kaix90

The block_composition function is obsolete. Please comment out the line and it should be fine. Sent from my iPhone > On Dec 9, 2020, at 5:18 AM, Luyuan Wang wrote:...

Sorry about the late replay. Regarding the questions: 1. We concatenate Y, Cb and Cr into 192 channels, which were then fed into the gate module in Figure 4. 2....

> 还有几点疑问是关于文中实验部分： > 第一点，请问后续实验提到的Gumbel采样出DCT-24, 48, 64三种网络结构进行处理。但是通过lambda系数怎么控制精确地通道数呢？这些通道数是人为指定的吧？ > 第二点，关于figure 5展示的YCbCr三通道的heat map，请问这个heat map是对Gumbel采样后得到的one-hot 或者probability tensor进行的数据集统计可视化吗？还是其他的tensor呢？ > 第三点，关于selection module的Gumbel采样问题，请问这里是返回one-hot采样结果还是近似one-hot（ont-hot like）的tensor呢？ > 第四点，关于tensor3点乘两个自适应数，能否请问下这两个数值的初始化怎么指定吗？ > > 十分期待您的回答！！以上都是我在复现过程中遇到的一部分不清楚的问题！谢谢！ 抱歉回复的不及时。 1. 通过lambda系数无法精确控制通道数。DCT-24，48，64三种模型的通道数是人为指定的。 2. heap map展示的Gumbel采样后得到的one-hot的结果。 3. selection...

That's a good point. Currently, I don't have data on hand. I will post the result as soon as a result is obtained. Please tolerate my delayed response because I...

You don't need to convert from the frequency domain back to RGB because the extracted features can be directly used for making decisions. Thanks.

The solution in https://github.com/torch/image/issues/171 may help. Thanks.

Thanks very much for your question. We use ILSVRC2012 with ~150Gb and 1000 classes.

All links are updated. Sorry for the inconvenience.

All links are updated. Sorry for the inconvenience.

Hi Feng, Thanks very much for your interest. The preprocessing pipeline for validation is defined in the function valloader_upscaled_static in datasets/dataloader_imagenet_dct.py. Please ignore transform4 and transform5 because they are used...