1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
|
// Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
#include <vespa/eval/eval/tensor_spec.h>
#include <vespa/eval/onnx/onnx_wrapper.h>
#include <vespa/vespalib/util/stringfmt.h>
#include <vespa/vespalib/gtest/gtest.h>
using namespace vespalib::eval;
using namespace vespalib::tensor;
using vespalib::make_string_short::fmt;
using TensorInfo = Onnx::TensorInfo;
using ElementType = Onnx::ElementType;
using DZ = Onnx::DimSize;
std::string get_source_dir() {
const char *dir = getenv("SOURCE_DIRECTORY");
return (dir ? dir : ".");
}
std::string source_dir = get_source_dir();
std::string simple_model = source_dir + "/simple.onnx";
std::string dynamic_model = source_dir + "/dynamic.onnx";
std::string int_types_model = source_dir + "/int_types.onnx";
std::string guess_batch_model = source_dir + "/guess_batch.onnx";
void dump_info(const char *ctx, const std::vector<TensorInfo> &info) {
fprintf(stderr, "%s:\n", ctx);
for (size_t i = 0; i < info.size(); ++i) {
fprintf(stderr, " %s[%zu]: '%s' %s\n", ctx, i, info[i].name.c_str(), info[i].type_as_string().c_str());
}
}
TEST(WirePlannerTest, known_dimension_sizes_must_match) {
Onnx::WirePlanner planner;
ValueType type1 = ValueType::from_spec("tensor<float>(a[5],b[10])");
ValueType type2 = ValueType::from_spec("tensor<float>(a[10],b[5])");
ValueType type3 = ValueType::from_spec("tensor<float>(a[5],b[5])");
TensorInfo info = TensorInfo{"info", {DZ(5),DZ(5)}, ElementType::FLOAT};
EXPECT_FALSE(planner.bind_input_type(type1, info));
EXPECT_FALSE(planner.bind_input_type(type2, info));
EXPECT_TRUE(planner.bind_input_type(type3, info));
}
TEST(WirePlannerTest, symbolic_dimension_sizes_must_match) {
Onnx::WirePlanner planner;
ValueType type1 = ValueType::from_spec("tensor<float>(a[5])");
ValueType type2 = ValueType::from_spec("tensor<float>(a[10])");
TensorInfo info = TensorInfo{"info", {DZ("dim")}, ElementType::FLOAT};
EXPECT_TRUE(planner.bind_input_type(type1, info)); // binds 'dim' to 5
EXPECT_FALSE(planner.bind_input_type(type2, info));
EXPECT_TRUE(planner.bind_input_type(type1, info));
}
TEST(WirePlannerTest, unknown_dimension_sizes_match_anything) {
Onnx::WirePlanner planner;
ValueType type1 = ValueType::from_spec("tensor<float>(a[5])");
ValueType type2 = ValueType::from_spec("tensor<float>(a[10])");
TensorInfo info = TensorInfo{"info", {DZ()}, ElementType::FLOAT};
EXPECT_TRUE(planner.bind_input_type(type1, info));
EXPECT_TRUE(planner.bind_input_type(type2, info));
}
TEST(WirePlannerTest, all_output_dimensions_must_be_bound) {
Onnx::WirePlanner planner;
ValueType type = ValueType::from_spec("tensor<float>(a[5],b[10])");
TensorInfo info1 = TensorInfo{"info", {DZ()}, ElementType::FLOAT};
TensorInfo info2 = TensorInfo{"info", {DZ("dim")}, ElementType::FLOAT};
TensorInfo info3 = TensorInfo{"info", {DZ("dim"),DZ()}, ElementType::FLOAT};
EXPECT_TRUE(planner.make_output_type(info1).is_error());
EXPECT_TRUE(planner.make_output_type(info2).is_error());
EXPECT_TRUE(planner.make_output_type(info3).is_error());
EXPECT_TRUE(planner.bind_input_type(type, info3)); // binds 'dim' to 5
EXPECT_TRUE(planner.make_output_type(info1).is_error());
EXPECT_EQ(planner.make_output_type(info2).to_spec(), "tensor<float>(d0[5])");
EXPECT_TRUE(planner.make_output_type(info3).is_error());
}
TEST(WirePlannerTest, dimensions_resolve_left_to_right) {
