273 lines
13 KiB
Java
273 lines
13 KiB
Java
|
/*******************************************************************************
|
||
|
|
* Copyright (c) 2015-2018 Skymind, Inc.
|
||
|
|
*
|
||
|
|
* This program and the accompanying materials are made available under the
|
||
|
|
* terms of the Apache License, Version 2.0 which is available at
|
||
|
|
* https://www.apache.org/licenses/LICENSE-2.0.
|
||
|
|
*
|
||
|
|
* Unless required by applicable law or agreed to in writing, software
|
||
|
|
* distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
|
||
|
|
* WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
|
||
|
|
* License for the specific language governing permissions and limitations
|
||
|
|
* under the License.
|
||
|
|
*
|
||
|
|
* SPDX-License-Identifier: Apache-2.0
|
||
|
|
******************************************************************************/
|
||
|
|
|
||
|
|
package org.deeplearning4j.nn.misc;
|
||
|
|
|
||
|
|
import org.apache.commons.io.FileUtils;
|
||
|
|
import org.deeplearning4j.BaseDL4JTest;
|
||
|
|
import org.deeplearning4j.nn.conf.CacheMode;
|
||
|
|
import org.deeplearning4j.nn.conf.ComputationGraphConfiguration;
|
||
|
|
import org.deeplearning4j.nn.conf.MultiLayerConfiguration;
|
||
|
|
import org.deeplearning4j.nn.conf.NeuralNetConfiguration;
|
||
|
|
import org.deeplearning4j.nn.conf.graph.*;
|
||
|
|
import org.deeplearning4j.nn.conf.graph.rnn.DuplicateToTimeSeriesVertex;
|
||
|
|
import org.deeplearning4j.nn.conf.graph.rnn.LastTimeStepVertex;
|
||
|
|
import org.deeplearning4j.nn.conf.inputs.InputType;
|
||
|
|
import org.deeplearning4j.nn.conf.layers.*;
|
||
|
|
import org.deeplearning4j.nn.conf.memory.MemoryReport;
|
||
|
|
import org.deeplearning4j.nn.conf.memory.MemoryType;
|
||
|
|
import org.deeplearning4j.nn.conf.memory.MemoryUseMode;
|
||
|
|
import org.deeplearning4j.nn.conf.preprocessor.FeedForwardToCnnPreProcessor;
|
||
|
|
import org.junit.Test;
|
||
|
|
import org.nd4j.linalg.activations.Activation;
|
||
|
|
import org.nd4j.linalg.api.buffer.DataBuffer;
|
||
|
|
import org.nd4j.linalg.api.buffer.DataType;
|
||
|
|
import org.nd4j.linalg.api.ndarray.INDArray;
|
||
|
|
import org.nd4j.linalg.factory.Nd4j;
|
||
|
|
import org.nd4j.linalg.io.ClassPathResource;
|
||
|
|
import org.nd4j.linalg.primitives.Pair;
|
||
|
|
|
||
|
|
import java.io.File;
|
||
|
|
import java.nio.charset.Charset;
|
||
|
|
import java.util.ArrayList;
|
||
|
|
import java.util.List;
|
||
|
|
|
||
|
|
import static org.junit.Assert.assertArrayEquals;
|
||
|
|
import static org.junit.Assert.assertEquals;
|
||
|
|
|
||
|
|
/**
|
||
|
|
* Created by Alex on 14/07/2017.
