PaddleVideo：onnx模型导出

本文节介绍 PP-TSM 模型如何转化为 ONNX 模型，并基于 ONNX 引擎预测。

1：环境准备

安装 Paddle2ONNX

python -m pip install paddle2onnx

安装 ONNXRuntime

# 建议安装 1.9.0 版本，可根据环境更换版本号
python -m pip install onnxruntime==1.9.0

2：模型转换

paddle2onnx \
--model_dir=output/ppTSMv2_inference \
--model_filename=ppTSMv2.pdmodel \
--params_filename=ppTSMv2.pdiparams \
--save_file=output/ppTSMv2_inference/ppTSMv2.onnx \
--opset_version=10 \
--enable_onnx_checker=True

3：模型预测

onnx 推理

python deploy/paddle2onnx/predict_onnx.py \
--input_file data/example.avi \
--config configs/recognition/pptsm/v2/pptsm_lcnet_ucf101_8frames_uniform.yaml \
--onnx_file=output/ppTSMv2_inference/ppTSMv2.onnx

paddle 推理

python tools/predict.py --input_file data/example.avi \
                           --config configs/recognition/pptsm/v2/pptsm_lcnet_ucf101_8frames_uniform.yaml \
                           --model_file output/ppTSMv2_inference/ppTSMv2.pdmodel \
                           --params_file output/ppTSMv2_inference/ppTSMv2.pdiparams \
                           --use_gpu=True \
                           --use_tensorrt=False

可以看出onnx结果与Paddle 推理结果完全一致。

参考链接

1：PaddleVideo/deploy/paddle2onnx at develop · PaddlePaddle/PaddleVideo · GitHubAwesome video understanding toolkits based on PaddlePaddle. It supports video data annotation tools, lightweight RGB and skeleton based action recognition model, practical applications for video tagging and sport action detection. - PaddleVideo/deploy/paddle2onnx at develop · PaddlePaddle/PaddleVideohttps://github.com/PaddlePaddle/PaddleVideo/tree/develop/deploy/paddle2onnx