A little question about training

[Qinluyi]

if s + 1 == stages: avg_factor, bias, norm = len(modes), 0, 1 x = round_func((pred / avg_factor) + bias) if phase == "train": x = x / 255.0

In the training stage，when it comes to the last stage, why the bias and norm are 0 and 1 respectively,instead of 127 and 255 in previous stages? Besides, why the avg_factor isn't len(modes)4 just like it is in previous stage? It rotate 4 times in every mode. I think len(modes)4 make sense. Could you please explain these for me?

[ddlee-cn]

if s + 1 == stages: avg_factor, bias, norm = len(modes), 0, 1 x = round_func((pred / avg_factor) + bias) if phase == "train": x = x / 255.0

In the training stage，when it comes to the last stage, why the bias and norm are 0 and 1 respectively,instead of 127 and 255 in previous stages? Besides, why the avg_factor isn't len(modes)4 just like it is in previous stage? It rotate 4 times in every mode. I think len(modes)4 make sense. Could you please explain these for me?

Thanks for your interest in our work.

We followed SR-LUT in this one. Empirically, it is indeed better than the variant with avg_factor. From my point of view, since there is a direct supervision from the ground truth pixel values (0-1) for the second stage, omitting averaging helps the variety of predictions across different rotations and modes, without worrying about training stability like in the intermediate stages. Hope it helps.

[Qinluyi]

Thank you for your answer, it helps a lot.

[Qinluyi]

Besides,I wonder if there is a little problem that the "self.conv6 = Conv(nf 5, 1 upscale upscale, 1) " should be "self.conv6 = Conv(nf 5, out_c upscale upscale, 1)" in the MuLUTUnit module in network.py , so that the input parameter out_c can be used. In the DMNet module, you use "self.model = MuLUTUnit('2x2', nf, upscale=2, out_c=3, dense=dense)"，the out_c is 3，so in the MuLUTUnit module, the self.conv6 should produce output with out_c upscale upscale channels, and in the next step you use pixelshuffle(upscale) and the output will have dimension: batchsize,out_c,Hupscale,Wupscale. Am I right? looking forward to your answer.

[ddlee-cn]

Besides,I wonder if there is a little problem that the "self.conv6 = Conv(nf 5, 1 upscale upscale, 1) " should be "self.conv6 = Conv(nf 5, out_c upscale upscale, 1)" in the MuLUTUnit module in network.py , so that the input parameter out_c can be used. In the DMNet module, you use "self.model = MuLUTUnit('2x2', nf, upscale=2, out_c=3, dense=dense)"，the out_c is 3，so in the MuLUTUnit module, the self.conv6 should produce output with out_c upscale upscale channels, and in the next step you use pixelshuffle(upscale) and the output will have dimension: batchsize,out_c,H_upscale,W_upscale. Am I right? looking forward to your answer.

Yeah, you can change it to support multi-channel output.

[Qinluyi]

Thank you for your reply.