For more dry goods, please pay attention to the WeChat public account "AI Frontline", (ID: ai-front)
Some cool Python projects went open source in April. From Deep Painterly Harmonization, a library that makes synthetic images realistic, to Swift for TensorFlow, this post covers the best open source projects of the past month.
Let's take a look at the hottest open source projects and most interesting discussions on Reddit for data science and machine learning in April.
Check out the most popular open source projects from the past three months at the links below:
February
https://www.analyticsvidhya.com/blog/2018/02/top-5-github-repositories-january-2018/
March
https://www.analyticsvidhya.com/blog/2018/03/top-5-github-repositories-february-2018/
April
https://www.analyticsvidhya.com/blog/2018/04/top-7-github-repositories-march-2018/
https://github.com/luanfujun/deep-painterly-harmonization
Technology to make composite images look more realistic has been around for a few years. Through deep learning, this task will become more efficient and the synthetic effect will be more realistic. Developers have come up with new algorithms that allow external elements to blend perfectly with the hand-painting, resulting in a final composite that is almost indistinguishable from the original drawing.
The three images above - the third frame is the final output, without the first two images, it's hard to know where the balloon was composited later. The algorithm is more refined than manual compositing, a level of sophistication that manual editing has so far been difficult to achieve.
You can learn more about Deep Painterly Harmonization here AVBytes.
https://www.analyticsvidhya.com/blog/2018/04/add-objects-paintings-images-seamlessly-amazing-python-script/
https://github.com/tensorflow/swift
Unveiled at last month's TensorFlow Developer Summit, it's now open sourced to the entire community on GitHub. The goal of the development team is to provide TensorFlow with a new platform that builds on the powerful capabilities of TensorFlow while taking its usability to a whole new level.
Since the project is still in its infancy, it is not yet suitable for building deep learning models. The team admits that as of open source for the community, the project is still far from its envisaged goals. But there is a lot of untapped potential in it.
We are here
https://www.analyticsvidhya.com/blog/2018/04/swift-tensorflow-now-open-sourced-github/ introduces Swift for TensorFlow for reference.
https://github.com/NVlabs/MUNIT
A research group at Cornell University has proposed the Multimodal Unsupervised Image-to-Image Translation (MUNIT) framework for converting images from one domain to another. Its purpose is to convert an image into a new image of a specified domain (eg, convert an image of a dog to a cat).
Preexisting methods are able to perform only one-to-one mapping of a given image and cannot produce different outputs for the same image. However, MUNIT is capable of producing multiple outputs. Get excited!
We are on AVBytes
https://www.analyticsvidhya.com/blog/2018/04/framework-for-unsupervised-image-to-image-translation/ describes how it works.
https://github.com/dmlc/gluon-nlp
Deep learning has made great strides in the field of natural language processing. There are many words that have appeared on the Internet going back centuries! GluonNLP is a toolkit designed to make NLP tasks easier. Word processing is made easier with massive amounts of data and deep learning neural models built. This makes NLP research more efficient.
The library has a good documentation with detailed usage examples. The library also has a 60-minute crash course for newbies.
https://github.com/eriklindernoren/PyTorch-GAN
This library is a gold mine. This library collects PyTorch implementations of published research papers on GANs (or generative adversarial networks). There are currently 24 different implementations listed, each of which is unique. This list contains implementations of Adversarial Autoencoders, CycleGAN, Least Squares GAN, Pix2Pix, etc.
https://www.reddit.com/r/MachineLearning/comments/8b4vi0/d_anyone_having_trouble_reading_a_particular/
If you're having trouble understanding any research paper, the Reddit machine learning community can help. This is a great idea, and it has helped a lot of people find the answer to their question and move on when they were about to give up.
But when you want to post, make sure to provide as much detail as possible, like a summary of the article, where you got stuck, what you found yourself, etc. This comment sums it up nicely - "Think of yourself as a member of the research team writing the paper, rather than waiting for someone else to give you an answer".
https://www.reddit.com/r/MachineLearning/comments/8fmtr9/d_statement_on_nature_machine_intelligence/
The debate over whether research should be open source or closed has raged for decades. Recently, the popular "Nature" magazine announced that it would publish a closed journal. This led to a major campaign against them, with many influencers (Jeff Dean, Ian Goodfellow, etc.) adding their signatures to a petition stating that they would not contribute to such a publication.
There are different opinions on whether the findings should be open sourced to the community. This is a topic worth knowing about, and reading the entire thread is highly recommended to see what the ML community has to say about the subject.
https://www.reddit.com/r/MachineLearning/comments/8daqki/d_very_sobering_presentation_on_the_current_state/
Michael Jordan, a distinguished professor at Berkeley, spoke at length in a recent talk about how far we are from true machine intelligence. This is a thoughtful speech that will make you think about the issue after listening.
The thread has generated over 100 comments, with users giving their thoughts on the state of AI in abundance. In-depth discussions from users made this topic even more interesting. This topic is still under discussion, join them.
https://www.reddit.com/r/MachineLearning/comments/8ekmqy/scientists_plan_huge_european_ai_hub_to_compete/
There is a lot of good stuff in this question. Data scientists and machine learning researchers from all over Europe and the US participated in a lively discussion about the structure of ML and salaries. Lots of perspectives can be gained here about ML projects and potential compensation structures.
https://www.reddit.com/r/MachineLearning/comments/8f9dvm/ruberai_measuring_the_intrinsic_dimension_of/
This theme was developed from Uber's video, which develops intrinsic dimensions as fundamental properties of neural networks. If you have any questions about the content provided in the video, the community has answered those questions in detail. The best part seems to be that people like to turn a research paper into a video, which makes research easier to understand.
For more dry goods, please pay attention to the WeChat public account "AI Frontline", (ID: ai-front)