模型在2080Ti上的推理耗时 FP32/wo TensorRT

• Linux or MacOS

• CUDA >= 10.0

• Python >= 3.6

• Pytorch >= 1.7

• experimental support Windows (Notice: Windows not support distributed training before pytorch1.7)

• Install requirements

pip install -r requirements.txt

• Setup NanoDet

python setup.py develop

1. Prepare dataset convert your dataset annotations to MS COCO format

2. Start training

   Baseline and DRDet are both now using pytorch lightning for training.

   For both single-GPU or multiple-GPUs, run:

   python tools/train.py CONFIG_FILE_PATH

3. SAHI

   slice dataset with coco format

   >> pip install sahi
   >> sahi coco slice --image_dir dir/to/images --dataset_json_path dataset.json

   slice the given images and COCO formatted annotations and export them to given output folder directory.

   Specify slice height/width size as --slice_size 512.

   Specify slice overlap ratio for height/width size as --overlap_ratio 0.2.

   If you want to ignore images with annotations set it add --ignore_negative_samples argument.

4. Test and Vis

   python demo/demo.py --demo image --config ./config/nanodet-plus-m-1.5x_416.yml --model ./nanodet-plus-m-1.5x_416_checkpoint.ckpt --path ./inferimgs