VAE

[ayushtues]

Ayush Mangal , Pytorch , MNIST , based on Carl Doresch VAE tutorial

[aniket-agarwal1999]

Yea you seem the first for VAE implementation. So yea it's a go-ahead from our side. Be sure to once go through the README on the structure of PR that is to be submitted. :smile:

[dakshitagrawal]

@ayushtues Any progress in this ticket?

[ayushtues]

@dakshitagrawal97 This is my first stand-alone project ever so I kind of followed the DCGAN implementation discussion and wanted to make things as similar as possible so I didn't make the PR yet :sweat_smile:

The following things are done

• The model is mainly an encoder which has conv layer based on Alex net (the conv_layer - relu - batch norm) some fully connected layers and then for the decoder some TransposeConv layers and some fully connected layers. Although I am using Leaky ReLU instead and Xavier initialization in the conv layers. Also dropout in the encoder but not in the decoder (something I found in aniket-agarwal1999 's repo :sweat_smile: )

• I can reconstruct training images(it's reconstructing fine ) & save them in a folder generated by the code itself

• Can save the model checkpoints in a seperate folder made by the code itself

• Applied t-sne visualization of the latent space and it has the 10 clusters needed although 1 cluster always seems to be far away from the other 9 which are sort of gaussian as expected

• The generated output from random gaussian inputs ..they well ...kind of look like numbers :sweat_smile:

• Did the external config file parsing thing

• Added progress bar

• Tried training on google colab once and it worked well although I usually train it on my CPU. Will do the final training in the end

Things to do:

• TensorboardX graphs (Kinda stuck in this)

• Making the typical VAE smooth transition between numbers grid

• Making the entire readme :sweat_smile:

• The generate.py file to make images from the pre-trained model

• Kind of confused about the exact filter size and channel numbers of the Convolution and TransposeConv part. So will play a little bit with it. Also might try to figure out if using Leaky ReLU and Xavier Init and batch_norm helps or makes it worse :sweat_smile:

• Might try to play with the relative weights of the recosntruction and KLD loss to get better results (kind of like beta-VAE)

• Other stuff I can't remember :sweat_smile:

Also sorry for the long comment ... it's my first time using GitHub so am not really familiar with the norms

[dakshitagrawal]

@ayushtues Thanks for the detailed update. No need to be sorry for writing a long comment, helps us know exactly where you are. :)

Feel free to make a PR whenever you feel the code is ready to be merged.

You might wanna keep a log of all your experiments, and note down your observations in the Readme.

[aarushgupta]

Great work @ayushtues! Feel free to shoot us a message in case you get stuck anywhere.

[ayushtues]

Merged! Closing.