ArmNN
 22.05.01
Conv3D.cpp
Go to the documentation of this file.
1 //
2 // Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
6 #include "ParserFlatbuffersFixture.hpp"
7 #include <sstream>
8 
9 // Conv3D support was added in TF 2.5, so for backwards compatibility a hash define is needed.
10 #if defined(ARMNN_POST_TFLITE_2_3)
11 TEST_SUITE("TensorflowLiteParser_Conv3D")
12 {
13 struct SimpleConv3DFixture : public ParserFlatbuffersFixture
14 {
15  explicit SimpleConv3DFixture()
16  {
17  m_JsonString = R"(
18  {
19  "version": 3,
20  "operator_codes": [ { "builtin_code": "CONV_3D" } ],
21  "subgraphs": [ {
22  "tensors": [
23  {
24  "shape": [ 1, 2, 3, 3, 1 ],
25  "type": "UINT8",
26  "buffer": 0,
27  "name": "inputTensor",
28  "quantization": {
29  "min": [ 0.0 ],
30  "max": [ 255.0 ],
31  "scale": [ 1.0 ],
32  "zero_point": [ 0 ],
33  }
34  },
35  {
36  "shape": [ 1, 1, 1, 1, 1 ],
37  "type": "UINT8",
38  "buffer": 1,
39  "name": "outputTensor",
40  "quantization": {
41  "min": [ 0.0 ],
42  "max": [ 511.0 ],
43  "scale": [ 2.0 ],
44  "zero_point": [ 0 ],
45  }
46  },
47  {
48  "shape": [ 2, 3, 3, 1, 1 ],
49  "type": "UINT8",
50  "buffer": 2,
51  "name": "filterTensor",
52  "quantization": {
53  "min": [ 0.0 ],
54  "max": [ 255.0 ],
55  "scale": [ 1.0 ],
56  "zero_point": [ 0 ],
57  }
58  }
59  ],
60  "inputs": [ 0 ],
61  "outputs": [ 1 ],
62  "operators": [
63  {
64  "opcode_index": 0,
65  "inputs": [ 0, 2 ],
66  "outputs": [ 1 ],
67  "builtin_options_type": "Conv3DOptions",
68  "builtin_options": {
69  "padding": "VALID",
70  "stride_d": 1,
71  "stride_w": 1,
72  "stride_h": 1,
73  "fused_activation_function": "NONE"
74  },
75  "custom_options_format": "FLEXBUFFERS"
76  }
77  ],
78  } ],
79  "buffers" : [
80  { },
81  { },
82  { "data": [ 2,1,0, 6,2,1, 4,1,2,
83  1,2,1, 2,0,2, 2,1,1 ], },
84  { },
85  ]
86  }
87  )";
88  SetupSingleInputSingleOutput("inputTensor", "outputTensor");
89  }
90 };
91 
92 TEST_CASE_FIXTURE(SimpleConv3DFixture, "ParseSimpleConv3D")
93 {
94  RunTest<5, armnn::DataType::QAsymmU8>(
95  0,
96  {
97  1, 2, 3,
98  4, 5, 6,
99  7, 8, 9,
100 
101  10, 11, 12,
102  13, 14, 15,
103  16, 17, 18,
104  },
105  // Due to the output scaling we need to half the values.
