This repository contains a transfer learning exercises on CIFAR-10 done in Keras.
We provided utilities to download, extract and visualise the data.
A number of models are fitted:
- baseline: HOG features + linear SVM
- SVM on top of CNN codes extracted using ResNet50 pretrained on ImageNet
- Fine tuning of ResNet50 (with discussion of suitability of Keras BN layer to fine tuning task)
- Fine tuning with data augmentation
Both development and training were conducted on Google Colab.
If you want to recreate Google Colab environment locally, pip install -r requirements.txt in your virtualenv.
In custom_resnet directory you can find code needed to perform transfer learning with ResNet50 in Keras. It was adapted from keras-applications by Keras Team and keras-resnet by Broad Institute.