• 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 // See docs in ../ops/candidate_sampling_ops.cc.
17 
18 #define EIGEN_USE_THREADS
19 
20 #include <cfloat>
21 #include <unordered_map>
22 #include <vector>
23 
24 #include "tensorflow/core/framework/op_kernel.h"
25 #include "tensorflow/core/framework/tensor_shape.h"
26 #include "tensorflow/core/kernels/range_sampler.h"
27 #include "tensorflow/core/platform/logging.h"
28 #include "tensorflow/core/util/guarded_philox_random.h"
29 
30 namespace tensorflow {
31 
32 class BaseCandidateSamplerOp : public OpKernel {
33  public:
BaseCandidateSamplerOp(OpKernelConstruction * context)34   explicit BaseCandidateSamplerOp(OpKernelConstruction* context)
35       : OpKernel(context) {
36     OP_REQUIRES_OK(context, context->GetAttr("num_sampled", &num_sampled_));
37     OP_REQUIRES_OK(context, context->GetAttr("num_true", &num_true_));
38     OP_REQUIRES_OK(context, context->GetAttr("unique", &unique_));
39     OP_REQUIRES_OK(context, generator_.Init(context));
40   }
41 
Compute(OpKernelContext * context)42   void Compute(OpKernelContext* context) override {
43     const Tensor& true_classes = context->input(0);
44     OP_REQUIRES(context, true_classes.dims() == 2,
45                 errors::InvalidArgument("true_classes must be a matrix"));
46     const int32 batch_size = true_classes.dim_size(0);
47     OP_REQUIRES(
48         context, true_classes.dim_size(1) == num_true_,
49         errors::InvalidArgument("true_classes must have "
50                                 "num_true columns, expected: ",
51                                 true_classes.dim_size(1), " was: ", num_true_));
52     CHECK(sampler_) << "CandidateSamplerOp did not set sampler_";
53 
54     if (unique_) {
55       OP_REQUIRES(context, num_sampled_ <= sampler_->range(),
56                   errors::InvalidArgument("Sampler's range is too small."));
57     }
58 
59     // Output candidates and expected_count.
60     Tensor* out_sampled_candidates = nullptr;
61     OP_REQUIRES_OK(context,
62                    context->allocate_output(0, TensorShape({num_sampled_}),
63                                             &out_sampled_candidates));
64 
65     Tensor* out_true_expected_count = nullptr;
66     OP_REQUIRES_OK(context, context->allocate_output(
67                                 1, TensorShape({batch_size, num_true_}),
68                                 &out_true_expected_count));
69     Tensor* out_sampled_expected_count = nullptr;
70     OP_REQUIRES_OK(context,
71                    context->allocate_output(2, TensorShape({num_sampled_}),
72                                             &out_sampled_expected_count));
73 
74     gtl::ArraySlice<int64> true_candidate(true_classes.matrix<int64>().data(),
75                                           batch_size * num_true_);
76     gtl::MutableArraySlice<int64> sampled_candidate(
77         out_sampled_candidates->vec<int64>().data(), num_sampled_);
78     gtl::MutableArraySlice<float> true_expected_count(
79         out_true_expected_count->matrix<float>().data(),
80         batch_size * num_true_);
81     gtl::MutableArraySlice<float> sampled_expected_count(
82         out_sampled_expected_count->vec<float>().data(), num_sampled_);
83 
84     // Approximately conservatively estimate the number of samples required.
85     // In cases where rejection sampling is used we may occasionally use more
86     // samples than expected, which will result in reused random bits.
87     const int64 samples32 = 2048 * num_sampled_;
88 
89     // Pick sampled candidates.
