Sampling frames, and what it costs

A recommender samples 8 frames out of 900. That is indexing along the time axis — the same operation as section 03, applied to a different axis.

What is lost is everything between the samples. Keep every 45th frame of 720 and you keep 16 frames: 97.8% of the recorded moments are not in the tensor at all.

kept = np.arange(0, 720, 45)
kept.shape          # (16,)
kept[1]             # 45 — source frame, not "the second moment"