Question: what is MODEL.WEIGHTS in d2go.exporter options?

[soid]

I'm running d2go.exporter as specified in the docs, and I'm getting the error that MODEL.WEIGHTS is not recognized as an export method:

KeyError: "No object named 'MODEL.WEIGHTS' found in 'ModelExportMethod' registry! Available names: ['torchscript_mobile_int8@tracing', 'torchscript_mobile-vulkan@tracing', 'torchscript_mobile-metal@tracing', 'torchscript_mobile@tracing', 'torchscript_int8@tracing', 'torchscript@tracing', 'torchscript_mobile_int8@scripting', 'torchscript_mobile-vulkan@scripting', 'torchscript_mobile-metal@scripting', 'torchscript_mobile@scripting', 'torchscript_int8@scripting', 'torchscript@scripting', 'torchscript_mobile_int8', 'torchscript_mobile-vulkan', 'torchscript_mobile-metal', 'torchscript_mobile', 'torchscript_int8', 'torchscript']"

The way I run the script:

d2go.exporter --config-file configs/faster_rcnn_fbnetv3a_C4.yaml \
--output-dir ./ --predictor-type torchscript_int8 \
MODEL.WEIGHTS https://mobile-cv.s3-us-west-2.amazonaws.com/d2go/models/246823121/model_0479999.pth

I was looking into the exporter code and can't figure out what's the meaning of this MODEL.WEIGHTS https://...pth options? It seems is part of --predictor-type parameter, which is parsed one by one as exporting types. But doesn't make much sense. What does it actually do?

[soid]

Ok, I figured that --output-dir should follow --predictor-type, not the other way, as in the example with torchscript. I created a PR to correct the doc.
I'm still curious of the meaning of MODEL.WEIGHTS http:/... argument.

[wat3rBro]

@soid The weird problem is because a conflict of nargs when using argparse: in d2go.exporter the --predictor-type is set with nargs="+" (https://github.com/facebookresearch/d2go/blob/main/tools/exporter.py#L123) while there's another opts argument is set with nargs=argparse.REMAINDER (https://github.com/facebookresearch/d2go/blob/main/d2go/setup.py#L111). Therefore --predictor-type can't be used as the last argument, for example putting --output-dir behind it would work. The opt argument is for overwriting the config file, MODEL.WEIGHTS http:/... overwrites the model weights in the config file configs/faster_rcnn_fbnetv3a_C4.yaml with a trained one.