[Bug]: AttributeError: 'NoneType' object has no attribute 'flatten'

[Septembit]

Describe the bug

when I tried to train a fastflow model by config file, I got this error.
the version is anomalib==2.0.0b2

Dataset

Folder

Model

FastFlow

Steps to reproduce the behavior

anomalib fit --config ./config.yaml

OS information

OS information:

• OS: [e.g. Ubuntu 20.04]
• Python version: [e.g. 3.10.0]
• Anomalib version: [e.g. 0.3.6]
• PyTorch version: [e.g. 1.9.0]
• CUDA/cuDNN version: [e.g. 11.1]
• GPU models and configuration: [e.g. 2x GeForce RTX 3090]
• Any other relevant information: [e.g. I'm using a custom dataset]

Expected behavior

NG

Screenshots

`

Pip/GitHub

pip

What version/branch did you use?

2.0.0b2

Configuration YAML

# anomalib==2.0.0b2
seed_everything: true

model:
  class_path: anomalib.models.Fastflow
  init_args:
    backbone: resnet18
    pre_trained: true
    flow_steps: 8
    conv3x3_only: false
    hidden_ratio: 1.0

trainer:
  max_epochs: 500
  callbacks:
    - class_path: lightning.pytorch.callbacks.EarlyStopping
      init_args:
        patience: 3
        monitor: pixel_AUROC
        mode: max
data:
  class_path: anomalib.data.Folder
  init_args:
    name: bottle_0528
    normal_dir: ok
    root: data/bottle-bottom-dataset
    abnormal_dir: ng
    normal_test_dir: null
    mask_dir: null
    normal_split_ratio: 0.2
    extensions: null
    train_batch_size: 1
    eval_batch_size: 1
    num_workers: 0
    train_augmentations: null
    val_augmentations: null
    test_augmentations: null
    augmentations: null
    test_split_mode: from_dir
    test_split_ratio: 0.2
    val_split_mode: same_as_test
    val_split_ratio: 0.2
    seed: null
logging:
  log_graph: false
default_root_dir: results
ckpt_path: null

Logs

NG

Code of Conduct

• I agree to follow this project's Code of Conduct

[alexriedel1]

If you don't provide a mask dir mask_dir: null of ground truth masks of anomalous samples, the trainer assumes you want to perform a classifaction task instead of an segmentation task. In an classification task, you cannot calculate pixel metrics (such as pixel_AUROC)

[Septembit]

@alexriedel1
Thanks for your reply, I tried to change the pixel_AUROC to image_AUROC in trainer filed. But I still meet this error.

# anomalib==2.0.0b2
seed_everything: true

model:
  class_path: anomalib.models.Fastflow
  init_args:
    backbone: resnet18
    pre_trained: true
    flow_steps: 8
    conv3x3_only: false
    hidden_ratio: 1.0

trainer:
  max_epochs: 500
  callbacks:
    - class_path: lightning.pytorch.callbacks.EarlyStopping
      init_args:
        patience: 3
        monitor: image_AUROC
        mode: max
data:
  class_path: anomalib.data.Folder
  init_args:
    name: bottle_0528
    normal_dir: ok
    root: data/bottle-bottom-dataset
    abnormal_dir: ng
    normal_test_dir: null
    mask_dir: null
    normal_split_ratio: 0.2
    extensions: null
    train_batch_size: 64
    eval_batch_size: 64
    num_workers: 0
    train_augmentations: null
    val_augmentations: null
    test_augmentations: null
    augmentations: null
    test_split_mode: from_dir
    test_split_ratio: 0.2
    val_split_mode: same_as_test
    val_split_ratio: 0.2
    seed: null
logging:
  log_graph: false
default_root_dir: results
ckpt_path: null

[alexriedel1]

Ah I see!

The pixel threshold is calculated based on ground truth masks, even if it is a classification task and there is no mask!

anomalib/src/anomalib/post_processing/one_class.py

Lines 107 to 108 in 8d14ae1

	if outputs.anomaly_map is not None:
	self._pixel_threshold.update(outputs.anomaly_map, outputs.gt_mask)

To fix this, I created this PR @samet-akcay #2513

The other problem you will face after this fix is, that the default evaluator will evaluate pixel and image metrics.

anomalib/src/anomalib/models/components/base/anomalib_module.py

Lines 375 to 380 in 8d14ae1

	image_auroc = AUROC(fields=["pred_score", "gt_label"], prefix="image_")
	image_f1score = F1Score(fields=["pred_label", "gt_label"], prefix="image_")
	pixel_auroc = AUROC(fields=["anomaly_map", "gt_mask"], prefix="pixel_")
	pixel_f1score = F1Score(fields=["pred_mask", "gt_mask"], prefix="pixel_")
	test_metrics = [image_auroc, image_f1score, pixel_auroc, pixel_f1score]
	return Evaluator(test_metrics=test_metrics)

So if you have a classification task, you need to configure a custom evaluator. I'm not very familiar with the current config file design, but with the API, you would do it like this:

datamodule = Folder(
        name="bottle_0528",
        root = "data/bottle-bottom-dataset",
        normal_dir="ok",
        abnormal_dir="ng",
    )
    evaluator = Evaluator(
        test_metrics=[
            F1Score(fields=["pred_label", "gt_label"], prefix="image_"),
        ]
    )
    model = Patchcore(evaluator=evaluator)
    engine = Engine()
    engine.fit(datamodule=datamodule, model=model)