Newer
Older
# Author : Jan Schlüter
from torch import nn
import torch
import numpy as np
def create_mel_filterbank(sample_rate, frame_len, num_bands, min_freq, max_freq,
norm=True, crop=False):
"""
Creates a mel filterbank of `num_bands` triangular filters, with the first
filter starting at `min_freq` and the last one stopping at `max_freq`.
Returns the filterbank as a matrix suitable for a dot product against
magnitude spectra created from samples at a sample rate of `sample_rate`
with a window length of `frame_len` samples. If `norm`, will normalize
each filter by its area. If `crop`, will exclude rows that exceed the
maximum frequency and are therefore zero.
"""
# mel-spaced peak frequencies
min_mel = 1127 * np.log1p(min_freq / 7000.0)
max_mel = 1127 * np.log1p(max_freq / 7000.0)
peaks_bin = peaks_hz * frame_len / sample_rate
# create filterbank
input_bins = (frame_len // 2) + 1
if crop:
input_bins = min(input_bins,
int(np.ceil(max_freq * frame_len /
float(sample_rate))))
x = torch.arange(input_bins, dtype=peaks_bin.dtype)[:, np.newaxis]
l, c, r = peaks_bin[0:-2], peaks_bin[1:-1], peaks_bin[2:]
# triangles are the minimum of two linear functions f(x) = a*x + b
# left side of triangles: f(l) = 0, f(c) = 1 -> a=1/(c-l), b=-a*l
tri_left = (x - l) / (c - l)
# right side of triangles: f(c) = 1, f(r) = 0 -> a=1/(c-r), b=-a*r
tri_right = (x - r) / (c - r)
# combine by taking the minimum of the left and right sides
tri = torch.min(tri_left, tri_right)
# and clip to only keep positive values
filterbank = torch.clamp(tri, min=0)
# normalize by area
if norm:
filterbank /= filterbank.sum(0)
return filterbank
class MelFilter(nn.Module):
def __init__(self, sample_rate, winsize, num_bands, min_freq, max_freq):
super(MelFilter, self).__init__()
melbank = create_mel_filterbank(sample_rate, winsize, num_bands,
min_freq, max_freq, crop=True)
self.register_buffer('bank', melbank)
def forward(self, x):
x = x.transpose(-1, -2) # put fft bands last
x = x[..., :self.bank.shape[0]] # remove unneeded fft bands
x = x.matmul(self.bank) # turn fft bands into mel bands
x = x.transpose(-1, -2) # put time last
return x
result = super(MelFilter, self).state_dict(destination, prefix, keep_vars)
# remove all buffers; we use them as cached constants
for k in self._buffers:
del result[prefix + k]
return result
# ignore stored buffers for backwards compatibility
for k in self._buffers:
state_dict.pop(prefix + k, None)
# temporarily hide the buffers; we do not want to restore them
buffers = self._buffers
self._buffers = {}
result = super(MelFilter, self)._load_from_state_dict(state_dict, prefix, *args, **kwargs)
self._buffers = buffers
return result
class STFT(nn.Module):
def __init__(self, winsize, hopsize, complex=False):
super(STFT, self).__init__()
self.winsize = winsize
self.hopsize = hopsize
self.register_buffer('window',
torch.hann_window(winsize, periodic=False))
self.complex = complex
result = super(STFT, self).state_dict(destination, prefix, keep_vars)
# remove all buffers; we use them as cached constants
for k in self._buffers:
del result[prefix + k]
return result
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
# ignore stored buffers for backwards compatibility
for k in self._buffers:
state_dict.pop(prefix + k, None)
# temporarily hide the buffers; we do not want to restore them
buffers = self._buffers
self._buffers = {}
result = super(STFT, self)._load_from_state_dict(state_dict, prefix, *args, **kwargs)
self._buffers = buffers
return result
def forward(self, x):
x = x.unsqueeze(1)
# we want each channel to be treated separately, so we mash
# up the channels and batch size and split them up afterwards
batchsize, channels = x.shape[:2]
x = x.reshape((-1,) + x.shape[2:])
# we apply the STFT
x = torch.stft(x, self.winsize, self.hopsize, window=self.window,
center=False, return_complex=False)
# we compute magnitudes, if requested
if not self.complex:
x = x.norm(p=2, dim=-1)
# restore original batchsize and channels in case we mashed them
x = x.reshape((batchsize, channels, -1) + x.shape[2:]) #if channels > 1 else x.reshape((batchsize, -1) + x.shape[2:])
return x
class TemporalBatchNorm(nn.Module):
"""
Batch normalization of a (batch, channels, bands, time) tensor over all but
the previous to last dimension (the frequency bands).
"""
def __init__(self, num_bands):
super(TemporalBatchNorm, self).__init__()
self.bn = nn.BatchNorm1d(num_bands)
def forward(self, x):
shape = x.shape
# squash channels into the batch dimension
x = x.reshape((-1,) + x.shape[-2:])
# pass through 1D batch normalization
x = self.bn(x)
# restore squashed dimensions
return x.reshape(shape)
class Log1p(nn.Module):
"""
Applies log(1 + 10**a * x), with scale fixed or trainable.
"""
def __init__(self, a=0, trainable=False):
super(Log1p, self).__init__()
if trainable:
a = nn.Parameter(torch.tensor(a, dtype=torch.get_default_dtype()))
self.a = a
self.trainable = trainable
def forward(self, x):
if self.trainable or self.a != 0:
x = torch.log1p(10 ** self.a * x)
return x
def extra_repr(self):
return 'trainable={}'.format(repr(self.trainable))