SthPhoenix

They are already supported for more than a year. This isn't a new release, InsightFace team just moved models from their own hosting to GitHub releases.

corresponding models are `scrfd_10g_bnkps` and `w600k_r50`

Looks like everything completed successfully, you just need to open localhost:18081 after some time, needed to download modules and build TRT engines. You can try running `docker attach insightface-rest-gpu0-trt` to...

> Thanks a lot for your answer. When i ran the command you said, i was able to reach error logs. The problem caused by the versions of 'pydantic' library,...

Sorry for late reply, If you are using custom model It was probably exported to ONNX without proper dynamic axes required for batch inference.

Unfortunately there is no simple way, this model still needs face detector to detect actual faces, so it should be implemented as module somewhere between detector and recognition module, also...

TRT has slight accuracy drop, though it's almost negligible in real world applications.

Hi! It's known issue with downloading models from some regions, where google drive might be inaccessible. Downloading models manually should fix the issue.

To select different models from supported list you should just replace following lines in `compose/.env`: ``` DET_NAME=yolov5l-face REC_NAME=w600k_r50 ``` Adding custom detector is a bit tricky since you need to...

Hi! I'm not sure If i'll be able to find raw torch checkpoints, but: 1. This repo provides batch processing using TensorRT, multiple GPU can be utilized by running multiple...