Model Export
SSCMA currently supports the following methods to convert and export models. You can refer to the corresponding tutorials to complete the model export, and then put the exported model into deployment.
TIP
By default, both ONNX and TFLite models are exported. If you only need to export one of them, you can use the --targets parameter to specify the exported model type, e.g. --targets onnx or --targets tflite.
PyTorch to ONNX: Converts PyTorch model and
.pthweights to ONNX model.onnxPyTorch to TFLite: Converts PyTorch model and
.pthweights to TFLite model.tflite
TIP
Before you can start exporting models, you need to complete the Training section and obtain model weights .pth file before start exporting.
Parameter Descriptions
For more parameters for model exporting, you can refer the code below.
python3 tools/export.py --help
# Convert and export PyTorch model to TFLite or ONNX models
# positional arguments:
# config the model config file path
# checkpoint the PyTorch checkpoint file path
# optional arguments:
# -h, --help show this help message and exit
# --targets TARGETS [TARGETS ...]
# the target type of model(s) to export e.g. tflite onnx
# --precisions PRECISIONS [PRECISIONS ...]
# the precisions exported model, e.g. 'int8', 'uint8', 'int16', 'float16' and 'float32'
# --work_dir WORK_DIR, --work-dir WORK_DIR
# the directory to save logs and models
# --output_stem OUTPUT_STEM, --output-stem OUTPUT_STEM
# the stem of output file name (with path)
# --device DEVICE the device used for convert & export
# --input_shape INPUT_SHAPE [INPUT_SHAPE ...], --input-shape INPUT_SHAPE [INPUT_SHAPE ...]
# the shape of input data, e.g. 1 3 224 224
# --input_type {audio,image,sensor}, --input-type {audio,image,sensor}
# the type of input data
# --cfg_options CFG_OPTIONS [CFG_OPTIONS ...], --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
# --simplify SIMPLIFY the level of graph simplification, 0 means disable, max: 5
# --opset_version OPSET_VERSION, --opset-version OPSET_VERSION
# ONNX: operator set version of exported model
# --dynamic_export, --dynamic-export
# ONNX: export with a dynamic input shape
# --algorithm {l2,kl} TFLite: conversion algorithm
# --backend {qnnpack,fbgemm}
# TFLite: converter backend
# --calibration_epochs CALIBRATION_EPOCHS, --calibration-epochs CALIBRATION_EPOCHS
# TFLite: max epoches for quantization calibration
# --mean MEAN [MEAN ...]
# TFLite: mean for model input (quantization), range: [0, 1], applied to all channels, using the average if multiple values are provided
# --mean_and_std MEAN_AND_STD [MEAN_AND_STD ...], --mean-and-std MEAN_AND_STD [MEAN_AND_STD ...]
# TFLite: mean and std for model input(s), default: [((0.0,), (1.0,))], calculated on normalized input(s), applied to all channel(s), using the average if multiple values are provided