90     auto local_gen = generator_.ReserveSamples32(samples32);
91     random::SimplePhilox random(&local_gen);
92     sampler_->SampleBatchGetExpectedCount(&random, unique_, sampled_candidate,
93                                           sampled_expected_count,
94                                           true_candidate, true_expected_count);
95 
96     if (sampler_->NeedsUpdates()) {
97       sampler_->Update(true_candidate);
98     }
99   }
100 
101  protected:
set_sampler(RangeSampler * sampler)102   void set_sampler(RangeSampler* sampler) { sampler_.reset(sampler); }
103 
104  private:
105   int32 num_true_;
106   int32 num_sampled_;
107   bool unique_;
108   std::unique_ptr<RangeSampler> sampler_;
109   GuardedPhiloxRandom generator_;
110 };
111 
112 template <class RangeSamplerType>
113 class SimpleCandidateSamplerOp : public BaseCandidateSamplerOp {
114  public:
SimpleCandidateSamplerOp(OpKernelConstruction * context)115   explicit SimpleCandidateSamplerOp(OpKernelConstruction* context)
116       : BaseCandidateSamplerOp(context) {
117     int64 range_max;
118     OP_REQUIRES_OK(context, context->GetAttr("range_max", &range_max));
119     set_sampler(new RangeSamplerType(range_max));
120   }
121 };
122 
123 REGISTER_KERNEL_BUILDER(Name("UniformCandidateSampler").Device(DEVICE_CPU),
124                         SimpleCandidateSamplerOp<UniformSampler>);
125 
126 REGISTER_KERNEL_BUILDER(Name("LogUniformCandidateSampler").Device(DEVICE_CPU),
127                         SimpleCandidateSamplerOp<LogUniformSampler>);
128 
129 REGISTER_KERNEL_BUILDER(
130     Name("LearnedUnigramCandidateSampler").Device(DEVICE_CPU),
131     SimpleCandidateSamplerOp<UnigramSampler>);
132 
133 REGISTER_KERNEL_BUILDER(
134     Name("ThreadUnsafeUnigramCandidateSampler").Device(DEVICE_CPU),
135     SimpleCandidateSamplerOp<ThreadUnsafeUnigramSampler>);
136 
137 class AllCandidateSamplerOp : public BaseCandidateSamplerOp {
138  public:
AllCandidateSamplerOp(OpKernelConstruction * context)139   explicit AllCandidateSamplerOp(OpKernelConstruction* context)
140       : BaseCandidateSamplerOp(context) {
141     int64 range_max;
142     OP_REQUIRES_OK(context, context->GetAttr("num_sampled", &range_max));
143     set_sampler(new AllSampler(range_max));
144   }
145 };
146 
147 REGISTER_KERNEL_BUILDER(Name("AllCandidateSampler").Device(DEVICE_CPU),
148                         AllCandidateSamplerOp);
149 
150 class FixedUnigramCandidateSamplerOp : public BaseCandidateSamplerOp {
151  public:
FixedUnigramCandidateSamplerOp(OpKernelConstruction * context)152   explicit FixedUnigramCandidateSamplerOp(OpKernelConstruction* context)
153       : BaseCandidateSamplerOp(context) {
154     int64 range_max;
155     OP_REQUIRES_OK(context, context->GetAttr("range_max", &range_max));
156     string vocab_file;
157     OP_REQUIRES_OK(context, context->GetAttr("vocab_file", &vocab_file));
158     std::vector<float> unigrams;
159     OP_REQUIRES_OK(context, context->GetAttr("unigrams", &unigrams));
160     OP_REQUIRES(
161         context, !vocab_file.empty() || !unigrams.empty(),
162         errors::InvalidArgument("Must provide either vocab_file or unigrams."));
163     OP_REQUIRES(context, vocab_file.empty() || unigrams.empty(),
164                 errors::InvalidArgument(
165                     "Must only provide one of vocab_file and unigrams."));
166     float distortion;
167     OP_REQUIRES_OK(context, context->GetAttr("distortion", &distortion));
168     int64 num_reserved_ids;
169     OP_REQUIRES_OK(context,
170                    context->GetAttr("num_reserved_ids", &num_reserved_ids));
171     int64 num_shards;
172     OP_REQUIRES_OK(context, context->GetAttr("num_shards", &num_shards));
173     int64 shard;
174     OP_REQUIRES_OK(context, context->GetAttr("shard", &shard));
175 
176     if (!vocab_file.empty()) {
