using anomalib for training my own dataset

[yousra1672]

Hi everyone, I want to train a model that can classify and detect defects in a product. I’m using Google Colab and created my dataset following the MVTec dataset structure. If anyone has done something similar, could you please show me how to train the model?

[abc-125]

Hello,

You can use Folder datamodule:

from anomalib.data import Folder

custom_datamodule = Folder(
  name="custom_folder",
  root="./datasets/custom",
  normal_dir="good",
  abnormal_dir="defect"
)

After this you can follow the default example for training:

from anomalib.models import Patchcore
from anomalib.engine import Engine

# Initialize components
model = Patchcore()
engine = Engine()

# Train the model
engine.fit(datamodule=custom_datamodule, model=model)

[abc-125]

You're welcome! A validation set is needed to calculate metric thresholds, so you can create it from the test set. You can use the default val_split_mode=ValSplitMode.FROM_TEST and val_split_ratio=0.2 (val_split_ratio shows a fraction of test data to use for validation in this case).

[AltekBilly]

why your Folder() can set "image_size" and "task" parameters?

[yousra1672]

idk, but is it necessary? Because the task will be defined by default as a segmentation task, right?

[samet-akcay]

if you are using v1, yes. With v2 api, you don't need to pass image_size and task parameters.
https://anomalib.readthedocs.io/en/v2.0.0/markdown/guides/reference/data/datamodules/image/folder.html

[yousra1672]

I'm currently using Anomalib V2 , and I would like to better understand the dataset structure, especially because I have many different types of objects that I want to classify as good or not good.
Additionally, I would like to locate the position of the defect or weakness on the objects.
I have already tried using PatchCore and DRAEM, but the results were not sufficient for my use case.
My questions are:
Is it necessary to train a separate model for each object type?
Do all object images need to be captured in the same orientation, position, and lighting conditions?
Is there any detailed documentation on how to prepare datasets correctly for Anomalib?
I’d appreciate any advice, documentation links, or tips you can share. Thank you very much!