Merge pull request #1225 from jennifermliu/dev

Split samples when building dictionary for COVER

lib/dictBuilder/cover.c

@@ -39,6 +39,7 @@
 *  Constants
 ***************************************/
 #define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((U32)-1) : ((U32)1 GB))
+#define DEFAULT_SPLITPOINT 1.0
 
 /*-*************************************
 *  Console display
@@ -203,6 +204,8 @@ typedef struct {
   size_t *offsets;
   const size_t *samplesSizes;
   size_t nbSamples;
+  size_t nbTrainSamples;
+  size_t nbTestSamples;
   U32 *suffix;
   size_t suffixSize;
   U32 *freqs;
@@ -222,7 +225,7 @@ static COVER_ctx_t *g_ctx = NULL;
  */
 static size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) {
   size_t sum = 0;
-  size_t i;
+  unsigned i;
   for (i = 0; i < nbSamples; ++i) {
     sum += samplesSizes[i];
   }
@@ -494,6 +497,10 @@ static int COVER_checkParameters(ZDICT_cover_params_t parameters,
   if (parameters.d > parameters.k) {
     return 0;
   }
+  /* 0 < splitPoint <= 1 */
+  if (parameters.splitPoint <= 0 || parameters.splitPoint > 1){
+    return 0;
+  }
   return 1;
 }
 
@@ -531,9 +538,14 @@ static void COVER_ctx_destroy(COVER_ctx_t *ctx) {
  */
 static int COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
                           const size_t *samplesSizes, unsigned nbSamples,
-                          unsigned d) {
+                          unsigned d, double splitPoint) {
   const BYTE *const samples = (const BYTE *)samplesBuffer;
   const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples);
+  /* Split samples into testing and training sets */
+  const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples;
+  const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples;
+  const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize;
+  const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize;
   /* Checks */
   if (totalSamplesSize < MAX(d, sizeof(U64)) ||
       totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) {
@@ -541,15 +553,29 @@ static int COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
                  (U32)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20));
     return 0;
   }
+  /* Check if there are at least 5 training samples */
+  if (nbTrainSamples < 5) {
+    DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid.", nbTrainSamples);
+    return 0;
+  }
+  /* Check if there's testing sample */
+  if (nbTestSamples < 1) {
+    DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.", nbTestSamples);
+    return 0;
+  }
   /* Zero the context */
   memset(ctx, 0, sizeof(*ctx));
-  DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbSamples,
+  DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples,
-               (U32)totalSamplesSize);
+               (U32)trainingSamplesSize);
+  DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples,
+               (U32)testSamplesSize);
   ctx->samples = samples;
   ctx->samplesSizes = samplesSizes;
   ctx->nbSamples = nbSamples;
+  ctx->nbTrainSamples = nbTrainSamples;
+  ctx->nbTestSamples = nbTestSamples;
   /* Partial suffix array */
-  ctx->suffixSize = totalSamplesSize - MAX(d, sizeof(U64)) + 1;
+  ctx->suffixSize = trainingSamplesSize - MAX(d, sizeof(U64)) + 1;
   ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
   /* Maps index to the dmerID */
   ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
@@ -563,7 +589,7 @@ static int COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
   ctx->freqs = NULL;
   ctx->d = d;
 
-  /* Fill offsets from the samlesSizes */
+  /* Fill offsets from the samplesSizes */
   {
     U32 i;
     ctx->offsets[0] = 0;
@@ -665,7 +691,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
   BYTE* const dict = (BYTE*)dictBuffer;
   COVER_ctx_t ctx;
   COVER_map_t activeDmers;
-
+  parameters.splitPoint = 1.0;
   /* Initialize global data */
   g_displayLevel = parameters.zParams.notificationLevel;
   /* Checks */
@@ -684,7 +710,7 @@ ZDICTLIB_API size_t ZDICT_trainFromBuffer_cover(
   }
   /* Initialize context and activeDmers */
   if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples,
-                      parameters.d)) {
+                      parameters.d, parameters.splitPoint)) {
     return ERROR(GENERIC);
   }
   if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
@@ -839,7 +865,7 @@ typedef struct COVER_tryParameters_data_s {
 } COVER_tryParameters_data_t;
 
