Available Models
Complete list of pretrained models — VGG, ResNet, AlexNet, LeNet, YOLO, UNet, and other architectures
All models are in the org.deeplearning4j.zoo.model package. Add the deeplearning4j-zoo dependency to your project to use them (see the Model Zoo Overview for setup instructions).
Input shapes follow NCHW convention: {channels, height, width}. Pretrained weights are downloaded automatically to the DL4J cache directory on first use and verified by checksum.
AlexNet
DL4J's implementation of AlexNet, based on the original ImageNet Classification paper. The architecture uses five convolutional layers followed by three fully connected layers with local response normalization and dropout.
Class
org.deeplearning4j.zoo.model.AlexNet
Input shape
3 × 224 × 224
Pretrained datasets
None
Use case
Image classification baseline, educational reference
Paper: ImageNet Classification with Deep Convolutional Neural Networks (Krizhevsky et al., NeurIPS 2012)
ZooModel model = AlexNet.builder()
.numClasses(1000)
.seed(123)
.build();
MultiLayerNetwork net = (MultiLayerNetwork) model.init();Darknet19
Darknet19 is the feature-extraction backbone used in YOLO v2. It uses 19 convolutional layers and 5 max-pooling layers. Two pretrained resolutions are available: 224 × 224 and 448 × 448. Call setInputShape() to select which resolution you need before initialization.
Class
org.deeplearning4j.zoo.model.Darknet19
Input shape
3 × 224 × 224 or 3 × 448 × 448
Pretrained datasets
ImageNet
Use case
Object detection backbone, image classification
Input images must have channels in RGB order (not BGR) with pixel values normalized to [0, 1]. Output labels match the ImageNet shortnames list.
Paper: YOLO9000: Better, Faster, Stronger (Redmon & Farhadi, 2016)
FaceNetNN4Small2
A compact variant of the FaceNet model that produces 128-dimensional face embeddings. Uses triplet loss for training. The "NN4Small2" name refers to the NN4 inception module variant at a smaller scale. Suitable for face verification and recognition tasks where a similarity metric between face embeddings is used at inference time.
Class
org.deeplearning4j.zoo.model.FaceNetNN4Small2
Input shape
3 × 96 × 96
Pretrained datasets
None
Use case
Face verification, face recognition, embedding extraction
Papers: FaceNet: A Unified Embedding for Face Recognition and Clustering; OpenFace: A general-purpose face recognition library
InceptionResNetV1
Inception-ResNet-V1 combines the inception module design with residual connections. This DL4J implementation is adapted for face recognition (trained with triplet loss on VGGFace data) rather than general ImageNet classification, which distinguishes it from the standard InceptionResNetV1 used for image classification.
Class
org.deeplearning4j.zoo.model.InceptionResNetV1
Input shape
3 × 160 × 160
Pretrained datasets
VGGFace
Use case
Face recognition, face verification
Paper: FaceNet: A Unified Embedding for Face Recognition and Clustering
LeNet
LeNet-5 is one of the earliest convolutional neural networks, developed by Yann LeCun for handwritten digit recognition. Despite its age and simplicity (two convolutional layers, two pooling layers, two fully connected layers), it remains a useful baseline and educational reference. MNIST pretrained weights are available.
Class
org.deeplearning4j.zoo.model.LeNet
Input shape
1 × 28 × 28
Parameters
~60,000
Pretrained datasets
MNIST
Use case
Digit recognition, educational baseline
Paper: Gradient-Based Learning Applied to Document Recognition (LeCun et al., 1998)
NASNet
Neural Architecture Search Network (NASNet-A) was designed by learning the model architecture directly on the dataset of interest using reinforcement learning. The DL4J implementation uses 1056 penultimate filters and the standard 224 × 224 input. ImageNet weights are available, converted from the Keras Applications implementation.
Class
org.deeplearning4j.zoo.model.NASNet
Input shape
3 × 224 × 224
Pretrained datasets
ImageNet
Use case
High-accuracy image classification
Paper: Learning Transferable Architectures for Scalable Image Recognition (Zoph et al., 2017)
ResNet50
ResNet-50 is a 50-layer residual network that uses skip connections to allow gradients to flow through very deep networks without vanishing. It achieved state-of-the-art accuracy on ImageNet in 2015 and remains a widely used backbone for both classification and feature extraction. ImageNet weights are available, converted from Keras Applications.
Class
org.deeplearning4j.zoo.model.ResNet50
Input shape
3 × 224 × 224
Parameters
~25 million
Pretrained datasets
ImageNet
Use case
Image classification, transfer learning backbone
Paper: Deep Residual Learning for Image Recognition (He et al., 2015)
SimpleCNN
A compact convolutional network for generic image classification, adapted from an open-source face classification implementation. Useful as a lightweight baseline or starting point when you need a small model that trains quickly. No pretrained weights are available; this model is intended to be trained on your own data.
