approx_torch_quantile

approx_torch_quantile#

ptyrad.utils.math_ops.approx_torch_quantile(t, q, sample_size=16000000)[source]#

Approximated quantile to prevent the 2^24 element (roughly 16.7M) limitation of torch.quantile as of now. See pytorch/pytorch#64947 RuntimeError: quantile() input tensor is too large Note that this approximated quantile would have some randomness.

Parameters:

  • t (torch.Tensor) – Input torch tensor

  • q (float) – Targeted quantile number [0,1]

  • sample_size (int, optional) – Number of randomly selected elements used to approximate the true quantile. Defaults to 16_000_000.

Returns:

The approximated quantile value for the input tensor

Return type:

float

Contents