1 /* Copyright (c) 2008-2011 Octasic Inc.
2 2012-2017 Jean-Marc Valin */
3 /*
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions
6 are met:
7
8 - Redistributions of source code must retain the above copyright
9 notice, this list of conditions and the following disclaimer.
10
11 - Redistributions in binary form must reproduce the above copyright
12 notice, this list of conditions and the following disclaimer in the
13 documentation and/or other materials provided with the distribution.
14
15 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
16 ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
17 LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
18 A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR
19 CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
20 EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
21 PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
22 PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
23 LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
24 NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
25 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 */
27
28 #ifdef HAVE_CONFIG_H
29 #include "config.h"
30 #endif
31
32 #include <math.h>
33 #include "opus_types.h"
34 #include "common.h"
35 #include "arch.h"
36 #include "tansig_table.h"
37 #include "rnn.h"
38 #include "rnn_data.h"
39 #include <stdio.h>
40
tansig_approx(float x)41 static OPUS_INLINE float tansig_approx(float x)
42 {
43 int i;
44 float y, dy;
45 float sign=1;
46 /* Tests are reversed to catch NaNs */
47 if (!(x<8))
48 return 1;
49 if (!(x>-8))
50 return -1;
51 #ifndef FIXED_POINT
52 /* Another check in case of -ffast-math */
53 if (celt_isnan(x))
54 return 0;
55 #endif
56 if (x<0)
57 {
58 x=-x;
59 sign=-1;
60 }
61 i = (int)floor(.5f+25*x);
62 x -= .04f*i;
63 y = tansig_table[i];
64 dy = 1-y*y;
65 y = y + x*dy*(1 - y*x);
66 return sign*y;
67 }
68
sigmoid_approx(float x)69 static OPUS_INLINE float sigmoid_approx(float x)
70 {
71 return .5 + .5*tansig_approx(.5*x);
72 }
73
relu(float x)74 static OPUS_INLINE float relu(float x)
75 {
76 return x < 0 ? 0 : x;
77 }
78
compute_dense(const DenseLayer * layer,float * output,const float * input)79 void compute_dense(const DenseLayer *layer, float *output, const float *input)
80 {
81 int i, j;
82 int N, M;
83 int stride;
84 M = layer->nb_inputs;
85 N = layer->nb_neurons;
86 stride = N;
87 for (i=0;i<N;i++)
88 {
89 /* Compute update gate. */
90 float sum = layer->bias[i];
91 for (j=0;j<M;j++)
92 sum += layer->input_weights[j*stride + i]*input[j];
93 output[i] = WEIGHTS_SCALE*sum;
94 }
95 if (layer->activation == ACTIVATION_SIGMOID) {
96 for (i=0;i<N;i++)
97 output[i] = sigmoid_approx(output[i]);
98 } else if (layer->activation == ACTIVATION_TANH) {
99 for (i=0;i<N;i++)
100 output[i] = tansig_approx(output[i]);
101 } else if (layer->activation == ACTIVATION_RELU) {
102 for (i=0;i<N;i++)
103 output[i] = relu(output[i]);
104 } else {
105 *(int*)0=0;
106 }
107 }
108
compute_gru(const GRULayer * gru,float * state,const float * input)109 void compute_gru(const GRULayer *gru, float *state, const float *input)
110 {
111 int i, j;
112 int N, M;
113 int stride;
114 float z[MAX_NEURONS];
115 float r[MAX_NEURONS];
116 float h[MAX_NEURONS];
117 M = gru->nb_inputs;
118 N = gru->nb_neurons;
119 stride = 3*N;
120 for (i=0;i<N;i++)
121 {
122 /* Compute update gate. */
123 float sum = gru->bias[i];
124 for (j=0;j<M;j++)
125 sum += gru->input_weights[j*stride + i]*input[j];
126 for (j=0;j<N;j++)
127 sum += gru->recurrent_weights[j*stride + i]*state[j];
128 z[i] = sigmoid_approx(WEIGHTS_SCALE*sum);
129 }
130 for (i=0;i<N;i++)
131 {
132 /* Compute reset gate. */
133 float sum = gru->bias[N + i];
134 for (j=0;j<M;j++)
135 sum += gru->input_weights[N + j*stride + i]*input[j];
136 for (j=0;j<N;j++)
137 sum += gru->recurrent_weights[N + j*stride + i]*state[j];
138 r[i] = sigmoid_approx(WEIGHTS_SCALE*sum);
139 }
140 for (i=0;i<N;i++)
141 {
142 /* Compute output. */
143 float sum = gru->bias[2*N + i];
144 for (j=0;j<M;j++)
145 sum += gru->input_weights[2*N + j*stride + i]*input[j];
146 for (j=0;j<N;j++)
147 sum += gru->recurrent_weights[2*N + j*stride + i]*state[j]*r[j];
148 if (gru->activation == ACTIVATION_SIGMOID) sum = sigmoid_approx(WEIGHTS_SCALE*sum);
149 else if (gru->activation == ACTIVATION_TANH) sum = tansig_approx(WEIGHTS_SCALE*sum);
150 else if (gru->activation == ACTIVATION_RELU) sum = relu(WEIGHTS_SCALE*sum);
151 else *(int*)0=0;
152 h[i] = z[i]*state[i] + (1-z[i])*sum;
153 }
154 for (i=0;i<N;i++)
155 state[i] = h[i];
156 }
157
158 #define INPUT_SIZE 42
159
compute_rnn(RNNState * rnn,float * gains,float * vad,const float * input)160 void compute_rnn(RNNState *rnn, float *gains, float *vad, const float *input) {
161 int i;
162 float dense_out[MAX_NEURONS];
163 float noise_input[MAX_NEURONS*3];
164 float denoise_input[MAX_NEURONS*3];
165 compute_dense(rnn->model->input_dense, dense_out, input);
166 compute_gru(rnn->model->vad_gru, rnn->vad_gru_state, dense_out);
167 compute_dense(rnn->model->vad_output, vad, rnn->vad_gru_state);
168 for (i=0;i<rnn->model->input_dense_size;i++) noise_input[i] = dense_out[i];
169 for (i=0;i<rnn->model->vad_gru_size;i++) noise_input[i+rnn->model->input_dense_size] = rnn->vad_gru_state[i];
170 for (i=0;i<INPUT_SIZE;i++) noise_input[i+rnn->model->input_dense_size+rnn->model->vad_gru_size] = input[i];
171 compute_gru(rnn->model->noise_gru, rnn->noise_gru_state, noise_input);
172
173 for (i=0;i<rnn->model->vad_gru_size;i++) denoise_input[i] = rnn->vad_gru_state[i];
174 for (i=0;i<rnn->model->noise_gru_size;i++) denoise_input[i+rnn->model->vad_gru_size] = rnn->noise_gru_state[i];
175 for (i=0;i<INPUT_SIZE;i++) denoise_input[i+rnn->model->vad_gru_size+rnn->model->noise_gru_size] = input[i];
176 compute_gru(rnn->model->denoise_gru, rnn->denoise_gru_state, denoise_input);
177 compute_dense(rnn->model->denoise_output, gains, rnn->denoise_gru_state);
178 }
179