tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

rnn.cc (3241B)

      1 /*
      2 *  Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
      3 *
      4 *  Use of this source code is governed by a BSD-style license
      5 *  that can be found in the LICENSE file in the root of the source
      6 *  tree. An additional intellectual property rights grant can be found
      7 *  in the file PATENTS.  All contributing project authors may
      8 *  be found in the AUTHORS file in the root of the source tree.
      9 */
     10 
     11 #include "modules/audio_processing/agc2/rnn_vad/rnn.h"
     12 
     13 #include "api/array_view.h"
     14 #include "modules/audio_processing/agc2/cpu_features.h"
     15 #include "modules/audio_processing/agc2/rnn_vad/common.h"
     16 #include "modules/audio_processing/agc2/rnn_vad/rnn_fc.h"
     17 #include "modules/audio_processing/agc2/rnn_vad/rnn_gru.h"
     18 #include "rtc_base/checks.h"
     19 #include "third_party/rnnoise/src/rnn_vad_weights.h"
     20 
     21 namespace webrtc {
     22 namespace rnn_vad {
     23 namespace {
     24 
     25 using ::rnnoise::kInputLayerInputSize;
     26 static_assert(kFeatureVectorSize == kInputLayerInputSize, "");
     27 using ::rnnoise::kInputDenseBias;
     28 using ::rnnoise::kInputDenseWeights;
     29 using ::rnnoise::kInputLayerOutputSize;
     30 static_assert(kInputLayerOutputSize <= kFullyConnectedLayerMaxUnits, "");
     31 
     32 using ::rnnoise::kHiddenGruBias;
     33 using ::rnnoise::kHiddenGruRecurrentWeights;
     34 using ::rnnoise::kHiddenGruWeights;
     35 using ::rnnoise::kHiddenLayerOutputSize;
     36 static_assert(kHiddenLayerOutputSize <= kGruLayerMaxUnits, "");
     37 
     38 using ::rnnoise::kOutputDenseBias;
     39 using ::rnnoise::kOutputDenseWeights;
     40 using ::rnnoise::kOutputLayerOutputSize;
     41 static_assert(kOutputLayerOutputSize <= kFullyConnectedLayerMaxUnits, "");
     42 
     43 }  // namespace
     44 
     45 RnnVad::RnnVad(const AvailableCpuFeatures& cpu_features)
     46    : input_(kInputLayerInputSize,
     47             kInputLayerOutputSize,
     48             kInputDenseBias,
     49             kInputDenseWeights,
     50             ActivationFunction::kTansigApproximated,
     51             cpu_features,
     52             /*layer_name=*/"FC1"),
     53      hidden_(kInputLayerOutputSize,
     54              kHiddenLayerOutputSize,
     55              kHiddenGruBias,
     56              kHiddenGruWeights,
     57              kHiddenGruRecurrentWeights,
     58              cpu_features,
     59              /*layer_name=*/"GRU1"),
     60      output_(kHiddenLayerOutputSize,
     61              kOutputLayerOutputSize,
     62              kOutputDenseBias,
     63              kOutputDenseWeights,
     64              ActivationFunction::kSigmoidApproximated,
     65              // The output layer is just 24x1. The unoptimized code is faster.
     66              NoAvailableCpuFeatures(),
     67              /*layer_name=*/"FC2") {
     68  // Input-output chaining size checks.
     69  RTC_DCHECK_EQ(input_.size(), hidden_.input_size())
     70      << "The input and the hidden layers sizes do not match.";
     71  RTC_DCHECK_EQ(hidden_.size(), output_.input_size())
     72      << "The hidden and the output layers sizes do not match.";
     73 }
     74 
     75 RnnVad::~RnnVad() = default;
     76 
     77 void RnnVad::Reset() {
     78  hidden_.Reset();
     79 }
     80 
     81 float RnnVad::ComputeVadProbability(
     82    ArrayView<const float, kFeatureVectorSize> feature_vector,
     83    bool is_silence) {
     84  if (is_silence) {
     85    Reset();
     86    return 0.f;
     87  }
     88  input_.ComputeOutput(feature_vector);
     89  hidden_.ComputeOutput(input_);
     90  output_.ComputeOutput(hidden_);
     91  RTC_DCHECK_EQ(output_.size(), 1);
     92  return output_.data()[0];
     93 }
     94 
     95 }  // namespace rnn_vad
     96 }  // namespace webrtc