torchaudio 0.2.0 → 0.2.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 9488128781f307b653965c253dcacbe246ebaf27bf4e6359b030d9a93bafc1b2
-  data.tar.gz: 5dc18368bdef8945ecaeb1ad2d60771df98bfdd8c356ad2fcdd2b343b8c02b51
+  metadata.gz: 9ed4c14921f1eee18f5e08ddabfae51e09a9b5a7ef408f1dd67fdf7bfe9622fe
+  data.tar.gz: 1e37d5b9abed9cab7bf56a8c30a769bc8ff8f8a3e15e78bbb772847c444571b2
 SHA512:
-  metadata.gz: ddabdfa32632e9d2af024a0f7b67f2add1b694a9a7ef9036f2a1c2f9338106069e6ec1afaea50a0c6f68a8aaa695e57c492d981fb7b3d704fd28ef420c3e2519
-  data.tar.gz: 83131487d0566bb957bab2388e888843ca9b108ec95644811fb702c57bfad628bb41a1ea2b9d1b728090230b48767175bc53af7f0601e8acf8c4d299b5a6de33
+  metadata.gz: 9ca5436d7e4309dd9659fdce7ee893b122e9da96e9f7b15bf00de5dea32c635e2828a99939f04ef5bf0d9494ab89957829a65002dc3e855fa8a66f54abbbd181
+  data.tar.gz: d62b2a137c19d3b24facb11eda5c1b81be5841120b505877b8617bee2b9f183dbe4b4d42a95af27447346a3d48476d7faec48b57cd89b88c0ddc9709f1b5d51b

data/CHANGELOG.md

@@ -1,3 +1,8 @@
+## 0.2.1 (2021-07-16)
+
+- Added `create_dct` method
+- Added `ComputeDeltas`, `Fade`, `MFCC`, and `Vol` transforms
+
 ## 0.2.0 (2021-05-23)
 
 - Updated to Rice 4

data/README.md

@@ -51,10 +51,16 @@ TorchAudio::Transforms::Spectrogram.new.call(waveform)
 
 Supported transforms are:
 
+- AmplitudeToDB
+- ComputeDeltas
+- Fade
+- MelScale
 - MelSpectrogram
+- MFCC
 - MuLawDecoding
 - MuLawEncoding
 - Spectrogram
+- Vol
 
 ## Functional
 
@@ -64,7 +70,11 @@ TorchAudio::Functional.lowpass_biquad(waveform, sample_rate, cutoff_freq)
 
 Supported functions are:
 
+- amplitude_to_DB
 - compute_deltas
+- create_dct
+- create_fb_matrix
+- DB_to_amplitude
 - dither
 - gain
 - highpass_biquad

data/ext/torchaudio/extconf.rb

@@ -22,7 +22,6 @@ apple_clang = RbConfig::CONFIG["CC_VERSION_MESSAGE"] =~ /apple clang/i
 
 # check omp first
 if have_library("omp") || have_library("gomp")
-  $CXXFLAGS += " -DAT_PARALLEL_OPENMP=1"
   $CXXFLAGS += " -Xclang" if apple_clang
   $CXXFLAGS += " -fopenmp"
 end

data/lib/torchaudio.rb

@@ -15,12 +15,16 @@ require "set"
 require "torchaudio/datasets/utils"
 require "torchaudio/datasets/yesno"
 require "torchaudio/functional"
+require "torchaudio/transforms/compute_deltas"
+require "torchaudio/transforms/fade"
 require "torchaudio/transforms/mel_scale"
 require "torchaudio/transforms/mel_spectrogram"
 require "torchaudio/transforms/mu_law_encoding"
 require "torchaudio/transforms/mu_law_decoding"
 require "torchaudio/transforms/spectrogram"
 require "torchaudio/transforms/amplitude_to_db"
+require "torchaudio/transforms/mfcc"
+require "torchaudio/transforms/vol"
 require "torchaudio/version"
 
 module TorchAudio

data/lib/torchaudio/functional.rb

@@ -12,9 +12,18 @@ module TorchAudio
         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
-        )
+        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])
@@ -240,6 +249,23 @@ module TorchAudio
         Torch.pow(Torch.pow(10.0, db * 0.1), power) * ref
       end
 
