| |
|
|
| import random |
| import torch |
|
|
| from .densepose_base import DensePoseBaseSampler |
|
|
|
|
| class DensePoseUniformSampler(DensePoseBaseSampler): |
| """ |
| Samples DensePose data from DensePose predictions. |
| Samples for each class are drawn uniformly over all pixels estimated |
| to belong to that class. |
| """ |
|
|
| def __init__(self, count_per_class: int = 8): |
| """ |
| Constructor |
| |
| Args: |
| count_per_class (int): the sampler produces at most `count_per_class` |
| samples for each category |
| """ |
| super().__init__(count_per_class) |
|
|
| def _produce_index_sample(self, values: torch.Tensor, count: int): |
| """ |
| Produce a uniform sample of indices to select data |
| |
| Args: |
| values (torch.Tensor): an array of size [n, k] that contains |
| estimated values (U, V, confidences); |
| n: number of channels (U, V, confidences) |
| k: number of points labeled with part_id |
| count (int): number of samples to produce, should be positive and <= k |
| |
| Return: |
| list(int): indices of values (along axis 1) selected as a sample |
| """ |
| k = values.shape[1] |
| return random.sample(range(k), count) |
|
|
