Model Zoo

WekaDeeplearning4J contains a wide range of popular architectures, ready to use either for training or as feature extractors. The table below outlines the different models included, whether pretrained weights are available, the types of pretrained weights, and the model variations (if any). WekaDeeplearning4j merges the model zoo of Deeplearning4j and Keras. Values in bold are the defaults.

Framework Model Input Shape Pretrained Implemented Weights Available Varieties
DL4J AlexNet 3, 224, 224 No - -
DL4J Darknet19 3, 224, 224 Yes ImageNet 224x224 or 448x448 input size
DL4J FaceNetNN4Small2 3, 96, 96 No - -
DL4J InceptionResNetV1 3, 160, 160 No - -
DL4J LeNet 1, 28, 28 Yes MNIST -
DL4J ResNet50 3, 224, 224 Yes ImageNet -
DL4J SqueezeNet 3, 227, 227 Yes ImageNet -
DL4J VGG 3, 224, 224 Yes ImageNet, VGGFace (VGG16 only) 16, 19
DL4J XCeption 3, 299, 299 Yes ImageNet -
Keras DenseNet 3, 224, 224 Yes ImageNet 121, 169, 201
Keras EfficientNet 3, 224, 224 Yes ImageNet B0 - B7
Keras InceptionV3 3, 299, 299 Yes ImageNet -
Keras NASNet 3, 224, 224 (Mobile) or 3, 331, 331 (Large) Yes ImageNet Mobile, Large
Keras ResNet 3, 224, 224 Yes ImageNet 50, 50V2, 101, 101V2, 152, 152V2
Keras VGG 3, 224, 224 Yes ImageNet 16, 19
Keras Xception 3, 299, 299 Yes ImageNet -

Finetuning some of these models can result in NaN errors (returns NaN for all classes when performing prediction on an instance). Currently, the only fix for this is to retune the network: DL4J Help


To set a predefined model, e.g. ResNet50, from the model zoo in the GUI is straightforward via the corresponding pop-up menu. To set a predefined model from the command-line or via the API, it is necessary to add the -zooModel ".Dl4jResNet50" option via commandline, or call the setZooModel(new ResNet50()) on the Dl4jMlpClassifier object.

Model names from Keras are prepended with Keras, i.e., KerasResNet, and similarly for Deeplearning4j models (e.g., DL4JDarknet19). In addition, some models support different variations. Again, it is straightforward to do this via the GUI. To do via command line you must add the -variation argument e.g.:

-ZooModel ".KerasResNet -variation RESNET152V2" 

If using the Java API, these can be set via .setVariation() e.g.:

KerasResNet kerasResNet = new KerasResNet();

View the featurizing tutorial and the finetuning tutorial for usage examples with the model zoo.

Image Instance Iterator

Pretrained zoo models have their input shape pre-specified. To avoid errors when the user-specified image dimensions don't match this input shape, the ImageInstanceIterator enforces image dimensions based on the selected zoo model's input shape.

Model Summaries

The following summaries are generated from the pretrained zoo models included in WekaDeeplearning4j. These may be useful as a reference for trying different feature extraction layers, or simply to investigate famous model architectures.