177       set_sampler(new FixedUnigramSampler(context->env(), range_max, vocab_file,
178                                           distortion, num_reserved_ids,
179                                           num_shards, shard));
180     } else {
181       set_sampler(new FixedUnigramSampler(range_max, unigrams, distortion,
182                                           num_reserved_ids, num_shards, shard));
183     }
184   }
185 };
186 
187 REGISTER_KERNEL_BUILDER(Name("FixedUnigramCandidateSampler").Device(DEVICE_CPU),
188                         FixedUnigramCandidateSamplerOp);
189 
190 class ComputeAccidentalHitsOp : public OpKernel {
191  public:
ComputeAccidentalHitsOp(OpKernelConstruction * context)192   explicit ComputeAccidentalHitsOp(OpKernelConstruction* context)
193       : OpKernel(context) {
194     OP_REQUIRES_OK(context, context->GetAttr("num_true", &num_true_));
195   }
196 
Compute(OpKernelContext * context)197   void Compute(OpKernelContext* context) override {
198     const Tensor& in_true_candidates = context->input(0);
199     const TensorShape& in_true_candidates_shape = in_true_candidates.shape();
200     OP_REQUIRES(context,
201                 TensorShapeUtils::IsMatrix(in_true_candidates_shape) &&
202                     in_true_candidates_shape.dim_size(1) == num_true_,
203                 errors::InvalidArgument(
204                     "true_candidates must be a batch_size * num_true matrix"));
205 
206     const int64 batch_size = in_true_candidates_shape.dim_size(0);
207 
208     const Tensor& in_sampled_candidates = context->input(1);
209     OP_REQUIRES(context,
210                 TensorShapeUtils::IsVector(in_sampled_candidates.shape()),
211                 errors::InvalidArgument(
212                     "sampled_candidates must be a vector, which is typically "
213                     "an output from CandidateSampler"));
214 
215     std::unordered_map<int64, int> sampled_candidate_to_pos;
216     for (int64 i = 0; i < in_sampled_candidates.dim_size(0); ++i) {
217       sampled_candidate_to_pos[in_sampled_candidates.vec<int64>()(i)] = i;
218     }
219 
220     // Produce output in the same format as UnpackSparseFeatures.
221     std::vector<int> indices;
222     std::vector<int64> ids;
223     std::vector<float> weights;
224 
225     for (int64 i = 0; i < batch_size; ++i) {
226       for (int64 j = 0; j < num_true_; ++j) {
227         const int64 true_candidate = in_true_candidates.matrix<int64>()(i, j);
228         const auto look = sampled_candidate_to_pos.find(true_candidate);
229         if (look != sampled_candidate_to_pos.end()) {
230           indices.push_back(i);
231           ids.push_back(look->second);
232           weights.push_back(-FLT_MAX);
233         }
234       }
235     }
236 
237     Tensor* out_indices = nullptr;
238     OP_REQUIRES_OK(
239         context,
240         context->allocate_output(
241             0, TensorShape({static_cast<int>(indices.size())}), &out_indices));
242     Tensor* out_ids = nullptr;
243     OP_REQUIRES_OK(
244         context, context->allocate_output(
245                      1, TensorShape({static_cast<int>(ids.size())}), &out_ids));
246     Tensor* out_weights = nullptr;
247     OP_REQUIRES_OK(
248         context,
249         context->allocate_output(
250             2, TensorShape({static_cast<int>(weights.size())}), &out_weights));
251 
252     for (size_t i = 0; i < indices.size(); ++i) {
253       out_indices->vec<int32>()(i) = indices[i];
254       out_ids->vec<int64>()(i) = ids[i];
255       out_weights->vec<float>()(i) = weights[i];
256     }
257   }
258 
259  private:
260   int64 num_true_;
261 };
262 
263 REGISTER_KERNEL_BUILDER(Name("ComputeAccidentalHits").Device(DEVICE_CPU),
264                         ComputeAccidentalHitsOp);
265 
266 }  // namespace tensorflow
267