Skip to content

Latest commit

 

History

History
166 lines (114 loc) · 5.8 KB

README.md

File metadata and controls

166 lines (114 loc) · 5.8 KB

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.

Requirements

Installation

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

  1. Install dependencies.
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.