SphinxBase  5prealpha
lda.c
1 /* -*- c-basic-offset: 4; indent-tabs-mode: nil -*- */
2 /* ====================================================================
3  * Copyright (c) 2006 Carnegie Mellon University. All rights
4  * reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  *
10  * 1. Redistributions of source code must retain the above copyright
11  * notice, this list of conditions and the following disclaimer.
12  *
13  * 2. Redistributions in binary form must reproduce the above copyright
14  * notice, this list of conditions and the following disclaimer in
15  * the documentation and/or other materials provided with the
16  * distribution.
17  *
18  * This work was supported in part by funding from the Defense Advanced
19  * Research Projects Agency and the National Science Foundation of the
20  * United States of America, and the CMU Sphinx Speech Consortium.
21  *
22  * THIS SOFTWARE IS PROVIDED BY CARNEGIE MELLON UNIVERSITY ``AS IS'' AND
23  * ANY EXPRESSED OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
24  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
25  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL CARNEGIE MELLON UNIVERSITY
26  * NOR ITS EMPLOYEES BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
27  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
28  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
29  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
30  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
31  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
32  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33  *
34  * ====================================================================
35  *
36  */
37 /*
38  * lda.c -- Read and apply LDA matrices to features.
39  *
40  * Author: David Huggins-Daines <dhuggins@cs.cmu.edu>
41  */
42 
43 #include <assert.h>
44 #include <string.h>
45 #ifdef HAVE_CONFIG_H
46 #include <config.h>
47 #endif
48 
49 #ifdef _MSC_VER
50 #pragma warning (disable: 4018)
51 #endif
52 
53 #include "sphinxbase/feat.h"
54 #include "sphinxbase/ckd_alloc.h"
55 #include "sphinxbase/bio.h"
56 #include "sphinxbase/err.h"
57 
58 #define MATRIX_FILE_VERSION "0.1"
59 
60 int32
61 feat_read_lda(feat_t *feat, const char *ldafile, int32 dim)
62 {
63  FILE *fh;
64  int32 byteswap;
65  uint32 chksum, i, m, n;
66  char **argname, **argval;
67 
68  assert(feat);
69  if (feat->n_stream != 1) {
70  E_ERROR("LDA incompatible with multi-stream features (n_stream = %d)\n",
71  feat->n_stream);
72  return -1;
73  }
74 
75  if ((fh = fopen(ldafile, "rb")) == NULL) {
76  E_ERROR_SYSTEM("Failed to open transform file '%s' for reading", ldafile);
77  return -1;
78  }
79 
80  if (bio_readhdr(fh, &argname, &argval, &byteswap) < 0) {
81  E_ERROR("Failed to read header from transform file '%s'\n", ldafile);
82  fclose(fh);
83  return -1;
84  }
85 
86  for (i = 0; argname[i]; i++) {
87  if (strcmp(argname[i], "version") == 0) {
88  if (strcmp(argval[i], MATRIX_FILE_VERSION) != 0)
89  E_WARN("%s: Version mismatch: %s, expecting %s\n",
90  ldafile, argval[i], MATRIX_FILE_VERSION);
91  }
92  }
93 
94  bio_hdrarg_free(argname, argval);
95  argname = argval = NULL;
96 
97  chksum = 0;
98 
99  if (feat->lda)
100  ckd_free_3d((void ***)feat->lda);
101 
102  {
103  /* Use a temporary variable to avoid strict-aliasing problems. */
104  void ***outlda;
105 
106  if (bio_fread_3d(&outlda, sizeof(float32),
107  &feat->n_lda, &m, &n,
108  fh, byteswap, &chksum) < 0) {
