• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 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 #include "tensorflow/lite/kernels/test_util.h"
16 
17 #include "tensorflow/lite/version.h"
18 #include "tensorflow/core/platform/logging.h"
19 
20 namespace tflite {
21 
22 using ::testing::FloatNear;
23 using ::testing::Matcher;
24 
ArrayFloatNear(const std::vector<float> & values,float max_abs_error)25 std::vector<Matcher<float>> ArrayFloatNear(const std::vector<float>& values,
26                                            float max_abs_error) {
27   std::vector<Matcher<float>> matchers;
28   matchers.reserve(values.size());
29   for (const float& v : values) {
30     matchers.emplace_back(FloatNear(v, max_abs_error));
31   }
32   return matchers;
33 }
34 
ArrayComplex64Near(const std::vector<std::complex<float>> & values,float max_abs_error)35 std::vector<Matcher<std::complex<float>>> ArrayComplex64Near(
36     const std::vector<std::complex<float>>& values, float max_abs_error) {
37   std::vector<Matcher<std::complex<float>>> matchers;
38   matchers.reserve(values.size());
39   for (const std::complex<float>& v : values) {
40     matchers.emplace_back(
41         AllOf(::testing::Property(&std::complex<float>::real,
42                                   FloatNear(v.real(), max_abs_error)),
43               ::testing::Property(&std::complex<float>::imag,
44                                   FloatNear(v.imag(), max_abs_error))));
45   }
46   return matchers;
47 }
48 
AddInput(const TensorData & t,bool is_variable)49 int SingleOpModel::AddInput(const TensorData& t, bool is_variable) {
50   int id = 0;
51   if (t.per_channel_quantization) {
52     id = AddTensorPerChannelQuant(t);
53   } else {
54     id = AddTensor<float>(t, {}, is_variable);
55   }
56   inputs_.push_back(id);
57   return id;
58 }
59 
AddNullInput()60 int SingleOpModel::AddNullInput() {
61   int id = kOptionalTensor;
62   inputs_.push_back(id);
63   return id;
64 }
65 
AddOutput(const TensorData & t)66 int SingleOpModel::AddOutput(const TensorData& t) {
67   int id = AddTensor<float>(t, {});
68   outputs_.push_back(id);
69   return id;
70 }
71 
SetBuiltinOp(BuiltinOperator type,BuiltinOptions builtin_options_type,flatbuffers::Offset<void> builtin_options)72 void SingleOpModel::SetBuiltinOp(BuiltinOperator type,
73                                  BuiltinOptions builtin_options_type,
74                                  flatbuffers::Offset<void> builtin_options) {
75   opcodes_.push_back(CreateOperatorCode(builder_, type, 0));
76   operators_.push_back(CreateOperator(
77       builder_, /*opcode_index=*/0, builder_.CreateVector<int32_t>(inputs_),
78       builder_.CreateVector<int32_t>(outputs_), builtin_options_type,
79       builtin_options,
80       /*custom_options=*/0, CustomOptionsFormat_FLEXBUFFERS));
81 }
82 
SetCustomOp(const string & name,const std::vector<uint8_t> & custom_option,const std::function<TfLiteRegistration * ()> & registration)83 void SingleOpModel::SetCustomOp(
84     const string& name, const std::vector<uint8_t>& custom_option,
85     const std::function<TfLiteRegistration*()>& registration) {
86   custom_registrations_[name] = registration;
87   opcodes_.push_back(
88       CreateOperatorCodeDirect(builder_, BuiltinOperator_CUSTOM, name.data()));
89   operators_.push_back(CreateOperator(
90       builder_, /*opcode_index=*/0, builder_.CreateVector<int32_t>(inputs_),
91       builder_.CreateVector<int32_t>(outputs_), BuiltinOptions_NONE, 0,
92       builder_.CreateVector<uint8_t>(custom_option),
93       CustomOptionsFormat_FLEXBUFFERS));
94 }
95 
BuildInterpreter(std::vector<std::vector<int>> input_shapes,bool allow_fp32_relax_to_fp16)96 void SingleOpModel::BuildInterpreter(std::vector<std::vector<int>> input_shapes,
97                                      bool allow_fp32_relax_to_fp16) {
98   auto opcodes = builder_.CreateVector(opcodes_);
99   auto operators = builder_.CreateVector(operators_);
100   auto tensors = builder_.CreateVector(tensors_);
101   auto inputs = builder_.CreateVector<int32_t>(inputs_);
102   auto outputs = builder_.CreateVector<int32_t>(outputs_);
103   // Create a single subgraph
104   std::vector<flatbuffers::Offset<SubGraph>> subgraphs;
105   auto subgraph = CreateSubGraph(builder_, tensors, inputs, outputs, operators);
106   subgraphs.push_back(subgraph);
107   auto subgraphs_flatbuffer = builder_.CreateVector(subgraphs);
108 
109   auto buffers = builder_.CreateVector(buffers_);
110   auto description = builder_.CreateString("programmatic model");
111   builder_.Finish(CreateModel(builder_, TFLITE_SCHEMA_VERSION, opcodes,
112                               subgraphs_flatbuffer, description, buffers));
113 
114   auto* model = GetModel(builder_.GetBufferPointer());
115 
116   if (!resolver_) {
117     auto resolver = new ops::builtin::BuiltinOpResolver();
118     for (const auto& reg : custom_registrations_) {
119       resolver->AddCustom(reg.first.data(), reg.second());
120     }
121     resolver_ = std::unique_ptr<OpResolver>(resolver);
122   }
123   CHECK(InterpreterBuilder(model, *resolver_)(&interpreter_) == kTfLiteOk);
124 
125   CHECK(interpreter_ != nullptr);
126 
127   for (size_t i = 0; i < input_shapes.size(); ++i) {
128     const int input_idx = interpreter_->inputs()[i];
129     if (input_idx == kOptionalTensor) continue;
130     const auto& shape = input_shapes[i];
131     if (shape.empty()) continue;
132     CHECK(interpreter_->ResizeInputTensor(input_idx, shape) == kTfLiteOk);
133   }
134 
135   interpreter_->SetAllowFp16PrecisionForFp32(allow_fp32_relax_to_fp16);
136 
137   CHECK(interpreter_->AllocateTensors() == kTfLiteOk)
138       << "Cannot allocate tensors";
139   interpreter_->ResetVariableTensors();
140 
141   // Modify delegate with function.
142   if (apply_delegate_fn_) {
143     apply_delegate_fn_(interpreter_.get());
144   }
145 }
146 
Invoke()147 void SingleOpModel::Invoke() { CHECK(interpreter_->Invoke() == kTfLiteOk); }
148 
GetTensorSize(int index) const149 int32_t SingleOpModel::GetTensorSize(int index) const {
150   TfLiteTensor* t = interpreter_->tensor(index);
151   CHECK(t);
152   int total_size = 1;
153   for (int i = 0; i < t->dims->size; ++i) {
154     total_size *= t->dims->data[i];
155   }
156   return total_size;
157 }
158 
159 template <>
ExtractVector(int index)160 std::vector<string> SingleOpModel::ExtractVector(int index) {
161   TfLiteTensor* tensor_ptr = interpreter_->tensor(index);
162   CHECK(tensor_ptr != nullptr);
163   const int num_strings = GetStringCount(tensor_ptr);
164   std::vector<string> result;
165   result.reserve(num_strings);
166   for (int i = 0; i < num_strings; ++i) {
167     const auto str = GetString(tensor_ptr, i);
168     result.emplace_back(str.str, str.len);
169   }
170   return result;
171 }
172 }  // namespace tflite
173