jasonrig/transfer-learning-demos

Demo code for transfer learning and CNN analysis

Transfer Learning Demo code

This package is companion code for my presentation on convolutional neural networks, downloadable here: https://doi.org/10.4225/03/5acdaeadb87d0

The package can be installed on your system using pip:

pip install -U git+https://github.com/jasonrig/transfer-learning-demos

Visualising network activations

python3 -m TransferLearningDemo.demos.vgg_19_activate_filters <conv_block> <conv_layer> <filter_index>

To visualise the second filter of the first convolutional layer of the fifth block, run:

python3 -m TransferLearningDemo.demos.vgg_19_activate_filters 5 1 1

Remember that filter numbers are indexed from zero, whereas the convolutional layers are indexed from one.

Training a retinal haemorrhage detection model

Train using:

python3 -m TransferLearningDemo.demos.vgg_19_retrain_fc train

Evaluate using:

python3 -m TransferLearningDemo.demos.vgg_19_retrain_fc evaluate

Predict image(s) using:

python3 -m TransferLearningDemo.demos.vgg_19_retrain_fc predict <image_name>

Add multiple files for more than one prediction.

Neural style transfer

To run using the default style and content images:

python3 -m TransferLearningDemo.demos.vgg_19_style_transfer

To run using your own (jpeg only):

python3 -m TransferLearningDemo.demos.vgg_19_style_transfer <style_image> <content_image>