View Example on GitHub

TensorFlow VGGNet16 Import

This walkthrough reconstructs VGGNet16 in Deep Netts, imports TensorFlow/Keras ImageNet weights, and classifies an image without retraining the model.

Problem: Classify an image as one of 1,000 ImageNet categories using pretrained TensorFlow weights.

Dataset: No training dataset is downloaded; the project bundles 1,000 ImageNet labels and small test images for inference.

At a glance

1. Configure the runtime

The program verifies Java 25 and the Vector API, disables CUDA, and prints the memory requirement. Use -Xmx8g; the epoch property does not apply because there is no training.

2. Reconstruct VGGNet16

ConvolutionalNetwork.builder()
        .addInputLayer(224, 224, 3)
        // five convolution/pooling blocks: 64, 128, 256, 512, 512 filters
        .addFullyConnectedLayer(4_096)
        .addFullyConnectedLayer(4_096)
        .addOutputLayer(1_000, ActivationType.SOFTMAX)
        .hiddenActivationFunction(ActivationType.RELU)
        .lossFunction(LossType.CROSS_ENTROPY)
        .build();

The layer sequence must exactly match the network that produced the exported weights. See the runnable class for all 13 convolutional layers.

3. Download, cache, and import weights

Path weights = VggWeightsDownloader.obtain(CACHE_DIRECTORY);
TensorflowUtils.importWeights(vggNet16, weights.toString());

The downloader reuses a valid cache, writes new downloads to a .part file, checks HTTP failure, atomically renames the completed archive, and prevents ZIP entries from escaping the cache directory. Generated archives and models remain outside Git.

4. Load ImageNet labels and save

String[] labels = Files.readString(LABELS_PATH).trim().split("\\s*,\\s*");
if (labels.length != 1_000) {
    throw new IllegalStateException("Expected 1,000 ImageNet labels");
}
vggNet16.setOutputLabels(labels);
vggNet16.save("models/vggnet16.dnet");

Only the 1,000 label names and small test images are bundled. The ImageNet training dataset is never downloaded.

5. Apply VGG preprocessing

VggNet16InputImage resizes an image to 224 × 224, changes RGB to BGR channel order, and subtracts the ImageNet means 103.939, 116.779, and 123.68. Imported models must receive inputs prepared exactly like their original training inputs.

6. Run inference

Path imagePath = args.length > 0 ? Path.of(args[0]) : DEFAULT_IMAGE;
VggNet16InputImage image = new VggNet16InputImage(imagePath);
float[] probabilities = vggNet16.predict(image.getInput()).getValues();
int predictedIndex = indexOfLargest(probabilities);
System.out.println(labels[predictedIndex]);

The complete example also warms up the JVM and reports confidence and measured inference time before shutting down the network thread pool.

Run the example

Pass these options to the Java process through your IDE's application run configuration or the command line:

--add-modules=jdk.incubator.vector -Xmx8g

Without arguments the example uses elephant.jpg. To classify another file, add its absolute path under Arguments. The first run requires internet access and can take significantly longer due to the weight download and import.

← Back to Image Classification Examples