Rewriting CNN

Tags: control and think
Personhours: 12
Rewriting CNN By Arjun and Abhi

Task: Begin rewriting the Convolutional Neural Network using Java and DL4J

While we were using Python and TensorFlow to train our convolutional neural network, we decided to attempt writing this in Java, as the code for our robot is entirely in Java, and before we can use our neural network, it must be written in java.

We also decided to try using DL4J, a competing library to TensorFlow, to write our neural network, to determine if it was easier to write a neural network using DL4J or TensorFlow. We found that both DL4J and TensorFlow were similarly easy to use, and while each had a different style, code written using both were equally easy to read and maintain.

 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
java
		//Download dataset
		DataDownloader downloader = new DataDownloader();
		File rootDir = downloader.downloadFilesFromGit("https://github.com/arjvik/RoverRuckusTrainingData.git", "data/RoverRuckusTrainingData", "TrainingData");
		
		//Read in dataset
		DataSetIterator iterator = new CustomDataSetIterator(rootDir, 1);
		
		//Normalization
		DataNormalization scaler = new ImagePreProcessingScaler(0, 1);
		scaler.fit(iterator);
		iterator.setPreProcessor(scaler);
		
		//Read in test dataset
		DataSetIterator testIterator = new CustomDataSetIterator(new File(rootDir, "Test"), 1);
			
		//Test Normalization
		DataNormalization testScaler = new ImagePreProcessingScaler(0, 1);
		testScaler.fit(testIterator);
		testIterator.setPreProcessor(testScaler);
		
		//Layer Configuration
		MultiLayerConfiguration conf = new NeuralNetConfiguration.Builder()
				.seed(SEED)
				.l2(0.005)
				.weightInit(WeightInit.XAVIER)
				.list()
				.layer(0, new ConvolutionLayer.Builder()
						.nIn(1)
						.kernelSize(3, 3)
						.stride(1, 1)
						.activation(Activation.RELU)
						.build())
				.layer(1, new ConvolutionLayer.Builder()
						.nIn(1)
						.kernelSize(3, 3)
						.stride(1, 1)
						.activation(Activation.RELU)
						.build())
				/* ...more layer code... */
				.build();

Next Steps

We still need to attempt to to fix the inaccuracy in the predictions made by our neural network.

Date | October 20, 2018