Forwarded from Data Science | Machine Learning | Artificial Intelligence
R Wavelet Feature Maps Compression for Image-to-Image CNNs
Our NeurIPS 2022 paper "Wavelet Feature Maps Compression for Image-to-Image CNNs" is now available.
In this paper, we propose a novel approach to compress CNNs using a modified wavelet compression technique.
Abstract:
>Convolutional Neural Networks (CNNs) are known for requiring extensive computational resources, and quantization is among the best and most common methods for compressing them. While aggressive quantization (i.e., less than 4-bits) performs well for classification, it may cause severe performance degradation in image-to-image tasks such as semantic segmentation and depth estimation. In this paper, we propose Wavelet Compressed Convolution (WCC) -- a novel approach for high-resolution activation maps compression integrated with point-wise convolutions, which are the main computational cost of modern architectures. To this end, we use an efficient and hardware-friendly Haar-wavelet transform, known for its effectiveness in image compression, and define the convolution on the compressed activation map. We experiment with various tasks that benefit from high-resolution input. By combining WCC with light quantization, we achieve compression rates equivalent to 1-4bit activation quantization with relatively small and much more graceful degradation in performance.
​
Paper: [https://arxiv.org/abs/2205.12268](https://arxiv.org/abs/2205.12268)
Code: https://github.com/BGUCompSci/WaveletCompressedConvolution/
​
Cityscapes semantic segmentation with different compressions.
​
KITTI depth prediction with different compressions.
https://redd.it/y22rk0
@artificialintelligence24x7
Our NeurIPS 2022 paper "Wavelet Feature Maps Compression for Image-to-Image CNNs" is now available.
In this paper, we propose a novel approach to compress CNNs using a modified wavelet compression technique.
Abstract:
>Convolutional Neural Networks (CNNs) are known for requiring extensive computational resources, and quantization is among the best and most common methods for compressing them. While aggressive quantization (i.e., less than 4-bits) performs well for classification, it may cause severe performance degradation in image-to-image tasks such as semantic segmentation and depth estimation. In this paper, we propose Wavelet Compressed Convolution (WCC) -- a novel approach for high-resolution activation maps compression integrated with point-wise convolutions, which are the main computational cost of modern architectures. To this end, we use an efficient and hardware-friendly Haar-wavelet transform, known for its effectiveness in image compression, and define the convolution on the compressed activation map. We experiment with various tasks that benefit from high-resolution input. By combining WCC with light quantization, we achieve compression rates equivalent to 1-4bit activation quantization with relatively small and much more graceful degradation in performance.
​
Paper: [https://arxiv.org/abs/2205.12268](https://arxiv.org/abs/2205.12268)
Code: https://github.com/BGUCompSci/WaveletCompressedConvolution/
​
Cityscapes semantic segmentation with different compressions.
​
KITTI depth prediction with different compressions.
https://redd.it/y22rk0
@artificialintelligence24x7
GitHub
GitHub - BGUCompSci/WaveletCompressedConvolution: Official implementation for Wavelet Feature Maps Compression for Image-to-Image…
Official implementation for Wavelet Feature Maps Compression for Image-to-Image CNNs, NeurIPS 2022. - GitHub - BGUCompSci/WaveletCompressedConvolution: Official implementation for Wavelet Feature M...
Forwarded from Data Science | Machine Learning | Artificial Intelligence
P Paperlib: An open-source and modern-designed academic paper management tool.
https://preview.redd.it/0v2r4ftf5yoc1.png?width=3104&format=png&auto=webp&s=99653d16c6d10587f3cb75a8381f53d49778624a
Github: https://github.com/Future-Scholars/paperlib
​
Website: https://paperlib.app/en/
Hi guys, I'm a computer vision PhD student. Conference papers are in major in my research community, which is different from other disciplines. Without DOI, ISBN, metadata of a lot of conference papers are hard to look up (e.g., NIPS, ICLR, ICML etc.). When I cite a publication in a draft paper, I need to manually check the publication information of it in Google Scholar or DBLP over and over again.
Why not Zotero, Mendely?
A good metadata scraping capability is one of the core functions of a paper management tool. Unfortunately, no software in this world does this well for conference papers, not even commercial software.
A modern UI/UX.
In Paperlib 3.0, I bring the Extension System. It allows you to use extensions from official and community, and publish your own extensions. I have provided some official extensions, such as connecting Paprlib with LLM!
Paperlib provides:
​
OPEN SOURCE
Scrape paper’s metadata and even source code links with many scrapers. Tailored especially for machine learning.
Fulltext and advanced search.
Smart filter.
Rating, flag, tag, folder and markdown/plain text note.
RSS feed subscription to follow the newest publications on your research topic.
