Available Models

Pretrained models available through OmniHub — DL4J and SameDiff model catalog

OmniHub currently hosts models in two frameworks: DL4J (ComputationGraph format) and SameDiff (FlatBuffers format). All models are accessed through the Pretrained class or directly via OmniHubUtils.

---

DL4J Models

DL4J models are loaded as ComputationGraph objects. They are stored as zip archives in the dl4j/ directory of the zoo repository.

VGG19 (no top)

Property	Value
Method	Pretrained.dl4j().vgg19noTop(forceDownload)
File	vgg19_weights_tf_dim_ordering_tf_kernels_notop.zip
Return type	ComputationGraph
Source	Keras Applications VGG19 (TensorFlow dimension ordering)
Input shape	[batch, 3, 224, 224] (channels-first)
Output	Feature map from the last pooling layer, shape [batch, 512, 7, 7]
Pretrained dataset	ImageNet

This model is the VGG19 convolutional base without the fully connected classifier head (hence "no top"). It is suitable as a feature extractor or as a base for transfer learning. Weights were converted from the Keras Applications VGG19 checkpoint using TensorFlow channel ordering (tf_dim_ordering).

Loading example:

import org.eclipse.deeplearning4j.omnihub.models.Pretrained;
import org.deeplearning4j.nn.graph.ComputationGraph;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.api.ndarray.INDArray;

ComputationGraph vgg19 = Pretrained.dl4j().vgg19noTop(false);

// Input: one RGB image at 224x224
INDArray input = Nd4j.rand(new int[]{1, 3, 224, 224});
INDArray[] features = vgg19.output(input);

System.out.println(features[0].shapeInfoToString());
// [1, 512, 7, 7]

Preprocessing. VGG19 expects pixel values in [0, 255] range with per-channel mean subtraction (ImageNet means: R=103.939, G=116.779, B=123.68 in BGR order). The Keras Applications preprocessing function is the reference. Apply the same preprocessing to your inputs when using this model for feature extraction.

---

SameDiff Models

SameDiff models are loaded as SameDiff objects. They are stored as FlatBuffers files (.fb) in the samediff/ directory of the zoo repository. Each was converted to SameDiff from an ONNX checkpoint or a PyTorch torchvision model via ONNX export.

AgeGoogLeNet

Property	Value
Method	Pretrained.samediff().ageGooglenet(forceDownload)
File	age_googlenet.fb
Return type	SameDiff
Source	ONNX Model Zoo — age_googlenet.onnx
Original source	ONNX Models on GitHub
Input name	data
Input shape	[1, 3, 224, 224] (NCHW)
Output name	loss3/loss3_Y
Output shape	[1, 8]
Output classes	8 age groups
Pretrained dataset	Adience benchmark

AgeGoogLeNet is a GoogLeNet/Inception variant trained to predict human age from a face image. The output is a softmax distribution over 8 age buckets: (0–2), (4–6), (8–13), (15–20), (25–32), (38–43), (48–53), (60+).

The model was originally published as part of the ONNX Model Zoo and converted to SameDiff format via the DL4J ONNX importer.

Loading and inference example:

import org.eclipse.deeplearning4j.omnihub.models.Pretrained;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.api.ndarray.INDArray;
import java.util.Map;

SameDiff ageModel = Pretrained.samediff().ageGooglenet(false);

// Input: face image normalized to [0,1], shape [1, 3, 224, 224]
INDArray faceImage = Nd4j.rand(new int[]{1, 3, 224, 224});

Map<String, INDArray> outputs = ageModel.output(
    null,
    new String[]{"loss3/loss3_Y"},
    Map.of("data", faceImage)
);

INDArray ageProbs = outputs.get("loss3/loss3_Y"); // shape [1, 8]
int predictedBucket = ageProbs.argMax(1).getInt(0);
String[] ageBuckets = {"0-2","4-6","8-13","15-20","25-32","38-43","48-53","60+"};
System.out.println("Predicted age group: " + ageBuckets[predictedBucket]);

---

ResNet18

Property	Value
Method	Pretrained.samediff().resnet18(forceDownload)
File	resnet18.fb
Return type	SameDiff
Source	PyTorch torchvision resnet18 — exported via ONNX
Original source	torchvision models
Input name	input.1
Input shape	[1, 3, 224, 224] (NCHW)
Output name	495
Output shape	[1, 1000]
Output classes	1000 ImageNet classes
Pretrained dataset	ImageNet (ILSVRC 2012)

ResNet18 is an 18-layer residual network. The torchvision pretrained weights achieve approximately 69.8% top-1 accuracy on the ImageNet validation set. The model was exported from PyTorch via ONNX and imported into SameDiff.

The output tensor contains raw logits (not softmax probabilities). Apply a softmax if you need class probabilities.

Loading and inference example:

import org.eclipse.deeplearning4j.omnihub.models.Pretrained;
import org.nd4j.autodiff.samediff.SameDiff;
import org.nd4j.linalg.factory.Nd4j;
import org.nd4j.linalg.api.ndarray.INDArray;
import org.nd4j.linalg.ops.transforms.Transforms;
import java.util.Map;

SameDiff resnet = Pretrained.samediff().resnet18(false);

// Input: ImageNet-normalized image, shape [1, 3, 224, 224]
// Normalization: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]
INDArray image = Nd4j.rand(new int[]{1, 3, 224, 224});

Map<String, INDArray> outputs = resnet.output(
    null,
    new String[]{"495"},
    Map.of("input.1", image)
);

INDArray logits = outputs.get("495"); // [1, 1000]
INDArray probs = Transforms.softmax(logits, false);
int topClass = probs.argMax(1).getInt(0);
System.out.println("Top ImageNet class index: " + topClass);

Preprocessing. The torchvision preprocessing pipeline normalizes each channel by subtracting the ImageNet mean and dividing by the standard deviation:

• Mean: [0.485, 0.456, 0.406] (RGB)

• Std: [0.229, 0.224, 0.225] (RGB)

Pixels should be in [0.0, 1.0] before normalization.

---

Downloading All Models Upfront

If you want to pre-populate the cache (for example, during a Docker image build), call each model with forceDownload = false. You can also call OmniHubUtils.downloadAndLoadFromZoo directly without loading the model:

import org.eclipse.deeplearning4j.omnihub.OmniHubUtils;

// Pre-download without loading into memory
OmniHubUtils.downloadAndLoadFromZoo("dl4j",
    "vgg19_weights_tf_dim_ordering_tf_kernels_notop.zip", false);
OmniHubUtils.downloadAndLoadFromZoo("samediff", "age_googlenet.fb", false);
OmniHubUtils.downloadAndLoadFromZoo("samediff", "resnet18.fb", false);

---

Model Count Summary

Framework	Count	Models
DL4J (ComputationGraph)	1	VGG19 (no top)
SameDiff	2	AgeGoogLeNet, ResNet18

The zoo is actively expanded. Models are added as new files to the zoo GitHub repository without requiring a library release. Check the omnihub-zoo repository for the current file listing.