AOGNet-v2

This project provides source code for AOGNets: Compositional Grammatical Architectures for Deep Learning (CVPR 2019) and Attentive Normalization.

Installation

1. Create a conda environment, and install all dependencies.

conda create -n aognet-v2 python=3.7
conda activate aognet-v2
git clone https://github.com/iVMCL/AOGNet-v2
cd AOGNet-v2
pip install -r requirements.txt

2. Install PyTorch 1.0+, follows https://pytorch.org/

3. Install apex, follows https://github.com/NVIDIA/apex

ImageNet dataset preparation

• Download the ImageNet dataset to YOUR_IMAGENET_PATH and move validation images to labeled subfolders

• The script may be helpful.

• Create a datasets subfolder under your cloned AOGNet-v2 and a symbolic link to the ImageNet dataset

$ cd AOGNet-v2
$ mkdir datasets
$ ln -s PATH_TO_YOUR_IMAGENET ./datasets/

Pretrained models for Evaluation

$ cd AOGNet-v2
mkdir pretrained_models

Download pretrained models from the links provided in the tables below, unzip it and place into the pretrained_models directory.

$ ./scripts/test_fp16.sh <pretrained model directory path>

• Models trained with normal training setup: 120 epochs, cosine lr scheduling, SGD, ... (MobileNet-v2 trained with 150 epochs). All models are trained with 8 Nvidia V100 GPUs.

* From original paper

• Models trained with advanced training setup: 200 epochs, 0.1 label smoothing, 0.2 mixup

• Remarks: The accuracy from the pretrained models might be slightly different than that reported during the training (used in our papers). The reason is still unclear.

Perform training on ImageNet dataset

$ cd AOGNet-v2
$ ./scripts/train_fp16.sh <config_fie>

e.g. For training AOGNet-12M with AN

$ ./scripts/train_fp16.sh configs/aognet-12m-an-imagenet.yaml

See more configuration files in configs. Change the GPU settings in scripts/train_fp16.sh

Object Detection and Instance Segmentation on COCO

We performed object detection and instance segmentation task on COCO with our models pretrained on ImageNet. We implement it based on the mmdetection framework. The code is released in https://github.com/iVMCL/AttentiveNorm_Detection/tree/master/configs/attn_norm.

Citations

Please consider citing the AOGNets or Attentive Normalization papers in your publications if it helps your research.

@inproceedings{li2019aognets,
  title={AOGNets: Compositional Grammatical Architectures for Deep Learning},
  author={Li, Xilai and Song, Xi and Wu, Tianfu},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={6220--6230},
  year={2019}
}

@article{li2019attentive,
  title={Attentive Normalization},
  author={Li, Xilai and Sun, Wei and Wu, Tianfu},
  journal={arXiv preprint arXiv:1908.01259},
  year={2019}
}

Contact

Please feel free to report issues and any related problems to Xilai Li (xli47 at ncsu dot edu), and Tianfu Wu (twu19 at ncsu dot edu).

License

AOGNets related codes are under RESEARCH ONLY LICENSE.