Module: TorchAudio::Functional
- Defined in:
- lib/torchaudio/functional.rb
Class Method Summary collapse
- .amplitude_to_DB(amp, multiplier, amin, db_multiplier, top_db: nil) ⇒ Object
- .biquad(waveform, b0, b1, b2, a0, a1, a2) ⇒ Object
- .complex_norm(complex_tensor, power: 1.0) ⇒ Object
- .compute_deltas(specgram, win_length: 5, mode: "replicate") ⇒ Object
- .create_fb_matrix(n_freqs, f_min, f_max, n_mels, sample_rate, norm: nil) ⇒ Object
- .DB_to_amplitude(db, ref, power) ⇒ Object
- .dither(waveform, density_function: "TPDF", noise_shaping: false) ⇒ Object
- .gain(waveform, gain_db: 1.0) ⇒ Object
- .highpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707) ⇒ Object
- .lfilter(waveform, a_coeffs, b_coeffs, clamp: true) ⇒ Object
- .lowpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707) ⇒ Object
- .mu_law_decoding(x_mu, quantization_channels) ⇒ Object
- .mu_law_encoding(x, quantization_channels) ⇒ Object
- .spectrogram(waveform, pad, window, n_fft, hop_length, win_length, power, normalized) ⇒ Object
Class Method Details
.amplitude_to_DB(amp, multiplier, amin, db_multiplier, top_db: nil) ⇒ Object
230 231 232 233 234 235 236 237 |
# File 'lib/torchaudio/functional.rb', line 230 def amplitude_to_DB(amp, multiplier, amin, db_multiplier, top_db: nil) db = Torch.log10(Torch.clamp(amp, min: amin)) * multiplier db -= multiplier * db_multiplier db = db.clamp(min: db.max.item - top_db) if top_db db end |
.biquad(waveform, b0, b1, b2, a0, a1, a2) ⇒ Object
136 137 138 139 140 141 142 143 144 145 146 |
# File 'lib/torchaudio/functional.rb', line 136 def biquad(waveform, b0, b1, b2, a0, a1, a2) device = waveform.device dtype = waveform.dtype output_waveform = lfilter( waveform, Torch.tensor([a0, a1, a2], dtype: dtype, device: device), Torch.tensor([b0, b1, b2], dtype: dtype, device: device) ) output_waveform end |
.complex_norm(complex_tensor, power: 1.0) ⇒ Object
54 55 56 |
# File 'lib/torchaudio/functional.rb', line 54 def complex_norm(complex_tensor, power: 1.0) complex_tensor.pow(2.0).sum(-1).pow(0.5 * power) end |
.compute_deltas(specgram, win_length: 5, mode: "replicate") ⇒ Object
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 |
# File 'lib/torchaudio/functional.rb', line 92 def compute_deltas(specgram, win_length: 5, mode: "replicate") device = specgram.device dtype = specgram.dtype # pack batch shape = specgram.size specgram = specgram.reshape(1, -1, shape[-1]) raise ArgumentError, "win_length must be >= 3" unless win_length >= 3 n = (win_length - 1).div(2) # twice sum of integer squared denom = n * (n + 1) * (2 * n + 1) / 3 specgram = Torch::NN::Functional.pad(specgram, [n, n], mode: mode) kernel = Torch.arange(-n, n + 1, 1, device: device, dtype: dtype).repeat([specgram.shape[1], 1, 1]) output = Torch::NN::Functional.conv1d(specgram, kernel, groups: specgram.shape[1]) / denom # unpack batch output = output.reshape(shape) end |
.create_fb_matrix(n_freqs, f_min, f_max, n_mels, sample_rate, norm: nil) ⇒ Object
58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 |
