shithub: opus

Info  •  Files  •  Log  •  Branches

Adding sampling

--- a/dnn/lpcnet.c

+++ b/dnn/lpcnet.c

@@ -33,6 +33,9 @@

 #define NB_FEATURES 38

+#define PITCH_GAIN_FEATURE 37

+#define PDF_FLOOR 0.002

+

 #define FRAME_INPUT_SIZE (NB_FEATURES + EMBED_PITCH_OUT_SIZE)

 #define SAMPLE_INPUT_SIZE (2*EMBED_SIG_OUT_SIZE + EMBED_EXC_OUT_SIZE + FEATURE_DENSE2_OUT_SIZE)

@@ -79,11 +82,10 @@

     RNN_CLEAR(lpc, LPC_ORDER);

 }

-int run_sample_network(NNetState *net, const float *condition, int last_exc, int last_sig, int pred)

+void run_sample_network(NNetState *net, float *pdf, const float *condition, int last_exc, int last_sig, int pred)

 {

     float in_a[SAMPLE_INPUT_SIZE];

     float in_b[GRU_A_STATE_SIZE+FEATURE_DENSE2_OUT_SIZE];

-    float pdf[DUAL_FC_OUT_SIZE];

     compute_embedding(&embed_sig, &in_a[0], last_sig);

     compute_embedding(&embed_sig, &in_a[EMBED_SIG_OUT_SIZE], pred);

     compute_embedding(&embed_exc, &in_a[2*EMBED_SIG_OUT_SIZE], last_exc);

@@ -93,8 +95,6 @@

     RNN_COPY(&in_b[GRU_A_STATE_SIZE], condition, FEATURE_DENSE2_OUT_SIZE);

     compute_gru(&gru_b, net->gru_b_state, in_b);

     compute_mdense(&dual_fc, pdf, net->gru_b_state);

-    /* FIXME: Do the actual sampling here. */

-    return 0;

 }

 void generate_samples(LPCNetState *lpcnet, short *output, const float *features, int pitch, int N)

@@ -102,6 +102,7 @@

     int i;

     float condition[FEATURE_DENSE2_OUT_SIZE];

     float lpc[LPC_ORDER];

+    float pdf[DUAL_FC_OUT_SIZE];

     run_frame_network(&lpcnet->nnet, condition, lpc, features, pitch);

     for (i=0;i<N;i++)

     {

@@ -115,7 +116,8 @@

         pred = (int)floor(.5f + sum);

         last_sig_ulaw = lin2ulaw(lpcnet->last_sig[0]);

         pred_ulaw = lin2ulaw(pred);

-        exc = run_sample_network(&lpcnet->nnet, condition, lpcnet->last_exc, last_sig_ulaw, pred_ulaw);

+        run_sample_network(&lpcnet->nnet, pdf, condition, lpcnet->last_exc, last_sig_ulaw, pred_ulaw);

+        exc = sample_from_pdf(pdf, DUAL_FC_OUT_SIZE, MAX16(0, 1.5f*features[PITCH_GAIN_FEATURE] - .5f), PDF_FLOOR);

         output[i] = pred + ulaw2lin(exc);

         RNN_MOVE(&lpcnet->last_sig[1], &lpcnet->last_sig[0], LPC_ORDER-1);

         lpcnet->last_sig[0] = output[i];

--- a/dnn/nnet.c

+++ b/dnn/nnet.c

@@ -30,6 +30,7 @@

 #include "config.h"

 #endif

+#include <stdlib.h>

 #include <math.h>

 #include "opus_types.h"

 #include "arch.h"

@@ -236,3 +237,32 @@

    }    

 }

+int sample_from_pdf(const float *pdf, int N, float exp_boost, float pdf_floor)

+{

+    int i;

+    float sum, norm;

+    float r;

+    float tmp[DUAL_FC_OUT_SIZE];

+    celt_assert(N <= DUAL_FC_OUT_SIZE);

+    sum = 0;

+    /* Decrease the temperature of the sampling. */

+    for (i=0;i<N;i++)

+    {

+        tmp[i] = pow(pdf[i], 1.f+exp_boost);

+        sum += tmp[i];

+    }

+    norm = 1.f/sum;

+    /* Convert tmp to a CDF while subtracting the floor */

+    tmp[0] = MAX16(0, norm*tmp[0] - pdf_floor);

+    for (i=1;i<N;i++)

+    {

+        tmp[i] = tmp[i-1] + MAX16(0, norm*tmp[i] - pdf_floor);

+    }

+    /* Do the sampling (from the cdf). */

+    r = tmp[N-1] * ((float)rand()/RAND_MAX);

+    for (i=0;i<N-1;i++)

+    {

+        if (r > tmp[i]) return r;

+    }

+    return N-1;

+}

--- a/dnn/nnet.h

+++ b/dnn/nnet.h

@@ -89,5 +89,6 @@

 void compute_embedding(const EmbeddingLayer *layer, float *output, int input);

+int sample_from_pdf(const float *pdf, int N, float exp_boost, float pdf_floor);

 #endif /* _MLP_H_ */

-- 

⑨