README.md

This repository is the official PyTorch implementation of training & evaluation code for Distilling Image Classifiers in Object Detectors, NeurIPS 2021.

Distilling Image Classifiers in Object Detectors

Code is in early release and may be subject to change. Please feel free to open an issue in case of questions.

We use PyTorch and MMDetection v2.10.0 as the codebase.

Overview

Paper: Distilling Image Classifiers in Object Detectors

Figure 1: Overview of our classifier-to-detector distillation framework. (a) Existing methods perform distillation across corresponding stages in the teacher and student, which restricts their applicability to detector-to-detector distillation. (b) By contrast, we introduce strategies to transfer the knowledge from an image classification teacher to an object detection student, improving both its recognition and localization accuracy.

Environment

We provide a docker file under the root directory ([root-dir]) for better reproducibility. Note that the base image relies access to nvcr.io. Then, you need to login to build the docker image. Please see nvidia-docker for more information to build and run docker container.

After the container is alive, please run the following commands to build up our developed MMDetection:

pip install -r requirements/build.txt
pip install -v -e .

We mainly use NVIDIA Tesla V100 (16 or 32GB) for our experiments.

Data

All experiments are conducted on MS COCO2017. Please download the dataset and make sure you can run a baseline model successfully. More details are here. The dataset should be organized as:

coco2017
 ├── annotations
 ├── train2017
 ├── val2017
 ├── test2017

Note that you need to set the data_root in configs_dev/ssd/ssd300_coco_cls_loc.py.

Pre-trained Classification Teacher

We provide the pre-trained classification teacher model for SSD300, which is ResNet50 trained on COCO2017 classification dataset. Please download it first from Drive and unzip it to [root-dir]/cls_teachers/ssd/. The details of building up the classification dataset and training the teachers are in the supplementary material Section S2.

Training

To train a model, run training script under the [root-dir].

SSD300 example

bash ./tools/dist_train_cls_loc.sh \
      configs_dev/ssd/ssd300_coco_cls_loc.py \ # configuration file
      8 # num of gpus

This will reproduce the result of SSD300 in Table 1 in the main paper:

Table A: Results of our classifier-to-detector distillation with SSD300 on the COCO2017 validation set.

Model	mAP
SSD300-VGG16	25.6
+ Ours	27.9

Evaluation

The trained SSD300 model can be evaluated as follow:

# single gpu
python ./tools/test.py \
       configs/ssd/ssd300.py \
       [checkpoint.pth] \
       --eval bbox

# multiple gpus
bash ./tools/dist_test.sh \
       configs/ssd/ssd300.py \
       [checkpoint.pth] \
       8 # num of gpus

License

Please check the LICENSE file. This work may be used non-commercially, meaning for research or evaluation purposes only. For business inquiries, please contact [email protected].

Citation

@article{guo2021distilling,
         title={Distilling Image Classifiers in Object Detectors}, 
         author={Shuxuan Guo and Jose M. Alvarez and Mathieu Salzmann},
         year={2021},
         eprint={2106.05209},
         archivePrefix={arXiv},
         primaryClass={cs.CV}
}