Autocare DLT

Autocare DeepLearning Toolkit은 SNUAILAB의 모델 개발 및 Autocare T의 학습을 지원하기 위한 pytorch 기반 deep learning toolkit입니다.

Updates

• v0.2
  • HPO 추가
  • Mutli-GPU 지원
  • inference 및 data_selection에서 coco input 지원

설치

Prerequisite

• Python >= 3.9
• CUDA == 11.3
• pytorch >= 1.12.1 (link)
  • torchvision >= 0.13.1

Install

• tx_model은 repo를 clone하여 CLI로 사용하는 방식 및 python package(*.whl) 파일을 통하여 설치하는 방법 2가지를 지원한다.

git clone

git clone git@github.com:snuailab/autocare_dlt.git
cd autocare_dlt
pip install -r requirements.txt

package 설치

pip install autocare_dlt

실행

Model config 준비

• 기본적인 template은 ./models참조
• 사용하고자 하는 Model에 맞춰서 config값 수정
  • 모듈에 따라 hyper-parameter값이 다양해지기 때문에 해당 모듈의 code를 참조하여 수정 할 것을 권장

Data config 준비

• 기본적인 template은 ./datasets 참조
• 사용하고자 하는 Dataset에 맞춰서 config값 수정
  • workers_per_gpu (int) : dataloader work 갯수
  • batch_size_per_gpu (int): GPU당 batch size
  • img_size (int): 모델의 image size (img_size, img_size) → 추후 업데이트 예정
  • train, val, test (dict): 각 dataset의 config
    • type: dataset의 type
    • data_root: data의 root path
    • ann: annotation 파일의 path
    • augmentation: data augmentation세팅
      • CV2 모듈들이 먼저 적용되고 pytorch(torchvision) 모듈 적용됨
      • top down 순서대로 적용

지원 하는 package Tools

• 해당 tool들을 import하여 사용 혹은 cli로 실행

autocare_dlt.tools.train.run(exp_name: str, model_cfg: str, data_cfg: str, gpus: str = '0', ckpt: ~typing.Union[str, dict] = None, world_size: int = 1, output_dir: str = 'outputs', resume: bool = False, fp16: bool = False, ema: bool = False)→ tooNone

Run training

Parameters

• exp_name (str) – experiment name. a folder with this name will be created in the output_dir, and the log files will be saved there.

• model_cfg (str) – path for model configuration file

• data_cfg (str) – path for dataset configuration file

• gpus (str, optional) – GPU IDs to use. Default to ‘0’

• ckpt (str, optional) – path for checkpoint file. Defaults to None.

• world_size (int, optional) – world size for ddp. Defaults to 1.

• output_dir (str, optional) – log output directory. Defaults to ‘outputs’.

• resume (bool, optional) – whether to resume the previous training or not. Defaults to False.

• fp16 (bool, optional) – whether to use float point 16 or not. Defaults to False.

• ema (bool, optional) – whether to use EMA(exponential moving average) or not. Defaults to False.

autocare_dlt.tools.inference.run(inputs: str, model_cfg: str, output_dir: str, gpus: str, ckpt: Union[str, dict], input_size: list = None, letter_box: bool = None, vis: bool = False, save_imgs: bool = False, root_dir: str = '')→ None

Run inference

Parameters

• inputs (str) – path for input - image, directory, or json

• model_cfg (str) – path for model configuration file

• output_dir (str) – path for inference results

• gpus (str) – GPU IDs to use

• ckpt (Union[str, dict]) – path for checkpoint file or state dict

• input_size (list, optional) – input size of model inference. Defaults to [640].

• letter_box (bool, optional) – whether to use letter box or not. Defaults to False.

• vis (bool, optional) – whether to visualize inference in realtime or not. Defaults to False.

• save_imgs (bool, optional) – whether to draw and save inference results as images or not. Defaults to False.

• root_dir (str, optional) – path for input image when using json input. Defaults to “”.

autocare_dlt.tools.eval.run(model_cfg: str, data_cfg: str, gpus: str, ckpt: Union[str, dict])→ None

Evaluate a model

Parameters

• model_cfg (str) – path for model configuration file

• data_cfg (str) – path for dataset configureation file

• gpus (str) – GPU IDs to use

• ckpt (Union[str, dict]) – path for checkpoint file or state dict

autocare_dlt.tools.export_onnx.run(output_name: str, model_cfg: str, ckpt: Union[str, dict], input_size: list = None, opset: int = 11, no_onnxsim: bool = False)→ None

Export onnx file

Parameters

• output_name (str) – file name for onnx output (.onnx)

• model_cfg (str) – path for model configuration file

• ckpt (Union[str, dict]) – path for checkpoint file or state dict

• input_size (list, optional) – input size of model. use model config value if input_size is None. Default to None.

