AttributeError: 'OurTrainingArguments' object has no attribute 'deepspeed_plugin'

[levente-murgas]

Describe the bug When trying to run main.py with my own config JSON file, as it is advised in the README, the execution is aborted with an error AttributeError: 'OurTrainingArguments' object has no attribute 'deepspeed_plugin'.

To Reproduce Steps to reproduce the behavior:

1. Create new virtualenv (thus we get a clean slate).
2. Clone repository containing main.py from GitHub.
3. Install the module with pip: pip install multimodal-transformers
4. Pass an appropriate JSON file to the scrip and run main.py: python /content/Multimodal-Toolkit/main.py /content/data/config.json
5. See error

Expected behavior The selected model should be trained with the configuration given in the JSON.

Desktop:

• OS: Windows 11
• Python: 3.10.12
  • Environment: Colab

[akashsaravanan-georgian]

Hi @levente-murgas, were you able to run the code using the sample configs that are present?

[levente-murgas]

Hi @akashsaravanan-georgian

I tried running the sample configs with $python ./Multimodal-Toolkit/main.py ./Multimodal-Toolkit/datasets/Melbourne_Airbnb_Open_Data/train_config.json, but it fails with the same error sequence as it did with my own config: first it complains that accelerate library is not installed: File \venv\lib\site-packages\transformers\trainer.py", line 3801, in create_accelerator_and_postprocess if version.parse(accelerate_version) > version.parse("0.20.3"): NameError: name 'accelerate_version' is not defined

Then I install accelerate with pip install accelerate.

Then I run it again and get the deepspeed_plugin error. Using the sample configs still would not change the fact that OurTrainingArguments class has no deepspeed_plugin attribute.

[akashsaravanan-georgian]

Hi @levente-murgas, I think I've figured out the issue. The latest version of transformers seems to have broken something in the library. Could you downgrade transformers to pip install transformers==4.26.1 and see if the same error occurs?

[levente-murgas]

Hi @akashsaravanan-georgian,

That's what I was thinking, too. Downgrading to 4.26.1 indeed solved the issue. I have a different question regarding the implementation of the MultimodalDataTrainingArguments class. In the implementation the num_classes attribute is used to tell the number of classes for the classification. If I use the default value (-1), will the program infer the number of classes for itself (e.g: from the label_col in column_info or from somewhere else) or do I have to explicitly provide the number of classes in num_classes for the program in order to work as expected?

[akashsaravanan-georgian]

I believe that's the case. If you pass in -1, it uses the number of unique labels in your training data.