Spaces:
Runtime error
Runtime error
| # Copyright (c) Facebook, Inc. and its affiliates. | |
| 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) | |