This is the homepage of PACE including tool implementation
, evaluation scripts
, studied DNN models
, corresponding testing sets
and experiment results
.
Before running PACE, please make sure you have installed various related packages, including keras, tensorflow, hdbscan and sklearn.
You can install hdbscan with the following command:
pip install hdbscan
Please use the following command to execute PACE:
python -u -m mnist_cifar_imagenet_svhn.selection --exp_id=lenet1 --select_layer_idx=-3 --dec_dim=8 --min_samples=4 --min_cluster_size=80
exp_id
: the id of the modelselect_layer_idx
: index of layer which is selected to extract featuredec_dim
: the dimension after reductionmin_samples
andmin_cluster_size
: the parameters required for hdbscan clustering
Also, we put the raw data results for all experiments in AllResult
.
We published all studied DNN models we utilized and you can find them in mnist_cifar_imagenet_svhn\model
.
Meanwhile, we released all corresponding testing sets in the mnist_cifar_imagenet_svhn\data
. The data of MNIST, CIFAR-10 and CIFAR-100 can be obtained directly from Keras API.
Regarding to Driving, the pre-trained models can be found in folder driving
, and the testing sets are in the driving\testing
.