SadTalkerTriton
SadTalkerTriton copied to clipboard
SadTalker Triton server demo
Env
- 服务镜像 >= tritonserver_22.11 差的包,参考*_reference.txt
- 客户端(或tritonclient 镜像) 差的包,参考*_reference.txt
Using
百度网盘链接 提取码: gogs
drive.google
g.onnx
-
download weights.zip and unzip weights.zip in onnx_weights and add generator.onnx and kp_detector.onnx
-
cd SadTalkerTriton
-
docker run -v $PWD:/models --gpus device=1 --shm-size 1g -it --name sadtalker_triton tritonserver_22.11_pypack(your image name)
- in server container:
tritonserve --model-repo ./
- in client container:
python client.py
I did
- Export all submodel to onnx and runtime with onnx_privider_trtfp16
- Removed the facexlib and replace the face detection module for simple and fast
- Clean the repeate functions in the pipeline
- Extract useful model weights and structure codes
- Removed unnecessary features
- Optimized Generator inference time during face render (time(onnx_privider_trtfp16)/2) and mutilprocess Poisson fusion.
You can do
- Dynamic batching
- More efficient organization compliant with triton
- int8 for generator network
Reminder
- generator.onnx and kp_detector.onnx you must do it by yourself,it's big and some licence.