unit_visualization

CNN Unit Visualization

This code adapts the reference code for the paper Object Detectors Emerge in Deep Scene CNNs (arXiv) to visualize the internal units of trained CNNs.

Usage

The generate_unitsegments.py script takes in a trained model's config file and a checkpoint epoch to use. The script uses the validation split of our DDSM data. To generate unit visualizations for the same models we used, please go to training, download the pretrained models, and then use the following commands:

# alexnet_2class
python generate_unitsegments.py --experiment_name alexnet_2class --config_path ../training/pretrained/alexnet_2class/config.yml --epoch 45

# vgg16_2class
python generate_unitsegments.py --experiment_name vgg16_2class --config_path ../training/pretrained/vgg16_2class/config.yml --epoch 12

# inception_v3_2class
python generate_unitsegments.py --experiment_name inception_v3_2class --config_path ../training/pretrained/inception_v3_2class/config.yml --epoch 7

# resnet152_2class
python generate_unitsegments.py --experiment_name resnet152_2class --config_path ../training/pretrained/resnet152_2class/config.yml --epoch 5

# resnet152_3class
python generate_unitsegments.py --experiment_name resnet152_3class --config_path ../training/pretrained/resnet152_3class/config.yml --epoch 5

The output visualizations will be written to output/images/, in a subdirectory specified by --experiment_name. Visualizations are created for all units in each convolutional layer. By default, the top 64 images for each unit are visualized. For example, output/images/alexnet/conv5/unit_0001 would contain 64 images corresponding to visualizations of the top 64 images for unit_0001 in the conv5 layer of a trained AlexNet.