tyomj / product_detection

Working with scale: 2nd place solution to Product Detection in Densely Packed Scenes
Apache License 2.0
48 stars 9 forks source link

Model weights #7

Open mrgiba opened 1 year ago

mrgiba commented 1 year ago

Hello,

Are the model weights available somewhere ?

Thank you

tyomj commented 1 year ago

Hi,

Yes, they are, please take a look at the comment here.

mrgiba commented 1 year ago

Awesome

And is there any caveat for doing transfer learning using that pre trained model as a base ?

Thanks

tyomj commented 1 year ago

There are built-in features in the mmdetection framework to use a pre-trained checkpoint, such as load_from or resume_from. You may find more info in the original documentation.

mrgiba commented 1 year ago

Is it needed to follow the mmdet v1 installation steps at https://github.com/open-mmlab/mmdetection/blob/v1.0.0/docs/INSTALL.md or should mmdet be installed differently ?

The conda install step there would install the latest pytorch version, but by the number of errors and deprecation messages I see here when running pip install -v -e ., I wonder if the installation steps should restrain pytorch to a specific version instead

mrgiba commented 1 year ago

Yes, by restraining to PyTorch 1.4 backwards and also restricting the version of a couple of dependencies, I was able to setup, do transfer learning and run inference.

I can contribute with the exact installation steps I did