001/*
002 * Copyright (c) 2021, Oracle and/or its affiliates. All rights reserved.
003 *
004 * Licensed under the Apache License, Version 2.0 (the "License");
005 * you may not use this file except in compliance with the License.
006 * You may obtain a copy of the License at
007 *
008 *     http://www.apache.org/licenses/LICENSE-2.0
009 *
010 * Unless required by applicable law or agreed to in writing, software
011 * distributed under the License is distributed on an "AS IS" BASIS,
012 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express implied.
013 * See the License for the specific language governing permissions and
014 * limitations under the License.
015 */
016
017package org.tribuo.util.onnx;
018
019import ai.onnx.proto.OnnxMl;
020import com.google.protobuf.ByteString;
021
022import java.nio.ByteBuffer;
023import java.nio.ByteOrder;
024import java.nio.DoubleBuffer;
025import java.nio.FloatBuffer;
026import java.nio.IntBuffer;
027import java.nio.LongBuffer;
028import java.util.Arrays;
029import java.util.Collections;
030import java.util.List;
031import java.util.function.Consumer;
032import java.util.stream.Collectors;
033
034/**
035 * Helper functions for building ONNX protos.
036 */
037public abstract class ONNXUtils {
038
039    /**
040     * Private constructor for abstract util class.
041     */
042    private ONNXUtils() {}
043
044    /**
045     * Builds a type proto for the specified shape and tensor type.
046     * @param shape The shape.
047     * @param type The tensor type.
048     * @return The type proto.
049     */
050    public static OnnxMl.TypeProto buildTensorTypeNode(ONNXShape shape, OnnxMl.TensorProto.DataType type) {
051        OnnxMl.TypeProto.Builder builder = OnnxMl.TypeProto.newBuilder();
052
053        OnnxMl.TypeProto.Tensor.Builder tensorBuilder = OnnxMl.TypeProto.Tensor.newBuilder();
054        tensorBuilder.setElemType(type.getNumber());
055        tensorBuilder.setShape(shape.getProto());
056        builder.setTensorType(tensorBuilder.build());
057
058        return builder.build();
059    }
060
061    /**
062     * Builds a TensorProto containing the scalar value.
063     * @param context The naming context.
064     * @param name The base name for the proto.
065     * @param value The value to store.
066     * @return A TensorProto containing the value as an int.
067     */
068    public static OnnxMl.TensorProto scalarBuilder(ONNXContext context, String name, int value) {
069        OnnxMl.TensorProto.Builder scalarBuilder = OnnxMl.TensorProto.newBuilder();
070        scalarBuilder.setName(context.generateUniqueName(name));
071        scalarBuilder.setDataType(OnnxMl.TensorProto.DataType.INT32.getNumber());
072        scalarBuilder.addInt32Data(value);
073        return scalarBuilder.build();
074    }
075
076    /**
077     * Builds a TensorProto containing the scalar value.
078     * @param context The naming context.
079     * @param name The base name for the proto.
080     * @param value The value to store.
081     * @return A TensorProto containing the value as a long.
082     */
083    public static OnnxMl.TensorProto scalarBuilder(ONNXContext context, String name, long value) {
084        OnnxMl.TensorProto.Builder scalarBuilder = OnnxMl.TensorProto.newBuilder();
085        scalarBuilder.setName(context.generateUniqueName(name));
086        scalarBuilder.setDataType(OnnxMl.TensorProto.DataType.INT64.getNumber());
087        scalarBuilder.addInt64Data(value);
088        return scalarBuilder.build();
089    }
090
091    /**
092     * Builds a TensorProto containing the scalar value.
093     * @param context The naming context.
094     * @param name The base name for the proto.
095     * @param value The value to store.
096     * @return A TensorProto containing the value as a float.