+      def create_dct(n_mfcc, n_mels, norm: nil)
+        n = Torch.arange(n_mels.to_f)
+        k = Torch.arange(n_mfcc.to_f).unsqueeze!(1)
+        dct = Torch.cos((n + 0.5) * k * Math::PI / n_mels.to_f)
+
+        if norm.nil?
+          dct *= 2.0
+        else
+          raise ArgumentError, "Invalid DCT norm value" unless norm == :ortho
+
+          dct[0] *= 1.0 / Math.sqrt(2.0)
+          dct *= Math.sqrt(2.0 / n_mels)
+        end
+
+        dct.t
+      end
+
       private
 
       def _apply_probability_distribution(waveform, density_function: "TPDF")

data/lib/torchaudio/transforms/compute_deltas.rb

@@ -0,0 +1,15 @@
+module TorchAudio
+  module Transforms
+    class ComputeDeltas < Torch::NN::Module
+      def initialize(win_length: 5, mode: "replicate")
+        super()
+        @win_length = win_length
+        @mode = mode
+      end
+
+      def forward(specgram)
+        F.compute_deltas(specgram, win_length: @win_length, mode: @mode)
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/fade.rb

@@ -0,0 +1,74 @@
+module TorchAudio
+  module Transforms
+    class Fade < Torch::NN::Module
+      def initialize(fade_in_len: 0, fade_out_len: 0, fade_shape: "linear")
+        super()
+        @fade_in_len = fade_in_len
+        @fade_out_len = fade_out_len
+        @fade_shape = fade_shape
+      end
+
+      def forward(waveform)
+        waveform_length = waveform.size[-1]
+        device = waveform.device
+        fade_in(waveform_length).to(device) * fade_out(waveform_length).to(device) * waveform
+      end
+
+      private
+
+      def fade_in(waveform_length)
+        fade = Torch.linspace(0, 1, @fade_in_len)
+        ones = Torch.ones(waveform_length - @fade_in_len)
+
+        if @fade_shape == "linear"
+          fade = fade
+        end
+
+        if @fade_shape == "exponential"
+          fade = Torch.pow(2, (fade - 1)) * fade
+        end
+
+        if @fade_shape == "logarithmic"
+          fade = Torch.log10(0.1 + fade) + 1
+        end
+
+        if @fade_shape == "quarter_sine"
+          fade = Torch.sin(fade * Math::PI / 2)
+        end
+
+        if @fade_shape == "half_sine"
+          fade = Torch.sin(fade * Math::PI - Math::PI / 2) / 2 + 0.5
+        end
+
+        Torch.cat([fade, ones]).clamp!(0, 1)
+      end
+
+      def fade_out(waveform_length)
+        fade = Torch.linspace(0, 1, @fade_out_len)
+        ones = Torch.ones(waveform_length - @fade_out_len)
+
+        if @fade_shape == "linear"
+          fade = - fade + 1
+        end
+
+        if @fade_shape == "exponential"
+          fade = Torch.pow(2, - fade) * (1 - fade)
+        end
+
+        if @fade_shape == "logarithmic"
+          fade = Torch.log10(1.1 - fade) + 1
+        end
+
+        if @fade_shape == "quarter_sine"
+          fade = Torch.sin(fade * Math::PI / 2 + Math::PI / 2)
+        end
+
+        if @fade_shape == "half_sine"
+          fade = Torch.sin(fade * Math::PI + Math::PI / 2) / 2 + 0.5
+        end
+
+        Torch.cat([ones, fade]).clamp!(0, 1)
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/mel_spectrogram.rb

@@ -1,6 +1,8 @@
 module TorchAudio
   module Transforms
     class MelSpectrogram < Torch::NN::Module
+      attr_reader :n_mels
+
       def initialize(
         sample_rate: 16000, n_fft: 400, win_length: nil, hop_length: nil, f_min: 0.0,
         f_max: nil, pad: 0, n_mels: 128, window_fn: Torch.method(:hann_window),

