Unsupervised learning with Autoencoders using Fastai

Part 1 (2019)
1

Hello everyone !

I just wrote a post on Medium on Autoencoders using Fastai. I have first written an article on the principles behind it, and added some code to play with it.

I am intending to write a part 2 with some maths behind it, and a part 3 showing how to customize Fastai to more easily write Autoencoders using Fastai.

I am open to suggestions and hope you will like it :slight_smile:

Here is the link to the Medium article : https://medium.com/@danielhuynh_48554/an-introduction-to-unsupervised-learning-with-fastai-a6dbd78eca2b?source=friends_link&sk=ffcdd8017888d837df64ac70e2e01908

6 Likes
2

Hi DanyWin thanks for producing an informative, instructive easy to read post on unsupervised learning.

I really enjoyed reading it.

I was a bit confused by this paragraph.

Just for the fun, we can also see how these methods work for Fashion MNIST which is a harder dataset than MNIST, where we also used only 512 images. We first get the following reconstruction :

I learned a lot from your post.

Cheers mrfabulous1 :smiley::smiley:

3

Hi @mrfabulous1 !
Glad you liked it :slight_smile:

Well I meant Fashion MNIST is harder than MNIST

I will try to implement cool papers with VAE to keep it coming :smiley:

4

Ah, this is awesome! I’m going to take the time to go through this properly soon. At least I think I need this to solve my problem :smiley: Where did you learn about it?

edit: also - perhaps a link here? http://wiki.fast.ai/index.php/Deep_Learning_Glossary#Unsupervised_Learning

1 Like
5

Took me some time to really understand how it works. It is in fact quite close to Bayesian Neural Networks.

I think this paper from Kingma, one of the original authors of the VAEs can give you a good introduction : https://arxiv.org/pdf/1906.02691.pdf

If you need some help to understand it better I can help you with it. The post gave the global idea, the paper gave the implementation and the maths and you can look at the code to see how it works, but if you need more explanatinons I can help ^^

I was supposed to write the follow up article but I got busy on my bs finder. Nonetheless, if you tell me what you want to do I can give you advice :slight_smile:

6

very cool, I’ll let you know where I get stuck :slight_smile: