According to Wikipedia:
PyTorch is an open source machine learning library based on the Torch library, used for applications such as computer vision and natural language processing. It is primarily developed by Facebook’s artificial intelligence research group.
Facebook recently released the much anticipated 1.3 version at PyTorch DevCon 2019, adding support for Google TPU, PyTorch Mobile and more. Also, **The Gradient** release a report on the current state of machine learning frameworks, stating that more and more researchers favor PyTorch to TensorFlow as their main framework. Another good reason if you are leaning/using the fast.ai library, it is based on PyTorch because it:
“use all of the flexibility and capability of regular python code to build and train neural networks”, and “we were able to tackle a much wider range of problems”
To me, it feels more natural working with PyTorch, being Pythonic and all. A better analogy is that developing machine learning models using TensorFlow is like wearing heavy armor: It’s powerful but very clunky. PyTorch gives you all the freedom and a smooth flow of actions.
Photo from GHYFY
Photo from deeplizard vlog — YouTube
But the flexibility of PyTorch comes with a price. Getting into PyTorch isn’t easy. More freedom means you have more factors that need considering and more nuances to balance.
That’s why a great tutorial will help you greatly smooth out the learning curve. There are many resources out there. Jeremy Howard’s wonderful tutorial on the PyTorch website is a good starting point. Yet if you want to delve down even deeper, I recommend you check out deeplizard’s PyTorch Tutorial Series on YouTube.
It’s to-the-point (respect viewer’s time by being concise), relevant (based on PyTorch 1.1) and most importantly, fun to watch. It uses a lot of neat animations/graphic editing techniques to make the video engaging and pleasant to watch. The production quality is very impressive.
Chris is really great at explaining a complicated concept in a very concise and clear way with the help of great animations and versatile form of short video clips. Their videos will keep you occupied from start to finish. Couple of things they really stand out:
deep lizard uses animation to explain how convolution works
Some YouTube videos offer exceptional content, but the aesthetic is not there, especially for the screencast videos. Viewers usually have to stare at one or two windows most of the time, which is visually boring and easy to get tired. Not for deeplizard though. The team is very good at creating subtle yet aesthetically pleasing animations. Even for backdrop images, which are usually static, they created some zoom in/zoom out effect to make it less boring. The motion is subtle enough so there is no concern of motion sickness. Plenty of eye-candy 🍬 I’d say.
Image from deeplizard @ YouTube
Embedding short video clips of relevant yet different styled content is another way to make the learning engaging. Our brain gets tired quickly if only one part gets stimulated. Looking at the same scene or listening to the same person talking, people won’t keep their focus long enough. At least can’t do it without some mental efforts. Embedding a variety of styled video clips solved the problem. Different parts of your brain get excited and you can keep the learning flow effortlessly.
Male, female and a special Sci-Fi style ‘AI’ character called ‘deeplizard’ voiceover to explain different type of problems, spice the content up
R2D2 and C-CPO from theverge.com
Other than the two lovely YouTubers of the channel Chris and Mandy, there is a virtual ‘AI creature’ they created as a third voice-over. It sounds like C-3PO in Star War movies, but female. It guides the viewer through the debugging process or asks some thought-provoking cryptic questions, etc. If you are into the Sci-Fi vibe, you’ll totally love it. Again, less boring, more engaging.
Besides the videos itself, they also have a membership website where you can find extra learning materials: Blogs, Quiz, Code snippets, and other extra resources. It’s behind a paywall but I’d say it’s a good supplement of the video.
The main takeaway? I felt that they’ve made PyTorch seems quite straightforward and easy to understand. I felt like I can totally do my ML project on PyTorch going forward. Though I’ve only finished their PyTorch tutorials, I’d guess their other contents are also good too. Feel free to explore a bit more and let me know your experience below. If you are learning fast.ai course, since it’s built on top of PyTorch, sooner or later you’ll have to beef up your PyTorch knowledge and deeplizard’s tutorial is a good place to start. Link here: