A issue about training

[Rainyliuliu]

Hi,when i run main.py ,there is an AttributeError: module 'model.wdsr_a' has no attribute 'make_model'

[JiahuiYu]

@Rainyliuliu Thanks for your interest. You can rename MODEL to make_model here to train. https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L26

[Rainyliuliu]

Thank you very much!And i have another question. i add : parser.add_argument('--r_mean', type=int,default=0.4488) parser.add_argument('--g_mean', type=int,default=0.4371) parser.add_argument('--b_mean', type=int,default=0.4040) parser.add_argument('--n_resblocks', type=int, default=32, help='number of residual blocks') parser.add_argument('--n_feats', type=int, default=256, help='number of feature maps') parser.add_argument('--block_feats', type=int,default=256) parser.add_argument('--patch_size', type=int, default=192, help='output patch size') to option.py ,and there is another error :+1: RuntimeError: input and target shapes do not match: input [16 x 3 x 176 x 176], target [16 x 3 x 192 x 192] at /pytorch/ate

[Rainyliuliu]

i think in there def forward(self, x): x = (x - self.rgb_mean.cuda()*255)/127.5 if not self.training: x = self.pad(x) s = self.skip(x) x = self.head(x)

    x = self.body(x)

    x = self.tail(x)
    x += s
    x = x*127.5 + self.rgb_mean.cuda()*255
    return x

s is 176*176 ,but i don't know why

[JiahuiYu]

@Rainyliuliu Sorry we missed it. You can either:

1. Remove https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L78-L79, change padding of https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L57 to 1, and https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L62 to 2.

or :

2. in training code of EDSR, crop the target hr to the shape of prediction sr.

[YongboLiang]

I think both paddings of https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L57 , and https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L62 should be changed to 2.

[JiahuiYu]

@YongboLiang We have explicit padding here. https://github.com/JiahuiYu/wdsr_ntire2018/blob/a4a7ade1b72bac4f49fd6e28d0c4c74d449c5e8b/wdsr_a.py#L79

[wangmenglu]

there is a RuntimeError: The expanded size of the tensor (188) must match the existing size (192) at non-singleton dimension 3

[JiahuiYu]

@wangmenglu Have you modified any thing in our models? It should be fine to run without modification.

[wangmenglu]

@wangmenglu Have you modified any thing in our models? It should be fine to run without modification.

I didn't modified anything.

[JiahuiYu]

Sorry it seems you will need to crop target HR patch to the output size in the main training loop. The reason is because we did not use padding in last convolution. Instead we only pad during testing.

The explicit padding has very close performance in our experiments. Thus we will remove it soon. Apologize for inconvenience and thanks for reporting.

[JiahuiYu]

We have removed explicit padding. Sorry again for the inconvenience.