# File 'lib/torchaudio/functional.rb', line 58 def create_fb_matrix(n_freqs, f_min, f_max, n_mels, sample_rate, norm: nil) if norm && norm != "slaney" raise ArgumentError, "norm must be one of None or 'slaney'" end # freq bins # Equivalent filterbank construction by Librosa all_freqs = Torch.linspace(0, sample_rate.div(2), n_freqs) # calculate mel freq bins # hertz to mel(f) is 2595. * math.log10(1. + (f / 700.)) m_min = 2595.0 * Math.log10(1.0 + (f_min / 700.0)) m_max = 2595.0 * Math.log10(1.0 + (f_max / 700.0)) m_pts = Torch.linspace(m_min, m_max, n_mels + 2) # mel to hertz(mel) is 700. * (10**(mel / 2595.) - 1.) f_pts = (Torch.pow(10, m_pts / 2595.0) - 1.0) * 700.0 # calculate the difference between each mel point and each stft freq point in hertz f_diff = f_pts[1..-1] - f_pts[0...-1] # (n_mels + 1) slopes = f_pts.unsqueeze(0) - all_freqs.unsqueeze(1) # (n_freqs, n_mels + 2) # create overlapping triangles zero = Torch.zeros(1) down_slopes = (slopes[0..-1, 0...-2] * -1.0) / f_diff[0...-1] # (n_freqs, n_mels) up_slopes = slopes[0..-1, 2..-1] / f_diff[1..-1] # (n_freqs, n_mels) fb = Torch.max(zero, Torch.min(down_slopes, up_slopes)) if norm && norm == "slaney" # Slaney-style mel is scaled to be approx constant energy per channel enorm = 2.0 / (f_pts[2...(n_mels + 2)] - f_pts[:n_mels]) fb *= enorm.unsqueeze(0) end fb end |
.DB_to_amplitude(db, ref, power) ⇒ Object
239 240 241 |
# File 'lib/torchaudio/functional.rb', line 239 def DB_to_amplitude(db, ref, power) Torch.pow(Torch.pow(10.0, db * 0.1), power) * ref end |
.dither(waveform, density_function: "TPDF", noise_shaping: false) ⇒ Object
125 126 127 128 129 130 131 132 133 134 |
# File 'lib/torchaudio/functional.rb', line 125 def dither(waveform, density_function: "TPDF", noise_shaping: false) dithered = _apply_probability_distribution(waveform, density_function: density_function) if noise_shaping raise "Not implemented yet" # _add_noise_shaping(dithered, waveform) else dithered end end |
.gain(waveform, gain_db: 1.0) ⇒ Object
117 118 119 120 121 122 123 |
# File 'lib/torchaudio/functional.rb', line 117 def gain(waveform, gain_db: 1.0) return waveform if gain_db == 0 ratio = 10 ** (gain_db / 20) waveform * ratio end |
.highpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707) ⇒ Object
148 149 150 151 152 153 154 155 156 157 158 159 |
# File 'lib/torchaudio/functional.rb', line 148 def highpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707) w0 = 2 * Math::PI * cutoff_freq / sample_rate alpha = Math.sin(w0) / 2.0 / q b0 = (1 + Math.cos(w0)) / 2 b1 = -1 - Math.cos(w0) b2 = b0 a0 = 1 + alpha a1 = -2 * Math.cos(w0) a2 = 1 - alpha biquad(waveform, b0, b1, b2, a0, a1, a2) end |
.lfilter(waveform, a_coeffs, b_coeffs, clamp: true) ⇒ Object
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 |
# File 'lib/torchaudio/functional.rb', line 174 def lfilter(waveform, a_coeffs, b_coeffs, clamp: true) # pack batch shape = waveform.size waveform = waveform.reshape(-1, shape[-1]) raise ArgumentError unless (a_coeffs.size(0) == b_coeffs.size(0)) raise ArgumentError unless (waveform.size.length == 2) raise ArgumentError unless (waveform.device == a_coeffs.device) raise ArgumentError unless (b_coeffs.device == a_coeffs.device) device = waveform.device dtype = waveform.dtype n_channel, n_sample = waveform.size n_order = a_coeffs.size(0) n_sample_padded = n_sample + n_order - 1 raise ArgumentError unless (n_order > 0) # Pad the input and create output padded_waveform = Torch.zeros(n_channel, n_sample_padded, dtype: dtype, device: device) padded_waveform[0..-1, (n_order - 1)..