Onnx::WirePlanner planner;
ValueType type1 = ValueType::from_spec("tensor<float>(a[5],b[10])");
ValueType type2 = ValueType::from_spec("tensor<float>(a[10],b[10])");
ValueType type3 = ValueType::from_spec("tensor<float>(a[5],b[5])");
TensorInfo info = TensorInfo{"info", {DZ("dim"),DZ("dim")}, ElementType::FLOAT};
EXPECT_FALSE(planner.bind_input_type(type1, info)); // binds 'dim' to 5, then fails (5 != 10)
EXPECT_FALSE(planner.bind_input_type(type2, info));
EXPECT_TRUE(planner.bind_input_type(type3, info));
}
TEST(OnnxTest, simple_onnx_model_can_be_inspected)
{
Onnx model(simple_model, Onnx::Optimize::DISABLE);
dump_info("inputs", model.inputs());
dump_info("outputs", model.outputs());
ASSERT_EQ(model.inputs().size(), 3);
ASSERT_EQ(model.outputs().size(), 1);
//-------------------------------------------------------------------------
EXPECT_EQ(model.inputs()[0].name, "query_tensor");
EXPECT_EQ(model.inputs()[0].type_as_string(), "float[1][4]");
//-------------------------------------------------------------------------
EXPECT_EQ(model.inputs()[1].name, "attribute_tensor");
EXPECT_EQ(model.inputs()[1].type_as_string(), "float[4][1]");
//-------------------------------------------------------------------------
EXPECT_EQ(model.inputs()[2].name, "bias_tensor");
EXPECT_EQ(model.inputs()[2].type_as_string(), "float[1][1]");
//-------------------------------------------------------------------------
EXPECT_EQ(model.outputs()[0].name, "output");
EXPECT_EQ(model.outputs()[0].type_as_string(), "float[1][1]");
}
TEST(OnnxTest, dynamic_onnx_model_can_be_inspected)
{
Onnx model(dynamic_model, Onnx::Optimize::DISABLE);
dump_info("inputs", model.inputs());
dump_info("outputs", model.outputs());
ASSERT_EQ(model.inputs().size(), 3);
ASSERT_EQ(model.outputs().size(), 1);
//-------------------------------------------------------------------------
EXPECT_EQ(model.inputs()[0].name, "query_tensor");
EXPECT_EQ(model.inputs()[0].type_as_string(), "float[batch][4]");
//-------------------------------------------------------------------------
EXPECT_EQ(model.inputs()[1].name, "attribute_tensor");
EXPECT_EQ(model.inputs()[1].type_as_string(), "float[4][1]");
//-------------------------------------------------------------------------
EXPECT_EQ(model.inputs()[2].name, "bias_tensor");
EXPECT_EQ(model.inputs()[2].type_as_string(), "float[batch][]");
//-------------------------------------------------------------------------
EXPECT_EQ(model.outputs()[0].name, "output");
EXPECT_EQ(model.outputs()[0].type_as_string(), "float[batch][1]");
}
TEST(OnnxTest, simple_onnx_model_can_be_evaluated)
{
Onnx model(simple_model, Onnx::Optimize::ENABLE);
Onnx::WirePlanner planner;
ValueType query_type = ValueType::from_spec("tensor<float>(a[1],b[4])");
std::vector<float> query_values({1.0, 2.0, 3.0, 4.0});
DenseValueView query(query_type, TypedCells(query_values));
EXPECT_TRUE(planner.bind_input_type(query_type, model.inputs()[0]));
ValueType attribute_type = ValueType::from_spec("tensor<float>(a[4],b[1])");
std::vector<float> attribute_values({5.0, 6.0, 7.0, 8.0});
DenseValueView attribute(attribute_type, TypedCells(attribute_values));
EXPECT_TRUE(planner.bind_input_type(attribute_type, model.inputs()[1]));
ValueType bias_type = ValueType::from_spec("tensor<float>(a[1],b[1])");
std::vector<float> bias_values({9.0});
DenseValueView bias(bias_type, TypedCells(bias_values));
