SENNA_nn.c

#include "SENNA_nn.h"
#include "SENNA_utils.h"

#include <string.h>
#include <float.h>

#ifdef USE_ATLAS_BLAS
#define USE_BLAS
#include "cblas.h"
#endif

#ifdef USE_MKL_BLAS
#define USE_BLAS
#include "mkl_cblas.h"
#endif

#ifdef USE_APPLE_BLAS
#define USE_BLAS
#include "Accelerate/Accelerate.h"
#endif

void SENNA_nn_lookup(float *dest, int dest_stride, const float *wordweights, int wordsize, int maxwordidx, const int *wordindices, int nword, int padidx, int npad)
{
  int i;

  if(padidx < 0 || padidx >= maxwordidx)
    SENNA_error("lookup: padding index out of range");

  for(i = 0; i < npad; i++)
    memcpy(dest+i*dest_stride, wordweights+padidx*wordsize, wordsize*sizeof(float));

  for(i = 0; i < nword; i++)
  {
    int wordidx = wordindices[i];
    if(wordidx < 0 || wordidx >= maxwordidx)
      SENNA_error("lookup: index out of range");

    memcpy(dest+(i+npad)*dest_stride, wordweights+wordidx*wordsize, wordsize*sizeof(float));
  }

  for(i = 0; i < npad; i++)
    memcpy(dest+(i+npad+nword)*dest_stride, wordweights+padidx*wordsize, wordsize*sizeof(float));
}

void SENNA_nn_hardtanh(float *output, float *input, int size)
{
  int i;

  for(i = 0; i < size; i++)
  {
    float z = input[i];
    if(z >= -1 && z <= 1)
      output[i] = z;
    else if(z < -1)
      output[i] = -1;
    else
      output[i] = 1;
  }
}

void SENNA_nn_linear(float *output, int output_size, float *weights, float *biases, float *input, int input_size)
{
#ifdef USE_BLAS
  if(biases)
    cblas_scopy(output_size, biases, 1, output, 1);
  cblas_sgemv(CblasColMajor, CblasTrans, input_size, output_size, 1.0, weights, input_size, input, 1, (biases ? 1.0 : 0.0), output, 1);
#else
  int i, j;
  
  for(i = 0; i < output_size; i++)
  {
    float z = (biases ? biases[i] : 0);
    float *weights_row = weights + i*input_size;
    for(j = 0; j < input_size; j++)
      z += input[j]*weights_row[j];
    output[i] = z;
  }
#endif
}

void SENNA_nn_max(float *value_, int *idx_, float *input, int input_size)
{
  float value = -FLT_MAX;
  int idx = -1;
  int i;

  for(i = 0; i < input_size; i++)
  {
    if(input[i] > value)
    {
      value = input[i];
      idx = i;
    }
  }

  if(value_)
    *value_ = value;

  if(idx_)
    *idx_ = idx;
}

void SENNA_nn_temporal_convolution(float *output, int output_frame_size, float *weights, float *biases, float *input, int input_frame_size, int n_frames, int k_w)
{
#ifdef USE_BLAS
  if(k_w == 1)
  {
    if(biases)
    {
      int t;
      for(t = 0; t < n_frames; t++)
        cblas_scopy(output_frame_size, biases, 1, output+t*output_frame_size, 1);
    }
    cblas_sgemm(CblasColMajor, CblasTrans, CblasNoTrans, output_frame_size, n_frames, input_frame_size, 1.0, weights, input_frame_size, input, input_frame_size, (biases ? 1.0 : 0.0), output, output_frame_size);
  }
  else
#endif
  {
    int t;

    for(t = 0; t < n_frames-k_w+1; t++)
      SENNA_nn_linear(output+t*output_frame_size, output_frame_size, weights, biases, input+t*input_frame_size, input_frame_size*k_w);
  }
}

void SENNA_nn_temporal_max_convolution(float *output, float *bias, float *input, int input_frame_size, int n_frames, int k_w)
{
  int i, j, k;
  int h_k_w = (k_w-1)/2;

  for(k = 0; k < n_frames; k++)
  {
    for(i = 0; i < input_frame_size; i++)
    {
      float maxval = -FLT_MAX;
      for(j = -k; j < n_frames-k; j++)
      {
        int jbias = j + h_k_w;
        int jinput = k+j;
        float z;
        
        if(jbias < 0)
          jbias = 0;
        if(jbias >= k_w)
          jbias = k_w-1;
        
        z = input[i+jinput*input_frame_size] + bias[i+jbias*input_frame_size];
        if(z > maxval)
          maxval = z;
      }
      output[i+k*input_frame_size] = maxval;
    }
  }
}

void SENNA_nn_temporal_max(float *output, float *input, int N, int T)
{
  int n, t;

  for(n = 0; n < N; n++)
  {
    float z = -FLT_MAX;
    for(t = 0; t < T; t++)
    {
      if(input[t*N+n] > z)
        z = input[t*N+n];
    }
    output[n] = z;
  }
}

#define NN_MIN(a,b) ((a) < (b) ? (a) : (b))
#define NN_MAX(a,b) ((a) > (b) ? (a) : (b))

void SENNA_nn_distance(int *dest, int idx, int max_idx, int sentence_size, int padding_size)
{
  int i;

  max_idx = (max_idx-1)/2;

  for(i = 0; i < padding_size; i++)
    dest[i] = NN_MAX( NN_MIN(i-padding_size-idx, max_idx), -max_idx) + max_idx;

  for(i = 0; i < sentence_size; i++)
    dest[i+padding_size] = NN_MAX( NN_MIN(i-idx, max_idx), -max_idx) + max_idx;

  for(i = 0; i < padding_size; i++)
    dest[i+padding_size+sentence_size] = NN_MAX( NN_MIN(i+sentence_size-idx, max_idx), -max_idx) + max_idx;
}

void SENNA_nn_viterbi(int *path, float *init, float *transition, float *emission, int N, int T)
{
  float *delta, *deltap;
  int *phi;
  int i, j, t;
     
  /* misc allocations */
  delta = SENNA_malloc(sizeof(float), N);
  deltap = SENNA_malloc(sizeof(float), N);
  phi = SENNA_malloc(sizeof(float), N*T);
  
  /* init */
  for(i = 0; i < N; i++)
    deltap[i] = init[i] + emission[i];
  
  /* recursion */
  for(t = 1; t < T; t++)
  {
    float *deltan = delta;
    for(j = 0; j < N; j++)
    {
      float maxValue = -FLT_MAX;
      int maxIndex = 0;
      for(i = 0; i < N; i++)
      {
        float z = deltap[i] + transition[i+j*N];
        if(z > maxValue)
        {
          maxValue = z;
          maxIndex = i;
        }
      }
      delta[j] = maxValue + emission[j+t*N];
      phi[j+t*N] = maxIndex;
    }
    delta  = deltap;
    deltap = deltan;
  }

  {
    float maxValue = -FLT_MAX;
    int maxIndex = 0;
    for(j = 0; j < N; j++)
    {
      if(deltap[j] > maxValue)
      {
        maxValue = deltap[j];
        maxIndex = j;
      }
    }
    path[T-1] = maxIndex;
  }

  for(t = T-2; t >= 0; t--)
    path[t] = phi[path[t+1]+(t+1)*N];
  
  SENNA_free(delta);
  SENNA_free(deltap);
  SENNA_free(phi);
}