SQ-DETR

SQ-DETR research code for DETR-style object detection experiments with query-dropping configurations.

This repository is a cleaned public snapshot of the implementation surface: model code, training entrypoint, and representative experiment configs. Datasets, checkpoints, generated outputs, notebooks, and machine-local helper notes are intentionally excluded.

Contents

• src/ — model, solver, data, optimization, and utility code
• configs/ — RT-DETR baselines and SQ-DETR experiment configs
• tools/train.py — train/evaluate/FLOPs entrypoint
• tools/experiments.sh — example commands

Setup

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

Install the PyTorch/CUDA build that matches your machine if the pinned wheel in requirements.txt is not appropriate for your environment.

Data layout

Configs use git-ignored local dataset paths by default:

datasets/
  coco2017/
    train2017/
    val2017/
    annotations/instances_train2017.json
    annotations/instances_val2017.json
  CrowdHuman/
    train/
    val/
    annotations/train.json
    annotations/val.json

Adjust img_folder, ann_file, and output_dir in the selected config if your layout differs.

Example commands

# Train SQ-DETR R18 on a COCO-style config
torchrun --nproc_per_node=2 tools/train.py -c configs/sqdetr/r18/r18_hc_50_50_50.yml

# Evaluate from a local checkpoint
torchrun --nproc_per_node=2 tools/train.py   -c configs/sqdetr/r50/r50_hc_25.yml   -r output/sq_detr_r50/hc_25/best.pth   --test-only

# Print FLOPs breakdown
torchrun --nproc_per_node=2 tools/train.py   -c configs/sqdetr/r50/r50_hc_25.yml   --flops

Local-only files

Do not commit datasets, checkpoints, generated runs, notebooks, local environment files, credentials, or personal command notes. The repository .gitignore covers the usual paths and artifact types.

License

Apache License 2.0.