106  {
107  (1*2 + 2*1 + 3*0 +
108  4*6 + 5*2 + 6*1 +
109  7*4 + 8*1 + 9*2 +
110 
111  10*1 + 11*2 + 12*1 +
112  13*2 + 14*0 + 15*2 +
113  16*2 + 17*1 + 18*1) /2
114  });
115 }
116 struct Conv3DWithBiasesFixture : public ParserFlatbuffersFixture
117 {
118  explicit Conv3DWithBiasesFixture(const std::string& inputShape,
119  const std::string& outputShape,
120  const std::string& filterShape,
121  const std::string& filterData,
122  const std::string& biasShape,
123  const std::string& biasData,
124  const std::string& strides,
125  const std::string& activation="NONE",
126  const std::string& filterScale="1.0",
127  const std::string& filterZeroPoint="0",
128  const std::string& outputScale="1.0",
129  const std::string& outputZeroPoint="0")
130  {
131  m_JsonString = R"(
132  {
133  "version": 3,
134  "operator_codes": [ { "builtin_code": "CONV_3D" } ],
135  "subgraphs": [ {
136  "tensors": [
137  {
138  "shape": )" + inputShape + R"(,
139  "type": "UINT8",
140  "buffer": 0,
141  "name": "inputTensor",
142  "quantization": {
143  "min": [ 0.0 ],
144  "max": [ 255.0 ],
145  "scale": [ 1.0 ],
146  "zero_point": [ 0 ],
147  }
148  },
149  {
150  "shape": )" + outputShape + R"(,
151  "type": "UINT8",
152  "buffer": 1,
153  "name": "outputTensor",
154  "quantization": {
155  "min": [ 0.0 ],
156  "max": [ 511.0 ],
157  "scale": [ )" + outputScale + R"( ],
158  "zero_point": [ )" + outputZeroPoint + R"( ],
159  }
160  },
161  {
162  "shape": )" + filterShape + R"( ,
163  "type": "UINT8",
164  "buffer": 2,
165  "name": "filterTensor",
166  "quantization": {
167  "min": [ 0.0 ],
168  "max": [ 255.0 ],
169  "scale": [ )" + filterScale + R"( ],
170  "zero_point": [ )" + filterZeroPoint + R"( ],
171  }
172  },
173  {
174  "shape": )" + biasShape + R"( ,
175  "type": "INT32",
176  "buffer": 3,
177  "name": "biasTensor",
178  "quantization": {
179  "min": [ 0.0 ],
180  "max": [ 255.0 ],
181  "scale": [ 1.0 ],
182  "zero_point": [ 0 ],
183  }
184  }
185  ],
186  "inputs": [ 0 ],
187  "outputs": [ 1 ],
188  "operators": [
189  {
190  "opcode_index": 0,
191  "inputs": [ 0, 2, 3 ],
192  "outputs": [ 1 ],
193  "builtin_options_type": "Conv3DOptions",
194  "builtin_options": {
195  "padding": "SAME",
196  "stride_d": )" + strides + R"(,
197  "stride_w": )" + strides + R"(,
198  "stride_h": )" + strides + R"(,
199  "fused_activation_function": )" + activation + R"(
200  },
201  "custom_options_format": "FLEXBUFFERS"
202  }
203  ],
204  } ],
205  "buffers" : [
206  { },
207  { },
208  { "data": )" + filterData + R"(, },
209  { "data": )" + biasData + R"(, },
210  ]
211  }
212  )";
213  SetupSingleInputSingleOutput("inputTensor", "outputTensor");
214  }
215 };
216 
217 struct SimpleConv3DWithBiasesFixture : Conv3DWithBiasesFixture
218 {
219  SimpleConv3DWithBiasesFixture()
220  : Conv3DWithBiasesFixture("[ 1, 2, 2, 2, 1 ]", // inputShape
221  "[ 1, 2, 2, 2, 1 ]", // outputShape