• opset (int, optional) – onnx opset version. Defaults to 11.

• no_onnxsim (bool, optional) – whether to use onnxsim or not. Defaults to False.

autocare_dlt.tools.data_selection.run(model_cfg: str, ckpt: Union[str, dict], inputs: str, num_outputs: int, output_dir: str, gpus: str, input_size: list = None, letter_box: bool = None, copy_img: bool = False, root_dir: str = '')→ None

Select active learning data

Parameters

• model_cfg (str) – path for model configuration file

• ckpt (Union[str, dict]) – path for checkpoint file or state dict

• inputs (str) – path for input - image, directory, or json

• num_outputs (int) – number of images to select

• output_dir (str) – path for output result

• gpus (str) – GPU IDs to use

• input_size (list, optional) – input size of model inference. Defaults to [640].

• letter_box (bool, optional) – whether to use letter box or not. Defaults to False.

• copy_img (bool, optional) – whether to copy images to output. Defaults to False.

• root_dir (str, optional) – path for input image when using json input. Defaults to “”.

autocare_dlt.tools.hpo.run(exp_name: str, model_cfg: str, data_cfg: str, hpo_cfg: str = None gpus: str = '0', ckpt: ~typing.Union[str, dict] = None, world_size: int = 1, output_dir: str = 'outputs', resume: bool = False, fp16: bool = False, ema: bool = False)→ None

Run Hyperparameter Optimization

Parameters

• exp_name (str) – experiment name. a folder with this name will be created in the output_dir, and the log files will be saved there.

• model_cfg (str) – path for model configuration file

• data_cfg (str) – path for dataset configuration file

• hpo_cfg (str, optional): path for hpo configuration file. Default to None.

• gpus (str, optional) – GPU IDs to use. Default to ‘0’

• ckpt (str, optional) – path for checkpoint file. Defaults to None.

• world_size (int, optional) – world size for ddp. Defaults to 1.

• output_dir (str, optional) – log output directory. Defaults to ‘outputs’.

• resume (bool, optional) – whether to resume the previous training or not. Defaults to False.

• fp16 (bool, optional) – whether to use float point 16 or not. Defaults to False.

• ema (bool, optional) – whether to use EMA(exponential moving average) or not. Defaults to False.

CLI 명령어 예시

Supervised Learning

python autocare_dlt/tools/train.py --exp_name {your_exp} --model_cfg {path}/{model}.json --data_cfg {path}/{data}.json} --ckpt {path}/{ckpt}.pth --gpus {gpu #}

Distributed training (Multi-GPU training)

Multi-GPU 훈련을 진행하기 위해서는 'python'이 아닌 'torchrun'을 이용해야 함

torchrun autocare_dlt/tools/train.py --exp_name {your_exp} --model_cfg {path}/{model}.json --data_cfg {path}/{data}.json} --ckpt {path}/{ckpt}.pth --gpus {gpu #,#,...} --multi_gpu True

[권장] 같은 서버에서 다수의 Multi-GPU 훈련을 하기 위해서는 아래 명령어를 이용해야 함

torchrun --rdzv_backend=c10d --rdzv_endpoint=localhost:0 --nnodes=1 autocare_dlt/tools/train.py --exp_name {your_exp} --model_cfg {path}/{model}.json --data_cfg {path}/{data}.json} --ckpt {path}/{ckpt}.pth --gpus {gpu #,#,...}

• training 결과는 outputs/{your_exp} 위치에 저장됨

run evaluation

python autocare_dlt/tools/eval.py --model_cfg {path}/{model}.json --data_cfg {path}/{data}.json} --ckpt {path}/{ckpt}.pth --gpus 0

export onnx

python autocare_dlt/tools/export_onnx.py --output_name {path}/{model_name}.onnx --model_cfg {path}/{model}.json --batch_size 1 --ckpt {path}/{ckpt}.pth

run inference

• OCR관련
  • Prerequest : 한글 폰트 파일 (ex. NanumPen.ttf)

python tools/inference.py --inputs {path}/{input_dir, img, video, coco json} --model_cfg {path}/{model}.json --output_dir {path}/{output dir name} --ckpt {path}/{model_name}.pth --input_size {width} {height} --gpus {gpu_id} (optional)--root_dir {root path of coco}

run data selection

python tools/data_selection.py --inputs {path}/{input_dir, cocojson} --model_cfg {path}/{model}.json --output_dir {path}/{output dir name} --ckpt {path}/{model_name}.pth --num_outputs {int} --input_size {width} {height} --letter_box {bool} --gpus {gpu_id} (optional)--root_dir {root path of coco}

References

This code is based on and inspired on those repositories (TBD)

• YOLOv5
• Detectron2
• MMCV
• YOLOX