Class
org.deeplearning4j.zoo.model.SimpleCNN
Input shape
3 × 224 × 224
Pretrained datasets
None
Use case
Custom image classification, fast prototyping
SqueezeNet
SqueezeNet achieves AlexNet-level accuracy on ImageNet with approximately 50× fewer parameters through the use of "fire modules" (1 × 1 squeeze convolutions followed by a mix of 1 × 1 and 3 × 3 expand convolutions). It is well-suited for deployment in memory-constrained environments. ImageNet weights are available, converted from the keras-squeezenet implementation.
Class
org.deeplearning4j.zoo.model.SqueezeNet
Input shape
3 × 227 × 227
Parameters
~1.2 million
Pretrained datasets
ImageNet
Use case
Embedded/mobile image classification
Paper: SqueezeNet: AlexNet-level accuracy with 50x fewer parameters (Iandola et al., 2016)
TextGenerationLSTM
An LSTM recurrent neural network designed for character-level text generation. The architecture follows the Keras LSTM text generation example. numClasses must be set to the size of the character vocabulary for your corpus. Pretrained weights trained on the complete works of Walt Whitman are available.
Class
org.deeplearning4j.zoo.model.TextGenerationLSTM
Input type
Character sequences
Pretrained datasets
Walt Whitman corpus
Use case
Character-level text generation, sequence modeling
TinyYOLO
Tiny YOLO is a compact, fast variant of the YOLO object detection architecture built on top of Darknet. It trades some accuracy for significantly faster inference, making it suitable for real-time applications on modest hardware. Pretrained weights covering ImageNet and PASCAL VOC object categories are available, converted from the original Darknet implementation via YAD2K.
Class
org.deeplearning4j.zoo.model.TinyYOLO
Input shape
3 × 416 × 416
Pretrained datasets
ImageNet + VOC (20 object categories)
Use case
Real-time object detection
Input images must have channels in RGB order (not BGR) with pixel values normalized to [0, 1].
Paper: YOLO9000: Better, Faster, Stronger (Redmon & Farhadi, 2016)
UNet
U-Net is a convolutional encoder-decoder architecture designed for biomedical image segmentation. Its "U"-shaped design uses skip connections between the encoder and decoder paths to preserve spatial information lost during downsampling. It set the state of the art on the ISBI cell-tracking challenge and is widely used in medical imaging. Pretrained weights trained on a synthetic segmentation dataset are available.
Class
org.deeplearning4j.zoo.model.UNet
Input shape
1 × 512 × 512
Pretrained datasets
Synthetic segmentation data
Use case
Image segmentation, medical imaging
Paper: U-Net: Convolutional Networks for Biomedical Image Segmentation (Ronneberger et al., 2015)
VGG16
VGG-16 is a 16-layer network from the Oxford Visual Geometry Group, notable for its uniform architecture of 3 × 3 convolutions throughout. Despite being over a decade old, it remains popular for transfer learning due to its straightforward structure. Three pretrained weight sets are available: ImageNet classification, CIFAR-10, and VGGFace (face recognition).
Class
org.deeplearning4j.zoo.model.VGG16
Input shape
3 × 224 × 224
Parameters
~138 million
Pretrained datasets
ImageNet, CIFAR-10, VGGFace
Use case
Image classification, transfer learning, face recognition
Papers: Very Deep Convolutional Networks for Large-Scale Image Recognition; Deep Face Recognition
VGG19
VGG-19 extends VGG-16 with three additional convolutional layers for a total of 19 weight layers. It has slightly higher capacity but also substantially more parameters. Due to its size, consider using WorkspaceMode.SINGLE on large machines to reduce memory overhead. ImageNet weights are available.
Class
org.deeplearning4j.zoo.model.VGG19
Input shape
3 × 224 × 224
Parameters
~143 million
Pretrained datasets
ImageNet
Use case
Image classification, style transfer
Paper: Very Deep Convolutional Networks for Large-Scale Image Recognition
Xception
Xception ("Extreme Inception") replaces standard Inception modules with depthwise separable convolutions, which decouple the mapping of cross-channel correlations from spatial correlations. This results in a more parameter-efficient architecture that outperforms Inception V3 on ImageNet while using fewer parameters. ImageNet weights are available, converted from Keras Applications.
Class
org.deeplearning4j.zoo.model.Xception
Input shape
3 × 299 × 299
Parameters
~22 million
Pretrained datasets
ImageNet
Use case
Image classification, efficient transfer learning
Paper: Xception: Deep Learning with Depthwise Separable Convolutions (Chollet, 2016)
YOLO2
YOLO v2 (full resolution) uses Darknet19 as its backbone and processes 608 × 608 input images for higher accuracy than TinyYOLO. Pretrained weights covering ImageNet pretraining followed by fine-tuning on COCO object detection categories are available, converted from the original Darknet implementation. A Yolo2OutputLayer is appended during fine-tuning on custom datasets.
Class
org.deeplearning4j.zoo.model.YOLO2
Input shape
3 × 608 × 608
Pretrained datasets
ImageNet + COCO (80 object categories)
Use case
Full-resolution object detection
Input images must have channels in RGB order (not BGR) with pixel values normalized to [0, 1].
Paper: YOLO9000: Better, Faster, Stronger (Redmon & Farhadi, 2016)
Last updated
Was this helpful?