109  E_ERROR_SYSTEM("%s: bio_fread_3d(lda) failed\n", ldafile);
110  fclose(fh);
111  return -1;
112  }
113  feat->lda = (void *)outlda;
114  }
115  fclose(fh);
116 
117 #ifdef FIXED_POINT
118  /* FIXME: This is a fragile hack that depends on mfcc_t and
119  * float32 being the same size (which they are, but...) */
120  for (i = 0; i < feat->n_lda * m * n; ++i) {
121  feat->lda[0][0][i] = FLOAT2MFCC(((float *)feat->lda[0][0])[i]);
122  }
123 #endif
124 
125  /* Note that SphinxTrain stores the eigenvectors as row vectors. */
126  if (n != feat->stream_len[0])
127  E_FATAL("LDA matrix dimension %d doesn't match feature stream size %d\n", n, feat->stream_len[0]);
128 
129  /* Override dim from file if it is 0 or greater than m. */
130  if (dim > m || dim <= 0) {
131  dim = m;
132  }
133  feat->out_dim = dim;
134 
135  return 0;
136 }
137 
138 void
139 feat_lda_transform(feat_t *fcb, mfcc_t ***inout_feat, uint32 nfr)
140 {
141  mfcc_t *tmp;
142  uint32 i, j, k;
143 
144  tmp = ckd_calloc(fcb->stream_len[0], sizeof(mfcc_t));
145  for (i = 0; i < nfr; ++i) {
146  /* Do the matrix multiplication inline here since fcb->lda
147  * is transposed (eigenvectors in rows not columns). */
148  /* FIXME: In the future we ought to use the BLAS. */
149  memset(tmp, 0, sizeof(mfcc_t) * fcb->stream_len[0]);
150  for (j = 0; j < feat_dimension(fcb); ++j) {
151  for (k = 0; k < fcb->stream_len[0]; ++k) {
152  tmp[j] += MFCCMUL(inout_feat[i][0][k], fcb->lda[0][j][k]);
153  }
154  }
155  memcpy(inout_feat[i][0], tmp, fcb->stream_len[0] * sizeof(mfcc_t));
156  }
157  ckd_free(tmp);
158 }
#define E_ERROR_SYSTEM(...)
Print error text; Call perror(&quot;&quot;);.
Definition: err.h:99
#define ckd_calloc(n, sz)
Macros to simplify the use of above functions.
Definition: ckd_alloc.h:248
SPHINXBASE_EXPORT int32 feat_read_lda(feat_t *feat, const char *ldafile, int32 dim)
Add an LDA transformation to the feature module from a file.
Definition: lda.c:61
#define E_ERROR(...)
Print error message to error log.
Definition: err.h:104
#define feat_dimension(f)
Total dimensionality of feature output.
Definition: feat.h:199
Sphinx&#39;s memory allocation/deallocation routines.
Cross platform binary IO to process files in sphinx3 format.
SPHINXBASE_EXPORT int32 bio_fread_3d(void ****arr, size_t e_sz, uint32 *d1, uint32 *d2, uint32 *d3, FILE *fp, uint32 swap, uint32 *chksum)
Read a 3-d array (set of matrices)
Definition: bio.c:430
Structure for describing a speech feature type Structure for describing a speech feature type (no...
SPHINXBASE_EXPORT void ckd_free(void *ptr)
Test and free a 1-D array.
Definition: ckd_alloc.c:244
compute the dynamic coefficients from the cepstral vector.
Implementation of logging routines.
#define E_WARN(...)
Print warning message to error log.
Definition: err.h:109
SPHINXBASE_EXPORT void ckd_free_3d(void *ptr)
Free a 3-D array (ptr) previously allocated by ckd_calloc_3d.
Definition: ckd_alloc.c:297
SPHINXBASE_EXPORT void feat_lda_transform(feat_t *fcb, mfcc_t ***inout_feat, uint32 nfr)
Transform a block of features using the feature module&#39;s LDA transform.
Definition: lda.c:139
#define E_FATAL(...)
Exit with non-zero status after error message.
Definition: err.h:81
SPHINXBASE_EXPORT void bio_hdrarg_free(char **name, char **val)
Free name and value strings previously allocated and returned by bio_readhdr.
Definition: bio.c:121
SPHINXBASE_EXPORT int32 bio_readhdr(FILE *fp, char ***name, char ***val, int32 *swap)
Read binary file format header: has the following format.
Definition: bio.c:187