-1] = waveform padded_output_waveform = Torch.zeros(n_channel, n_sample_padded, dtype: dtype, device: device) # Set up the coefficients matrix # Flip coefficients' order a_coeffs_flipped = a_coeffs.flip([0]) b_coeffs_flipped = b_coeffs.flip([0]) # calculate windowed_input_signal in parallel # create indices of original with shape (n_channel, n_order, n_sample) window_idxs = Torch.arange(n_sample, device: device).unsqueeze(0) + Torch.arange(n_order, device: device).unsqueeze(1) window_idxs = window_idxs.repeat([n_channel, 1, 1]) window_idxs += (Torch.arange(n_channel, device: device).unsqueeze(-1).unsqueeze(-1) * n_sample_padded) window_idxs = window_idxs.long # (n_order, ) matmul (n_channel, n_order, n_sample) -> (n_channel, n_sample) input_signal_windows = Torch.matmul(b_coeffs_flipped, Torch.take(padded_waveform, window_idxs)) input_signal_windows.div!(a_coeffs[0]) a_coeffs_flipped.div!(a_coeffs[0]) input_signal_windows.t.each_with_index do |o0, i_sample| windowed_output_signal = padded_output_waveform[0..-1, i_sample...(i_sample + n_order)] o0.addmv!(windowed_output_signal, a_coeffs_flipped, alpha: -1) padded_output_waveform[0..-1, i_sample + n_order - 1] = o0 end output = padded_output_waveform[0..-1, (n_order - 1)..-1] if clamp output = Torch.clamp(output, -1.0, 1.0) end # unpack batch output = output.reshape(shape[0...-1] + output.shape[-1..-1]) output end |
.lowpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707) ⇒ Object
161 162 163 164 165 166 167 168 169 170 171 172 |
# File 'lib/torchaudio/functional.rb', line 161 def lowpass_biquad(waveform, sample_rate, cutoff_freq, q: 0.707) w0 = 2 * Math::PI * cutoff_freq / sample_rate alpha = Math.sin(w0) / 2 / q b0 = (1 - Math.cos(w0)) / 2 b1 = 1 - Math.cos(w0) b2 = b0 a0 = 1 + alpha a1 = -2 * Math.cos(w0) a2 = 1 - alpha biquad(waveform, b0, b1, b2, a0, a1, a2) end |
.mu_law_decoding(x_mu, quantization_channels) ⇒ Object
43 44 45 46 47 48 49 50 51 52 |
# File 'lib/torchaudio/functional.rb', line 43 def mu_law_decoding(x_mu, quantization_channels) mu = quantization_channels - 1.0 if !x_mu.floating_point? x_mu = x_mu.to(dtype: :float) end mu = Torch.tensor(mu, dtype: x_mu.dtype) x = ((x_mu) / mu) * 2 - 1.0 x = Torch.sign(x) * (Torch.exp(Torch.abs(x) * Torch.log1p(mu)) - 1.0) / mu x end |
.mu_law_encoding(x, quantization_channels) ⇒ Object
32 33 34 35 36 37 38 39 40 41 |
# File 'lib/torchaudio/functional.rb', line 32 def mu_law_encoding(x, quantization_channels) mu = quantization_channels - 1.0 if !x.floating_point? x = x.to(dtype: :float) end mu = Torch.tensor(mu, dtype: x.dtype) x_mu = Torch.sign(x) * Torch.log1p(mu * Torch.abs(x)) / Torch.log1p(mu) x_mu = ((x_mu + 1) / 2 * mu + 0.5).to(dtype: :int64) x_mu end |
.spectrogram(waveform, pad, window, n_fft, hop_length, win_length, power, normalized) ⇒ Object
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 |
# File 'lib/torchaudio/functional.rb', line 4 def spectrogram(waveform, pad, window, n_fft, hop_length, win_length, power, normalized) if pad > 0 # TODO add "with torch.no_grad():" back when JIT supports it waveform = Torch::NN::Functional.pad(waveform, [pad, pad], "constant") end # pack batch shape = waveform.size waveform = waveform.reshape(-1, shape[-1]) # default values are consistent with librosa.core.spectrum._spectrogram spec_f = Torch.stft( waveform, n_fft, hop_length: hop_length, win_length: win_length, window: window, center: true, pad_mode: "reflect", normalized: false, onesided: true ) # unpack batch spec_f = spec_f.reshape(shape[0..-2] + spec_f.shape[-3..-1]) if normalized spec_f.div!(window.pow(2.0).sum.sqrt) end if power spec_f = complex_norm(spec_f, power: power) end spec_f end |