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 <stdint.h> |
16 | |
17 | #include "tensorflow/lite/c/builtin_op_data.h" |
18 | #include "tensorflow/lite/c/common.h" |
19 | #include "tensorflow/lite/kernels/internal/tensor.h" |
20 | #include "tensorflow/lite/kernels/internal/tensor_ctypes.h" |
21 | #include "tensorflow/lite/kernels/kernel_util.h" |
22 | |
23 | namespace tflite { |
24 | namespace ops { |
25 | namespace builtin { |
26 | namespace one_hot { |
27 | |
28 | constexpr int kIndicesTensor = 0; |
29 | constexpr int kDepthTensor = 1; |
30 | constexpr int kOnValueTensor = 2; |
31 | constexpr int kOffValueTensor = 3; |
32 | constexpr int kOutputTensor = 0; |
33 | |
34 | // Convenience utility for destructuring a node into the appropriate tensors and |
35 | // data for the op. Note that this destructuring is quite cheap, so we can avoid |
36 | // allocating op-specific, persistent data on the heap. |
37 | struct OneHotContext { |
38 | OneHotContext(TfLiteContext* context, TfLiteNode* node) { |
39 | indices = GetInput(context, node, kIndicesTensor); |
40 | depth = GetInput(context, node, kDepthTensor); |
41 | on_value = GetInput(context, node, kOnValueTensor); |
42 | off_value = GetInput(context, node, kOffValueTensor); |
43 | output = GetOutput(context, node, kOutputTensor); |
44 | |
45 | const auto* params = |
46 | reinterpret_cast<TfLiteOneHotParams*>(node->builtin_data); |
47 | const int indices_dims = indices->dims->size; |
48 | axis = (params->axis == -1) ? indices_dims : params->axis; |
49 | output_dims = indices_dims + 1; |
50 | dtype = on_value->type; |
51 | } |
52 | |
53 | const TfLiteTensor* indices; |
54 | const TfLiteTensor* depth; |
55 | const TfLiteTensor* on_value; |
56 | const TfLiteTensor* off_value; |
57 | TfLiteTensor* output; |
58 | int axis; |
59 | int output_dims; |
60 | TfLiteType dtype; |
61 | }; |
62 | |
63 | template <typename T, typename TI> |
64 | void OneHotComputeImpl(const OneHotContext& op_context) { |
65 | // prefix_dim_size == # of elements before the axis |
66 | // depth == # of elements per axis |
67 | // suffix_dim_size == # of elements after the axis |
68 | int prefix_dim_size = 1; |
69 | for (int i = 0; i < op_context.axis; ++i) { |
70 | prefix_dim_size *= op_context.indices->dims->data[i]; |
71 | } |
72 | if (prefix_dim_size == 0) { |
73 | // If indices tensor is degenerate, return a degenerate tensor, just like |
74 | // TensorFlow does. |
75 | return; |
76 | } |
77 | const int suffix_dim_size = NumElements(op_context.indices) / prefix_dim_size; |
78 | const int depth = *op_context.depth->data.i32; |
79 | |
80 | const T on_value = *GetTensorData<T>(op_context.on_value); |
81 | const T off_value = *GetTensorData<T>(op_context.off_value); |
82 | |
83 | // View the indices as a matrix of size: |
84 | // prefix_dim_size x suffix_dim_size |
85 | // View the output as a matrix of size: |
86 | // prefix_dim_size x depth x suffix_dim_size |
87 | // Then the output is: |
88 | // output(i, j, k) == (indices(i, k) == j) ? on : off |
89 | T* output = GetTensorData<T>(op_context.output); |
90 | const TI* indices = GetTensorData<TI>(op_context.indices); |
91 | for (int i = 0; i < prefix_dim_size; ++i) { |
92 | for (int j = 0; j < depth; ++j) { |
93 | for (int k = 0; k < suffix_dim_size; ++k, ++output) { |
94 | *output = static_cast<int>(indices[i * suffix_dim_size + k]) == j |
95 | ? on_value |
96 | : off_value; |
97 | } |
98 | } |
99 | } |
100 | } |
101 | |
102 | template <typename T> |
103 | void OneHotCompute(const OneHotContext& op_context) { |
104 | if (op_context.indices->type == kTfLiteInt64) { |
105 | OneHotComputeImpl<T, int64_t>(op_context); |
106 | } else { |