097     */
098    public static OnnxMl.TensorProto scalarBuilder(ONNXContext context, String name, float value) {
099        OnnxMl.TensorProto.Builder scalarBuilder = OnnxMl.TensorProto.newBuilder();
100        scalarBuilder.setName(context.generateUniqueName(name));
101        scalarBuilder.setDataType(OnnxMl.TensorProto.DataType.FLOAT.getNumber());
102        scalarBuilder.addFloatData(value);
103        return scalarBuilder.build();
104    }
105
106    /**
107     * Builds a TensorProto containing the scalar value.
108     * @param context The naming context.
109     * @param name The base name for the proto.
110     * @param value The value to store.
111     * @return A TensorProto containing the value as a double.
112     */
113    public static OnnxMl.TensorProto scalarBuilder(ONNXContext context, String name, double value) {
114        OnnxMl.TensorProto.Builder scalarBuilder = OnnxMl.TensorProto.newBuilder();
115        scalarBuilder.setName(context.generateUniqueName(name));
116        scalarBuilder.setDataType(OnnxMl.TensorProto.DataType.DOUBLE.getNumber());
117        scalarBuilder.addDoubleData(value);
118        return scalarBuilder.build();
119    }
120
121    /**
122     * Generic method to create float {@link OnnxMl.TensorProto} instances.
123     *
124     * @param context the naming context.
125     * @param name the base name for the proto.
126     * @param dims the dimensions of the input data.
127     * @param dataPopulator a method to populate a {@link FloatBuffer} that will be written into the TensorProto's rawData field.
128     * @return a float-typed TensorProto representation of the data.
129     */
130    public static OnnxMl.TensorProto floatTensorBuilder(ONNXContext context, String name, List<Integer> dims, Consumer<FloatBuffer> dataPopulator) {
131        int size = dims.stream().reduce((a, b) -> a * b).orElse(0);
132        ByteBuffer buffer = ByteBuffer.allocate(size * 4).order(ByteOrder.LITTLE_ENDIAN);
133        FloatBuffer floatBuffer = buffer.asFloatBuffer();
134        dataPopulator.accept(floatBuffer);
135        floatBuffer.rewind();
136        return OnnxMl.TensorProto.newBuilder()
137                .setName(context.generateUniqueName(name))
138                .setDataType(OnnxMl.TensorProto.DataType.FLOAT.getNumber())
139                .addAllDims(dims.stream().map(Integer::longValue).collect(Collectors.toList()))
140                .setRawData(ByteString.copyFrom(buffer))
141                .build();
142    }
143
144    /**
145     * Generic method to create double {@link OnnxMl.TensorProto} instances.
146     * <p>
147     * Note that ONNX fp64 support is poor compared to fp32.
148     * @param context the naming context.
149     * @param name the base name for the proto.
150     * @param dims the dimensions of the input data.
151     * @param dataPopulator a method to populate a {@link DoubleBuffer} that will be written into the TensorProto's rawData field.
152     * @return a double-typed TensorProto representation of the data.
153     */
154    public static OnnxMl.TensorProto doubleTensorBuilder(ONNXContext context, String name, List<Integer> dims, Consumer<DoubleBuffer> dataPopulator) {
155        int size = dims.stream().reduce((a, b) -> a * b).orElse(0);
156        ByteBuffer buffer = ByteBuffer.allocate(size * 8).order(ByteOrder.LITTLE_ENDIAN);
157        DoubleBuffer doubleBuffer = buffer.asDoubleBuffer();
158        dataPopulator.accept(doubleBuffer);
159        doubleBuffer.rewind();
160        return OnnxMl.TensorProto.newBuilder()
161                .setName(context.generateUniqueName(name))
162                .setDataType(OnnxMl.TensorProto.DataType.DOUBLE.getNumber())
163                .addAllDims(() -> dims.stream().map(Integer::longValue).iterator())
164                .setRawData(ByteString.copyFrom(buffer))
165                .build();
166    }
167
168    /**
169     * Builds a TensorProto containing the array.
170     * @param context The naming context.
171     * @param name The base name for the proto.
172     * @param parameters The array to store in the proto.
173     * @return A TensorProto containing the array as floats.