EXPECT_TRUE(planner.bind_input_type(bias_type, model.inputs()[2]));
EXPECT_EQ(planner.make_output_type(model.outputs()[0]).to_spec(),
"tensor<float>(d0[1],d1[1])");
Onnx::WireInfo wire_info = planner.get_wire_info(model);
Onnx::EvalContext ctx(model, wire_info);
const Value &output = ctx.get_result(0);
EXPECT_EQ(output.type().to_spec(), "tensor<float>(d0[1],d1[1])");
//-------------------------------------------------------------------------
ctx.bind_param(0, query);
ctx.bind_param(1, attribute);
ctx.bind_param(2, bias);
ctx.eval();
auto cells = output.cells();
EXPECT_EQ(cells.type, CellType::FLOAT);
EXPECT_EQ(cells.size, 1);
EXPECT_EQ(cells.typify<float>()[0], 79.0);
//-------------------------------------------------------------------------
std::vector<float> new_bias_values({10.0});
DenseValueView new_bias(bias_type, TypedCells(new_bias_values));
ctx.bind_param(2, new_bias);
ctx.eval();
EXPECT_EQ(output.cells().typify<float>()[0], 80.0);
//-------------------------------------------------------------------------
}
TEST(OnnxTest, dynamic_onnx_model_can_be_evaluated)
{
Onnx model(dynamic_model, Onnx::Optimize::ENABLE);
Onnx::WirePlanner planner;
ValueType query_type = ValueType::from_spec("tensor<float>(a[1],b[4])");
std::vector<float> query_values({1.0, 2.0, 3.0, 4.0});
DenseValueView query(query_type, TypedCells(query_values));
EXPECT_TRUE(planner.bind_input_type(query_type, model.inputs()[0]));
ValueType attribute_type = ValueType::from_spec("tensor<float>(a[4],b[1])");
std::vector<float> attribute_values({5.0, 6.0, 7.0, 8.0});
DenseValueView attribute(attribute_type, TypedCells(attribute_values));
EXPECT_TRUE(planner.bind_input_type(attribute_type, model.inputs()[1]));
ValueType bias_type = ValueType::from_spec("tensor<float>(a[1],b[2])");
std::vector<float> bias_values({4.0, 5.0});
DenseValueView bias(bias_type, TypedCells(bias_values));
EXPECT_TRUE(planner.bind_input_type(bias_type, model.inputs()[2]));
EXPECT_EQ(planner.make_output_type(model.outputs()[0]).to_spec(),
"tensor<float>(d0[1],d1[1])");
Onnx::WireInfo wire_info = planner.get_wire_info(model);
Onnx::EvalContext ctx(model, wire_info);
const Value &output = ctx.get_result(0);
EXPECT_EQ(output.type().to_spec(), "tensor<float>(d0[1],d1[1])");
//-------------------------------------------------------------------------
ctx.bind_param(0, query);
ctx.bind_param(1, attribute);
ctx.bind_param(2, bias);
ctx.eval();
auto cells = output.cells();
EXPECT_EQ(cells.type, CellType::FLOAT);
EXPECT_EQ(cells.size, 1);
EXPECT_EQ(cells.typify<float>()[0], 79.0);
//-------------------------------------------------------------------------
std::vector<float> new_bias_values({5.0,6.0});
DenseValueView new_bias(bias_type, TypedCells(new_bias_values));
ctx.bind_param(2, new_bias);
ctx.eval();
EXPECT_EQ(output.cells().typify<float>()[0], 81.0);
//-------------------------------------------------------------------------
}
TEST(OnnxTest, int_types_onnx_model_can_be_evaluated)
{
Onnx model(int_types_model, Onnx::Optimize::ENABLE);
Onnx::WirePlanner planner;
ValueType query_type = ValueType::from_spec("tensor<float>(a[1],b[4])");
std::vector<float> query_values({1.0, 2.0, 3.0, 4.0});
DenseValueView query(query_type, TypedCells(query_values));
EXPECT_TRUE(planner.bind_input_type(query_type, model.inputs()[0]));
ValueType attribute_type = ValueType::from_spec("tensor<double>(a[4],b[1])");
