2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 #ifndef __ARM_COMPUTE_GRAPH_UTILS_H__
25 #define __ARM_COMPUTE_GRAPH_UTILS_H__
27 #include "arm_compute/core/PixelValue.h"
28 #include "arm_compute/graph/Graph.h"
29 #include "arm_compute/graph/ITensorAccessor.h"
30 #include "arm_compute/graph/Types.h"
41 /** Preprocessor interface **/
45 virtual ~IPreprocessor() = default;
46 virtual void preprocess(ITensor &tensor) = 0;
49 /** Caffe preproccessor */
50 class CaffePreproccessor : public IPreprocessor
53 /** Default Constructor
55 * @param mean Mean array in RGB ordering
56 * @param bgr Boolean specifying if the preprocessing should assume BGR format
58 CaffePreproccessor(std::array<float, 3> mean = std::array<float, 3> { { 0, 0, 0 } }, bool bgr = true);
59 void preprocess(ITensor &tensor) override;
62 std::array<float, 3> _mean;
66 /** TF preproccessor */
67 class TFPreproccessor : public IPreprocessor
70 void preprocess(ITensor &tensor) override;
73 /** PPM writer class */
74 class PPMWriter : public graph::ITensorAccessor
79 * @param[in] name PPM file name
80 * @param[in] maximum Maximum elements to access
82 PPMWriter(std::string name, unsigned int maximum = 1);
83 /** Allows instances to move constructed */
84 PPMWriter(PPMWriter &&) = default;
86 // Inherited methods overriden:
87 bool access_tensor(ITensor &tensor) override;
90 const std::string _name;
91 unsigned int _iterator;
92 unsigned int _maximum;
95 /** Dummy accessor class */
96 class DummyAccessor final : public graph::ITensorAccessor
101 * @param[in] maximum Maximum elements to write
103 DummyAccessor(unsigned int maximum = 1);
104 /** Allows instances to move constructed */
105 DummyAccessor(DummyAccessor &&) = default;
107 // Inherited methods overriden:
108 bool access_tensor(ITensor &tensor) override;
111 unsigned int _iterator;
112 unsigned int _maximum;
115 /** PPM accessor class */
116 class PPMAccessor final : public graph::ITensorAccessor
121 * @param[in] ppm_path Path to PPM file
122 * @param[in] bgr (Optional) Fill the first plane with blue channel (default = false)
123 * @param[in] preprocessor (Optional) PPM pre-processing object
125 PPMAccessor(std::string ppm_path, bool bgr = true, std::unique_ptr<IPreprocessor> preprocessor = nullptr);
126 /** Allow instances of this class to be move constructed */
127 PPMAccessor(PPMAccessor &&) = default;
129 // Inherited methods overriden:
130 bool access_tensor(ITensor &tensor) override;
133 const std::string _ppm_path;
135 std::unique_ptr<IPreprocessor> _preprocessor;
138 /** Result accessor class */
139 class TopNPredictionsAccessor final : public graph::ITensorAccessor
144 * @param[in] labels_path Path to labels text file.
145 * @param[in] top_n (Optional) Number of output classes to print
146 * @param[out] output_stream (Optional) Output stream
148 TopNPredictionsAccessor(const std::string &labels_path, size_t top_n = 5, std::ostream &output_stream = std::cout);
149 /** Allow instances of this class to be move constructed */
150 TopNPredictionsAccessor(TopNPredictionsAccessor &&) = default;
151 /** Prevent instances of this class from being copied (As this class contains pointers) */
152 TopNPredictionsAccessor(const TopNPredictionsAccessor &) = delete;
153 /** Prevent instances of this class from being copied (As this class contains pointers) */
154 TopNPredictionsAccessor &operator=(const TopNPredictionsAccessor &) = delete;
156 // Inherited methods overriden:
157 bool access_tensor(ITensor &tensor) override;
160 template <typename T>
161 void access_predictions_tensor(ITensor &tensor);
163 std::vector<std::string> _labels;
164 std::ostream &_output_stream;
168 /** Random accessor class */
169 class RandomAccessor final : public graph::ITensorAccessor
174 * @param[in] lower Lower bound value.
175 * @param[in] upper Upper bound value.
176 * @param[in] seed (Optional) Seed used to initialise the random number generator.