174     */
175    public static OnnxMl.TensorProto arrayBuilder(ONNXContext context, String name, float[] parameters) {
176        return floatTensorBuilder(context, name, Collections.singletonList(parameters.length),
177                (FloatBuffer fb) -> fb.put(parameters));
178    }
179
180    /**
181     * Builds a TensorProto containing the array.
182     * <p>
183     * Downcasts the doubles into floats as ONNX's fp64 support is poor compared to fp32.
184     * @param context The naming context.
185     * @param name The base name for the proto.
186     * @param parameters The array to store in the proto.
187     * @return A TensorProto containing the array as floats.
188     */
189    public static OnnxMl.TensorProto arrayBuilder(ONNXContext context, String name, double[] parameters) {
190        return arrayBuilder(context,name,parameters,true);
191    }
192
193    /**
194     * Builds a TensorProto containing the array.
195     * <p>
196     * Optionally downcasts the doubles into floats.
197     * @param context The naming context.
198     * @param name The base name for the proto.
199     * @param parameters The array to store in the proto.
200     * @param downcast Downcasts the doubles into floats.
201     * @return A TensorProto containing the array as either floats or doubles.
202     */
203    public static OnnxMl.TensorProto arrayBuilder(ONNXContext context, String name, double[] parameters, boolean downcast) {
204        if(downcast) {
205            return floatTensorBuilder(context, name, Collections.singletonList(parameters.length),
206                    (FloatBuffer fb) -> Arrays.stream(parameters).forEachOrdered(d -> fb.put((float)d)));
207        } else {
208            return doubleTensorBuilder(context, name, Collections.singletonList(parameters.length),
209                    (DoubleBuffer db) -> Arrays.stream(parameters).forEachOrdered(db::put));
210        }
211    }
212
213    /**
214     * Builds a TensorProto containing the array.
215     * @param context The naming context.
216     * @param name The base name for the proto.
217     * @param parameters The array to store in the proto.
218     * @return A TensorProto containing the array as ints.
219     */
220    public static OnnxMl.TensorProto arrayBuilder(ONNXContext context, String name, int[] parameters) {
221        OnnxMl.TensorProto.Builder arrBuilder = OnnxMl.TensorProto.newBuilder();
222        arrBuilder.setName(context.generateUniqueName(name));
223        arrBuilder.addDims(parameters.length);
224        int capacity = parameters.length * 4;
225        ByteBuffer buffer = ByteBuffer.allocate(capacity).order(ByteOrder.LITTLE_ENDIAN);
226        arrBuilder.setDataType(OnnxMl.TensorProto.DataType.INT32.getNumber());
227        IntBuffer intBuffer = buffer.asIntBuffer();
228        intBuffer.put(parameters);
229        intBuffer.rewind();
230        arrBuilder.setRawData(ByteString.copyFrom(buffer));
231        return arrBuilder.build();
232    }
233
234    /**
235     * Builds a TensorProto containing the array.
236     * @param context The naming context.
237     * @param name The base name for the proto.
238     * @param parameters The array to store in the proto.
239     * @return A TensorProto containing the array as longs.
240     */
241    public static OnnxMl.TensorProto arrayBuilder(ONNXContext context, String name, long[] parameters) {
242        OnnxMl.TensorProto.Builder arrBuilder = OnnxMl.TensorProto.newBuilder();
243        arrBuilder.setName(context.generateUniqueName(name));
244        arrBuilder.addDims(parameters.length);
245        int capacity = parameters.length * 8;
246        ByteBuffer buffer = ByteBuffer.allocate(capacity).order(ByteOrder.LITTLE_ENDIAN);
247        arrBuilder.setDataType(OnnxMl.TensorProto.DataType.INT64.getNumber());
248        LongBuffer longBuffer = buffer.asLongBuffer();
249        longBuffer.put(parameters);
250        longBuffer.rewind();
251        arrBuilder.setRawData(ByteString.copyFrom(buffer));
252        return arrBuilder.build();
253    }
254}