Transfer Learning: CIFAR100

This project deals with the multi-class images classification of images from the CIFAR-100 Dataset.

Dataset:

The CIFAR100 dataset has over 60000 32x32 images from 100 different classes. This dataset is split into 5:1 ratio for train and validation.
I have performed data augmentation using "torchvision.transforms" library. In this i performed HorizontalFlipping and Normalization

Models:

1. I first started by building a custom VGG-16 Network with 13 Convolution Layers and a classifier Network.
   
2. I used the EfficientNetB0 Model trained on 1 million+ images from ImageNet Dataset and finetuned it on our data.
   
3. Lastly I applied transfer learning on the pretrained ResNet50 Model
   

Results:

Test Accuracy :

1. VGG-16: 49%
   
2. EfficientNetB0 : 79%
   

3) ResNet50 : 97%