shithub: opus

Info  •  Files  •  Log  •  Branches

Sampling directly from the logit

Avoids having to compute a sigmoid

--- a/dnn/lpcnet.c

+++ b/dnn/lpcnet.c

@@ -77,7 +77,7 @@

     if (lpcnet->frame_count < 1000) lpcnet->frame_count++;

 }

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

+int run_sample_network(NNetState *net, const float *condition, const float *gru_a_condition, int last_exc, int last_sig, int pred, const float *sampling_logit_table)

 {

     float gru_a_input[3*GRU_A_STATE_SIZE];

     float in_b[GRU_A_STATE_SIZE+FEATURE_DENSE2_OUT_SIZE];

@@ -94,7 +94,7 @@

     RNN_COPY(in_b, net->gru_a_state, GRU_A_STATE_SIZE);

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

     compute_gru2(&gru_b, net->gru_b_state, in_b);

-    return sample_mdense(&dual_fc, net->gru_b_state);

+    return sample_mdense(&dual_fc, net->gru_b_state, sampling_logit_table);

 }

 LPCNET_EXPORT int lpcnet_get_size()

@@ -104,8 +104,13 @@

 LPCNET_EXPORT int lpcnet_init(LPCNetState *lpcnet)

 {

+    int i;

     memset(lpcnet, 0, lpcnet_get_size());

     lpcnet->last_exc = lin2ulaw(0.f);

+    for (i=0;i<256;i++) {

+        float prob = .025+.95*i/255.;

+        lpcnet->sampling_logit_table[i] = -log((1-prob)/prob);

+    }

     return 0;

 }

@@ -155,7 +160,7 @@

         for (j=0;j<LPC_ORDER;j++) pred -= lpcnet->last_sig[j]*lpc[j];

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

         pred_ulaw = lin2ulaw(pred);

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

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

         pcm = pred + ulaw2lin(exc);

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

         lpcnet->last_sig[0] = pcm;

--- a/dnn/lpcnet_private.h

+++ b/dnn/lpcnet_private.h

@@ -29,6 +29,7 @@

     float old_input[FEATURES_DELAY][FEATURE_CONV2_OUT_SIZE];

     float old_lpc[FEATURES_DELAY][LPC_ORDER];

     float old_gain[FEATURES_DELAY];

+    float sampling_logit_table[256];

     int frame_count;

     float deemph_mem;

 };

--- a/dnn/nnet.c

+++ b/dnn/nnet.c

@@ -141,7 +141,7 @@

    compute_activation(output, output, N, layer->activation);

 }

-int sample_mdense(const MDenseLayer *layer, const float *input)

+int sample_mdense(const MDenseLayer *layer, const float *input, const float *sampling_logit_table)

 {

    int b, j, N, M, C, stride;

    M = layer->nb_inputs;

@@ -152,7 +152,12 @@

    celt_assert(N <= DUAL_FC_OUT_SIZE);

    int val=0;

-    

+   float thresholds[8];

+

+   /* Computing all the random thresholds in advance. These thresholds are directly

+      based on the logit to avoid computing the sigmoid.*/

+   for (b=0;b<8;b++) thresholds[b] = sampling_logit_table[rand()&0xFF];

+

    for (b=0;b<8;b++)

    {

       int bit;

@@ -171,9 +176,12 @@

       sum2 = layer->factor[N + i]*tanh_approx(sum2);

       sum1 += sum2;

       //sum1 = 1.f/(1 + exp(-sum1));

+#if 1 /* Sample the decision based on the logit. */

+      bit = thresholds[b] < sum1;

+#else

       sum1 = sigmoid_approx(sum1);

-      

       bit = .025+.95*((rand()+.5f)/(RAND_MAX+1.f)) < sum1;

+#endif

       val = (val << 1) | bit;

    }

    return val;

--- a/dnn/nnet.h

+++ b/dnn/nnet.h

@@ -97,7 +97,7 @@

 void compute_mdense(const MDenseLayer *layer, float *output, const float *input);

-int sample_mdense(const MDenseLayer *layer,  const float *input);

+int sample_mdense(const MDenseLayer *layer,  const float *input, const float *sampling_logit_table);

 void compute_gru(const GRULayer *gru, float *state, const float *input);

-- 

⑨