Skip to content

Conversation

fpertl
Copy link
Contributor

@fpertl fpertl commented May 11, 2021

Issues:

Feature:

The MPF component Dockerfile was updated, but to portion which instantiates an inference server on the fly and copies over files to run for tests, was never executed. Development with vscode used separate .devcontainer/Dockerfile and .devcontainer/docker-dompose.yml not included on github. These can be provide if that is helpful for the review.


This change is Reviewable

Franz A. Pertl added 5 commits May 5, 2021 19:13
communicationg with triton server via async client
Updated Dockerfile from .devcontainer development Dockerfile
@fpertl fpertl requested a review from jrobble May 11, 2021 18:58
@jrobble jrobble changed the title Feature/triton support parallel Upgrade trtis-detection component to use the NVIDIA Triton Inference Server Sep 7, 2021
@jrobble
Copy link
Member

jrobble commented Mar 8, 2022

From a user of the COCO model:

Hey!!! Heads up, Triton server, object detection, triton will crash if you send it an image that is like 3x > in either dimension than the other, like 300x100 = crash, 100x300 = crash

I'm not sure what the exact threshold is, but I use 3x as the threshold then pad the short side to be 1/2 of the longest. crash threshold might be like 3.5x, I tried 2.9x and it didn't crash.. something that was 3.8x crashed it

442x113 specifically

We should verify this isn't a problem, and if it is, consider padding the input image.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

Development

Successfully merging this pull request may close these issues.

2 participants