Locate and download PDF files from the web.
macOS spotlight-like plugin to copy-paste references easily when writing a draft paper. Also supports MS Word.
Cloud sync (self managed), supports macOS, Linux, and Windows.
Beautiful and clean UI.
Extensible. You can publish your own extensions.
Import from Zotero.
Here are some GIFs introducing the main features of Paperlib.
​
Scrape metadata for conference papers. You can also get the source code link!
[Scrape metadata for conference papers. You can also get the source code link!](https://i.redd.it/wnak205l1yoc1.gif)
​
Organize your library with tags, folders and smart filters!
Organize your library with tags, folders and smart filters!
​
Three view mode.
[Three view mode.](https://i.redd.it/uehvs6sx1yoc1.gif)
​
Summarize your papers by LLM. Tag your papers by LLM.
Summarize your papers by LLM. Tag your papers by LLM.
​
Smooth paper writing integration with any editors.
[Smooth paper writing integration with any editors.](https://i.redd.it/haul51n82yoc1.gif)
Extensions
https://preview.redd.it/7132inxe3yoc1.png?width=1480&format=png&auto=webp&s=b1cb03dec3285eaa14940efb42b4cb13656b64d9
​
​
​
https://redd.it/1bh63c1
@artificialintelligence24x7
https://preview.redd.it/0v2r4ftf5yoc1.png?width=3104&format=png&auto=webp&s=99653d16c6d10587f3cb75a8381f53d49778624a
Github: https://github.com/Future-Scholars/paperlib
​
Website: https://paperlib.app/en/
Hi guys, I'm a computer vision PhD student. Conference papers are in major in my research community, which is different from other disciplines. Without DOI, ISBN, metadata of a lot of conference papers are hard to look up (e.g., NIPS, ICLR, ICML etc.). When I cite a publication in a draft paper, I need to manually check the publication information of it in Google Scholar or DBLP over and over again.
Why not Zotero, Mendely?
A good metadata scraping capability is one of the core functions of a paper management tool. Unfortunately, no software in this world does this well for conference papers, not even commercial software.
A modern UI/UX.
In Paperlib 3.0, I bring the Extension System. It allows you to use extensions from official and community, and publish your own extensions. I have provided some official extensions, such as connecting Paprlib with LLM!
Paperlib provides:
​
OPEN SOURCE
Scrape paper’s metadata and even source code links with many scrapers. Tailored especially for machine learning.
Fulltext and advanced search.
Smart filter.
Rating, flag, tag, folder and markdown/plain text note.
RSS feed subscription to follow the newest publications on your research topic.
Locate and download PDF files from the web.
macOS spotlight-like plugin to copy-paste references easily when writing a draft paper. Also supports MS Word.
Cloud sync (self managed), supports macOS, Linux, and Windows.
Beautiful and clean UI.
Extensible. You can publish your own extensions.
Import from Zotero.
Here are some GIFs introducing the main features of Paperlib.
​
Scrape metadata for conference papers. You can also get the source code link!
[Scrape metadata for conference papers. You can also get the source code link!](https://i.redd.it/wnak205l1yoc1.gif)
​
Organize your library with tags, folders and smart filters!
Organize your library with tags, folders and smart filters!
​
Three view mode.
[Three view mode.](https://i.redd.it/uehvs6sx1yoc1.gif)
​
Summarize your papers by LLM. Tag your papers by LLM.
Summarize your papers by LLM. Tag your papers by LLM.
​
Smooth paper writing integration with any editors.
[Smooth paper writing integration with any editors.](https://i.redd.it/haul51n82yoc1.gif)
Extensions
https://preview.redd.it/7132inxe3yoc1.png?width=1480&format=png&auto=webp&s=b1cb03dec3285eaa14940efb42b4cb13656b64d9
​
​
​
https://redd.it/1bh63c1
@artificialintelligence24x7
Forwarded from Data Science | Machine Learning | Artificial Intelligence
How AI already changed my life
I feel like most of the public is not at all aware what AI already can do. They just think like:"Oh yea AI, you can make cute pics with it" Or the youngsters using it to swap out faces of people.
But most people do not realize it can already improve your life in a big way. All you gotta do is ... ask AI :P
So to further explain what it actually did for me you need to know a lil bit of my background first. So I am kinda a guy which in RPG would you call a jack of all traits in the field of IT but master of none.
I cannot code or program anything, but I have enough knowledge to make use of Wordpress to start a website. If there were problems with the code in my wordpress or when I messed something up I was kinda screwed. I remember it taking weeks for me to repair kinda simple problems, or sometimes it was just above my capability, I had not enough knowledge to fix it. I even remember paying a guy at Fiverr to fix some programming problem.
