Comments (7)
Sorry about the late replay. Regarding the questions:
- We concatenate Y, Cb and Cr into 192 channels, which were then fed into the gate module in Figure 4.
- In the segmentation experiment, we skip the input layers and directly feed the DCT input into the residual block.
Thanks very much for your interest.
from dctnet.
十分谢谢你的回复!!
还有一点疑问是,在我复现的过程中,文中提到的tensor4的每一个通道点都需要normalize,请问这一步是怎么操作的呢?我试过用sigmoid做,但效果不好,期待你的答复。
from dctnet.
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!
from dctnet.
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。
- 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。
- heap map展示的Gumbel采样后得到的one-hot的结果。
- selection module返回的tensor是one-hot。
- 初始化的原则是使tensor4中大部分的通道被选择,例如>90%的channel被选中。
from dctnet.
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。
- 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。
- heap map展示的Gumbel采样后得到的one-hot的结果。
- selection module返回的tensor是one-hot。
- 初始化的原则是使tensor4中大部分的通道被选择,例如>90%的channel被选中。
谢谢回复!!!
我用unet跑192通道结果发现并不好(或者说很差)。我统计了数据的mean以及std进行了norm。
即使没有去掉一些通道,结果依旧不好。
from dctnet.
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。
- 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。
- heap map展示的Gumbel采样后得到的one-hot的结果。
- selection module返回的tensor是one-hot。
- 初始化的原则是使tensor4中大部分的通道被选择,例如>90%的channel被选中。
谢谢回复!!!
我用unet跑192通道结果发现并不好(或者说很差)。我统计了数据的mean以及std进行了norm。
即使没有去掉一些通道,结果依旧不好。
我似乎遇到了跟你一样的问题,loss不下降。麻烦问一下,你有没有解决这个问题呢?(分割任务,tf1.13)
from dctnet.
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。
- 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。
- heap map展示的Gumbel采样后得到的one-hot的结果。
- selection module返回的tensor是one-hot。
- 初始化的原则是使tensor4中大部分的通道被选择,例如>90%的channel被选中。
谢谢回复!!!
我用unet跑192通道结果发现并不好(或者说很差)。我统计了数据的mean以及std进行了norm。
即使没有去掉一些通道,结果依旧不好。我似乎遇到了跟你一样的问题,loss不下降。麻烦问一下,你有没有解决这个问题呢?(分割任务,tf1.13)
没有。后面我就开始做另外的事情了。loss收敛还是依旧比较大。
from dctnet.
Related Issues (20)
- could you provide a dockerfile? libjpeg62 makes me crazy.
- would you please let me know why return "img, F.upscale(img, self.upscale_factor, self.interpolation)" with 2 images? thanks HOT 2
- look for "block_composition" HOT 10
- Do you transform the images from BGR to YCbCr color space? HOT 2
- pretrained model can't be downloaded HOT 1
- The mobilenetv2 cannot be evaluated with your pretrained models. HOT 1
- the link to Training code is not available now ,could you update it again? HOT 5
- Cifar model not found
- About mean and std in transfrom HOT 1
- pretrained model can't be downloaded HOT 1
- How do you group the components with same frequency into one channel? HOT 1
- The code is so bad HOT 2
- pretrained model can't open HOT 5
- Questions about the dataloader and missing files HOT 1
- Missing transforms (ToYCrCb, ChromaSubsample, UpsampleDCT, CenterCropDCT) in cvtransforms.py HOT 3
- update training code
- how to decompress the pretrained model?
- could you share the env config?
- ModuleNotFoundError: No module named '_dctfromjpg_wrapper'
- RuntimeError: Wrong JPEG library version: library is 62, caller expects 90 solved
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from dctnet.