Model Training

SSCMA offers a variety of algorithms, and you can choose the appropriate one based on your needs. Then, you can solve practical problems by training, exporting, and deploying models. This chapter will further introduce how to use SSCMA to train, export, and deploy models.

Training Parameter Description

You need to replace the following parts of the parameters according to the actual situation. The specific descriptions of each parameter are as follows:

python3 tools/train.py --help

usage: train.py [-h] [--amp] [--auto-scale-lr] [--resume] [--work_dir WORK_DIR] [--cfg-options CFG_OPTIONS [CFG_OPTIONS ...]]
                [--launcher {none,pytorch,slurm,mpi}] [--local_rank LOCAL_RANK]
                config

Train a detector

positional arguments:
  config                train config file path

options:
  -h, --help            show this help message and exit
  --amp                 enable automatic-mixed-precision training
  --auto-scale-lr       enable automatically scaling LR.
  --resume              resume from the latest checkpoint in the work_dir automatically
  --work_dir WORK_DIR, --work-dir WORK_DIR
                        the dir to save logs and models
  --cfg-options CFG_OPTIONS [CFG_OPTIONS ...]
                        override some settings in the used config, the key-value pair in xxx=yyy format will be merged into
                        config file. If the value to be overwritten is a list, it should be like key="[a,b]" or key=a,b It also
                        allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" Note that the quotation marks are necessary and
                        that no white space is allowed.
  --launcher {none,pytorch,slurm,mpi}
                        job launcher
  --local_rank LOCAL_RANK, --local-rank LOCAL_RANK

Export Parameter Description

You need to replace the following parts of the parameters according to the actual situation. The specific descriptions of each parameter are as follows:

python3 tools/export.py --help

usage: export.py [-h] [--work-dir WORK_DIR] [--out OUT] [--device DEVICE] [--img-size IMG_SIZE [IMG_SIZE ...]] [--simplify] [--opset OPSET]
                 [--image_path IMAGE_PATH] [--format [{onnx,tflite,vela,savemodel,torchscript,hailo} ...]]
                 [--arch {hailo8,hailo8l,hailo15,hailo15l}] [--verify] [--cfg-options CFG_OPTIONS [CFG_OPTIONS ...]]
                 [--launcher {none,pytorch,slurm,mpi}] [--local_rank LOCAL_RANK]
                 config checkpoint

test (and eval) a model

positional arguments:
  config                test config file path
  checkpoint            checkpoint file

options:
  -h, --help            show this help message and exit
  --work-dir WORK_DIR   the directory to save the file containing evaluation metrics
  --out OUT             dump predictions to a pickle file for offline evaluation
  --device DEVICE       the device to execute the model, options are 'cpu', 'cuda', etc.
  --img-size IMG_SIZE [IMG_SIZE ...], --img_size IMG_SIZE [IMG_SIZE ...], --imgsz IMG_SIZE [IMG_SIZE ...]
                        Image size of height and width
  --simplify            Simplify onnx model by onnx-sim
  --opset OPSET         ONNX opset version
  --image_path IMAGE_PATH
                        Used to export verification data of tflite
  --format [{onnx,tflite,vela,savemodel,torchscript,hailo} ...]
                        Model format to be exported
  --arch {hailo8,hailo8l,hailo15,hailo15l}
                        hailo hardware type
  --verify              Verify whether the exported tflite results are aligned with the tflitemicro results
  --cfg-options CFG_OPTIONS [CFG_OPTIONS ...]
                        override some settings in the used config, the key-value pair in xxx=yyy format will be merged into config file. If
                        the value to be overwritten is a list, it should be like key="[a,b]" or key=a,b It also allows nested list/tuple
                        values, e.g. key="[(a,b),(c,d)]" Note that the quotation marks are necessary and that no white space is allowed.
  --launcher {none,pytorch,slurm,mpi}
                        job launcher
  --local_rank LOCAL_RANK

These commands and their options provide a detailed guide on how to train and export models using SSCMA. By customizing these parameters according to your specific requirements, you can effectively utilize SSCMA for your machine learning tasks.

QAT Parameter Description

You need to replace the following parts of the parameters according to the actual situation. The specific descriptions of each parameter are as follows:

python3 tools/quantization.py --help

usage: quantization.py [-h] [--work-dir WORK_DIR] [--test] [--out OUT] [--cfg-options CFG_OPTIONS [CFG_OPTIONS ...]] [--launcher {none,pytorch,slurm,mpi}]
                       [--local_rank LOCAL_RANK]
                       config model

quantizer a model

positional arguments:
  config                quantization config file path
  model                 checkpoint file

options:
  -h, --help            show this help message and exit
  --work-dir WORK_DIR   the directory to save the file containing evaluation metrics
  --test                Whether to evaluate inference results
  --out OUT             dump predictions to a pickle file for offline evaluation
  --cfg-options CFG_OPTIONS [CFG_OPTIONS ...]
                        override some settings in the used config, the key-value pair in xxx=yyy format will be merged into config file. If the value to be overwritten is a
                        list, it should be like key="[a,b]" or key=a,b It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" Note that the quotation marks are
                        necessary and that no white space is allowed.
  --launcher {none,pytorch,slurm,mpi}
                        job launcher
  --local_rank LOCAL_RANK, --local-rank LOCAL_RANK