 /**
- * Tries a set of parameters and upates the COVER_best_t with the results.
+ * Tries a set of parameters and updates the COVER_best_t with the results.
  * This function is thread safe if zstd is compiled with multithreaded support.
  * It takes its parameters as an *OWNING* opaque pointer to support threading.
  */
@@ -870,7 +896,7 @@ static void COVER_tryParameters(void *opaque) {
                                               dictBufferCapacity, parameters);
     dictBufferCapacity = ZDICT_finalizeDictionary(
         dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
-        ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbSamples,
+        ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbTrainSamples,
         parameters.zParams);
     if (ZDICT_isError(dictBufferCapacity)) {
       DISPLAYLEVEL(1, "Failed to finalize dictionary\n");
@@ -889,7 +915,8 @@ static void COVER_tryParameters(void *opaque) {
     /* Allocate dst with enough space to compress the maximum sized sample */
     {
       size_t maxSampleSize = 0;
-      for (i = 0; i < ctx->nbSamples; ++i) {
+      i = parameters.splitPoint < 1.0 ? ctx->nbTrainSamples : 0;
+      for (; i < ctx->nbSamples; ++i) {
         maxSampleSize = MAX(ctx->samplesSizes[i], maxSampleSize);
       }
       dstCapacity = ZSTD_compressBound(maxSampleSize);
@@ -904,7 +931,8 @@ static void COVER_tryParameters(void *opaque) {
     }
     /* Compress each sample and sum their sizes (or error) */
     totalCompressedSize = dictBufferCapacity;
-    for (i = 0; i < ctx->nbSamples; ++i) {
+    i = parameters.splitPoint < 1.0 ? ctx->nbTrainSamples : 0;
+    for (; i < ctx->nbSamples; ++i) {
       const size_t size = ZSTD_compress_usingCDict(
           cctx, dst, dstCapacity, ctx->samples + ctx->offsets[i],
           ctx->samplesSizes[i], cdict);
@@ -941,6 +969,8 @@ ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
     ZDICT_cover_params_t *parameters) {
   /* constants */
   const unsigned nbThreads = parameters->nbThreads;
+  const double splitPoint =
+      parameters->splitPoint <= 0.0 ? DEFAULT_SPLITPOINT : parameters->splitPoint;
   const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
   const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
   const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
@@ -958,6 +988,10 @@ ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
   POOL_ctx *pool = NULL;
 
   /* Checks */
+  if (splitPoint <= 0 || splitPoint > 1) {
+    LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
+    return ERROR(GENERIC);
+  }
   if (kMinK < kMaxD || kMaxK < kMinK) {
     LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
     return ERROR(GENERIC);
@@ -988,7 +1022,7 @@ ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
     /* Initialize the context for this value of d */
     COVER_ctx_t ctx;
     LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d);
-    if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d)) {
+    if (!COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint)) {
       LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n");
       COVER_best_destroy(&best);
       POOL_free(pool);
@@ -1013,6 +1047,7 @@ ZDICTLIB_API size_t ZDICT_optimizeTrainFromBuffer_cover(
       data->parameters = *parameters;
       data->parameters.k = k;
       data->parameters.d = d;
+      data->parameters.splitPoint = splitPoint;
       data->parameters.steps = kSteps;
       data->parameters.zParams.notificationLevel = g_displayLevel;
       /* Check the parameters */

lib/dictBuilder/zdict.h

@@ -86,6 +86,7 @@ typedef struct {
     unsigned d;                  /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
     unsigned steps;              /* Number of steps : Only used for optimization : 0 means default (32) : Higher means more parameters checked */
     unsigned nbThreads;          /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
+    double splitPoint;           /* Percentage of samples used for training: the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (1.0), 1.0 when all samples are used for both training and testing */
     ZDICT_params_t zParams;
 } ZDICT_cover_params_t;
 

programs/README.md

@@ -150,7 +150,7 @@ Advanced arguments :
 
