Supported Networks, Layers, and Classes

GPU Coder™ supports code generation for series and directed acyclic graph (DAG) neural networks. This page provides a list of supported deep learning networks, layers and classes.

Note that the code generator represents characters in an 8-bit ASCII codeset that the locale setting determines. Therefore, the use of non-ASCII characters in class names, layer names, layer description, or network names might result in errors. For more information, see Encoding of Characters in Code Generation.

Supported Pretrained Networks

You can train a neural network on either a CPU, a GPU, or multiple GPUs by using the Deep Learning Toolbox™ or use one of the pretrained networks listed in the table and generate CUDA^® code.

These pretrained networks, available in Deep Learning Toolbox, are supported for code generation. You can use imagePretrainedNetwork (Deep Learning Toolbox) function to load these pretrained neural networks as dlnetwork (Deep Learning Toolbox) objects for code generation. Note that code generation does not support setting the name-value argument Weights of imagePretrainedNetwork function to "none". For example, use this code to load a pretrained GoogLeNet neural network.

net = imagePretrainedNetwork("googlenet")

`imagePretrainedNetwork` Model Name Argument	CUDA	TensorRT and cuDNN
`"alexnet"`
`"darknet19"`
`"darknet53"`
`"densenet201"`
`"efficientnetb0"`
`"googlenet"`
`"inceptionresnetv2"`
`"inceptionv3"`
`"mobilenetv2"`
`"nasnetlarge"`
`"nasnetmobile"`
`"resnet18"`
`"resnet50"`
`"resnet101"`
`"shufflenet"`
`"squeezenet"`
`"vgg16"`
`"vgg19"`
`"xception"`

Additionally, you can use these functions to create neural networks for code generation.

Network Name CUDA TensorRT and cuDNN

Caffe network

See importCaffeNetwork (Deep Learning Toolbox).

DeepLab v3+ network

deeplabv3plus (Computer Vision Toolbox)

Supported Layers

These layers are supported for code generation by GPU Coder for these target deep learning libraries.

Note

In the following tables, the information icon indicates that the network, layer, or class has limited code generation capabilities. You might see errors and unexpected behavior. For more information, see the Extended Capabilities section on the page for that network, layer, or class.

Input Layers

Layer Name	CUDA	TensorRT	cuDNN
`imageInputLayer` (Deep Learning Toolbox)
`sequenceInputLayer` (Deep Learning Toolbox)
`featureInputLayer` (Deep Learning Toolbox)

Convolution and Fully Connected Layers

Layer Name	CUDA	TensorRT and cuDNN
`convolution1dLayer` (Deep Learning Toolbox)
`convolution2dLayer` (Deep Learning Toolbox)
`fullyConnectedLayer` (Deep Learning Toolbox)
`groupedConvolution2dLayer` (Deep Learning Toolbox)
`transposedConv2dLayer` (Deep Learning Toolbox)

Sequence Layers

Layer Name	CUDA	TensorRT and cuDNN
`sequenceInputLayer` (Deep Learning Toolbox)
`bilstmLayer` (Deep Learning Toolbox)
`flattenLayer` (Deep Learning Toolbox)
`gruLayer` (Deep Learning Toolbox)
`lstmLayer` (Deep Learning Toolbox)
`wordEmbeddingLayer` (Text Analytics Toolbox)

Activation Layers

Layer Name	CUDA	TensorRT and cuDNN
`clippedReluLayer` (Deep Learning Toolbox)
`eluLayer` (Deep Learning Toolbox)
`leakyReluLayer` (Deep Learning Toolbox)
`preluLayer` (Deep Learning Toolbox)
`reluLayer` (Deep Learning Toolbox)
`gelu` (Deep Learning Toolbox)
`softplusLayer` (Reinforcement Learning Toolbox)
`swishLayer` (Deep Learning Toolbox)
`tanhLayer` (Deep Learning Toolbox)

