rohansh-tty/ForgottenPlutonium

TinyImageNet Classifier using ResNet18(MMDA)

ForgottenPlutonium

*Why this name? Because I think TinyImageNet Dataset is underrated and I wanted to make it cool and radioactive :sunglasses:

Notebook:

Model Features:

1. Used GPU as Device.

2. CNN Type: ResNet18

3. Total Params: 11,271,432

4. Implemented MMDA, used Albumentations since it's easy to integrate with PyTorch.

5. Also Trained the model a bit harder by adding Image Augmentation Techniques like RandomCrop, Flip & Cutout.

6. Max Learning Rate: 0.01

7. Used NLLLoss() to calculate loss value.

8. Ran the model for 50 Epochs

    * Highest Validation Accuracy: 51.20%
   

9. GradCam for 25 Misclassified Images.

Library Documentation:

1. : Applies required image transformation to both Train & Test dataset using Albumentations library.

2.: Consists of Custom DataSet Class and some helper functions to apply transformations, extract classID etc.

3.: Consists of main ResNet model

4.: Scripts to Test & Train the model.

5.: Scripts to load the datasets.

6.: Consists of helper functions to plot images from dataset & misclassified images.

7.: Consists of Gradcam class & other related functions.

8.: LR finder using FastAI Approach.

9.: Consists helper functions related to CycliclR.

Plots & Curves

Model Performance

Misclassified Images

GradCam for Misclassified Images

Model Logs

•