import torch def propose_spans(frame_probability, mask, threshold=0.5, max_spans=4, padding=2, return_bounds=False): if frame_probability.ndim != 2 or frame_probability.shape != (mask.numel(), 3): raise ValueError('expected [T,3] membership/start/end probabilities') active = ((frame_probability[:, 0] >= threshold) & mask).tolist() intervals = [] start = None for index, enabled in enumerate(active + [False]): if enabled and start is None: start = index if not enabled and start is not None: intervals.append((start, index)) start = None intervals.sort(key=lambda bounds: float(frame_probability[bounds[0]:bounds[1], 0].max()), reverse=True) intervals = sorted(intervals[:max_spans]) weights = frame_probability.new_zeros((max_spans, mask.numel())) bounds = torch.full((max_spans, 2), -1, dtype=torch.long, device=mask.device) for slot, (start, stop) in enumerate(intervals): bounds[slot] = bounds.new_tensor([start, stop]) left, right = (max(0, start - padding), min(mask.numel(), stop + padding)) weights[slot, left:right] = frame_probability[left:right, 0].clamp_min(0.05) weights = weights * mask[None] return (weights, bounds) if return_bounds else weights