fairseq icon indicating copy to clipboard operation
fairseq copied to clipboard

Empty 'args' value in Neural Language Modeling "Training a transformer language model with the CLI tools" example model

Open lancioni opened this issue 11 months ago • 2 comments

🐛 Bug

The model trained (in Colab) according to instructions in Neural Language Modeling "Training a transformer language model with the CLI tools" example model has an empty 'args' value resulting .pt model. This contrasts with pretrained models downloable in the same page and creates problems with conversion by CTranslate2, since the converter expects keys in 'args'.

To Reproduce

Follow closely instructions in https://github.com/facebookresearch/fairseq/blob/main/examples/language_model/README.md

Expected behavior

The resulting model should have filled 'args' fields.

Environment

  • fairseq Version (installed through pip):
  • How you installed fairseq (`pip):
  • Free Google colab environment

lancioni avatar Mar 26 '24 08:03 lancioni

@lancioni This issue is only with the CLI version right? Does the Python code version work fine?

SarthakNikhal avatar Mar 29 '24 23:03 SarthakNikhal

I didn't try actually. But I understand the CLI version is just calling the Python code, isn't it?

lancioni avatar Apr 09 '24 10:04 lancioni