 Dictionary builder :
 --train ## : create a dictionary from a training set of files
---train-cover[=k=#,d=#,steps=#] : use the cover algorithm with optional args
+--train-cover[=k=#,d=#,steps=#,split=#] : use the cover algorithm with optional args
 --train-legacy[=s=#] : use the legacy algorithm with selectivity (default: 9)
  -o file : `file` is dictionary name (default: dictionary)
 --maxdict=# : limit dictionary to specified size (default: 112640)

programs/dibio.c

@@ -323,7 +323,8 @@ int DiB_trainFromFiles(const char* dictFileName, unsigned maxDictSize,
                                                            srcBuffer, sampleSizes, fs.nbSamples,
                                                            coverParams);
             if (!ZDICT_isError(dictSize)) {
-                DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\n", coverParams->k, coverParams->d, coverParams->steps);
+                unsigned splitPercentage = (unsigned)(coverParams->splitPoint * 100);
+                DISPLAYLEVEL(2, "k=%u\nd=%u\nsteps=%u\nsplit=%u\n", coverParams->k, coverParams->d, coverParams->steps, splitPercentage);
             }
         } else {
             dictSize = ZDICT_trainFromBuffer_cover(dictBuffer, maxDictSize, srcBuffer,

programs/zstd.1

@@ -217,8 +217,9 @@ Split input files in blocks of size # (default: no split)
 A dictionary ID is a locally unique ID that a decoder can use to verify it is using the right dictionary\. By default, zstd will create a 4\-bytes random number ID\. It\'s possible to give a precise number instead\. Short numbers have an advantage : an ID < 256 will only need 1 byte in the compressed frame header, and an ID < 65536 will only need 2 bytes\. This compares favorably to 4 bytes default\. However, it\'s up to the dictionary manager to not assign twice the same ID to 2 different dictionaries\.
 .
 .TP
-\fB\-\-train\-cover[=k#,d=#,steps=#]\fR
+\fB\-\-train\-cover[=k#,d=#,steps=#,split=#]\fR
-Select parameters for the default dictionary builder algorithm named cover\. If \fId\fR is not specified, then it tries \fId\fR = 6 and \fId\fR = 8\. If \fIk\fR is not specified, then it tries \fIsteps\fR values in the range [50, 2000]\. If \fIsteps\fR is not specified, then the default value of 40 is used\. Requires that \fId\fR <= \fIk\fR\.
+Select parameters for the default dictionary builder algorithm named cover\. If \fId\fR is not specified, then it tries \fId\fR = 6 and \fId\fR = 8\. If \fIk\fR is not specified, then it tries \fIsteps\fR values in the range [50, 2000]\. If \fIsteps\fR is not specified, then the default value of 40 is used\. If \fIsplit\fR is not specified or \fIsplit\fR <= 0, then the default value of 100 is used\. If \fIsplit\fR is 100, all input samples are used for both training and testing
+to find optimal _d_ and _k_ to build dictionary.Requires that \fId\fR <= \fIk\fR\.
 .
 .IP
 Selects segments of size \fIk\fR with highest score to put in the dictionary\. The score of a segment is computed by the sum of the frequencies of all the subsegments of size \fId\fR\. Generally \fId\fR should be in the range [6, 8], occasionally up to 16, but the algorithm will run faster with d <= \fI8\fR\. Good values for \fIk\fR vary widely based on the input data, but a safe range is [2 * \fId\fR, 2000]\. Supports multithreading if \fBzstd\fR is compiled with threading support\.

programs/zstd.1.md

@@ -223,11 +223,12 @@ Compression of small files similar to the sample set will be greatly improved.
     This compares favorably to 4 bytes default.
     However, it's up to the dictionary manager to not assign twice the same ID to
     2 different dictionaries.
-* `--train-cover[=k#,d=#,steps=#]`:
+* `--train-cover[=k#,d=#,steps=#,split=#]`:
     Select parameters for the default dictionary builder algorithm named cover.
     If _d_ is not specified, then it tries _d_ = 6 and _d_ = 8.
     If _k_ is not specified, then it tries _steps_ values in the range [50, 2000].
     If _steps_ is not specified, then the default value of 40 is used.
+    If _split_ is not specified or split <= 0, then the default value of 100 is used.
     Requires that _d_ <= _k_.
 
     Selects segments of size _k_ with highest score to put in the dictionary.
@@ -237,6 +238,8 @@ Compression of small files similar to the sample set will be greatly improved.
     algorithm will run faster with d <= _8_.
     Good values for _k_ vary widely based on the input data, but a safe range is
     [2 * _d_, 2000].
+    If _split_ is 100, all input samples are used for both training and testing
+    to find optimal _d_ and _k_ to build dictionary.
     Supports multithreading if `zstd` is compiled with threading support.
 
     Examples:
@@ -249,6 +252,8 @@ Compression of small files similar to the sample set will be greatly improved.
 