data/lib/torchaudio/transforms/mfcc.rb

@@ -0,0 +1,43 @@
+module TorchAudio
+  module Transforms
+    class MFCC < Torch::NN::Module
+
+      SUPPORTED_DCT_TYPES = [2]
+
+      def initialize(sample_rate: 16000, n_mfcc: 40, dct_type: 2, norm: :ortho, log_mels: false, melkwargs: {})
+        super()
+
+        raise ArgumentError, "DCT type not supported: #{dct_type}" unless SUPPORTED_DCT_TYPES.include?(dct_type)
+
+        @sample_rate = sample_rate
+        @n_mfcc = n_mfcc
+        @dct_type = dct_type
+        @norm = norm
+        @top_db = 80.0
+        @amplitude_to_db = TorchAudio::Transforms::AmplitudeToDB.new(stype: :power, top_db: @top_db)
+
+        @melspectrogram = TorchAudio::Transforms::MelSpectrogram.new(sample_rate: @sample_rate, **melkwargs)
+
+        raise ArgumentError, "Cannot select more MFCC coefficients than # mel bins" if @n_mfcc > @melspectrogram.n_mels
+
+        dct_mat = F.create_dct(@n_mfcc, @melspectrogram.n_mels, norm: @norm)
+        register_buffer('dct_mat', dct_mat)
+
+        @log_mels = log_mels
+      end
+
+      def forward(waveform)
+        mel_specgram = @melspectrogram.(waveform)
+        if @log_mels
+          mel_specgram = Torch.log(mel_specgram + 1e-6)
+        else
+          mel_specgram = @amplitude_to_db.(mel_specgram)
+        end
+
+        Torch
+          .matmul(mel_specgram.transpose(-2, -1), @dct_mat)
+          .transpose(-2, -1)
+      end
+    end
+  end
+end

data/lib/torchaudio/transforms/vol.rb

@@ -0,0 +1,31 @@
+module TorchAudio
+  module Transforms
+    class Vol < Torch::NN::Module
+      def initialize(gain, gain_type: "amplitude")
+        super()
+        @gain = gain
+        @gain_type = gain_type
+
+        if ["amplitude", "power"].include?(gain_type) && gain < 0
+          raise ArgumentError, "If gain_type = amplitude or power, gain must be positive."
+        end
+      end
+
+      def forward(waveform)
+        if @gain_type == "amplitude"
+          waveform = waveform * @gain
+        end
+
+        if @gain_type == "db"
+          waveform = F.gain(waveform, @gain)
+        end
+
+        if @gain_type == "power"
+          waveform = F.gain(waveform, 10 * Math.log10(@gain))
+        end
+
+        Torch.clamp(waveform, -1, 1)
+      end
+    end
+  end
+end

data/lib/torchaudio/version.rb

@@ -1,3 +1,3 @@
 module TorchAudio
-  VERSION = "0.2.0"
+  VERSION = "0.2.1"
 end

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: torchaudio
 version: !ruby/object:Gem::Version
-  version: 0.2.0
+  version: 0.2.1
 platform: ruby
 authors:
 - Andrew Kane
 autorequire:
 bindir: bin
 cert_chain: []
-date: 2021-05-23 00:00:00.000000000 Z
+date: 2021-07-16 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: torch-rb
@@ -64,11 +64,15 @@ files:
 - lib/torchaudio/datasets/yesno.rb
 - lib/torchaudio/functional.rb
 - lib/torchaudio/transforms/amplitude_to_db.rb
+- lib/torchaudio/transforms/compute_deltas.rb
+- lib/torchaudio/transforms/fade.rb
 - lib/torchaudio/transforms/mel_scale.rb
 - lib/torchaudio/transforms/mel_spectrogram.rb
+- lib/torchaudio/transforms/mfcc.rb
 - lib/torchaudio/transforms/mu_law_decoding.rb
 - lib/torchaudio/transforms/mu_law_encoding.rb
 - lib/torchaudio/transforms/spectrogram.rb
+- lib/torchaudio/transforms/vol.rb
 - lib/torchaudio/version.rb
 homepage: https://github.com/ankane/torchaudio
 licenses:
@@ -89,7 +93,7 @@ required_rubygems_version: !ruby/object:Gem::Requirement
     - !ruby/object:Gem::Version
       version: '0'
 requirements: []
-rubygems_version: 3.2.3
+rubygems_version: 3.2.22
 signing_key:
 specification_version: 4
 summary: Data manipulation and transformation for audio signal processing