178 RandomAccessor(PixelValue lower, PixelValue upper, const std::random_device::result_type seed = 0);
179 /** Allows instances to move constructed */
180 RandomAccessor(RandomAccessor &&) = default;
182 // Inherited methods overriden:
183 bool access_tensor(ITensor &tensor) override;
186 template <typename T, typename D>
187 void fill(ITensor &tensor, D &&distribution);
190 std::random_device::result_type _seed;
193 /** Numpy Binary loader class*/
194 class NumPyBinLoader final : public graph::ITensorAccessor
197 /** Default Constructor
199 * @param filename Binary file name
201 NumPyBinLoader(std::string filename);
202 /** Allows instances to move constructed */
203 NumPyBinLoader(NumPyBinLoader &&) = default;
205 // Inherited methods overriden:
206 bool access_tensor(ITensor &tensor) override;
209 const std::string _filename;
212 /** Generates appropriate random accessor
214 * @param[in] lower Lower random values bound
215 * @param[in] upper Upper random values bound
216 * @param[in] seed Random generator seed
218 * @return A ramdom accessor
220 inline std::unique_ptr<graph::ITensorAccessor> get_random_accessor(PixelValue lower, PixelValue upper, const std::random_device::result_type seed = 0)
222 return arm_compute::support::cpp14::make_unique<RandomAccessor>(lower, upper, seed);
225 /** Generates appropriate weights accessor according to the specified path
227 * @note If path is empty will generate a DummyAccessor else will generate a NumPyBinLoader
229 * @param[in] path Path to the data files
230 * @param[in] data_file Relative path to the data files from path
232 * @return An appropriate tensor accessor
234 inline std::unique_ptr<graph::ITensorAccessor> get_weights_accessor(const std::string &path, const std::string &data_file)
238 return arm_compute::support::cpp14::make_unique<DummyAccessor>();
242 return arm_compute::support::cpp14::make_unique<NumPyBinLoader>(path + data_file);
246 /** Generates appropriate input accessor according to the specified ppm_path
248 * @note If ppm_path is empty will generate a DummyAccessor else will generate a PPMAccessor
250 * @param[in] ppm_path Path to PPM file
251 * @param[in] preprocessor Preproccessor object
252 * @param[in] bgr (Optional) Fill the first plane with blue channel (default = true)
254 * @return An appropriate tensor accessor
256 inline std::unique_ptr<graph::ITensorAccessor> get_input_accessor(const std::string &ppm_path,
257 std::unique_ptr<IPreprocessor> preprocessor = nullptr,
262 return arm_compute::support::cpp14::make_unique<DummyAccessor>();
266 return arm_compute::support::cpp14::make_unique<PPMAccessor>(ppm_path, bgr, std::move(preprocessor));
270 /** Utility function to return the TargetHint
272 * @param[in] target Integer value which expresses the selected target. Must be 0 for NEON, 1 for OpenCL or 2 for OpenCL with Tuner
274 * @return the TargetHint
276 inline graph::TargetHint set_target_hint(int target)
278 ARM_COMPUTE_ERROR_ON_MSG(target > 2, "Invalid target. Target must be 0 (NEON), 1 (OpenCL) or 2 (OpenCL with Tuner)");
279 if((target == 1 || target == 2) && graph::Graph::opencl_is_available())
281 // If type of target is OpenCL, check if OpenCL is available and initialize the scheduler
282 return graph::TargetHint::OPENCL;
286 return graph::TargetHint::NEON;
290 /** Generates appropriate output accessor according to the specified labels_path
292 * @note If labels_path is empty will generate a DummyAccessor else will generate a TopNPredictionsAccessor
294 * @param[in] labels_path Path to labels text file
295 * @param[in] top_n (Optional) Number of output classes to print
296 * @param[out] output_stream (Optional) Output stream
298 * @return An appropriate tensor accessor
300 inline std::unique_ptr<graph::ITensorAccessor> get_output_accessor(const std::string &labels_path, size_t top_n = 5, std::ostream &output_stream = std::cout)
302 if(labels_path.empty())
304 return arm_compute::support::cpp14::make_unique<DummyAccessor>(0);
308 return arm_compute::support::cpp14::make_unique<TopNPredictionsAccessor>(labels_path, top_n, output_stream);
311 } // namespace graph_utils
312 } // namespace arm_compute
314 #endif /* __ARM_COMPUTE_GRAPH_UTILS_H__ */