Cuda Out of memory error while try train patchcore with gpu

[nghiakthp2401]

I have tried to train patch core with gpu but got these error. I wonder how to edit param like batch size to reduce that? Where should i read about setting config params for training and predict?

This is terminal output when problem occur.
INFO:anomalib.models.image.patchcore.lightning_model:Applying core-set subsampling to get the embedding.
Selecting Coreset Indices.: 100%|███████████████████████████████████████████████████████████████████████████| 22892/22892 [02:28<00:00, 153.75it/s]
Backend qtagg is interactive backend.
Turning interactive mode on. | 0/5 [00:00<?, ?it/s]

This is my code

# Import the datamodule
from anomalib.data import Folder
import torch    
from anomalib.models import Patchcore
from anomalib.engine import Engine


# Create the datamodule
datamodule = Folder(
    name="AnomalibTestData",
    root=r"Dataset",
    normal_dir="good",
    abnormal_dir="defect",
    #task="classification",
)

if __name__ == '__main__':

    # Setup the datamodule
    datamodule.setup()

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Device: {device}")
    # Create the model and engine
    model = Patchcore().to(device)
    engine = Engine()

    # Train a Patchcore model on the given datamodule
    engine.train(datamodule=datamodule, model=model)

    torch.save(model.state_dict(), "PatchCore.pth")

[FedericoDeBona]

To fix the out of memory error you can reduce the batch size like this:

datamodule = Folder(
	name=...,
	root=...,
	normal_dir="good",
	abnormal_dir="defected",
	train_batch_size=32,
	eval_batch_size=32,
)

Also if you want to export models, use

from anomalib.deploy import ExportType

engine.export(model, export_type=ExportType.OPENVINO)
engine.export(model, export_type=ExportType.TORCH)
engine.export(model, export_type=ExportType.ONNX)

[abc-125]

I have tried to train patch core with gpu but got these error. I wonder how to edit param like batch size to reduce that?

Hello, with PatchCore, the required memory depends on the number of training images and their resolution. To avoid this error, you can reduce the number of training images or train on CPU instead.

Where should i read about setting config params for training and predict?

For a quick overview, you can try the documentation. For a better understanding of how PatchCore works, I recommend their paper.