Image Classifier Using Pre-Trained Models in Keras

This repository contains the example code for our article on pre-trained deep learning models with Keras.

Train, predict, visualize, and produce class-activation map animations for deep learning models in Keras using pre-trained models as their basis.

Dependencies

• Python 3.5+
• Imagemagick 7+

Running the Example

1. Download the example dataset

2. Preprocess the data

python convert_data.py --data-dir {path-to-data}

3. Train the model

python train.py --pretrained_model {model} \
                --data-dir {path-to-data} \
                --weight-directory {path-to-weight-directory} \
                --tensorboard-directory {path-to-tensorboard-logdir} \
                --epochs {max_epochs}

4. Visualize model predictions

python visualize.py --weight-file {path-to-weight-file} \
                    --data-directory {path-to-data} \
                    --output-directory {path-to-output-directory} \
                    --image-path {path-to-image-to-visualize}

5. Generate a CAM plot

python cam_animation.py --weight-directory {path-to-weight-directory} \
                        --data-directory {path-to-data-directory} \
                        --image-path {path-to-image-to-visualize} \
                        --cam-path {output-path-for-cam-images} \
                        --weight-limit {max-weights-to-plot}

convert -delay 30 -size 256x256 {output-path-for-cam-images}/*.png -loop 0 {final-gif-name}

To make the generation of CAM plots easier, you can use the ./generate_cam_gifs script. This assumes:

• Data directory is ../data_dir/simpsons_dataset
• Weight directory is ../data_dir/weights
• CAM output path is ../data_dir/cam_output/{model}/{character}
• All names passed into the script are basenames

# Generate a single CAM plot
./generate_cam_gifs {model} {character} {npz-file}
# Generate CAM plots for the first 100 images of a character
./generate_cam_gifs {model} {character}

About Innolitics

Innolitics is a team of talented software developers with medical and engineering backgrounds. Our mission is to accelerate progress in medical imaging by sharing knowledge, creating tools, and providing quality services to our clients, with the ultimate purpose of improving patient health. If you are working on a project that requires image processing or deep learning expertise, let us know! We offer consulting and development services.