• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2017 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 "EmbeddingLookup.h"
18 
19 #include "NeuralNetworksWrapper.h"
20 #include "gmock/gmock-matchers.h"
21 #include "gtest/gtest.h"
22 
23 using ::testing::FloatNear;
24 using ::testing::Matcher;
25 
26 namespace android {
27 namespace nn {
28 namespace wrapper {
29 
30 namespace {
31 
ArrayFloatNear(const std::vector<float> & values,float max_abs_error=1.e-6)32 std::vector<Matcher<float>> ArrayFloatNear(const std::vector<float>& values,
33                                            float max_abs_error=1.e-6) {
34   std::vector<Matcher<float>> matchers;
35   matchers.reserve(values.size());
36   for (const float& v : values) {
37     matchers.emplace_back(FloatNear(v, max_abs_error));
38   }
39   return matchers;
40 }
41 
42 }  // namespace
43 
44 using ::testing::ElementsAreArray;
45 
46 #define FOR_ALL_INPUT_AND_WEIGHT_TENSORS(ACTION) \
47   ACTION(Value, float)                           \
48   ACTION(Lookup, int)
49 
50 // For all output and intermediate states
51 #define FOR_ALL_OUTPUT_TENSORS(ACTION) \
52   ACTION(Output, float)
53 
54 class EmbeddingLookupOpModel {
55  public:
EmbeddingLookupOpModel(std::initializer_list<uint32_t> index_shape,std::initializer_list<uint32_t> weight_shape)56   EmbeddingLookupOpModel(std::initializer_list<uint32_t> index_shape,
57                          std::initializer_list<uint32_t> weight_shape) {
58     auto it = weight_shape.begin();
59     rows_ = *it++;
60     columns_ = *it++;
61     features_ = *it;
62 
63     std::vector<uint32_t> inputs;
64 
65     OperandType LookupTy(Type::TENSOR_INT32, index_shape);
66     inputs.push_back(model_.addOperand(&LookupTy));
67 
68     OperandType ValueTy(Type::TENSOR_FLOAT32, weight_shape);
69     inputs.push_back(model_.addOperand(&ValueTy));
70 
71     std::vector<uint32_t> outputs;
72 
73     OperandType OutputOpndTy(Type::TENSOR_FLOAT32, weight_shape);
74     outputs.push_back(model_.addOperand(&OutputOpndTy));
75 
76     auto multiAll = [](const std::vector<uint32_t> &dims) -> uint32_t {
77         uint32_t sz = 1;
78         for (uint32_t d : dims) { sz *= d; }
79         return sz;
80     };
81 
82     Value_.insert(Value_.end(), multiAll(weight_shape), 0.f);
83     Output_.insert(Output_.end(), multiAll(weight_shape), 0.f);
84 
85     model_.addOperation(ANEURALNETWORKS_EMBEDDING_LOOKUP, inputs, outputs);
86     model_.identifyInputsAndOutputs(inputs, outputs);
87 
88     model_.finish();
89   }
90 
Invoke()91   void Invoke() {
92     ASSERT_TRUE(model_.isValid());
93 
94     Compilation compilation(&model_);
95     compilation.finish();
96     Execution execution(&compilation);
97 
98 #define SetInputOrWeight(X, T)                                               \
99   ASSERT_EQ(execution.setInput(EmbeddingLookup::k##X##Tensor, X##_.data(),   \
100                                sizeof(T) * X##_.size()),                     \
101             Result::NO_ERROR);
102 
103     FOR_ALL_INPUT_AND_WEIGHT_TENSORS(SetInputOrWeight);
104 
105 #undef SetInputOrWeight
106 
107 #define SetOutput(X, T)                                                       \
108   ASSERT_EQ(execution.setOutput(EmbeddingLookup::k##X##Tensor, X##_.data(),   \
109                                 sizeof(T) * X##_.size()),                     \
110             Result::NO_ERROR);
111 
112     FOR_ALL_OUTPUT_TENSORS(SetOutput);
113 
114 #undef SetOutput
115 
116     ASSERT_EQ(execution.compute(), Result::NO_ERROR);
117   }
118 
119 #define DefineSetter(X, T)                       \
120   void Set##X(const std::vector<T>& f) {         \
121     X##_.insert(X##_.end(), f.begin(), f.end()); \
122   }
123 
124   FOR_ALL_INPUT_AND_WEIGHT_TENSORS(DefineSetter);
125 
126 #undef DefineSetter
127 
Set3DWeightMatrix(const std::function<float (int,int,int)> & function)128   void Set3DWeightMatrix(const std::function<float(int, int, int)>& function) {
129     for (uint32_t i = 0; i < rows_; i++) {
130       for (uint32_t j = 0; j < columns_; j++) {
131         for (uint32_t k = 0; k < features_; k++) {
132           Value_[(i * columns_ + j) * features_ + k] = function(i, j, k);
133         }
134       }
135     }
136   }
137 
GetOutput() const138   const std::vector<float> &GetOutput() const { return Output_; }
139 
140  private:
141   Model model_;
142   uint32_t rows_;
143   uint32_t columns_;
144   uint32_t features_;
145 
146 #define DefineTensor(X, T) std::vector<T> X##_;
147 
148   FOR_ALL_INPUT_AND_WEIGHT_TENSORS(DefineTensor);
149   FOR_ALL_OUTPUT_TENSORS(DefineTensor);
150 
151 #undef DefineTensor
152 };
153 
154 // TODO: write more tests that exercise the details of the op, such as
155 // lookup errors and variable input shapes.
TEST(EmbeddingLookupOpTest,SimpleTest)156 TEST(EmbeddingLookupOpTest, SimpleTest) {
157   EmbeddingLookupOpModel m({3}, {3, 2, 4});
158   m.SetLookup({1, 0, 2});
159   m.Set3DWeightMatrix(
160       [](int i, int j, int k) { return i + j / 10.0f + k / 100.0f; });
161 
162   m.Invoke();
163 
164   EXPECT_THAT(m.GetOutput(),
165               ElementsAreArray(ArrayFloatNear({
166                   1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13,  // Row 1
167                   0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13,  // Row 0
168                   2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13,  // Row 2
169               })));
170 }
171 
172 }  // namespace wrapper
173 }  // namespace nn
174 }  // namespace android
175