GPU not working for Detectron2 - Needs changing 'cpu' to 'cuda' inside config.json

[AliButtarPG]

GPU wasn't working for predictions made using SAHI even though I was using device as cuda. I thought this setting would override the setting inside the configuration file regardless of whatever it was but once I changed the device to cuda inside the config.json then it worked like a charm.

Something you could fix in the next iteration maybe. Thanks for the amazing tool.

[fcakyon]

Hello @AliButtarPG, thanks a lot for the bug report!

Can you be more specific on the issue?

Did I understand it correctly:

1. you did not set config_path param
2. you set device param as cuda
3. then SAHI performs Detectron2 prediction on CPU?

[AliButtarPG]

Hi @fcakyon,

Sorry for not providing details, wasn't sure what more to write.

1. I did set config_path param and just gave path to the config of my custom model
2. Yes I set the device param as cuda like this:

detection_model = Detectron2DetectionModel(
    model_path=model_path,
    config_path=config_path,
    confidence_threshold=0.1,
    device="cuda",  # or 'cuda:0'
)

3. Yes, until I changed the device param to cuda directly inside the config.json

[fcakyon]

@AliButtarPG we implemented SAHI Detectron2 support as: params in config file overwrites the params provided in Detectron2DetectionModel. This may be a bad practice, thanks for pointing that out! It would be better if the params in Detectron2DetectionModel would have higher priority 👍

[fcakyon]

This PR should fix this issue: https://github.com/obss/sahi/issues/478