DEtection TRansformer

A minimal pytorch implementation of DEtection TRansformer(DETR). Validated on VOC 2012 and VOC 2007 datasets.

Configuration is managed using Hydra, with optional experiment tracking via Weights & Biases (wandb).

Setup

Setup environment

Install uv and run

uv sync

Download dataset from kaggle

./download-voc-dataset.sh

Train

uv run train.py

Changing model

python train.py model=detr-resnet34

Resuming from checkpoint

python train.py init_from=logs/<run_name>/detr-resnet50.pt

Evaluation

uv run eval.py <checkpoint_file>

To calculate mAP

uv run eval.py <checkpoint_file> --calculate_map