mokka.functional

Functions which do not fit anywhere else.

Functions that do not fit anywhere else - in PyTorch.

mokka.functional.torch.convolve(signal, kernel, mode='full')

Calculate the 1-D convolution using the torch.conv1d function.

This function is a convenience to assist using the torch convolution function within a signal processing context. The convolution function provided by PyTorch has image processing in mind and therefore some parameters of this function can be simplfied.

Parameters:

  • signal – signal in the time domain.

  • kernel – second signal in the time domain, must be shorter.

  • mode – either “full”, “same” or “valid”.

Returns:

convolution between signal and kernel.

mokka.functional.torch.convolve_overlap_save(signal, kernel, mode='full')

Calculate the 1-D convolution using FFT overlap-save method.

This is only efficient for very long signals, where signal length >> 10*kernel length

mokka.functional.torch.distribution_quant_gumbel_softmax(probs, n, temp=0.1)

Compute a quantized distribution for given probabilities with Gumbel-Softmax.

This uses the Gumbel-Softmax sampling method to provide a differentiable way to embed the information about the distribution into the sequence. We use the straight-through trick to retain the gradients of softmax but actually use the values from a argmax for calculation.

Parameters:

  • probs – probabilities of occurence for each element

  • n – length of the sequence to sample

  • temp – temperature of Softmax.

Returns:

sampled indices in as sequence of length n

mokka.functional.torch.distribution_quantization(probs, n)

Compute a quantized distribution for given probabilities.

This function uses an algorithm introduced for calculating a discrete distribution minimizing the KL divergence for CCDM.

Parameters:

  • probs – probabilities of occurence for each element.

  • n – length of the full sequence.

Returns:

number of ocurrences for each element in the sequence.

mokka.functional.torch.unwrap(p, discont=3.141592653589793, dim=-1, period=6.283185307179586)

Perform phase unwrapping in PyTorch.

This implementation follows the NumPy implementation closely.

Parameters:

  • p – tensor with angle values

  • discont – discontinuity to unwrap

  • dim – dimension to unwrap

  • period – which periodicity the signal has

mokka.functional.torch.unwrap_torch(*args)

Perform phase unwrapping in PyTorch.

This implementation follows the NumPy implementation closely.

Parameters:

  • p – tensor with angle values

  • discont – discontinuity to unwrap

  • dim – dimension to unwrap

  • period – which periodicity the signal has