深度学习4j(DL4J)精度低，召回和F1

Question

我将使用DL4J来针对条件矩阵找到一个好的模型。 我已经准备了类似CSV的数据集(示例如下(，在微调超参数并多次训练模型后，我仍然无法获得合理的精度，召回率和F1结果。 请问我是否实施错了什么？

示例数据集：

## 基本上，每列定义每个样本的条件是否存在 (1( 或不存在 (0(。第一列是标签分类只有 2 个输出，即 1/0

[1, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 0, 1]
[1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0]
[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]
[1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
[0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]
...........

DataVec 部分：

int OUTPUT_NEURONS = 2;  // Only 2 classes for output
int CLASS_INDEX = 0;     // First column is the label
int FILE_SIZE = 0;       // FILE_SIZE will be calculated while preparing the datavecRecords below
List<List<Writable>> datavecRecords = new ArrayList<>();
......
Prepare the datavecRecords using above csv data 
......
CollectionRecordReader crr = new CollectionRecordReader(datavecRecords);
RecordReaderDataSetIterator iter = new RecordReaderDataSetIterator(crr, FILE_SIZE, CLASS_INDEX, OUTPUT_NEURONS);
allData = iter.next();
SplitTestAndTrain testAndTrain = allData.splitTestAndTrain(0.6);
DataSet trainingData = testAndTrain.getTrain();
DataSet testData = testAndTrain.getTest();
DataNormalization normalizer = new NormalizerStandardize();
normalizer.fit(trainingData);
normalizer.transform(trainingData);
normalizer.transform(testData);
// For early escaping use
DataSetIterator trainSetIterator = new ListDataSetIterator(trainingData.asList()); 
DataSetIterator testSetIterator = new ListDataSetIterator(testData.asList()); 
// sortedKeys is the calculated number of input columns
INPUT_NEURONS = sortedKeys.size() - 1; 
HIDDEN_NEURONS = FILE_SIZE / (2 * (INPUT_NEURONS + OUTPUT_NEURONS));
HIDDEN_NEURONS = HIDDEN_NEURONS <= 0 ? 1 : HIDDEN_NEURONS;

型：

int n=0;
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
.seed(12345)
.iterations(1)
.learningRate(0.001)
.weightInit(WeightInit.XAVIER)
.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
.regularization(true).l2(1e-4)
.updater(new Nesterovs(0.001,0.9))
.list()
.layer(n++, new DenseLayer.Builder()
.nIn(INPUT_NEURONS)
.nOut(HIDDEN_NEURONS)
.activation(Activation.RELU)
.build())
.layer(n++, new OutputLayer.Builder(LossFunction.NEGATIVELOGLIKELIHOOD)
.nIn(HIDDEN_NEURONS)
.nOut(OUTPUT_NEURONS)
.activation(Activation.SOFTMAX)
.build())
.pretrain(false).backprop(true).build();            
EarlyStoppingConfiguration esConf = new EarlyStoppingConfiguration.Builder()
.epochTerminationConditions(
new MaxEpochsTerminationCondition(10000), 
new ScoreImprovementEpochTerminationCondition(50))
.iterationTerminationConditions(new MaxTimeIterationTerminationCondition(5, TimeUnit.MINUTES))
.scoreCalculator(new DataSetLossCalculator(testSetIterator, true))
.evaluateEveryNEpochs(1)
.modelSaver(saver)
.build();

训练和测试代码

StatsStorage statsStorage = new InMemoryStatsStorage();
MultiLayerNetwork networkModel = new MultiLayerNetwork(conf);
networkModel.setListeners(new StatsListener(statsStorage), new ScoreIterationListener(10));

IEarlyStoppingTrainer trainer = new EarlyStoppingTrainer(esConf, networkModel, trainSetIterator);
EarlyStoppingResult<MultiLayerNetwork> result = trainer.fit();

// -------------------------- Evaluation trained model and print results --------------------------
System.out.println("Termination reason: " + result.getTerminationReason());
System.out.println("Termination details: " + result.getTerminationDetails());
System.out.println("Total epochs: " + result.getTotalEpochs());
System.out.println("Best epoch number: " + result.getBestModelEpoch());
System.out.println("Score at best epoch: " + result.getBestModelScore());
MultiLayerNetwork bestNetwork = result.getBestModel();
Evaluation eval1 = new Evaluation(OUTPUT_NEURONS);
testSetIterator.reset();
for (int i = 0; i < testData.numExamples(); i++) {
DataSet t = testData.get(i);
INDArray features = t.getFeatureMatrix();
INDArray labels = t.getLabels();
INDArray output = bestNetwork.output(features, false);
eval1.eval(labels, output);
}
M.messageln(eval1.stats());

结果：

Termination reason: EpochTerminationCondition
Termination details: ScoreImprovementEpochTerminationCondition(maxEpochsWithNoImprovement=50, minImprovement=0.0)
Total epochs: 55
Best epoch number: 4
Score at best epoch: 0.6579822991097982
Examples labeled as 0 classified by model as 0: 397 times
Examples labeled as 0 classified by model as 1: 58 times
Examples labeled as 1 classified by model as 0: 190 times
Examples labeled as 1 classified by model as 1: 55 times

==========================Scores========================================
# of classes:    2
Accuracy:        0.6457
Precision:       0.5815
Recall:          0.5485
F1 Score:        0.3073
========================================================================
Pattern1 :      Accuracy: 0.6457142857142857 | Precision: 0.5815229681446081 | Recall: 0.54850863422292 | F1: 0.3072625698324022

无论我如何调整学习率、输入和输出激活方法、更新器、调节等，我仍然无法得到满意的结果。 如果您能帮我如何更好地操纵DL4J，我将不胜感激。 我正在研究仲裁者，但没有运气。不确定我是否正在使用 0.9.1 稳定版本。

谢谢十亿！

Answer 1

您可以在用户社区中找到问题的答案：https://gitter.im/deeplearning4j/deeplearning4j/tuninghelp

Answer 2

代码看起来不错。 你在那里没有做错什么。 我认为只需要调整它。

您拥有的 60/40 测试/训练拆分高于我通常看到的 70/30 或 80/20。 您想保留更多数据进行测试的原因是什么？

当你说合理时，你的意思是与基线相比吗？如果没有，我会从一个本质上是逻辑回归的输出层开始，你可以使用它作为基线。

您是否尝试过添加更多图层？ 再增加一层辍学，可能会有用。