search

georgian-io / Multimodal-Toolkit

Multimodal model for text and tabular data with HuggingFace transformers as building block for text data
https://multimodal-toolkit.readthedocs.io
Apache License 2.0
591 stars 85 forks source link

Preparing data for inferencing #56

Closed insdrs480 closed 10 months ago

insdrs480 commented 1 year ago

Hi there,

I have been using the load_data function to create my datasets for HuggingFace's Trainer class, which has worked really well so far. However, I am not sure how I should approach inferencing without a label column which load_data requires. Would really appreciate some advice as I am rather new to this. Thank you!

akashsaravanan-georgian commented 1 year ago

Hi @insdrs480, for creating datasets we expect labeled data since it follows a train-eval-test loop. If you don't have labels, you can assign dummy labels to that particular column and discard it later. Please let me know if you need any other help!

insdrs480 commented 1 year ago

Thank you for the response! Are we able to specify the model's behaviour when it encounters unknown categorical data?

I'm getting the following error, which I believe is due to my test dataset containing values that are different from what I trained the model on: RuntimeError: mat1 and mat2 shapes cannot be multiplied (256x22 and 28x768)

akashsaravanan-georgian commented 11 months ago

Not at the moment unfortunately. This is something we'd like to add in though. Do you have any suggestions as to what you would like to see?

DougTrajano commented 10 months ago

Hi folks, I had the same issue and I suggested a solution in #61 to fix it in the load_data() function.

akashsaravanan-georgian commented 10 months ago

Thanks Doug!

I'm getting the following error, which I believe is due to my test dataset containing values that are different from what I trained the model on: RuntimeError: mat1 and mat2 shapes cannot be multiplied (256x22 and 28x768)

@insdrs480 Apologies I missed this bit! The model will expect the same set of columns that were used for training. You might be getting this error if they don't match. I would advice either setting up default values for missing columns or removing them from the training process entirely.

Closing this issue as Doug's fix solves this. Feel free to re-open/create a new one if you run into any issues!