Skip to content

Latest commit

 

History

History
97 lines (87 loc) · 2.64 KB

README.md

File metadata and controls

97 lines (87 loc) · 2.64 KB

Installation

We introduce the environment set up for evaluation and training for object detection on COCONut. More details refer to DETA repo.

Requirements

  • Linux or macOS with Python ≥ 3.7
  • pip install -r requirements.txt

Example virtualenv environment setup for kMaX-DeepLab

tested on torch==1.13.0+cu11.7

pip3 install virtualenv
python3 -m virtualenv deta --python=python3
source deta/bin/activate

pip install torch==1.13.1 torchvision==0.14.1 
git clone https://github.com/jozhang97/DETA.git
pip install -r requirements.txt

# Compiling CUDA operators
cd ./models/ops
sh ./make.sh
# unit test (should see all checking is True)
python test.py

Example test model checkpoint of DETA

  1. Prepare dataset, the dataset structure should be as follow:
datasets
└── coco
    ├── annotations 
    │   └── instances_val2017.json # relabeled_coco_val.json
    ├── val2017 # original COCO dataset val set images
  1. Use the script below to evaluate the model.
GPUS_PER_NODE=8 ./tools/run_dist_launch.sh 8 ./configs/deta_swin_ft.sh \
--eval --coco_path YOUR_COCO_DATASET_PATH \ # for example, ./datasets/coco
--resume YOUR_MODEL_CHECKPOINT

Distributed training

Need to set up your environment variables to run the training script below.

MASTER_ADDR=$WORKER_0_HOST_IP NODE_RANK=$NODE_ID GPUS_PER_NODE=8 \
./tools/run_dist_launch.sh 16 ./configs/deta_swin_ft.sh \
--coco_path YOUR_COCO_DATASET_PATH

Model zoo

COCO-val relabeled COCO-val COCONut-val
backbone training set AP_box AP_box AP_box model
Swin-L COCO 59.1 58.6 56.1 download
COCONut-S 54.5 61.3 58.9 download
COCONut-B 59.3 62.2 60.1 download