Mistake: the array's second frame is the video's second frame

“clip holds the video, so clip[1] is the second thing that happened.”

kept = np.arange(0, 720, 45)
assert kept.shape == (16,)
assert kept[1] == 45
assert round(100 * len(kept) / 720, 1) == 2.2

clip keeps every 45th frame of 720, so entry 1 is source frame 45, and 97.8% of the recorded moments are not in the tensor at all.

An index into a sampled axis is a position in the array, not a moment in time. To answer “when was clip[1]?” in seconds you have to store the frame rate and the kept indices beside the tensor.