Example from quick start fails with 'TrainingArguments' object has no attribute 'eval_strategy'

[salvador-dali]

Tried to follow https://huggingface.co/docs/setfit/main/en/quickstart#end-to-end

This notebook is just a copypaste from quickstart: https://colab.research.google.com/drive/1N9HPmqhaqBm_rL13nTuV3wC7rdIJseve#scrollTo=0NybX59O5o98

The train fails with

/usr/local/lib/python3.10/dist-packages/transformers/utils/notebook.py in on_train_begin(self, args, state, control, **kwargs)
    293 
    294     def on_train_begin(self, args, state, control, **kwargs):
--> 295         self.first_column = "Epoch" if args.eval_strategy == IntervalStrategy.EPOCH else "Step"
    296         self.training_loss = 0
    297         self.last_log = 0

AttributeError: 'TrainingArguments' object has no attribute 'eval_strategy'

[tomaarsen]

The most recent version of transformers has renamed evaluation_strategy into eval_strategy. I'd recommend downgrading transformers to a slightly older version. You can see your current version with pip show transformers.

• Tom Aarsen

[salvador-dali]

Thanks, I was able to fix this with !pip install transformers==4.39.0

[tomaarsen]

Excellent, thanks for sharing! I intend to fix this more properly in early June.

• Tom Aarsen

[giorgionicoli]

It seems to work up to transformers==4.40.2, but it fails with transformers>=4.41.0, in case useful :)

[splevine]

I was facing this issue yesterday, I will try that particular downgrade to transformers this week. Thank you!

[XeniaZerweck]

Hello, I am facing the same issue and downgrading to transformers==4.39.0 did not fix the problem. Neither did the workaround args.eval_strategy = args.evaluation_strategy. Are there any updates or any other insights? Thanks in advance!

[splevine]

The transformers downgrade worked for me

If in notebook you can run:

pip install setfit transformers==4.39.0

[XeniaZerweck]

Yes it worked for me too now, I forgot to restart my kernel. Apologies and thanks!