leaplabthu/adaptivenn-jittor

This repository is a port of the AdaptiveNN image classification model from PyTorch to a different AI computing framework called Jittor. Jittor is a deep learning framework developed by Tsinghua University that uses just-in-time compilation to speed up neural network operations. The original AdaptiveNN was published as research from the same university and appeared in the journal Nature Machine Intelligence. AdaptiveNN is described as a vision model that adapts how much computation it spends on different parts of an image, similar to how human eyes pay more attention to certain areas. This repository provides code for running that model and training it on a large image classification dataset called ImageNet-1K, which contains about 1.2 million images across 1,000 categories. The port reproduces the original model's accuracy: roughly 82 percent top-1 accuracy on the ImageNet validation set, matching what the PyTorch version achieves. Inference speed under the best configuration is about 88 percent of PyTorch's speed. Training is currently only stable when run in full 32-bit floating point mode, mixed precision training (which speeds things up on modern GPUs) is experimental and may produce numerical errors. The repository includes shell scripts for benchmarking inference speed across different batch sizes and precision settings, a training script, and instructions for setting up the environment using conda and pip. Users need to supply their own copy of the ImageNet dataset and download the pre-trained model weights separately, following a companion evaluation guide included in the repository. The code is licensed under MIT.