shithub: opus

Info  •  Files  •  Log  •  Branches

Cleaning up the sparse GRU

It no longer overwrites its input vector

--- a/dnn/nnet.c

+++ b/dnn/nnet.c

@@ -68,7 +68,7 @@

    }

 }

-void compute_activation(float *output, float *input, int N, int activation)

+void compute_activation(float *output, const float *input, int N, int activation)

 {

    int i;

    if (activation == ACTIVATION_SIGMOID) {

@@ -322,8 +322,8 @@

       state[i] = h[i];

 }

-/* WARNING: for efficiency reasons, this function overwrites the input vector. */

-void compute_sparse_gru(const SparseGRULayer *gru, float *state, float *input)

+/* The input of this GRU is after the input matrix multiply. */

+void compute_sparse_gru(const SparseGRULayer *gru, float *state, const float *input)

 {

    int i, k;

    int N;

@@ -333,9 +333,9 @@

    float *h;

    const float *bias;

    N = gru->nb_neurons;

-   z = input;

-   r = &input[N];

-   h = &input[2*N];

+   z = recur;

+   r = &recur[N];

+   h = &recur[2*N];

    celt_assert(gru->nb_neurons <= MAX_RNN_NEURONS);

    celt_assert(input != state);

    celt_assert(gru->reset_after);

@@ -344,17 +344,20 @@

 #else

    bias = &gru->bias[3*N];   

 #endif

-   for (k=0;k<3;k++)

+   for (k=0;k<2;k++)

    {

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

+         recur[k*N + i] = bias[k*N + i] + gru->diag_weights[k*N + i]*state[i] + input[k*N + i];

+   }

+   for (;k<3;k++)

+   {

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

          recur[k*N + i] = bias[k*N + i] + gru->diag_weights[k*N + i]*state[i];

    }

    sparse_sgemv_accum8x4(recur, gru->recurrent_weights, 3*N, N, gru->idx, state);

-   for (i=0;i<2*N;i++)

-      input[i] += recur[i];

-   compute_activation(input, input, 2*N, ACTIVATION_SIGMOID);

+   compute_activation(recur, recur, 2*N, ACTIVATION_SIGMOID);

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

-      h[i] += recur[2*N+i]*r[i];

+      h[i] = h[i]*r[i] + input[2*N+i];

    compute_activation(h, h, N, gru->activation);

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

       state[i] = z[i]*state[i] + (1-z[i])*h[i];

--- a/dnn/nnet.h

+++ b/dnn/nnet.h

@@ -91,7 +91,7 @@

   int dim;

 } EmbeddingLayer;

-void compute_activation(float *output, float *input, int N, int activation);

+void compute_activation(float *output, const float *input, int N, int activation);

 void compute_dense(const DenseLayer *layer, float *output, const float *input);

@@ -105,7 +105,7 @@

 void compute_gru3(const GRULayer *gru, float *state, const float *input);

-void compute_sparse_gru(const SparseGRULayer *gru, float *state, float *input);

+void compute_sparse_gru(const SparseGRULayer *gru, float *state, const float *input);

 void compute_conv1d(const Conv1DLayer *layer, float *output, float *mem, const float *input);

-- 

⑨