ValueError: Input contains infinity or a value too large for dtype('float32').

[aytek03]

When I implement train.py with 'HR-Crime' dataset (https://dataverse.nl/dataset.xhtml?persistentId=doi:10.34894/IRRDJE), I got error in training set : (Training Normal Videos)

ValueError: Input contains infinity or a value too large for dtype('float32').

I deleted and extracted some files among training set and I got no error.

Is it possible that I don't get this error without deleting any files?

[stdrr]

Waiting for the answer of the authors, who I thank for their work, I can tell that the error is generated by some divisions by zero while computing the coordinates of the bounding boxes. Precisely, in the function

@staticmethod
def _from_image_to_centre_bounding_box(coordinates, video_resolution):
    # TODO: Better implementation
    # coordinates = np.where(coordinates == 0, np.nan, coordinates)
    # bounding_boxes = np.apply_along_axis(compute_bounding_box, axis=1, arr=coordinates,
    #                                      video_resolution=video_resolution)
    # centre_x = (bounding_boxes[:, 0] + bounding_boxes[:, 1]) / 2
    # centre_y = (bounding_boxes[:, 2] + bounding_boxes[:, 3]) / 2
    for idx, kps in enumerate(coordinates):
        if any(kps):
            left, right, top, bottom = compute_bounding_box(kps, video_resolution=video_resolution)
            centre_x, centre_y = (left + right) / 2, (top + bottom) / 2
            xs, ys = np.hsplit(kps.reshape(-1, 2), indices_or_sections=2)
            xs, ys = np.where(xs == 0.0, centre_x, xs) - centre_x, np.where(ys == 0.0, centre_y, ys) - centre_y
            left, right, top, bottom = left - centre_x, right - centre_x, top - centre_y, bottom - centre_y
            width, height = right - left, bottom - top
            xs, ys = xs / width, ys / height
            kps = np.hstack((xs, ys)).ravel()

        coordinates[idx] = kps

    return coordinates

located at line 131 of skeleton_based_anomaly_detection/tbad/autoencoder/data.py, the division

xs, ys = xs / width, ys / height

may yield infinite results due to width==0 and/or height==0.

A way to bypass this issue is to change the previous line to

xs = xs / width if width != 0 else np.zeros_like(xs)
ys = ys / height if height != 0 else np.zeros_like(ys)

I can't tell if this leads to the expected behaviour of the function, but at least it solves the subsequent error ValueError: Input contains infinity or a value too large for dtype('float32') raised by the scaler on that (bad) input and lets one train the model. One may check if the model's results are meaningful afterwards.

[aytek03]

Hi sir, @stdrr

Thank you for your answering.

I solve this problem like this code.

            if(width ==0 or height==0):
              continue

            xs, ys = xs / width, ys / height
            kps = np.hstack((xs, ys)).ravel()

Thanks a lot by the way.