https://github.com/princeofpuppers/phys490hw4

https://github.com/princeofpuppers/phys490hw4

Last synced: 10 days ago
JSON representation

• Host: GitHub
• URL: https://github.com/princeofpuppers/phys490hw4
• Owner: PrinceOfPuppers
• Created: 2020-03-19T04:26:28.000Z (almost 5 years ago)
• Default Branch: master
• Last Pushed: 2020-03-22T14:15:52.000Z (almost 5 years ago)
• Last Synced: 2024-11-08T03:44:26.746Z (2 months ago)
• Language: Python
• Size: 2.21 MB
• Stars: 0
• Watchers: 3
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        

# Assignment *number 4*

- name: Joshua McPherson

- student ID: 20687868

## Dependencies

- pytorch

- numpy

- tqdm

- matplotlib

## Running `main.py`

To run `main.py`, place even_mnist.csv into the data directory and type the following command

```sh

python main.py -o result_dir -n 100

```

## Notes

Network will occasionally get stuck in local minima (roughly once every 10 times training it), Learning Rate aneeling is used to avoid this however it still occurs. The network will give a warning if this happens, please re-run main.py if it does so.

Training loss may increase near the end, this is because KLD weight is aneeled so it increases over epochs, the loss plot displays the actual reconstruction loss and KLD (un-weighted), which in testing always converges.

network has 2 latent nodes for mu and log sigma, and has a single convolutional layer in its encoder as required.