fully reproduce MDV5 in Torchserve environment - Githubissues

tnc-ca-geo / animl-ml

Machine Learning resources for camera trap data processing

Other

4 stars 1 forks source link

fully reproduce MDV5 in Torchserve environment #98

Closed rbavery closed 1 year ago

rbavery commented 1 year ago

This PR

updates the dockerfile to install most dependencies from conda instead of pip to address a jpeg decoding issue
fixes resizing to use the letterbox transform to match the original MDV5 inference steps
switches to depending on yolov5 using torch hub load by moving yolov5 code base into the torch model archive
removes a stale folder for the endpoint work

This is what we are currently using for deployments