Creating a hand written digit classifier based on MNIST dataset from scratch (without using any available network architectures).
- Purpose of this code is to understand the various components and inner workings of a simple deep learning model.
- This uses pure python, Pytorch and minimal fastai functions (only DataLoader for dividing the data into batches) to implement the MNIST classifier from scratch.
- The code does not use any pre-trained or available network architectures like resent etc. However, for illustration purpose one epoch training using resnet18 has been shown which generates much higher accuracy owing to deeper architecture than what is implemeted in the manual way.