Transfer Learning Lab

This is my implementation for the Transfer learning lab in the Deep Learning course taught in Zewail City.

Dataset

The dataset is the large scale fish dataset from kaggle :https://www.kaggle.com/crowww/a-large-scale-fish-dataset

In it, there are images for 10 classes of fish. We are required to build an image classifier (10 classes), using transfer learning.

Steps taken

I used

• VGG19 convolutional layers architecture, with pretrained weights on imagenet
• Then I used much simpler Fully connected layers, than those in the VGG19 architecture, and at the end used a softmax activated layer
• I also used Batch normalization for the Fully connected layers

Results:

The results were really impressive, the test accuracy is about 99% with only 5 epochs of training. This is mainly due to the weights learned from the imagenet data.

Loss and accuracy across epochs

Confusion matrix: