AvishkarArjan / DeepVOLite

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Lets not mess this up

This project is my second try at replicating the results of the paper DeepVO. The first attempt was so useless that I had to abandon weeks of coding and analysis altogether.

CURRENT UPDATE - So trying to reduce the number of parameters randomly within the model causes blunders. The least MSE loss i was able to get (training on my cpu) was around 0.10 - which is ridiculous in this case. I might have to try some more things to try to create a lighter version, maybe a different normalization . Idk.

Differences from Original paper

The original model architecture mentioned in the paper

I have limited computing resources - thus I drastically reduced the parameters by omitting out several layers from the model - to atleast me able to train the model. Based on some research, the max parameters of a model than can be trained free on Google Colab is under 1 million.

Some resources that may or may not have contributed to the development of this project:

https://github.com/thedavekwon/DeepVO

About


Languages

Language:Python 86.9%Language:Jupyter Notebook 13.1%