Leffa / densepose /data /samplers /densepose_uniform.py
franciszzj's picture
init code
b213d84
raw
history blame
1.29 kB
# 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)