lcj_MNIST
A slight change from https://github.com/peikexin9/deepxplore/tree/master/MNIST
Generate images with shape (1,28,28,1) to activate a set of neurons (instead of an individual neuron as in current implementation).
What I changed?
- init_coverage_tables --> init_set_coverage_tables
- init_dict --> init_set_dict
- neuron_to_cover --> set_to_cover
- neuron_covered --> set_covered
- update_coverage --> set_update_coverage
Change about loss
# loss1_neuron = K.mean(model1.get_layer(layer_name1).output[..., index1])
loss1_neuron = K.mean(K.stack([K.mean(model1.get_layer(location[0]).output[..., location[1]]) for location in location_list1]))
How to use
python gen_diff.py -h
An example:
python gen_diff.py 'occl' 0.5 0.1 20 20 30 0