Skip to content

Transfer learning on VGG16 using Keras with Caltech256 and Urban Tribes dataset. Dark knowledge in transfer learning.

License

Notifications You must be signed in to change notification settings

aliasvishnu/Keras-VGG16-TransferLearning

Repository files navigation

Keras-VGG16-TransferLearning

Introduction

The aim of this project is to understand ConvNets, use transfer learning to solve (kinda) the challenging problem of image recognition over 2 different datasets - Caltech256[1] and Urban tribes[2]. Caltech256 is a dataset with 256 classes of items, whereas urban tribes is dataset with 11 classes of people like - bikers, goth etc. There are 3-4 files in this repo. Here's a description of what they contain.

  • VGG Transfer Learning.ipynb - Shows how to create the transfer learning model, build the dataset from Caltech 256 images and train the model.
  • Urban Tribes.ipynb - Same thing as above, but applied on Urban Tribes dataset.
  • Caltech256 - Dark knowledge.ipynb - Shows how the temperature parameter can help uncover the knowledge in a network and use it to transfer knowledge.

[1] http://www.vision.caltech.edu/Image_Datasets/Caltech256/ [2] http://www.bmva.org/bmvc/2013/Papers/paper0014/abstract0014.pdf