Spaces:
Running
on
Zero
Running
on
Zero
# 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) | |