Ambiguity about Coverage

peikexin9 / deepxplore

DeepXplore code release

MIT License

393 stars 146 forks source link

Ambiguity about Coverage #15

Closed alipourm closed 6 years ago

alipourm commented 6 years ago

It's not clear how covered is calculated.

Why are the neurons scaled?
Why are they averaged to decide to decide if they are fired?

https://github.com/peikexin9/deepxplore/blob/dc8b405cb58bdc8e26d7ed0beeaf784115bc3ee7/PDF/utils.py#L93

peikexin9 commented 6 years ago

Sometimes one layer's output is in different order of magnitude with another. So to enable setting a global threshold to determine if a neuron is fired, I make each layer in same scale. For a convolution layer, I treat the neuron's output as the average of values in feature map, so it's averaged.