Problem fitting a model for QA with "deepset/bert-base-cased-squad2" and FARMReader

[Al3xisco]

My problem is that the model takes a long time to fit (more than 8 hours), perform QA with the "annotation tool".
I use a GPU in Colab, but it takes a long time and I have thought about the possibility that it is not possible.
I follow this guide: https://haystack.deepset.ai/tutorials/fine-tuning-a-model
I've searched for any related topics but haven't found any.

Can it be done with that base model?

[bglearning]

Hi @Al3xisco ,

How many epochs are you training for? Also, what is the GPU-usage like?
And to confirm, the 8 hours is on your training data (if so, how many QA pairs are there?) or the example in the tutorial?

[Al3xisco]

Hi @bglearning ,
I am training for 1 epoch, I have done the Q&A pairs with the annotation tool and got about 360 pairs. I am using a Testa T4 GPU with 15 of RAM, the result of the colab training is as shown in the image (it doesn't finish, nor show me any progress after 8 hours).
I attach the code, it is the same of the tutorial.

reader = FARMReader(model_name_or_path="deepset/bert-base-cased-squad2", use_gpu=True) data_dir = "/content/drive/MyDrive/Colab Notebooks/Ajuste_Pytorch/" reader.train(data_dir=data_dir, train_filename="/content/drive/MyDrive/Colab Notebooks/Ajuste_Pytorch/answers.json", use_gpu=True, n_epochs=1, `save_dir="/content/drive/MyDrive/BASE_CONOCIMIENTO/my_model")

[Al3xisco]

@bglearning, thank you very much, that parameter allowed to continue with the training and it was done very fast.
I used the latest version.
I leave the code that helped me in case it can be of help to someone else.

reader.train(data_dir=data_dir, train_filename="/content/drive/MyDrive/Colab Notebooks/Ajuste_Pytorch/answers.json", use_gpu=True, n_epochs=1, num_processes=1, save_dir="/content/drive/MyDrive/my_model")

[Al3xisco]

@bglearning
I have one more doubt, I need to test the model and I wanted to upload it to huggingface but it gave me an error at the time of implanting it, is it necessary to follow some procedure or something like that?

[bglearning]

Hi,
What error are you receiving?

[Al3xisco]

Hi, when I try to load it for use with QuestionAnsweringPipeline, it returns the following error: "Haystack_01 does not appear to have a file named config.json".

[Al3xisco]

Hola, cuando intento cargarlo para usarlo con QuestionAnsweringPipeline, devuelve el siguiente error: "Haystack_01 no parece tener un archivo llamado config.json".
I found the solution in a past post,
https://github.com/deepset-ai/haystack/discussions/2176#discussioncomment-2172195
If it helps anyone.