     `zstd --train-cover=k=50 FILEs`
 
+    `zstd --train-cover=k=50,split=60 FILEs`
+
 * `--train-legacy[=selectivity=#]`:
     Use legacy dictionary builder algorithm with the given dictionary
     _selectivity_ (default: 9).

programs/zstdcli.c

@@ -84,6 +84,7 @@ static U32 g_ldmMinMatch = 0;
 static U32 g_ldmHashEveryLog = LDM_PARAM_DEFAULT;
 static U32 g_ldmBucketSizeLog = LDM_PARAM_DEFAULT;
 
+#define DEFAULT_SPLITPOINT 1.0
 
 /*-************************************
 *  Display Macros
@@ -170,7 +171,7 @@ static int usage_advanced(const char* programName)
     DISPLAY( "\n");
     DISPLAY( "Dictionary builder : \n");
     DISPLAY( "--train ## : create a dictionary from a training set of files \n");
-    DISPLAY( "--train-cover[=k=#,d=#,steps=#] : use the cover algorithm with optional args\n");
+    DISPLAY( "--train-cover[=k=#,d=#,steps=#,split=#] : use the cover algorithm with optional args\n");
     DISPLAY( "--train-legacy[=s=#] : use the legacy algorithm with selectivity (default: %u)\n", g_defaultSelectivityLevel);
     DISPLAY( " -o file : `file` is dictionary name (default: %s) \n", g_defaultDictName);
     DISPLAY( "--maxdict=# : limit dictionary to specified size (default: %u) \n", g_defaultMaxDictSize);
@@ -282,10 +283,15 @@ static unsigned parseCoverParameters(const char* stringPtr, ZDICT_cover_params_t
         if (longCommandWArg(&stringPtr, "k=")) { params->k = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
         if (longCommandWArg(&stringPtr, "d=")) { params->d = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
         if (longCommandWArg(&stringPtr, "steps=")) { params->steps = readU32FromChar(&stringPtr); if (stringPtr[0]==',') { stringPtr++; continue; } else break; }
+        if (longCommandWArg(&stringPtr, "split=")) {
+          unsigned splitPercentage = readU32FromChar(&stringPtr);
+          params->splitPoint = (double)splitPercentage / 100.0;
+          if (stringPtr[0]==',') { stringPtr++; continue; } else break;
+        }
         return 0;
     }
     if (stringPtr[0] != 0) return 0;
-    DISPLAYLEVEL(4, "cover: k=%u\nd=%u\nsteps=%u\n", params->k, params->d, params->steps);
+    DISPLAYLEVEL(4, "cover: k=%u\nd=%u\nsteps=%u\nsplit=%u\n", params->k, params->d, params->steps, (unsigned)(params->splitPoint * 100));
     return 1;
 }
 
@@ -310,6 +316,7 @@ static ZDICT_cover_params_t defaultCoverParams(void)
     memset(&params, 0, sizeof(params));
     params.d = 8;
     params.steps = 4;
+    params.splitPoint = DEFAULT_SPLITPOINT;
     return params;
 }
 #endif

tests/playTests.sh

@@ -412,7 +412,7 @@ $ECHO "\n===>  cover dictionary builder : advanced options "
 TESTFILE=../programs/zstdcli.c
 ./datagen > tmpDict
 $ECHO "- Create first dictionary"
-$ZSTD --train-cover=k=46,d=8 *.c ../programs/*.c -o tmpDict
+$ZSTD --train-cover=k=46,d=8,split=80 *.c ../programs/*.c -o tmpDict
 cp $TESTFILE tmp
 $ZSTD -f tmp -D tmpDict
 $ZSTD -d tmp.zst -D tmpDict -fo result
@@ -426,6 +426,11 @@ cmp tmpDict tmpDict1 && die "dictionaries should have different ID !"
 $ECHO "- Create dictionary with size limit"
 $ZSTD --train-cover=steps=8 *.c ../programs/*.c -o tmpDict2 --maxdict=4K
 rm tmp*
+$ECHO "- Compare size of dictionary from 90% training samples with 80% training samples"
+$ZSTD --train-cover=split=90 -r *.c ../programs/*.c
+$ZSTD --train-cover=split=80 -r *.c ../programs/*.c
+$ECHO "- Create dictionary using all samples for both training and testing"
+$ZSTD --train-cover=split=100 -r *.c ../programs/*.c
 
 $ECHO "\n===>  legacy dictionary builder "