std::vector<double> attribute_values({5.0, 6.0, 7.0, 8.0});
DenseValueView attribute(attribute_type, TypedCells(attribute_values));
EXPECT_TRUE(planner.bind_input_type(attribute_type, model.inputs()[1]));
ValueType bias_type = ValueType::from_spec("tensor<double>(a[1],b[1])");
std::vector<double> bias_values({9.0});
DenseValueView bias(bias_type, TypedCells(bias_values));
EXPECT_TRUE(planner.bind_input_type(bias_type, model.inputs()[2]));
EXPECT_EQ(planner.make_output_type(model.outputs()[0]),
ValueType::from_spec("tensor<double>(d0[1],d1[1])"));
Onnx::WireInfo wire_info = planner.get_wire_info(model);
Onnx::EvalContext ctx(model, wire_info);
const Value &output = ctx.get_result(0);
EXPECT_EQ(output.type(), ValueType::from_spec("tensor<double>(d0[1],d1[1])"));
//-------------------------------------------------------------------------
ctx.bind_param(0, query);
ctx.bind_param(1, attribute);
ctx.bind_param(2, bias);
ctx.eval();
auto cells = output.cells();
EXPECT_EQ(cells.type, CellType::DOUBLE);
EXPECT_EQ(cells.size, 1);
EXPECT_EQ(cells.typify<double>()[0], 79.0);
//-------------------------------------------------------------------------
std::vector<double> new_bias_values({10.0});
DenseValueView new_bias(bias_type, TypedCells(new_bias_values));
ctx.bind_param(2, new_bias);
ctx.eval();
EXPECT_EQ(output.cells().typify<double>()[0], 80.0);
//-------------------------------------------------------------------------
}
TEST(OnnxTest, we_guess_batch_dimension_size_when_inference_fails) {
Onnx model(guess_batch_model, Onnx::Optimize::ENABLE);
Onnx::WirePlanner planner_3;
Onnx::WirePlanner planner_4;
ValueType in_3_type = ValueType::from_spec("tensor<float>(a[3])");
std::vector<float> in_3_values({1.0, 2.0, 3.0});
DenseValueView in_3(in_3_type, TypedCells(in_3_values));
EXPECT_TRUE(planner_3.bind_input_type(in_3_type, model.inputs()[0]));
EXPECT_TRUE(planner_3.bind_input_type(in_3_type, model.inputs()[1]));
ValueType in_4_type = ValueType::from_spec("tensor<float>(a[4])");
std::vector<float> in_4_values({1.0, 2.0, 3.0, 4.0});
DenseValueView in_4(in_4_type, TypedCells(in_4_values));
EXPECT_TRUE(planner_4.bind_input_type(in_4_type, model.inputs()[0]));
EXPECT_TRUE(planner_4.bind_input_type(in_4_type, model.inputs()[1]));
EXPECT_EQ(planner_3.make_output_type(model.outputs()[0]).to_spec(), "tensor<float>(d0[3])");
EXPECT_EQ(planner_4.make_output_type(model.outputs()[0]).to_spec(), "tensor<float>(d0[4])");
Onnx::WireInfo wire_info_3 = planner_3.get_wire_info(model);
Onnx::WireInfo wire_info_4 = planner_4.get_wire_info(model);
Onnx::EvalContext ctx_3(model, wire_info_3);
Onnx::EvalContext ctx_4(model, wire_info_4);
//-------------------------------------------------------------------------
ctx_3.bind_param(0, in_3);
ctx_3.bind_param(1, in_3);
ctx_3.eval();
ctx_4.bind_param(0, in_4);
ctx_4.bind_param(1, in_4);
ctx_4.eval();
//-------------------------------------------------------------------------
auto out_3 = TensorSpec::from_value(ctx_3.get_result(0));
auto out_4 = TensorSpec::from_value(ctx_4.get_result(0));
auto expect_3 = TensorSpec::from_expr("tensor<float>(d0[3]):[2,4,6]");
auto expect_4 = TensorSpec::from_expr("tensor<float>(d0[4]):[2,4,6,8]");
EXPECT_EQ(out_3, expect_3);
EXPECT_EQ(out_4, expect_4);
//-------------------------------------------------------------------------
}
GTEST_MAIN_RUN_ALL_TESTS()
|
