A numpy based python package with pytorch-like interface for training a Convolution Neural Network.
To train and evaluate a model, please refer to trainer.py and evalute.py. We also show that we can also perform fine-tuning in trainer.py.
We implemented some image augmentation based on numpy only, please refer to NumpyTorch/transforms.py. The interface of data augmentation is like torchvision.transforms.
We provide some configs and pre-trained models in config/ and checkpoints/. In our experiment, we show that the model trained in this package outperforms pytorch's implementation.
If you have any problem about this implementation, please let me know by opening an issue.