How two checkpoints are selected in TransCoder?

facebookresearch / CodeGen

Reference implementation of code generation projects from Facebook AI Research. General toolkit to apply machine learning to code, from dataset creation to model training and evaluation. Comes with pretrained models.

MIT License

710 stars 144 forks source link

How two checkpoints are selected in TransCoder? #31

Closed wasiahmad closed 3 years ago

wasiahmad commented 3 years ago

For the TransCoder model, two separate checkpoints are provided for the following language directions.

TransCoder_model_1 for C++ -> Java, Java -> C++ and Java -> Python
TransCoder_model_2 for C++ -> Python, Python -> C++ and Python -> Java

Did you train TransCoder individually for the above two different directions setting? Or, the checkpoint is selected based on validation performances for those directions.

baptisteroziere commented 3 years ago

Hi. The checkpoints were selected based on the validation performances on a single run. We tried to select two checkpoints with close to the best validation performance for several pairs to avoid having to one model per direction.