Spaces:
Running
Running
File size: 2,645 Bytes
9791162 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 |
import librosa
import numpy as np
import torch
import crepe
###############################################################################
# Probability sequence decoding methods
###############################################################################
def argmax(logits):
"""Sample observations by taking the argmax"""
bins = logits.argmax(dim=1)
# Convert to frequency in Hz
return bins, crepe.convert.bins_to_frequency(bins)
def weighted_argmax(logits):
"""Sample observations using weighted sum near the argmax"""
# Find center of analysis window
bins = logits.argmax(dim=1)
# Find bounds of analysis window
start = torch.max(torch.tensor(0, device=logits.device), bins - 4)
end = torch.min(torch.tensor(logits.size(1), device=logits.device), bins + 5)
# Mask out everything outside of window
for batch in range(logits.size(0)):
for time in range(logits.size(2)):
logits[batch, :start[batch, time], time] = -float('inf')
logits[batch, end[batch, time]:, time] = -float('inf')
# Construct weights
if not hasattr(weighted_argmax, 'weights'):
weights = crepe.convert.bins_to_cents(torch.arange(360))
weighted_argmax.weights = weights[None, :, None]
# Ensure devices are the same (no-op if they are)
weighted_argmax.weights = weighted_argmax.weights.to(logits.device)
# Convert to probabilities
with torch.no_grad():
probs = torch.sigmoid(logits)
# Apply weights
cents = (weighted_argmax.weights * probs).sum(dim=1) / probs.sum(dim=1)
# Convert to frequency in Hz
return bins, crepe.convert.cents_to_frequency(cents)
def viterbi(logits):
"""Sample observations using viterbi decoding"""
# Create viterbi transition matrix
if not hasattr(viterbi, 'transition'):
xx, yy = np.meshgrid(range(360), range(360))
transition = np.maximum(12 - abs(xx - yy), 0)
transition = transition / transition.sum(axis=1, keepdims=True)
viterbi.transition = transition
# Normalize logits
with torch.no_grad():
probs = torch.nn.functional.softmax(logits, dim=1)
# Convert to numpy
sequences = probs.cpu().numpy()
# Perform viterbi decoding
bins = np.array([
librosa.sequence.viterbi(sequence, viterbi.transition).astype(np.int64)
for sequence in sequences])
# Convert to pytorch
bins = torch.tensor(bins, device=probs.device)
# Convert to frequency in Hz
return bins, crepe.convert.bins_to_frequency(bins)
|