This repo constains the pytorch implementation for the paper Towards Realistic Predictors on ECCV2018.
- python 3.6
- pytorch = 0.4
- other common modules
Since our code is not an end-to-end trainable model (for a improved new version, please turn to an end to end implementation for realistic predictors), please run
train_HPnet_dataset_net_net.py
to get the hardness predictor after replacing 'dataset' and 'net' with 'imagenet' or 'indoor' and 'res' or 'vgg'.
Then run
train_rp_dataset_net_net.py
to get realistic predictors.
We used the custom data load form, so before training, please first make a train and test sample list. Each item is formed as
imagepath target index
For questions, feel free to reach
Pei Wang: [email protected]