222  "[ 2, 2, 2, 1, 1 ]", // filterShape
223  "[ 2,1, 1,0, 0,1, 1,1 ]", // filterData
224  "[ 1 ]", // biasShape
225  "[ 5, 0, 0, 0 ]", // biasData
226  "1") // stride d, w and h
227  {}
228 };
229 
230 TEST_CASE_FIXTURE(SimpleConv3DWithBiasesFixture, "ParseConv3DWithBias")
231 {
232  RunTest<5,
233  armnn::DataType::QAsymmU8>(0,
234  { 1, 2, 3, 4, 5, 6, 7, 8 },
235  { 33, 21, 23, 13, 28, 25, 27, 21 });
236 }
237 
238 TEST_CASE_FIXTURE(SimpleConv3DWithBiasesFixture, "ParseDynamicConv3DWithBias")
239 {
240  RunTest<5,
241  armnn::DataType::QAsymmU8,
242  armnn::DataType::QAsymmU8>(0,
243  { { "inputTensor", { 2, 4, 6, 8, 10, 12, 14, 16 } } },
244  { { "outputTensor", { 61, 37, 41, 21, 51, 45, 49, 37 } } },
245  true);
246 }
247 
248 struct Relu6Conv3DWithBiasesFixture : Conv3DWithBiasesFixture
249 {
250  Relu6Conv3DWithBiasesFixture()
251  : Conv3DWithBiasesFixture("[ 1, 2, 2, 2, 1 ]", // inputShape
252  "[ 1, 2, 2, 2, 1 ]", // outputShape
253  "[ 2, 2, 2, 1, 1 ]", // filterShape
254  "[ 2,1, 1,0, 0,1, 1,1 ]", // filterData
255  "[ 1 ]", // biasShape
256  "[ 0, 0, 0, 0 ]", // biasData
257  "1", // stride d, w, and h
258  "RELU6", // activation
259  "1.0", // filter scale
260  "0", // filter zero point
261  "2.0", // output scale
262  "0") // output zero point
263  {}
264 };
265 
266 TEST_CASE_FIXTURE(Relu6Conv3DWithBiasesFixture, "ParseConv3DAndRelu6WithBias")
267 {
268  uint8_t relu6Min = 6 / 2; // Divide by output scale
269 
270  RunTest<5, armnn::DataType::QAsymmU8>(
271  0,
272  {
273  1, 2, 3, 4, 5, 6, 7, 8
274  },
275  // RELU6 cuts output values at +6
276  {
277  std::min(relu6Min, static_cast<uint8_t>((1*2 + 2*1 + 3*1 + 4*0 + 5*0 + 6*1 + 7*1 + 8*1)/2)),
278  std::min(relu6Min, static_cast<uint8_t>((2*2 + 0*1 + 0*1 + 0*0 + 0*0 + 0*1 + 8*1 + 0*1)/2)),
279  std::min(relu6Min, static_cast<uint8_t>((3*2 + 0*1 + 0*1 + 0*0 + 0*0 + 8*1 + 0*1 + 0*1)/2)),
280  std::min(relu6Min, static_cast<uint8_t>((4*2 + 0*1 + 0*1 + 0*0 + 8*0 + 0*1 + 0*1 + 0*1)/2)),
281  std::min(relu6Min, static_cast<uint8_t>((5*2 + 0*1 + 0*1 + 8*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)),
282  std::min(relu6Min, static_cast<uint8_t>((6*2 + 0*1 + 8*1 + 0*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)),
283  std::min(relu6Min, static_cast<uint8_t>((7*2 + 8*1 + 0*1 + 0*0 + 0*0 + 0*1 + 0*1 + 0*1)/2)),
284  std::min(relu6Min, static_cast<uint8_t>((8*2 + 0*1 + 0*1 + 0*0 + 0*0 + 0*1 + 0*1 + 0*1)/2))
285  });
286 }
287 
288 }
289 #endif
armnn::TEST_SUITE
TEST_SUITE("TestConstTensorLayerVisitor")
Definition: ConstTensorLayerVisitor.cpp:110
TEST_CASE_FIXTURE
TEST_CASE_FIXTURE(ClContextControlFixture, "CopyBetweenNeonAndGpu")
Definition: MemCopyTests.cpp:89
ParserFlatbuffersFixture.hpp
ParserFlatbuffersFixture::SetupSingleInputSingleOutput
void SetupSingleInputSingleOutput(const std::string &inputName, const std::string &outputName)
Definition: ParserFlatbuffersFixture.hpp:152