Handwritten digit classification with Pytorch

This project uses the MNIST dataset for training. It has a total of 70000 handwritten digits split into train set and test set of 60000 and 10000 images respectively. The images are 28x28 pixelated grayscale images of single handwritten digits between 0 and 9.

The objective of this project is to classify a given image of handwritten digit into a integer from 0 to 9.

The process will be broken down into the following steps:

• Exploring the dataset
• Defining a neural network architecture
• Hyper-parameter search and Training the model
  • without validation
  • with validation
• Evaluating the model’s results and plotting graphs
• Experiment how RNNs do on image data
  • defining various RNN architectures
  • training those models

Results:

	Train Accuracy	Validation Accuracy	Test Accuracy
`Training without validation	99.30%	-	99.19%
`Training with validation	*99.34%	99.06%	99.14%
^Training with Vanilla RNN	*95.18%	-	95.86%
^Training with GRU	*99.42%	-	98.97%
^Training with LSTM	*99.24%	-	98.85%
^Training with Bidirectional LSTM	*99.16%	-	98.89%

_{* - running accuracy;
` - trained for 4 epochs;
^ - trained for 8 epochs;}

Todo

• data exploration
• train model with validation
• experiment how RNNs do on image data
• add data augmentation
• deploy with flask