add classification implementation

Author: hunto

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "classification"]
+	path = classification
+	url = https://github.com/hunto/image_classification_sota.git

README.md

@@ -1,21 +1,32 @@
 # Knowledge Diffusion for Distillation (DiffKD)
-Official implementation for paper "[Knowledge Diffusion for Distillation](https://arxiv.org/abs/2305.15712)" (DiffKD).
+Official implementation for paper "[Knowledge Diffusion for Distillation](https://arxiv.org/abs/2305.15712)" (DiffKD), NeurIPS 2023
 
 ---
 
-## Updates
-### June 9, 2023
-We released the core code of DiffKD. See `example.py` for usage.
+## Reproducing our results
 
----
+```shell
+git clone https://github.com/hunto/DiffKD.git --recurse-submodules
+cd DiffKD
+```
 
-## Reproducing our results
+**The implementation of DiffKD is in** [classification/lib/models/losses/diffkd](https://github.com/hunto/image_classification_sota/tree/main/lib/models/losses/diffkd).
 
-**We are refactoring our code and reproducing the results.**
+* classification: prepare your environment and datasets following the `README.md` in `classification`.
 
 ### ImageNet
 
-* Baseline settings (`R34-R101` and `R50-MBV1`):  
+```
+cd classification
+sh tools/dist_train.sh 8 ${CONFIG} ${MODEL} --teacher-model ${T_MODEL} --experiment ${EXP_NAME}
+```
+
+Example script for reproducing DiffKD on ResNet-34 teacher and ResNet-18 student with B1 baseline setting:
+```
+sh tools/dist_train.sh 8 configs/strategies/distill/diffkd/diffkd_b1.yaml tv_resnet18 --teacher-model tv_resnet34 --experiment diffkd_res34_res18
+```
+
+* Baseline settings (`R34-R18` and `R50-MBV1`):  
     ```
     CONFIG=configs/strategies/distill/TODO
     ```
@@ -24,3 +35,16 @@ We released the core code of DiffKD. See `example.py` for usage.
     |ResNet-18 (69.76)|ResNet-34 (73.31)|72.20|`tv_resnet18`|`tv_resnet34`|[log](https://github.com/hunto/DiffKD/releases/download/v0.0.1/log_b1_diffkd_res34_res18.txt)|[ckpt](https://drive.google.com/file/d/19Wy5RCfpkAg9oUIiDUhfWtl8DjD712DN/view?usp=sharing)|
     |MobileNet V1 (70.13)|ResNet-50 (76.16)|73.24|`mobilenet_v1`|`tv_resnet50`|to be reproduced||
 
+
+## License  
+This project is released under the [Apache 2.0 license](LICENSE).
+
+## Citation 
+```
+@article{huang2023knowledge,
+  title={Knowledge Diffusion for Distillation},
+  author={Huang, Tao and Zhang, Yuan and Zheng, Mingkai and You, Shan and Wang, Fei and Qian, Chen and Xu, Chang},
+  journal={arXiv preprint arXiv:2305.15712},
+  year={2023}
+}
+```