OutOfMemoryError while training the cross-encoder

[kunalr97]

train_args = CrossEncoderTrainingArgs(num_train_epochs = 5)

rr = CrossEncoderReranker()
output_dir = f'../outputs/{label2dict[label]}_index/cross_encoder_training/'

rr.fit(
    train_dataset = train,
    val_dataset = val,
    output_dir= output_dir,
    training_args = train_args,
    show_progress_bar = False
)

When i try to train the cross encoder on the BRONCO dataset for prediciting the ICD code for the diagnoses entities. I get this error:

OutOfMemoryError: CUDA out of memory. Tried to allocate 768.00 MiB (GPU 0; 15.77 GiB total capacity; 14.34 GiB already allocated; 379.12 MiB free; 15.03 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

I tried running this line and it does not seem to work. Also there are not any other processes running on the GPU.

import torch
torch.cuda.empty_cache()

Thanks in advance for your help.

[phlobo]

Hello!

The cross-encoder is indeed quite memory intensive (I tested everything with 48GB GPU memory). Two things that might work:

1) I'm not sure if all memory allocated by SapBERT would be cleared by empty_cache(), you might instead want to save the candidate dataset to disk and restart the process / notebook to make sure CUDA memory is entirely freed up.

2) You can reduce the memory footprint of the cross-encoder by reducing the number of candidates subject to re-ranking (which equals the batch size) to something like 16 instead of 64.

[phlobo]

Another thing that might work (though I have not tested the performance), would be to use a smaller BERT model, i.e.,

train_args = CrossEncoderTrainingArgs(model_name="distilbert-base-multilingual-cased")

[kunalr97]

Hi,
Thanks for your quick response. I will try this and hope that it works. Where exactly do i need to do this ?

2. You can reduce the memory footprint of the cross-encoder by reducing the number of candidates subject to re-ranking (which equals the batch size) to something like 16 instead of 64.

Thanks in advance

[phlobo]

There are multiple steps at which you can reduce the number of candidates. However, if you follow this notebook (https://github.com/hpi-dhc/xmen/blob/main/examples/02_BRONCO.ipynb), then setting K_RERANKING = 16 just before calling CrossEncoderReranker.prepare_data should do the trick.

Note: I assume that this will cost you a bit of recall@1, but it might actually increase precision. To get precision, recall and F1 scores at the end, use evaluate instead of evaluate_at_k

[kunalr97]

Thanks a lot! I don't get that error now.

[phlobo]

Thank you for pointing this issue out, I have linked this thread in the README