This is the code of paper deepinfo.
run the following command:
python main.py --datasets_name=CUB --net_name=vgg16 --step_name=net;
python main.py --datasets_name=CUB --net_name=vgg16 --step_name=x;important parameters:
- datasets_name:to select the dataset，used in tools.get_input.
- out_planes/image_size：parameters depending on datasets_name。
- net_name:to select the evaluated neural network，used in tools.get_input。
- net_lr/net_batch_size/...：parameters to train the evaluated neural network.
- step_name: "net" for network training, x for SID compuation, decoder for decoder training, y for RU computation.
- capability_batch_size: to control the batch size of noise in SID/RU computation. If it's very small, the result will not be stable. It can't be very large, limited to GPU memory.
- capability_lambda_init_x/capability_lambda_init_y: It's the lambda parameter shown in our paper, which can balance two kinds of loss function. It can't be either very large or very small.
TODO: add datasets and trained net/decoder
TODO: give more explanations to all the parameters.