MNIST

Playing around with the MNIST dataset using ML

CNN Model

Create a CNN model to classify handwritten images of digits (MNIST dataset)

1. Train the model Currently 10 epochs, which takes 1 min to train

py main.py train_model

2. Test the model

py main.py test_model

Accuracy is 98%

Fine Tuning ViT

Fine tune ViT on the MNIST dataset

py main.py fine_tune  --saved_model atancoder/ViT_fine_tuned_mnist

Accuracy is 99.6%

Generative Eneergy Model

Create a simple energy model that can generate images

Linear model works better than CNN. Images generated not as good as VAE. Had to subset data to a particular class label to get it working decently.

cd energy_model
py main.py train_model
py main.py gen_images

For linear model

• LR: 1e-5
• LD step size: 1e-3

For CNN model

• LR: 1e-4
• LD step size: 1e-1

VAE

Works great! Trained for 50 epochs

cd VAE
py main.py train_model
py main.py gen_images