1 /*
2 * Copyright (C) 2018 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "Multinomial.h"
18
19 #include "HalInterfaces.h"
20 #include "NeuralNetworksWrapper.h"
21 #include "philox_random.h"
22 #include "simple_philox.h"
23
24 #include <gmock/gmock-matchers.h>
25 #include <gtest/gtest.h>
26 #include <unsupported/Eigen/CXX11/Tensor>
27
28 namespace android {
29 namespace nn {
30 namespace wrapper {
31
32 using ::testing::FloatNear;
33
34 constexpr int kFixedRandomSeed1 = 37;
35 constexpr int kFixedRandomSeed2 = 42;
36
37 class MultinomialOpModel {
38 public:
MultinomialOpModel(uint32_t batch_size,uint32_t class_size,uint32_t sample_size)39 MultinomialOpModel(uint32_t batch_size, uint32_t class_size, uint32_t sample_size)
40 : batch_size_(batch_size), class_size_(class_size), sample_size_(sample_size) {
41 std::vector<uint32_t> inputs;
42 OperandType logitsType(Type::TENSOR_FLOAT32, {batch_size_, class_size_});
43 inputs.push_back(model_.addOperand(&logitsType));
44 OperandType samplesType(Type::INT32, {});
45 inputs.push_back(model_.addOperand(&samplesType));
46 OperandType seedsType(Type::TENSOR_INT32, {2});
47 inputs.push_back(model_.addOperand(&seedsType));
48
49 std::vector<uint32_t> outputs;
50 OperandType outputType(Type::TENSOR_INT32, {batch_size_, sample_size_});
51 outputs.push_back(model_.addOperand(&outputType));
52
53 model_.addOperation(ANEURALNETWORKS_RANDOM_MULTINOMIAL, inputs, outputs);
54 model_.identifyInputsAndOutputs(inputs, outputs);
55 model_.finish();
56 }
57
Invoke()58 void Invoke() {
59 ASSERT_TRUE(model_.isValid());
60
61 Compilation compilation(&model_);
62 compilation.finish();
63 Execution execution(&compilation);
64
65 tensorflow::random::PhiloxRandom rng(kFixedRandomSeed1);
66 tensorflow::random::SimplePhilox srng(&rng);
67 const int sample_count = batch_size_ * class_size_;
68 for (int i = 0; i < sample_count; ++i) {
69 input_.push_back(srng.RandDouble());
70 }
71 ASSERT_EQ(execution.setInput(Multinomial::kInputTensor, input_.data(),
72 sizeof(float) * input_.size()),
73 Result::NO_ERROR);
74 ASSERT_EQ(execution.setInput(Multinomial::kSampleCountParam, &sample_size_,
75 sizeof(sample_size_)),
76 Result::NO_ERROR);
77
78 std::vector<uint32_t> seeds{kFixedRandomSeed1, kFixedRandomSeed2};
79 ASSERT_EQ(execution.setInput(Multinomial::kRandomSeedsTensor, seeds.data(),
80 sizeof(uint32_t) * seeds.size()),
81 Result::NO_ERROR);
82
83 output_.insert(output_.end(), batch_size_ * sample_size_, 0);
84 ASSERT_EQ(execution.setOutput(Multinomial::kOutputTensor, output_.data(),
85 sizeof(uint32_t) * output_.size()),
86 Result::NO_ERROR);
87
88 ASSERT_EQ(execution.compute(), Result::NO_ERROR);
89 }
90
GetInput() const91 const std::vector<float>& GetInput() const { return input_; }
GetOutput() const92 const std::vector<uint32_t>& GetOutput() const { return output_; }
93
94 private:
95 Model model_;
96
97 const uint32_t batch_size_;
98 const uint32_t class_size_;
99 const uint32_t sample_size_;
100
101 std::vector<float> input_;
102 std::vector<uint32_t> output_;
103 };
104
TEST(MultinomialOpTest,ProbabilityDeltaWithinTolerance)105 TEST(MultinomialOpTest, ProbabilityDeltaWithinTolerance) {
106 constexpr int kBatchSize = 8;
107 constexpr int kNumClasses = 10000;
108 constexpr int kNumSamples = 128;
109 constexpr float kMaxProbabilityDelta = 0.025;
110
111 MultinomialOpModel multinomial(kBatchSize, kNumClasses, kNumSamples);
112 multinomial.Invoke();
113
114 std::vector<uint32_t> output = multinomial.GetOutput();
115 std::vector<int> class_counts;
116 class_counts.resize(kNumClasses);
117 for (auto index : output) {
118 class_counts[index]++;
119 }
120
121 std::vector<float> input = multinomial.GetInput();
122 for (int b = 0; b < kBatchSize; ++b) {
123 float probability_sum = 0;
124 const int batch_index = kBatchSize * b;
125 for (int i = 0; i < kNumClasses; ++i) {
126 probability_sum += expf(input[batch_index + i]);
127 }
128 for (int i = 0; i < kNumClasses; ++i) {
129 float probability =
130 static_cast<float>(class_counts[i]) / static_cast<float>(kNumSamples);
131 float probability_expected = expf(input[batch_index + i]) / probability_sum;
132 EXPECT_THAT(probability, FloatNear(probability_expected, kMaxProbabilityDelta));
133 }
134 }
135 }
136
137 } // namespace wrapper
138 } // namespace nn
139 } // namespace android
140