|
||
|
|
*/
|
||
|
|
public class TestMemoryReports extends BaseDL4JTest {
|
||
|
|
|
||
|
|
public static List<Pair<? extends Layer, InputType>> getTestLayers() {
|
||
|
|
List<Pair<? extends Layer, InputType>> l = new ArrayList<>();
|
||
|
|
l.add(new Pair<>(new ActivationLayer.Builder().activation(Activation.TANH).build(), InputType.feedForward(20)));
|
||
|
|
l.add(new Pair<>(new DenseLayer.Builder().nIn(20).nOut(20).build(), InputType.feedForward(20)));
|
||
|
|
l.add(new Pair<>(new DropoutLayer.Builder().nIn(20).nOut(20).build(), InputType.feedForward(20)));
|
||
|
|
l.add(new Pair<>(new EmbeddingLayer.Builder().nIn(1).nOut(20).build(), InputType.feedForward(20)));
|
||
|
|
l.add(new Pair<>(new OutputLayer.Builder().nIn(20).nOut(20).build(), InputType.feedForward(20)));
|
||
|
|
l.add(new Pair<>(new LossLayer.Builder().build(), InputType.feedForward(20)));
|
||
|
|
|
||
|
|
//RNN layers:
|
||
|
|
l.add(new Pair<>(new GravesLSTM.Builder().nIn(20).nOut(20).build(), InputType.recurrent(20, 30)));
|
||
|
|
l.add(new Pair<>(new LSTM.Builder().nIn(20).nOut(20).build(), InputType.recurrent(20, 30)));
|
||
|
|
l.add(new Pair<>(new GravesBidirectionalLSTM.Builder().nIn(20).nOut(20).build(), InputType.recurrent(20, 30)));
|
||
|
|
l.add(new Pair<>(new RnnOutputLayer.Builder().nIn(20).nOut(20).build(), InputType.recurrent(20, 30)));
|
||
|
|
|
||
|
|
return l;
|
||
|
|
}
|
||
|
|
|
||
|
|
public static List<Pair<? extends GraphVertex, InputType[]>> getTestVertices() {
|
||
|
|
|
||
|
|
List<Pair<? extends GraphVertex, InputType[]>> out = new ArrayList<>();
|
||
|
|
out.add(new Pair<>(new ElementWiseVertex(ElementWiseVertex.Op.Add),
|
||
|
|
new InputType[] {InputType.feedForward(10), InputType.feedForward(10)}));
|
||
|
|
out.add(new Pair<>(new ElementWiseVertex(ElementWiseVertex.Op.Add),
|
||
|
|
new InputType[] {InputType.recurrent(10, 10), InputType.recurrent(10, 10)}));
|
||
|
|
out.add(new Pair<>(new L2NormalizeVertex(), new InputType[] {InputType.feedForward(10)}));
|
||
|
|
out.add(new Pair<>(new L2Vertex(), new InputType[] {InputType.recurrent(10, 10), InputType.recurrent(10, 10)}));
|
||
|
|
out.add(new Pair<>(new MergeVertex(),
|
||
|
|
new InputType[] {InputType.recurrent(10, 10), InputType.recurrent(10, 10)}));
|
||
|
|
out.add(new Pair<>(new PreprocessorVertex(new FeedForwardToCnnPreProcessor(1, 10, 1)),
|
||
|
|
new InputType[] {InputType.convolutional(1, 10, 1)}));
|
||
|
|
out.add(new Pair<>(new ScaleVertex(1.0), new InputType[] {InputType.recurrent(10, 10)}));
|
||
|
|
out.add(new Pair<>(new ShiftVertex(1.0), new InputType[] {InputType.recurrent(10, 10)}));
|
||
|
|
out.add(new Pair<>(new StackVertex(),
|
||
|
|
new InputType[] {InputType.recurrent(10, 10), InputType.recurrent(10, 10)}));
|
||
|
|
out.add(new Pair<>(new UnstackVertex(0, 2), new InputType[] {InputType.recurrent(10, 10)}));
|
||
|
|
|
||
|
|
out.add(new Pair<>(new DuplicateToTimeSeriesVertex("0"),
|
||
|
|
new InputType[] {InputType.recurrent(10, 10), InputType.feedForward(10)}));
|
||
|
|
out.add(new Pair<>(new LastTimeStepVertex("0"), new InputType[] {InputType.recurrent(10, 10)}));
|
||
|
|
|
||
|
|
return out;
|
||
|
|
}
|
||
|
|
|
||
|
|
@Test
|
||
|
|
public void testMemoryReportSimple() {
|
||
|
|
|
||
|
|
List<Pair<? extends Layer, InputType>> l = getTestLayers();
|
||
|
|
|
||
|
|
|
||
|
|
for (Pair<? extends Layer, InputType> p : l) {
|
||
|
|
|
||
|
|
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list().layer(0, p.getFirst().clone())
|
||
|
|
.layer(1, p.getFirst().clone()).validateOutputLayerConfig(false).build();
|
||
|
