Skip to content

This is a application of a conditional Variational Autoencoder similar to the problem of "2 Bird 1 Neural Network"

Notifications You must be signed in to change notification settings

Mehul0x/BH-24-2Y

Repository files navigation

BH-24-2Y

This is an application of a conditional Variational Autoencoder similar to the problem of "2 Bird 1 Neural Network"

The objective is to use a generative model to find new functions (x0 & x1) that can fit certain functions, like cos(4 x0 + 8 x1)

The obvious choices were to use a CVAE or a GAN, I chose to go with a CVAE, since and if time permits experiment with a GAN.

Approaching the challenge in a structured challenge, I decided to first set up my hyperparameter tuning for the already given model, which would be useful in the future with any changes.

While setting up optuna I encountered some errors.

After correctly setting up my optuna, I got these results image image

After this, I varied the parameters of the model and then optimized its hyperparameter, through a lot of iterations I got the best result image image

This was the test loss on this iteration image

The same architecture with different hyperparameters on data2 performed as

image

I tried replacing a convolutional layer with a linear layer but the results had a significant drop, so I decided to add other convolution layers instead, I encountered some errors and time didn't permit me to fix them.

For the number of epochs, I noticed that increasing them beyond 40-45 was not affecting the y_loss, so I only used 50 epochs for each training. 1 weird thing I noticed was that my KLD loss was -ve in some epochs even though theoretically it wasn't possible

About

This is a application of a conditional Variational Autoencoder similar to the problem of "2 Bird 1 Neural Network"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published