Loading Data

Load tabular data with DataSets.readCsv(...). For a CNN, load consistently sized and labeled images into an ImageSet instead.

The tabular CSV recipe assumes that:

• every data row represents one sample;
• input columns come first;
• target columns come last;
• values use a consistent delimiter;
• the data is already numerical.

CSV structure

For a regression task with two inputs and one target, organize the file like this:

feature_1,feature_2,target
12.5,3.2,48.1
10.0,4.7,45.6

NUM_INPUTS = 2 tells Deep Netts to use the first two columns as model inputs. NUM_OUTPUTS = 1 makes the final column the prediction target.

Load the dataset

Add the required imports and describe the expected schema:

import deepnetts.data.DataSets;
import deepnetts.data.MLDataItem;
import deepnetts.data.TabularDataSet;

private static final String DATASET_PATH =
        "src/main/resources/datasets/data.csv";
private static final int NUM_INPUTS = 2;
private static final int NUM_OUTPUTS = 1;

Load the CSV:

TabularDataSet<MLDataItem> dataSet =
        DataSets.readCsv(
                DATASET_PATH,
                NUM_INPUTS,
                NUM_OUTPUTS,
                true,
                ","
        );

The arguments are:

Use false instead of true when the file does not contain a header. Change the final argument when the file uses another delimiter, such as ";".

Load labeled images for a CNN

The image-classification projects include an ImageData helper that reads an index file, validates each image, resizes it, optionally converts it to grayscale, and creates the ExampleImage objects stored in an ImageSet:

import com.deepnetts.examples.image.ImageData;
import deepnetts.data.ImageSet;

int width = 64;
int height = 64;

ImageSet images = ImageData.load(
        "src/main/resources/datasets/duke-logo",
        "index.txt",
        width,
        height,
        false,
        false
);

The final two arguments select grayscale conversion and value inversion. Keep both false for ordinary RGB input. Set grayscale to true for one-channel input, and enable inversion only when the dataset requires light and dark values to be reversed. The resulting CNN input has shape width × height × 3 for RGB or width × height × 1 for grayscale.

The index file maps each relative image path to its label, while labels.txt defines the stable output-class order. Inspect the loaded set before continuing:

System.out.println("Images: " + images.size());
System.out.println("Images by class: " + images.countByClasses());

Check that every expected class is present, labels use one consistent spelling, and class counts are plausible. A typo in a label creates another class rather than a harmless display difference.

Verify the tabular result

Do not continue directly to training. First confirm that Deep Netts loaded the expected number of samples and columns:

import java.util.Arrays;

System.out.println("Samples: " + dataSet.size());
System.out.println(
        "Columns: " + Arrays.toString(dataSet.getColumnNames())
);

Compare this output with the CSV file you intended to load. This small check catches incorrect paths, unexpected headers, and schema mistakes early.