How to Train and Predict with Supersimplenet without masks

[YunusEmreSayin]

Hi everyone,

I'm trying to train the Supersimplenet model in Anomalib using my own dataset, which contains only image-level labels — that is, I have folders for good and defective images, but no ground-truth masks for pixel-level anomalies.

My goal is to train the model only on the good images, and then run inference on the defective samples — similar to how other models like Padim or Uninet are used in the unsupervised setting.

Here is the exception: /site-packages/anomalib/models/image/supersimplenet/torch_model.py", line 116, in downsample_mask
masks = masks.type(torch.float32)
^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'type'
Epoch 0: 0%| | 0/40 [00:00<?, ?it/s]

Here is the training code I'm using successfully with other models(Padim,UniNet etc):
from anomalib.data import Folder
from anomalib.loggers import AnomalibTensorBoardLogger
from anomalib.models.image import Supersimplenet
from anomalib.engine import Engine

from colorama import Fore, Style, init
from torchvision.transforms import v2

init(autoreset=True)
tensorboard_logger = AnomalibTensorBoardLogger("logs")

augmentations = v2.Compose([
v2.RandomHorizontalFlip(p=0.5),
v2.RandomVerticalFlip(p=0.5),
v2.ColorJitter(brightness=0.1),
v2.GaussianBlur(kernel_size=(3, 3), sigma=(0.1, 0.5)),
v2.RandomApply([v2.RandomErasing(scale=(0.0001, 0.001), ratio=(1.0, 1.0))], p=0.5)
])

dataset_name = "dataset_name"
datamodule = Folder(
name=dataset_name,
root=f"datasets/customs/{dataset_name}",
normal_dir="train/good",
train_augmentations=augmentations,
train_batch_size=4,
eval_batch_size=4
)

datamodule.setup()
print(Fore.LIGHTCYAN_EX + f"dataset_length:{len(datamodule.train_dataloader().dataset)}")

model = Supersimplenet(visualizer=True)

engine = Engine(max_epochs=1, accelerator="gpu", logger=tensorboard_logger, log_every_n_steps=1)
engine.fit(model, datamodule)

engine.export(export_type="torch", export_root=f"exports-featres/SuperSimplenet/{dataset_name}", model=model, model_file_name=dataset_name)

[rajeshgangireddy]

Duplicate of #2868 .
I would like to close this and keep the issue open to track PRs.