to_onnx.md

Convert to ONNX model

After training in AI Studio or local, we can get a parameter file of dynamic graph from the saved folder, which is *.pdparams. But if it needs to be used in buildseg, we need to convert it into a ONNX model.

Preparation

Prepare the dynamic graph model (*.Pdparams) and install the third-party dependency library:

cd scripts
pip install -r requirements.txt

Convert

Executing the following command:

python to_onnx.py \
       --model_name model \
       --params_path model.pdparams

Parameter description

Parameter's name	Purpose	Optional	Default	Values
model_name	Name of model	No	-	ocrnet / segformer / bisenet
params_path	Path of dynamic graph model's parameter file	No	-	-
save_path	Path of folder where the static graph model is saved	Yes	../onnx_weight	-

Check model

Executing the following command:

python check_onnx.py

Parameter description

Parameter's name	Purpose	Optional	Default
save_path	Path of folder where the static graph model is saved	Yes	../onnx_weight