Choosing right default batch size for pathcore based on gpu memory configuration

[udasinnayan]

In the pathcore configuration file, how to increase the batch size based on the GPU memory usage. Right now in the dataset configuration file config.yaml there are two options which are train_batch_size: 32 and test_batch_size: 32. In the trainer section to choose batch size limit_train_batches: 1.0, limit_val_batches: 1.0, limit_test_batches: 1.0, and limit_predict_batches: 1.0.

If I want to use custom data format based model patchcore algorithm, by default it can chose batch size 2 for training and testing. But if I would like to increase the testing batch size to 4 what part in the config file needs to be updated to use 100% of test dataset with custom batch size based on the gpu memory usage? I tried to use batch size 4 in config file dataset section test_batch_size: 4 but it always chose 2 as a testing batch size? How do I choose a custom batch size? I am using lightning_inference.py for testing the model.

Do I have to change any script if I wanna use only gpu for testing? I mean the value of precision or something?

TIA

[samet-akcay]

@udasinnayan, apologies for my late reply to this. To change the batch size of validation, test and predict dataloaders, you need to configure eval_batch_size parameter in the config file.
https://github.com/openvinotoolkit/anomalib/blob/cec86bfb1174c7c1cedd39e8b900384e41553b41/src/anomalib/models/patchcore/config.yaml#L8

Regarding the limit_<split>_batches params in Trainer, they are more for debugging purposes, which subsamples the corresponding dataloader. For more information, you could refer to this link:
https://lightning.ai/docs/pytorch/stable/common/trainer.html#limit-train-batches

[samet-akcay]

Let us know if you have any other follow-up questions. Thanks!