107 | OneHotComputeImpl<T, int>(op_context); |
108 | } |
109 | } |
110 | |
111 | TfLiteStatus ResizeOutputTensor(TfLiteContext* context, |
112 | const OneHotContext& op_context) { |
113 | TF_LITE_ENSURE(context, *op_context.depth->data.i32 >= 0); |
114 | TfLiteIntArray* output_size = TfLiteIntArrayCreate(op_context.output_dims); |
115 | for (int i = 0; i < op_context.output_dims; ++i) { |
116 | if (i < op_context.axis) { |
117 | output_size->data[i] = op_context.indices->dims->data[i]; |
118 | } else if (i == op_context.axis) { |
119 | output_size->data[i] = *op_context.depth->data.i32; |
120 | } else { |
121 | output_size->data[i] = op_context.indices->dims->data[i - 1]; |
122 | } |
123 | } |
124 | return context->ResizeTensor(context, op_context.output, output_size); |
125 | } |
126 | |
127 | TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { |
128 | TF_LITE_ENSURE_EQ(context, NumInputs(node), 4); |
129 | TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); |
130 | |
131 | OneHotContext op_context{context, node}; |
132 | switch (op_context.dtype) { |
133 | // TODO(b/111744875): Support uint8 and quantization. |
134 | case kTfLiteFloat32: |
135 | case kTfLiteInt16: |
136 | case kTfLiteInt32: |
137 | case kTfLiteInt64: |
138 | case kTfLiteInt8: |
139 | case kTfLiteUInt8: |
140 | case kTfLiteBool: |
141 | op_context.output->type = op_context.dtype; |
142 | break; |
143 | default: |
144 | TF_LITE_KERNEL_LOG(context, "Unknown output data type: %s" , |
145 | TfLiteTypeGetName(op_context.dtype)); |
146 | return kTfLiteError; |
147 | } |
148 | |
149 | TF_LITE_ENSURE(context, op_context.indices->type == kTfLiteInt32 || |
150 | op_context.indices->type == kTfLiteInt64); |
151 | TF_LITE_ENSURE(context, op_context.axis >= 0 && |
152 | op_context.axis < op_context.output_dims); |
153 | TF_LITE_ENSURE_EQ(context, NumElements(op_context.depth), 1); |
154 | TF_LITE_ENSURE_EQ(context, NumElements(op_context.on_value), 1); |
155 | TF_LITE_ENSURE_EQ(context, NumElements(op_context.off_value), 1); |
156 | TF_LITE_ENSURE_TYPES_EQ(context, op_context.on_value->type, op_context.dtype); |
157 | TF_LITE_ENSURE_TYPES_EQ(context, op_context.off_value->type, |
158 | op_context.dtype); |
159 | |
160 | if (!IsConstantTensor(op_context.depth)) { |
161 | SetTensorToDynamic(op_context.output); |
162 | return kTfLiteOk; |
163 | } |
164 | |
165 | return ResizeOutputTensor(context, op_context); |
166 | } |
167 | |
168 | TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { |
169 | OneHotContext op_context{context, node}; |
170 | |
171 | if (IsDynamicTensor(op_context.output)) { |
172 | ResizeOutputTensor(context, op_context); |
173 | } |
174 | |
175 | switch (op_context.output->type) { |
176 | case kTfLiteFloat32: |
177 | OneHotCompute<float>(op_context); |
178 | break; |
179 | case kTfLiteInt32: |
180 | OneHotCompute<int>(op_context); |
181 | break; |
182 | case kTfLiteInt64: |
183 | OneHotCompute<int64_t>(op_context); |
184 | break; |
185 | case kTfLiteInt8: |
186 | OneHotCompute<int8_t>(op_context); |
187 | break; |
188 | case kTfLiteUInt8: |
189 | OneHotCompute<uint8_t>(op_context); |
190 | break; |
191 | case kTfLiteBool: |
192 | OneHotCompute<bool>(op_context); |
193 | break; |
194 | default: |
195 | return kTfLiteError; |
196 | } |
197 | |
198 | return kTfLiteOk; |
199 | } |
200 | |
201 | } // namespace one_hot |
202 | |
203 | TfLiteRegistration* Register_ONE_HOT() { |
204 | static TfLiteRegistration r = { |
205 | nullptr, |
206 | nullptr, |
207 | one_hot::Prepare, |
208 | one_hot::Eval, |
209 | }; |
210 | return &r; |
211 | } |
212 | |
213 | } // namespace builtin |
214 | } // namespace ops |
215 | } // namespace tflite |
216 | |