Modify 'run_glue.py' to train model from scratch and eval + danqi mv setup #1
Conversation
Danqi7
commented
Mar 15, 2021
Danqi7
commented
- Modified 'text-classification/run_glue.py' to take in argument '--from_scratch' to enable train random model from scratch and evaluate on selected task
- Danqi MV setup script
- script to train random models from scratch and eval on selected GLUE task
HF/examples/text-classification/run_glue.py to be able to train/eval on glue using random models (do not load pre-trained weights)
| export RUN_NO=_random | ||
| export OUTPUT_DIR=../../../bucket/model_outputs/glue/$TASK_NAME$RUN_NO/ | ||
|
|
||
| # Random: --from_scratch + --config_name + --tokenizer_name |
There was a problem hiding this comment.
Could we use a new file for scripts which are different from before? This way we can just keep adding scripts.
There was a problem hiding this comment.
I think "eval_random_glue.sh" is a new added script/file?
| @@ -114,7 +114,12 @@ class ModelArguments: | |||
| Arguments pertaining to which model/config/tokenizer we are going to fine-tune from. | |||
| """ | |||
|
|
|||
| from_scratch: bool = field( | |||
There was a problem hiding this comment.
Since this change is small I think it can stay in run_glue.py. But in the future I think we can just make a new copy of run_glue.py and run_mlm.py so that we don't touch the original files.
There was a problem hiding this comment.
that's a good suggestion! I will make a new copy so we don't ever touch the original files.
|
Reminder to add changes to |
|
@ameet-1997 Updated "CHANGELOG.md" and made a copy of the "run_glue.py" to avoid changing the original script. |
