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
|
// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package ai.vespa.modelintegration.evaluator;
import ai.onnxruntime.OnnxTensor;
import ai.onnxruntime.OnnxValue;
import ai.onnxruntime.OrtEnvironment;
import ai.onnxruntime.OrtException;
import ai.onnxruntime.OrtSession;
import com.yahoo.tensor.Tensor;
import com.yahoo.tensor.TensorType;
import java.util.Collections;
import java.util.HashMap;
import java.util.Map;
/**
* Evaluates an ONNX Model by deferring to ONNX Runtime.
*
* @author lesters
*/
public class OnnxEvaluator {
private final OrtEnvironment environment;
private final OrtSession session;
public OnnxEvaluator(String modelPath) {
this(modelPath, null);
}
public OnnxEvaluator(String modelPath, OnnxEvaluatorOptions options) {
environment = OrtEnvironment.getEnvironment();
session = createSession(modelPath, environment, options, true);
}
public Tensor evaluate(Map<String, Tensor> inputs, String output) {
Map<String, OnnxTensor> onnxInputs = null;
try {
output = mapToInternalName(output);
onnxInputs = TensorConverter.toOnnxTensors(inputs, environment, session);
try (OrtSession.Result result = session.run(onnxInputs, Collections.singleton(output))) {
return TensorConverter.toVespaTensor(result.get(0));
}
} catch (OrtException e) {
throw new RuntimeException("ONNX Runtime exception", e);
} finally {
if (onnxInputs != null) {
onnxInputs.values().forEach(OnnxTensor::close);
}
}
}
public Map<String, Tensor> evaluate(Map<String, Tensor> inputs) {
Map<String, OnnxTensor> onnxInputs = null;
try {
onnxInputs = TensorConverter.toOnnxTensors(inputs, environment, session);
Map<String, Tensor> outputs = new HashMap<>();
try (OrtSession.Result result = session.run(onnxInputs)) {
for (Map.Entry<String, OnnxValue> output : result) {
String mapped = TensorConverter.asValidName(output.getKey());
outputs.put(mapped, TensorConverter.toVespaTensor(output.getValue()));
}
return outputs;
}
} catch (OrtException e) {
throw new RuntimeException("ONNX Runtime exception", e);
} finally {
if (onnxInputs != null) {
onnxInputs.values().forEach(OnnxTensor::close);
}
}
}
public Map<String, TensorType> getInputInfo() {
try {
return TensorConverter.toVespaTypes(session.getInputInfo());
} catch (OrtException e) {
throw new RuntimeException("ONNX Runtime exception", e);
}
}
public Map<String, TensorType> getOutputInfo() {
try {
return TensorConverter.toVespaTypes(session.getOutputInfo());
} catch (OrtException e) {
throw new RuntimeException("ONNX Runtime exception", e);
}
}
private static OrtSession createSession(String modelPath, OrtEnvironment environment, OnnxEvaluatorOptions options, boolean tryCuda) {
if (options == null) {
options = new OnnxEvaluatorOptions();
}
try {
return environment.createSession(modelPath, options.getOptions(tryCuda && options.requestingGpu()));
} catch (OrtException e) {
if (e.getCode() == OrtException.OrtErrorCode.ORT_NO_SUCHFILE) {
throw new IllegalArgumentException("No such file: " + modelPath);
}
if (tryCuda && isCudaError(e) && !options.gpuDeviceRequired()) {
// Failed in CUDA native code, but GPU device is optional, so we can proceed without it
return createSession(modelPath, environment, options, false);
}
if (isCudaError(e)) {
throw new IllegalArgumentException("GPU device is required, but CUDA initialization failed", e);
}
throw new RuntimeException("ONNX Runtime exception", e);
}
}
private static boolean isCudaError(OrtException e) {
return switch (e.getCode()) {
case ORT_FAIL -> e.getMessage().contains("cudaError");
case ORT_EP_FAIL -> e.getMessage().contains("Failed to find CUDA");
default -> false;
};
}
public static boolean isRuntimeAvailable() {
return isRuntimeAvailable("");
}
public static boolean isRuntimeAvailable(String modelPath) {
try {
new OnnxEvaluator(modelPath);
return true;
} catch (IllegalArgumentException e) {
if (e.getMessage().equals("No such file: ")) {
return true;
}
return false;
} catch (UnsatisfiedLinkError | RuntimeException | NoClassDefFoundError e) {
return false;
}
}
private String mapToInternalName(String outputName) throws OrtException {
var info = session.getOutputInfo();
var internalNames = info.keySet();
for (String name : internalNames) {
if (name.equals(outputName)) {
return name;
}
}
for (String name : internalNames) {
String mapped = TensorConverter.asValidName(name);
if (mapped.equals(outputName)) {
return name;
}
}
// Probably will not work, but give the correct error from session.run
return outputName;
}
}
|
