Rewind smallPos in MultinomialSampler when every bigPos removed.

* Also refine unittest to multiple iteration to prevent luckily random number. * Remove unused unittest before.
PaddlePaddle · Sep 22, 2016 · f63e641 · f63e641
1 parent 02c6cdf
commit f63e641
Show file tree

Hide file tree

Showing 2 changed files with 34 additions and 59 deletions.
diff --git a/paddle/gserver/layers/MultinomialSampler.cpp b/paddle/gserver/layers/MultinomialSampler.cpp
@@ -58,10 +58,10 @@ MultinomialSampler::MultinomialSampler(const real* prob, int size)
     }
 
     if (intervals_[bigPos].thresh < 1) {
-      bigPos = nextBigPos(0);
-    } else {  // the big interval becomes a small interval.
-      smallPos = nextSmallPos(0);
+      // the big interval becomes a small interval.
+      bigPos = nextBigPos(bigPos + 1);
     }
+    smallPos = nextSmallPos(0);
   }
 
   // Handle the inaccuracy caused by finite-precision arithmetic which

diff --git a/paddle/gserver/tests/test_MultinomialSampler.cpp b/paddle/gserver/tests/test_MultinomialSampler.cpp
@@ -17,7 +17,6 @@ limitations under the License. */
 
 #include <gtest/gtest.h>
 #include <vector>
-#include <stdlib.h>
 
 #undef PADDLE_DISABLE_TIMER
 #include "paddle/utils/Stat.h"
@@ -43,69 +42,45 @@ TEST(MultinomialSampler, gen) {
   int numGrids = 1024 * 1024;
   int size = 1024 * 4;
   default_random_engine reng;
-  uniform_int_distribution<int> rand(1, numGrids / size * 1.8);
-  vector<real> prob;
-  int sum = 0;
-  for (int i = 0; i < size; ++i) {
-    prob.push_back(rand(reng));
-    sum += prob.back();
-  }
 
-  CHECK_LE(sum, numGrids);
-  prob.back() += numGrids - sum;
+  for (size_t iter=0; iter < 256; ++iter) {
+    uniform_int_distribution<int> rand(1, numGrids / size * 1.8);
+    vector<real> prob;
+    int sum = 0;
+    for (int i = 0; i < size; ++i) {
+      prob.push_back(rand(reng));
+      sum += prob.back();
+    }
 
-  vector<int> counts(size);
-  MultinomialSamplerTester sampler(&prob[0], size);
-  counts.assign(size, 0);
-  {
-    double s = (double)size / (double)numGrids;
-    REGISTER_TIMER("MultinomialSampler");
-    for (double i = 0; i < numGrids; ++i) {
-      int ret = sampler.testGen([i, s]() { return s * i; });
-      if (ret < 0 || ret >= size) {
-        EXPECT_GE(ret, 0);
-        EXPECT_LT(ret, size);
-        break;
+    CHECK_LE(sum, numGrids);
+    prob.back() += numGrids - sum;
+
+    vector<int> counts(size);
+    MultinomialSamplerTester sampler(&prob[0], size);
+    counts.assign(size, 0);
+    {
+      double s = (double)size / (double)numGrids;
+      REGISTER_TIMER("MultinomialSampler");
+      for (double i = 0; i < numGrids; ++i) {
+        int ret = sampler.testGen([i, s]() { return s * i; });
+        if (ret < 0 || ret >= size) {
+          EXPECT_GE(ret, 0);
+          EXPECT_LT(ret, size);
+          break;
+        }
+        ++counts[ret];
       }
-      ++counts[ret];
     }
-  }
-  for (int i = 0; i < size; ++i) {
-    if (prob[i] != counts[i]) {
-      EXPECT_EQ(prob[i], counts[i]);
-      break;
+    for (int i = 0; i < size; ++i) {
+      if (prob[i] != counts[i]) {
+        EXPECT_EQ(prob[i], counts[i]);
+        LOG(INFO) << iter;
+        break;
+      }
     }
   }
 }
 
-
-TEST(MultinomialSampler, larger_then_1) {
-  std::vector<int> probs = { 1, 100, 100, 1, 1};
-  std::vector<real> fProbs;
-  std::transform(probs.begin(), probs.end(),
-                 std::back_insert_iterator<std::vector<real>>(fProbs),
-                 [](int a){
-    return (real)a;
-  });
-
-  MultinomialSamplerTester sampler(fProbs.data(), probs.size());
-
-  int sum = std::accumulate(probs.begin(), probs.end(), 0);
-
-  std::vector<int> cnt(probs.size(), 0);
-
-
-  double divides = (double)probs.size() / (double) sum;
-
-  for (int i = 0; i < sum; ++i) {
-    ++cnt[sampler.testGen([&] {return i * divides;})];
-  }
-
-  for (size_t i=0; i < probs.size(); ++i) {
-    CHECK_LE(std::abs(cnt[i] - probs[i]), 1);
-  }
-}
-
 void benchmarkRandom() {
   int n = 1024 * 1024;