Resize bug in inference

[QQR1]

• When input the width and height of the image are not equal, sizes of input arguments that do not match will occur

Line 152 in ./anomalib/deploy/inferencers/torch.py should be anomaly_map = cv2.resize(anomaly_map, (meta_data["image_shape"][1], meta_data["image_shape"][0]))

meta_data["image_shape"] is (image_width, image_height)
-# cv2.resize(img, (image_height，image_width ),interpolation=cv2.INTER_CUBIC)

[samet-akcay]

Thanks @QQR1! Good catch!

[raviv]

@QQR1 Did you manage to train when width and height of the image are not equal?
I'm getting an error in anomaly_module.validation_epoch_end()

[samet-akcay]

Hi @raviv , is it the same error reported by @QQR1 ?

[raviv]

@samet-akcay No, I'm getting an error while training, not inference.
You can easily reproduce by setting image size in config.yaml to something like below and trying to train

  image_size:
    - 512
    - 256

This is the error trace:

  File "/home/raviv/dvlp/anomalib/anomalib/models/components/base/anomaly_module.py", line 151, in _compute_adaptive_threshold
    self.pixel_threshold.compute()
  File "/home/raviv/anaconda3/envs/anomalib_env/lib/python3.8/site-packages/torchmetrics/metric.py", line 380, in wrapped_func
    value = compute(*args, **kwargs)
  File "/home/raviv/dvlp/anomalib/anomalib/utils/metrics/adaptive_threshold.py", line 38, in compute
    precision, recall, thresholds = self.precision_recall_curve.compute()
  File "/home/raviv/anaconda3/envs/anomalib_env/lib/python3.8/site-packages/torchmetrics/metric.py", line 380, in wrapped_func
    value = compute(*args, **kwargs)
  File "/home/raviv/anaconda3/envs/anomalib_env/lib/python3.8/site-packages/torchmetrics/classification/precision_recall_curve.py", line 149, in compute
    return _precision_recall_curve_compute(preds, target, self.num_classes, self.pos_label)
  File "/home/raviv/anaconda3/envs/anomalib_env/lib/python3.8/site-packages/torchmetrics/functional/classification/precision_recall_curve.py", line 259, in _precision_recall_curve_compute
    return _precision_recall_curve_compute_single_class(preds, target, pos_label, sample_weights)
  File "/home/raviv/anaconda3/envs/anomalib_env/lib/python3.8/site-packages/torchmetrics/functional/classification/precision_recall_curve.py", line 139, in _precision_recall_curve_compute_single_class
    fps, tps, thresholds = _binary_clf_curve(
  File "/home/raviv/anaconda3/envs/anomalib_env/lib/python3.8/site-packages/torchmetrics/functional/classification/precision_recall_curve.py", line 39, in _binary_clf_curve
    target = target[desc_score_indices]
IndexError: index 978935 is out of bounds for dimension 0 with size 524288

Process finished with exit code 1

[samet-akcay]

which algorithm is this? Patchcore?

[raviv]

@samet-akcay PADIM

[samet-akcay]

Thanks, we'll reproduce the issue, and give you an update here.

[raviv]

@djdameln Please verify with other algorithms as well. Thanks :)

[djdameln]

@raviv #135 should solve the image_size problems for both training and inference for most models. Patchcore and GANomaly still do not allow changing the image size, but this will be addressed separately, see #136 and #40.

[raviv]

@djdameln got latest dev branch and the fix @QQR1 suggested is not there.
Training works but inference fails on the original bug in this thread (resizing anonmaly map)

[djdameln]

@raviv Sorry, I misinterpreted the original post and thought the problem was for non-square image_size parameter in the config.yaml. I implemented the fix suggested by @QQR1 in a new PR #140.

[samet-akcay]

@raviv PR #140 has now been merged, and the issue should be fixed. Feel free to re-open it if you think there is still an issue