I was also a very basic linux users, just barely able to install it, not using custom partitions since I had no clue how. I have had multiple instances where something in linux broke, which I could not fix and ended up reinstalling the entire system again, wasting a lot of time. These are even things like black screen caused by nvidia driver issues, which is easily fixed if you know how.
Ok that was then. So lets go ahead and see what my capabilities are now shall we ;)
So with the help of AI I have fixed complex issues on my linux system. And by doing a lot of commands in terminals, even though I just copied stuff from the AI, it also learned me a lot of commands. I can now perform a lot of commands in terminals I couldnt before.
That was just the beginning though ;) Once I understood how powerful AI can really be, I tried to seek its and mine limits of what I was now able to do.
Where I at first had trouble installing a new OS like linux, I now have a triple boot system with full disk encryption (because its cool) :P running windows, and 2 different linux distros. All with a custom made Grub launcher with a cool theme.
I still cannot program really as I do not know any programming language, but I was able to create several programs with the help of AI! I never thought this to be possible, me creating my own programs. It was still not easy, since I did not even know where to begin, but AI told me all I needed to know. Practical things like which program do I use to type the code in? How do I save the file?
I even tried making my own videogame which I think I could do, but I need to learn a lot more to do that. Since I will also need to learn something like stable diffusion to generate visual content for that game. This is more something for the long run though, I feel I need to learn more first.
The AI makes me feel so confident now to tackle all IT problems facing me. Although I do admit I do not always know what I am doing exactly.
I just feel it opened up a whole new world for me. Its so cool I can now create entire programs, like right now I am editing a GUI in Qt designer. I never even knew this existed, but AI told me about it and now im using it.
Another thing I find cool about AI is that rarely sells me a:"No we cant do that". It does not really matter how complex my question is, it always knows of a way to do something.
Btw this is pretty much all done with chatGPT 3.5 free version. I dont even know how good it can really be.
​
https://redd.it/1cc04x8
@artificialintelligence24x7
I feel like most of the public is not at all aware what AI already can do. They just think like:"Oh yea AI, you can make cute pics with it" Or the youngsters using it to swap out faces of people.
But most people do not realize it can already improve your life in a big way. All you gotta do is ... ask AI :P
So to further explain what it actually did for me you need to know a lil bit of my background first. So I am kinda a guy which in RPG would you call a jack of all traits in the field of IT but master of none.
I cannot code or program anything, but I have enough knowledge to make use of Wordpress to start a website. If there were problems with the code in my wordpress or when I messed something up I was kinda screwed. I remember it taking weeks for me to repair kinda simple problems, or sometimes it was just above my capability, I had not enough knowledge to fix it. I even remember paying a guy at Fiverr to fix some programming problem.
I was also a very basic linux users, just barely able to install it, not using custom partitions since I had no clue how. I have had multiple instances where something in linux broke, which I could not fix and ended up reinstalling the entire system again, wasting a lot of time. These are even things like black screen caused by nvidia driver issues, which is easily fixed if you know how.
Ok that was then. So lets go ahead and see what my capabilities are now shall we ;)
So with the help of AI I have fixed complex issues on my linux system. And by doing a lot of commands in terminals, even though I just copied stuff from the AI, it also learned me a lot of commands. I can now perform a lot of commands in terminals I couldnt before.
That was just the beginning though ;) Once I understood how powerful AI can really be, I tried to seek its and mine limits of what I was now able to do.
Where I at first had trouble installing a new OS like linux, I now have a triple boot system with full disk encryption (because its cool) :P running windows, and 2 different linux distros. All with a custom made Grub launcher with a cool theme.
I still cannot program really as I do not know any programming language, but I was able to create several programs with the help of AI! I never thought this to be possible, me creating my own programs. It was still not easy, since I did not even know where to begin, but AI told me all I needed to know. Practical things like which program do I use to type the code in? How do I save the file?
I even tried making my own videogame which I think I could do, but I need to learn a lot more to do that. Since I will also need to learn something like stable diffusion to generate visual content for that game. This is more something for the long run though, I feel I need to learn more first.
The AI makes me feel so confident now to tackle all IT problems facing me. Although I do admit I do not always know what I am doing exactly.
I just feel it opened up a whole new world for me. Its so cool I can now create entire programs, like right now I am editing a GUI in Qt designer. I never even knew this existed, but AI told me about it and now im using it.
Another thing I find cool about AI is that rarely sells me a:"No we cant do that". It does not really matter how complex my question is, it always knows of a way to do something.
Btw this is pretty much all done with chatGPT 3.5 free version. I dont even know how good it can really be.
​
https://redd.it/1cc04x8
@artificialintelligence24x7
Reddit
From the ArtificialInteligence community on Reddit
Explore this post and more from the ArtificialInteligence community