Instance-conditional Knowledge Distillation for Object Detection

This is a MegEngine implementation of the paper “Instance-conditional Knowledge Distillation for Object Detection”, based on MegEngine Models.

The pytorch implementation based on detectron2 will be released soon.

Instance-Conditional Knowledge Distillation for Object Detection,
Zijian Kang, Peizhen Zhang, Xiangyu Zhang, Jian Sun, Nanning Zheng
In: Proc. Advances in Neural Information Processing Systems (NeurIPS), 2021
[arXiv]

Requirements

Installation

In order to run the code, please prepare a CUDA environment with:

  1. Install dependancies.

pip3 install --upgrade pip
pip3 install -r requirements.txt
  1. Prepare MS-COCO 2017 dataset，put it to a proper directory with the following structures:

/path/to/
    |->coco
    |    |annotations
    |    |train2017
    |    |val2017

Microsoft COCO: Common Objects in Context Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. European Conference on Computer Vision (ECCV), 2014.

Usage

Train baseline models

Following MegEngine Models:

python3 train.py -f distill_configs/retinanet_res50_coco_1x_800size.py -n 8 \
                       -d /data/Datasets

train.py arguments：

  • -f, config file for the network.
  • -n, required devices(gpu).
  • -w, pretrained backbone weights.
  • -b, training batch size, default is 2.
  • -d, dataset root，default is /data/datasets.

Train with distillation

python3 train_distill_icd.py -f distill_configs/retinanet_res50_coco_1x_800size.py \ 
    -n 8 -l -d /data/Datasets -tf configs/retinanet_res101_coco_3x_800size.py \
    -df distill_configs/ICD.py \
    -tw _model_zoo/retinanet_res101_coco_3x_800size_41dot4_73b01887.pkl

train_distill_icd.py arguments：

  • -f, config file for the student network.
  • -w, pretrained backbone weights.
  • -tf, config file for the teacher network.
  • -tw, pretrained weights for the teacher.
  • -df, config file for the distillation module, distill_configs/ICD.py by default.
  • -l, use the inheriting strategy, load pretrained parameters.
  • -n, required devices(gpu).
  • -b, training batch size, default is 2.
  • -d, dataset root，default is /data/datasets.

Note that we set backbone_pretrained in distill configs, where backbone weights will be loaded automatically, that -w can be omitted. Checkpoints will be saved to a log-xxx directory.

Evaluate

python3 test.py -f distill_configs/retinanet_res50_coco_3x_800size.py -n 8 \
     -w log-of-xxx/epoch_17.pkl -d /data/Datasets/

test.py arguments：

  • -f, config file for the network.
  • -n, required devices(gpu).
  • -w, pretrained weights.
  • -d, dataset root，default is /data/datasets.

Examples and Results

Steps

  1. Download the pretrained teacher model to _model_zoo directory.
  2. Train baseline or distill with ICD.
  3. Evaluate checkpoints (use the last checkpoint by default).

Example of Common Detectors

RetinaNet

Command:

python3 train_distill_icd.py -f distill_configs/retinanet_res50_coco_1x_800size.py \
    -n 8 -l -d /data/Datasets -tf configs/retinanet_res101_coco_3x_800size.py \
    -df distill_configs/ICD.py \
    -tw _model_zoo/retinanet_res101_coco_3x_800size_41dot4_73b01887.pkl

FCOS

Command:

python3 train_distill_icd.py -f distill_configs/fcos_res50_coco_1x_800size.py \
    -n 8 -l -d /data/Datasets -tf configs/fcos_res101_coco_3x_800size.py \
    -df distill_configs/ICD.py \
    -tw _model_zoo/fcos_res101_coco_3x_800size_44dot3_f38e8df1.pkl

ATSS

Command:

python3 train_distill_icd.py -f distill_configs/atss_res50_coco_1x_800size.py \
    -n 8 -l -d /data/Datasets -tf configs/atss_res101_coco_3x_800size.py \
    -df distill_configs/ICD.py \
    -tw _model_zoo/atss_res101_coco_3x_800size_44dot7_9181687e.pkl

Results of AP in MS-COCO:

Model Baseline +ICD
Retinanet 36.8 40.3
FCOS 40.0 43.3
ATSS 39.6 43.0

Notice

  • Results of this implementation are mainly for demonstration, please refer to the Detectron2 version for reproduction.

  • We simply adopt the hyperparameter from Detectron2 version, further tunning could be helpful.

  • There is a known CUDA memory issue related to MegEngine: the actual memory consumption will be much larger than the theoretical value, due to the memory fragmentation. This is expected to be fixed in a future version of MegEngine.

Acknowledgement

This repo is modified from MegEngine Models. We also refer to Pytorch, DETR and Detectron2 for some implementations.

License

This repo is licensed under the Apache License, Version 2.0 (the “License”).

Citation

@inproceedings{kang2021icd,
    title={Instance-conditional Distillation for Object Detection},
    author={Zijian Kang, Peizhen Zhang, Xiangyu Zhang, Jian Sun, Nanning Zheng},
    year={2021},
    booktitle={NeurIPS},
}

GitHub

View Github