What pretrained model is used to generate the global feature?

luoyunan / ECNet

An evolutionary context-integrated deep learning framework for protein engineering

BSD 3-Clause "New" or "Revised" License

60 stars 16 forks source link

What pretrained model is used to generate the global feature? #7

Closed squareRoot3 closed 2 years ago

squareRoot3 commented 2 years ago

ECNet uses Type-Protein to encode global features but does not specify any pretrained weights to initialize the Transformer (`from_pretrained=None'). If so, the global features seem to be generated by a randomly initialized Transformer encoder.

luoyunan commented 2 years ago

Thanks for reporting the issue. TAPE now provides auto-downloading option which does not require specifying the path to the pretrained weights. We did not set the argument correctly when release the GitHub code. This issue has been fixed in https://github.com/luoyunan/ECNet/commit/a8c3cc754306578761c9ef030c9c5eef0cb29496. You don't need to pre-download the TAPE pretrained model before running ECNet -- it will be automatically downloaded to ~/.cache/torch/protein_models if it is not there.

squareRoot3 commented 2 years ago

I see. Thank you for your quick response!