Package ai.djl.basicdataset.cv.classification

Class ImageFolder

java.lang.Object
ai.djl.training.dataset.RandomAccessDataset
ai.djl.basicdataset.cv.ImageDataset
ai.djl.basicdataset.cv.classification.ImageClassificationDataset
ai.djl.basicdataset.cv.classification.AbstractImageFolder
ai.djl.basicdataset.cv.classification.ImageFolder
All Implemented Interfaces:
ai.djl.training.dataset.Dataset
public final class ImageFolder extends AbstractImageFolder
A dataset for loading image files stored in a folder structure.

Below is an example directory layout for the image folder: 

  The image folder should be structured as follows:
       root/shoes/Aerobic Shoes1.png
       root/shoes/Aerobic Shose2.png
       ...
       root/boots/Black Boots.png
       root/boots/White Boots.png
       ...
       root/pumps/Red Pumps.png
       root/pumps/Pink Pumps.png
       ...

  here shoes, boots, pumps are your labels

Here, the dataset will take the folder names (shoes, boots, bumps) in sorted order as your labels. Nested folder structures are not currently supported.

Then, you can create your instance of the dataset as follows: 

 // set the image folder path
 Repository repository = Repository.newInstance("folder", Paths.get("/path/to/imagefolder/root");
 ImageFolder dataset =
     ImageFolder.builder()
         .setRepository(repository)
         .addTransform(new Resize(100, 100)) // Use image transforms as necessary for your data
         .addTransform(new ToTensor()) // Usually required as the last transform to convert images to tensors
         .setSampling(batchSize, true)
         .build();

 // call prepare before using
 dataset.prepare();

 // to get the synset or label names
 List>String< synset = dataset.getSynset();