CNTK Examples: Image/TransferLearning

Overview

Data:	A data set containing images of 102 different types of flowers (website).
Purpose	Demonstrate how to perform transfer learning in CNTK.
Network	Pre-trained ResNet_18 model, which is modified to fit the flowers data set.
Training	In this example all layers (old and new) are trained with the same learning rate.

Running the example

Getting the data

We use the Flowers data set (Examples/Image/DataSets/Flowers) and a pre-trained ResNet_18 model PretrainedModels/ResNet18_ImageNet_CNTK.model. To download both run

python install_data_and_model.py

Details

Run python TransferLearning.py to train and evaluate the transfer learning model. The model achieves 93% accuracy on the Flowers data set after training for 20 epochs. A detailed walk through is provided in the 'Build your own image classifier using Transfer Learning' tutorial on the CNTK github wiki.