• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #include "tensorflow/core/lib/random/distribution_sampler.h"
17 
18 #include <memory>
19 #include <vector>
20 
21 namespace tensorflow {
22 namespace random {
23 
DistributionSampler(const gtl::ArraySlice<float> & weights)24 DistributionSampler::DistributionSampler(
25     const gtl::ArraySlice<float>& weights) {
26   DCHECK(!weights.empty());
27   int n = weights.size();
28   num_ = n;
29   data_.reset(new std::pair<float, int>[n]);
30 
31   std::unique_ptr<double[]> pr(new double[n]);
32 
33   double sum = 0.0;
34   for (int i = 0; i < n; i++) {
35     sum += weights[i];
36     set_alt(i, -1);
37   }
38 
39   // These are long/short items - called high/low because of reserved keywords.
40   std::vector<int> high;
41   high.reserve(n);
42   std::vector<int> low;
43   low.reserve(n);
44 
45   // compute proportional weights
46   for (int i = 0; i < n; i++) {
47     double p = (weights[i] * n) / sum;
48     pr[i] = p;
49     if (p < 1.0) {
50       low.push_back(i);
51     } else {
52       high.push_back(i);
53     }
54   }
55 
56   // Now pair high with low.
57   while (!high.empty() && !low.empty()) {
58     int l = low.back();
59     low.pop_back();
60     int h = high.back();
61     high.pop_back();
62 
63     set_alt(l, h);
64     DCHECK_GE(pr[h], 1.0);
65     double remaining = pr[h] - (1.0 - pr[l]);
66     pr[h] = remaining;
67 
68     if (remaining < 1.0) {
69       low.push_back(h);
70     } else {
71       high.push_back(h);
72     }
73   }
74   // Transfer pr to prob with rounding errors.
75   for (int i = 0; i < n; i++) {
76     set_prob(i, pr[i]);
77   }
78   // Because of rounding errors, both high and low may have elements, that are
79   // close to 1.0 prob.
80   for (size_t i = 0; i < high.size(); i++) {
81     int idx = high[i];
82     set_prob(idx, 1.0);
83     // set alt to self to prevent rounding errors returning 0
84     set_alt(idx, idx);
85   }
86   for (size_t i = 0; i < low.size(); i++) {
87     int idx = low[i];
88     set_prob(idx, 1.0);
89     // set alt to self to prevent rounding errors returning 0
90     set_alt(idx, idx);
91   }
92 }
93 
94 }  // namespace random
95 }  // namespace tensorflow
96