booooooooooo / csd Goto Github PK

In EDSR paper, the EDSR model (32 layer, 256 channel) is about 32.46. But in your paper, the performance is 32.60. I wonder the training setting about your teacher.

baseline student是从头开始训练的（scratch），而csd student用了教师部分channel的预训练权重+CSD，个人觉得这两者比较并不公平。

所以我们复现了student同样载入教师部分权重后与hr 直接进行l1作为baseline，没有用任何蒸馏。初步训练结果如下：

0.25 Set5 32.34541767076197 39.64636575838763 0.8969442045190894
0.25 Set14 28.727600156122755 23.11639569902273 0.7848623259477279
0.25 B100 27.661651828426916 19.849210476012832 0.738650274225103
0.25 Urban100 26.30338385572442 12.644811790998192 0.792405167459895

论文中CSDx4 student指标为：

Set5 32.34 0.8974 
Set14 28.72 0.7856 
B100 27.68 0.7396 
Urban100 26.34 0.7948

从指标看，用不用蒸馏的结果相差并不大。个人觉得很难有说服力吧。

We do not find 'trainer' in code, could you share a complete code ?

我复习代码 在bic_sample = lr[torch.randperm(self.neg_num), :, :, :] 为什么会出现类似tensor溢出的问题

Hey there! looking very interesting.

Is it only works on existing models, or eventually it outputs a pertained model based on a reference one?

Roi

#mindspore版---请问可以提供vgg19_ImageNet.ckpt文件吗？

请问例如0.25x的网络分支的参数量如何计算得到的？
且想问一下该自蒸馏方法训练保存的还是整个的模型，只是在进行模型推理的时候用的局部分支的参数。不能直接保存局部子分支的模型？

您好，首先非常感谢您的论文以及开源代码，给我的学习带来了很大帮助。
在反复研读了您的论文之后，如果我没理解错的话这个模型的目的是想创建两条支路，一条teacher，一条student。在测试时只加载一个pretrained model，比如edsr_x4_0.25student.pth。然后模型根据资源动态分配，决定使用teacher还是student。我有点不理解这个动态分配是如何实现的，因为我看了代码，test时给了--stu_width_mult 0.25 参数，那就是相当于只加载了student那条支路啊？

作者您好，请问报错这个的原因是什么，是不是window无法运行呀

Your paper use EDSR，but the code seems is RDN

您好，我们在一些工程上验证了该论文的方法，发现效果不错，且验证了CSD方法 x4（T&S），与论文一致，感谢您的开源。
但是baseline验证的时候有一些问题：下载的edsr_x4_baseline.pth 在--stu_width_mult 1 时 与论文指标是相同的，即teacher推理正确。

但是在0.25时set5的psnr 只有14，与论文32.23差距很大，该情况在其他数据集中也差的很大。
请问edsr_x4_baseline.pth支持0.25的推理吗？

Thanks for your job. It help me a lot. I have some question about speed up to ask.

in rw= 0.5x. in figure (a) .Because input channel and output channel will reduce by half. So the parameters only use 1/4 is easy to understand. But in figure (b), Why the speed time only reduce by half (not 1/4)? Can you explain a little bit? :)

Thanks for your great works! After reading the papers and codes，I have a question about Contrastive loss。In your codes, using the unrelated samples in a batch as negative samples. How to select negative samples really interests me. As we all know, the unrelated samples in a batch is far away from the outputs of student networks. Could using this as negative samples really be a good lower bounds? Have you ever consider using other images as negative samples? eg. upsample(self.lq) or gaussianBlur(upsample(self.lq))...
Hope for your reply, sincerely!

Thanks for sharing your code online! I have some question about construting negtive sample. In [CSD/PyTorch version/trainer/slim_contrast_trainer.py#L100]

bic_sample = lr[torch.randperm(self.neg_num), :, :, :]

(https://github.com/Booooooooooo/CSD/blob/main/PyTorch%20version/trainer/slim_contrast_trainer.py#L100). I think some negtive sample actually is a positive sample. For example, Why not set batchsize = old_batch_size+neg_num. In each iterate, use old_batch_size sample constructing sample+postive samples, and use the remaining neg_num samples as the negative samples

您好！看了您的论文以及代码实现，我有一个疑问，为什么不直接采用hr作为负样本，而是要对lr进行双线性插值上采样后作为负样本？这样做有什么好处么？