我是 Deeplearning4j 的初学者,并打算对 Cifar-10 图像分类进行测试。我只是从 DL4j 示例(AnimalsClassification.java)复制 Alexnet,例如:
MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
.seed(seed)
.weightInit(WeightInit.DISTRIBUTION)
.dist(new NormalDistribution(0.0, 0.01))
.activation(Activation.RELU)
.updater(Updater.NESTEROVS)
.iterations(iterations)
.gradientNormalization(GradientNormalization.RenormalizeL2PerLayer) // normalize to prevent vanishing or exploding gradients
.optimizationAlgo(OptimizationAlgorithm.STOCHASTIC_GRADIENT_DESCENT)
.learningRate(1e-2)
.biasLearningRate(1e-2*2)
.learningRateDecayPolicy(LearningRatePolicy.Step)
.lrPolicyDecayRate(0.1)
.lrPolicySteps(100000)
.regularization(true)
.l2(5 * 1e-4)
.momentum(0.9)
.miniBatch(false)
.list()
.layer(0, convInit("cnn1", channels, 96, new int[]{11, 11}, new int[]{4, 4}, new int[]{3, 3}, 0))
.layer(1, new LocalResponseNormalization.Builder().name("lrn1").build())
.layer(2, maxPool("maxpool1", new int[]{3,3}))
.layer(3, conv5x5("cnn2", 256, new int[] {1,1}, new int[] {2,2}, nonZeroBias))
.layer(4, new LocalResponseNormalization.Builder().name("lrn2").build())
.layer(5, maxPool("maxpool2", new int[]{3,3}))
.layer(6,conv3x3("cnn3", 384, 0))
.layer(7,conv3x3("cnn4", 384, nonZeroBias))
.layer(8,conv3x3("cnn5", 256, nonZeroBias))
.layer(9, maxPool("maxpool3", new int[]{3,3}))
.layer(10, fullyConnected("ffn1", 4096, nonZeroBias, dropOut, new GaussianDistribution(0, 0.005)))
.layer(11, fullyConnected("ffn2", 4096, nonZeroBias, dropOut, new GaussianDistribution(0, 0.005)))
.layer(12, new OutputLayer.Builder(LossFunctions.LossFunction.NEGATIVELOGLIKELIHOOD)
.name("output")
.nOut(numLabels)
.activation(Activation.SOFTMAX)
.build())
.backprop(true)
.pretrain(false)
.setInputType(InputType.convolutional(height, width, channels))
.build();
当我运行代码时,它抛出了一个异常,说新 int[]{3,3} 上的“layer-9”配置存在一些问题,它应该大于 0 并且小于 pHeight + 2*padH。在 java 代码中将 weight*height 从 32 * 32 更改为 100*100 时,它运行正常,但我不应该结果是好的。所以我对alexnet处理32 * 32图像的层配置有点困惑。