Architecture is given, learning based on Deep Learning module.
Best accuracy found is 99.06% based on these methods and architecture given.
Experimented with LeNet5 implemented in PyTorch, using dataloader from dataset files,
Main experiments include:
-
Normalising dataset with mean and std of training dataset.
-
Applying data augmentations of 35 degrees rotation and affine.
-
Xavier Initialisation of Parameters.
-
Increasing and Descreasing Angles
-
Appling inverse Laplacian filter to enhance image.
-
Not sure if model is overfitting thus need graph per training.
-
Developed by: Michael Peres
-
Model type: LeNet5
-
Language(s) (NLP): English
-
License: MIT
-
Finetuned from model: LeNet5
