❓ <title>Guidance on Multi-Category Training for Dinomaly Model

[luminous33]

Your Question

Hello Anomalib Team,

I am trying to train the Dinomaly model on multiple categories of the Visa dataset simultaneously (e.g., pcb1, pcb2, pcb3, and pcb4) in a single training run.

Based on the discussion in issue #2782, it seems that training on multiple categories is a supported feature. However, I have encountered errors with both the CLI (config.yaml) and a direct Python script approach. I would appreciate some guidance on the correct procedure.

Here is what I have tried so far:

1. CLI Attempt with config.yaml
   I tried to run the training using a config.yaml file that specifies a list of categories.

Command:

Bash

anomalib train --config sample.yaml
sample.yaml content:

YAML

data:
class_path: anomalib.data.Visa
init_args:
root: /data/data/visa/visa_dataset/
category:
- pcb1
- pcb2
- pcb3
- pcb4

model:
class_path: anomalib.models.Dinomaly
Error:
This approach fails with a parser error, indicating that the category key expects a string, not a list.

error: Parser key "data":
Does not validate against any of the Union subtypes
Subtypes: [<class 'NoneType'>, <class 'anomalib.data.datamodules.base.image.AnomalibDataModule'>]
Errors:
- Expected a <class 'NoneType'>
- Problem with given class_path 'anomalib.data.Visa':
Parser key "category":
Expected a <class 'str'>. Got value: ['pcb1', 'pcb2', 'pcb3', 'pcb4']
Given value type: <class 'dict'>
Given value: {'class_path': 'anomalib.data.Visa', 'init_args': {'root': '/data/data/visa/visa_pytorch/', 'category': ['pcb1', 'pcb2', 'pcb3', 'pcb4']}}
2. Python Script Attempt
Next, I tried using a Python script to instantiate the VisaDataModule with a list of categories.

train.py script:

Python

from anomalib.data import VisaDataModule
from anomalib.models import Dinomaly
from anomalib.engine import Engine

if name == "main":
datamodule = VisaDataModule(
root="/data/data/visa/visa_dataset/",
category=["pcb1", "pcb2", "pcb3", "pcb4"],
image_size=(256, 256),
train_batch_size=32
)

model = Dinomaly()
engine = Engine()

engine.fit(model=model, datamodule=datamodule)

Error:
This script fails with a TypeError when the Engine tries to create a results directory, as it cannot append a list to a Path object.

Traceback (most recent call last):
File "/data/model/anomalib/dino_train_gemini.py", line 38, in
engine.fit(model=model, datamodule=datamodule)
File "/data/model/anomalib/src/anomalib/engine/engine.py", line 404, in fit
self._setup_workspace(
File "/data/model/anomalib/src/anomalib/engine/engine.py", line 297, in _setup_workspace
root_dir = Path(self._cache.args["default_root_dir"]) / model.name / dataset_name / category

---

TypeError: unsupported operand type(s) for /: 'PosixPath' and 'list'
My Question
Both errors suggest that the category parameter is designed to accept a single string. Could you please clarify the correct way to configure anomalib to train a single model on multiple categories at once, as suggested in issue #2782?

Is there a different configuration key, a specific data module (e.g., a multi-class variant), or a different approach I should be using for this purpose? Or have I misunderstood the scope of the multi-category feature?

Thank you for your time and help!

Environment
Anomalib Version: 2.1.0
torch Version:
torch 2.5.1+cu124
torchaudio 2.5.1+cu124
torchmetrics 1.8.1
torchvision 0.20.1+cu124

Python Version: 3.11

OS: Ubuntu 22.04

Forum Check

• I have searched the discussions forum for an answer to my question.

[github-actions[bot]]

Thanks for submitting this issue! It has been added to our triage queue. A maintainer will review it shortly.

[rajeshgangireddy]

Hi @luminous33

Currently Dinomaly model in Anomalib supports only single-class anomaly detection. Adding multi-class support is planned but no one is actively working on it at the moment.

[waschsalz]

Could we use a non-optimized approach for multi-class training? It is certainly not possible to create multiple thresholds automatically for now but adding multiple classes into one directory and then training Dinomaly or UniNet on it should work or am I missing anything? After that we have different test directories that we can use to calculate thresholds on our own (for those use cases in which we know before inference what class / category we're about to see).