|
|
||
|
|
MemoryReport mr = conf.getMemoryReport(p.getSecond());
|
||
|
|
// System.out.println(mr.toString());
|
||
|
|
// System.out.println("\n\n");
|
||
|
|
|
||
|
|
//Test to/from JSON + YAML
|
||
|
|
String json = mr.toJson();
|
||
|
|
String yaml = mr.toYaml();
|
||
|
|
|
||
|
|
MemoryReport fromJson = MemoryReport.fromJson(json);
|
||
|
|
MemoryReport fromYaml = MemoryReport.fromYaml(yaml);
|
||
|
|
|
||
|
|
assertEquals(mr, fromJson);
|
||
|
|
assertEquals(mr, fromYaml);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
@Test
|
||
|
|
public void testMemoryReportSimpleCG() {
|
||
|
|
|
||
|
|
List<Pair<? extends Layer, InputType>> l = getTestLayers();
|
||
|
|
|
||
|
|
|
||
|
|
for (Pair<? extends Layer, InputType> p : l) {
|
||
|
|
|
||
|
|
ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder().graphBuilder().addInputs("in")
|
||
|
|
.addLayer("0", p.getFirst().clone(), "in").addLayer("1", p.getFirst().clone(), "0")
|
||
|
|
.setOutputs("1").validateOutputLayerConfig(false).build();
|
||
|
|
|
||
|
|
MemoryReport mr = conf.getMemoryReport(p.getSecond());
|
||
|
|
// System.out.println(mr.toString());
|
||
|
|
// System.out.println("\n\n");
|
||
|
|
|
||
|
|
//Test to/from JSON + YAML
|
||
|
|
String json = mr.toJson();
|
||
|
|
String yaml = mr.toYaml();
|
||
|
|
|
||
|
|
MemoryReport fromJson = MemoryReport.fromJson(json);
|
||
|
|
MemoryReport fromYaml = MemoryReport.fromYaml(yaml);
|
||
|
|
|
||
|
|
assertEquals(mr, fromJson);
|
||
|
|
assertEquals(mr, fromYaml);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
@Test
|
||
|
|
public void testMemoryReportsVerticesCG() {
|
||
|
|
List<Pair<? extends GraphVertex, InputType[]>> l = getTestVertices();
|
||
|
|
|
||
|
|
for (Pair<? extends GraphVertex, InputType[]> p : l) {
|
||
|
|
List<String> inputs = new ArrayList<>();
|
||
|
|
for (int i = 0; i < p.getSecond().length; i++) {
|
||
|
|
inputs.add(String.valueOf(i));
|
||
|
|
}
|
||
|
|
|
||
|
|
String[] layerInputs = inputs.toArray(new String[inputs.size()]);
|
||
|
|
if (p.getFirst() instanceof DuplicateToTimeSeriesVertex) {
|
||
|
|
layerInputs = new String[] {"1"};
|
||
|
|
}
|
||
|
|
|
||
|
|
ComputationGraphConfiguration conf = new NeuralNetConfiguration.Builder().graphBuilder().addInputs(inputs)
|
||
|
|
.allowDisconnected(true)
|
||
|
|
.addVertex("gv", p.getFirst(), layerInputs).setOutputs("gv").build();
|
||
|
|
|
||
|
|
MemoryReport mr = conf.getMemoryReport(p.getSecond());
|
||
|
|
// System.out.println(mr.toString());
|
||
|
|
// System.out.println("\n\n");
|
||
|
|
|
||
|
|
//Test to/from JSON + YAML
|
||
|
|
String json = mr.toJson();
|
||
|
|
String yaml = mr.toYaml();
|
||
|
|
|
||
|
|
MemoryReport fromJson = MemoryReport.fromJson(json);
|
||
|
|
MemoryReport fromYaml = MemoryReport.fromYaml(yaml);
|
||
|
|
|
||
|
|
assertEquals(mr, fromJson);
|
||
|
|
assertEquals(mr, fromYaml);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
@Test
|
||
|
|
public void testInferInputType() {
|
||
|
|
List<Pair<INDArray[], InputType[]>> l = new ArrayList<>();
|
||
|
|
l.add(new Pair<>(new INDArray[] {Nd4j.create(10, 8)}, new InputType[] {InputType.feedForward(8)}));
|
||
|
|
l.add(new Pair<>(new INDArray[] {Nd4j.create(10, 8), Nd4j.create(10, 20)},
|
||
|
|
new InputType[] {InputType.feedForward(8), InputType.feedForward(20)}));
|
||
|
|
l.add(new Pair<>(new INDArray[] {Nd4j.create(10, 8, 7)}, new InputType[] {InputType.recurrent(8, 7)}));
|
||
|
|
l.add(new Pair<>(new INDArray[] {Nd4j.create(10, 8, 7), Nd4j.create(10, 20, 6)},
|
||
|
|
new InputType[] {InputType.recurrent(8, 7), InputType.recurrent(20, 6)}));
