Buckets:
| """Lossless, pre-selection capture of decoded YOLO oriented predictions.""" | |
| import numpy as np | |
| import torch | |
| def capture_decoded(decoded, class_logits, class_names, tile_origin, valid_size, *, regression_logits=None): | |
| """Return every tile-local xywhr prediction and class logit, before selection. | |
| ``decoded`` is the model head's one-branch output: B x (4 + classes + 1) | |
| x anchors, with class probabilities and the original angle channel. Input | |
| batch size must be one. Tile origin and valid size are recorded separately; | |
| coordinates are not clipped or transformed. | |
| """ | |
| if decoded.ndim != 3 or class_logits.ndim != 3 or decoded.shape[0] != 1 or class_logits.shape[0] != 1: | |
| raise ValueError('expected single-image, three-dimensional tensors') | |
| count = class_logits.shape[1] | |
| if count != len(class_names) or decoded.shape[1] != count + 5 or decoded.shape[2] != class_logits.shape[2] or not decoded.shape[2]: | |
| raise ValueError('decoded, logit and class dimensions do not agree') | |
| if len(set(class_names)) != count or any(not isinstance(n,str) or not n for n in class_names): | |
| raise ValueError('class names must be nonempty and unique') | |
| if (len(tile_origin) != 2 or len(valid_size) != 2 or | |
| any(type(v) is not int or v < 0 for v in tile_origin) or | |
| any(type(v) is not int or not 0 < v <= 1024 for v in valid_size)): | |
| raise ValueError('invalid tile origin or valid size') | |
| if not torch.isfinite(decoded).all() or not torch.isfinite(class_logits).all(): | |
| raise ValueError('nonfinite raw model output') | |
| if not torch.allclose(decoded[:,4:4+count], class_logits.sigmoid(), rtol=1e-5, atol=1e-6): | |
| raise ValueError('decoded probabilities do not match class logits') | |
| boxes = torch.cat((decoded[:, :4], decoded[:, -1:]), dim=1)[0].transpose(0,1) | |
| if regression_logits is not None: | |
| if (regression_logits.ndim != 3 or regression_logits.shape[0] != 1 or | |
| regression_logits.shape[2] != decoded.shape[2] or not torch.isfinite(regression_logits).all()): | |
| raise ValueError('invalid raw box regression output') | |
| result = {'xywhr': boxes.detach().cpu().numpy().copy(), | |
| 'class_logits': class_logits[0].transpose(0,1).detach().cpu().numpy().copy(), | |
| 'class_names': np.asarray(class_names), 'tile_origin': np.asarray(tile_origin,dtype=np.int64), | |
| 'valid_size': np.asarray(valid_size,dtype=np.int64)} | |
| if regression_logits is not None: | |
| result['regression_logits'] = regression_logits[0].transpose(0,1).detach().cpu().numpy().copy() | |
| return result | |
Xet Storage Details
- Size:
- 2.63 kB
- Xet hash:
- 01d19a5e01d4eeb0fee8a493433cfc52c903144faf2291dce39fe861e26fd68b
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.