity", and etc. "36 blocks" is also the result of a trade-off between performance and latency.

These results are not set in stone either, and we haven't fully explored the effects of these quantities, so this is an open question.

[Amy-214]

Can enc_blk_nums: [1, 1, 1, 28] be understood as having four encoder modules, the first, second and third encoder modules contain 1 NAFBlock, and the fourth encoder module contains 28 NAFBlocks? What is the reason for this design?