|
||
|
|
|
||
|
|
//Activations order: [m,d,h,w]
|
||
|
|
l.add(new Pair<>(new INDArray[] {Nd4j.create(10, 8, 7, 6)},
|
||
|
|
new InputType[] {InputType.convolutional(7, 6, 8)}));
|
||
|
|
l.add(new Pair<>(new INDArray[] {Nd4j.create(10, 8, 7, 6), Nd4j.create(10, 4, 3, 2),},
|
||
|
|
new InputType[] {InputType.convolutional(7, 6, 8), InputType.convolutional(3, 2, 4)}));
|
||
|
|
|
||
|
|
for (Pair<INDArray[], InputType[]> p : l) {
|
||
|
|
InputType[] act = InputType.inferInputTypes(p.getFirst());
|
||
|
|
|
||
|
|
assertArrayEquals(p.getSecond(), act);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
|
||
|
|
@Test
|
||
|
|
public void validateSimple() {
|
||
|
|
|
||
|
|
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder().list()
|
||
|
|
.layer(0, new DenseLayer.Builder().nIn(10).nOut(20).build())
|
||
|
|
.layer(1, new DenseLayer.Builder().nIn(20).nOut(27).build()).build();
|
||
|
|
|
||
|
|
MemoryReport mr = conf.getMemoryReport(InputType.feedForward(10));
|
||
|
|
|
||
|
|
int numParams = (10 * 20 + 20) + (20 * 27 + 27); //787 -> 3148 bytes
|
||
|
|
int actSize = 20 + 27; //47 -> 188 bytes
|
||
|
|
int total15Minibatch = numParams + 15 * actSize;
|
||
|
|
|
||
|
|
//Fixed: should be just params
|
||
|
|
long fixedBytes = mr.getTotalMemoryBytes(0, MemoryUseMode.INFERENCE, CacheMode.NONE, DataType.FLOAT);
|
||
|
|
long varBytes = mr.getTotalMemoryBytes(1, MemoryUseMode.INFERENCE, CacheMode.NONE, DataType.FLOAT)
|
||
|
|
- fixedBytes;
|
||
|
|
|
||
|
|
assertEquals(numParams * 4, fixedBytes);
|
||
|
|
assertEquals(actSize * 4, varBytes);
|
||
|
|
|
||
|
|
long minibatch15 = mr.getTotalMemoryBytes(15, MemoryUseMode.INFERENCE, CacheMode.NONE, DataType.FLOAT);
|
||
|
|
assertEquals(total15Minibatch * 4, minibatch15);
|
||
|
|
|
||
|
|
// System.out.println(fixedBytes + "\t" + varBytes);
|
||
|
|
// System.out.println(mr.toString());
|
||
|
|
|
||
|
|
assertEquals(actSize * 4, mr.getMemoryBytes(MemoryType.ACTIVATIONS, 1, MemoryUseMode.TRAINING, CacheMode.NONE,
|
||
|
|
DataType.FLOAT));
|
||
|
|
assertEquals(actSize * 4, mr.getMemoryBytes(MemoryType.ACTIVATIONS, 1, MemoryUseMode.INFERENCE, CacheMode.NONE,
|
||
|
|
DataType.FLOAT));
|
||
|
|
|
||
|
|
int inputActSize = 10 + 20;
|
||
|
|
assertEquals(inputActSize * 4, mr.getMemoryBytes(MemoryType.ACTIVATION_GRADIENTS, 1, MemoryUseMode.TRAINING,
|
||
|
|
CacheMode.NONE, DataType.FLOAT));
|
||
|
|
assertEquals(0, mr.getMemoryBytes(MemoryType.ACTIVATION_GRADIENTS, 1, MemoryUseMode.INFERENCE, CacheMode.NONE,
|
||
|
|
DataType.FLOAT));
|
||
|
|
|
||
|
|
//Variable working memory - due to preout during backprop. But not it's the MAX value, as it can be GC'd or workspaced
|
||
|
|
int workingMemVariable = 27;
|
||
|
|
assertEquals(workingMemVariable * 4, mr.getMemoryBytes(MemoryType.WORKING_MEMORY_VARIABLE, 1,
|
||
|
|
MemoryUseMode.TRAINING, CacheMode.NONE, DataType.FLOAT));
|
||
|
|
assertEquals(0, mr.getMemoryBytes(MemoryType.WORKING_MEMORY_VARIABLE, 1, MemoryUseMode.INFERENCE,
|
||
|
|
CacheMode.NONE, DataType.FLOAT));
|
||
|
|
}
|
||
|
|
|
||
|
|
@Test
|
||
|
|
public void testPreprocessors() throws Exception {
|
||
|
|
//https://github.com/deeplearning4j/deeplearning4j/issues/4223
|
||
|
|
File f = new ClassPathResource("4223/CompGraphConfig.json").getTempFileFromArchive();
|
||
|
|
String s = FileUtils.readFileToString(f, Charset.defaultCharset());
|
||
|
|
|
||
|
|
ComputationGraphConfiguration conf = ComputationGraphConfiguration.fromJson(s);
|
||
|
|
|
||
|
|
conf.getMemoryReport(InputType.convolutional(17,19,19));
|
||
|
|
}
|
||
|
|
}
|