Utility Layers

Layer Name	CUDA	TensorRT and cuDNN
`batchNormalizationLayer` (Deep Learning Toolbox)
`crossChannelNormalizationLayer` (Deep Learning Toolbox)
`groupNormalizationLayer` (Deep Learning Toolbox)
`instanceNormalizationLayer` (Deep Learning Toolbox)
`layerNormalizationLayer` (Deep Learning Toolbox)
`crop2dLayer` (Deep Learning Toolbox)
`dropoutLayer` (Deep Learning Toolbox)
`scalingLayer` (Reinforcement Learning Toolbox)
`spatialDropoutLayer` (Deep Learning Toolbox)

Pooling and Unpooling Layers

Layer Name	CUDA	TensorRT and cuDNN
`adaptiveAveragePooling2dLayer` (Deep Learning Toolbox)
`averagePooling1dLayer` (Deep Learning Toolbox)
`averagePooling2dLayer` (Deep Learning Toolbox)
`globalAveragePooling1dLayer` (Deep Learning Toolbox)
`globalAveragePooling2dLayer` (Deep Learning Toolbox)
`globalMaxPooling1dLayer` (Deep Learning Toolbox)
`globalMaxPooling2dLayer` (Deep Learning Toolbox)
`maxPooling1dLayer` (Deep Learning Toolbox)
`maxPooling2dLayer` (Deep Learning Toolbox)
`maxUnpooling2dLayer` (Deep Learning Toolbox)

Combination Layers

Layer Name	CUDA	TensorRT and cuDNN
`additionLayer` (Deep Learning Toolbox)
`concatenationLayer` (Deep Learning Toolbox)
`depthConcatenationLayer` (Deep Learning Toolbox)

Transformer Layers

Layer Name	CUDA	TensorRT and cuDNN
`AttentionLayer` (Deep Learning Toolbox)
`embeddingConcatenationLayer` (Deep Learning Toolbox)
`indexing1dLayer` (Deep Learning Toolbox)
`patchEmbeddingLayer` (Computer Vision Toolbox)
`PositionEmbeddingLayer` (Deep Learning Toolbox)
`selfAttentionLayer` (Deep Learning Toolbox)

Object Detection Layers

Layer Name	CUDA	TensorRT and cuDNN
`depthToSpace2dLayer` (Image Processing Toolbox)
`spaceToDepthLayer` (Image Processing Toolbox)
`ssdMergeLayer` (Computer Vision Toolbox)
`yolov2TransformLayer` (Computer Vision Toolbox)

Output Layers

Layer Name	CUDA	TensorRT and cuDNN
`classificationLayer` (Deep Learning Toolbox)
`regressionLayer` (Deep Learning Toolbox)
`sigmoidLayer` (Deep Learning Toolbox)
`softmaxLayer` (Deep Learning Toolbox)
`Custom Output Layer` (Deep Learning Toolbox) More information An output layer including custom classification or regression output layers created by `nnet.layer.ClassificationLayer` or `nnet.layer.RegressionLayer` functions. For an example showing how to define a custom classification output layer and specify a loss function, see Define Custom Classification Output Layer (Deep Learning Toolbox). For an example showing how to define a custom regression output layer and specify a loss function, see Define Custom Regression Output Layer (Deep Learning Toolbox).

Custom Keras Layers

Layer Name	CUDA	TensorRT and cuDNN
`nnet.keras.layer.ClipLayer` (Deep Learning Toolbox)
`nnet.keras.layer.FlattenCStyleLayer` (Deep Learning Toolbox)
`nnet.keras.layer.GlobalAveragePooling2dLayer` (Deep Learning Toolbox)
`nnet.keras.layer.PreluLayer` (Deep Learning Toolbox)
`nnet.keras.layer.SigmoidLayer` (Deep Learning Toolbox)
`nnet.keras.layer.TanhLayer` (Deep Learning Toolbox)
`nnet.keras.layer.TimeDistributedFlattenCStyleLayer` (Deep Learning Toolbox)
`nnet.keras.layer.ZeroPadding2dLayer` (Deep Learning Toolbox)

Custom ONNX Layers

Layer Name	CUDA	TensorRT and cuDNN
`nnet.onnx.layer.ClipLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.ElementwiseAffineLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.FlattenInto2dLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.FlattenLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.GlobalAveragePooling2dLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.IdentityLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.PreluLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.SigmoidLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.TanhLayer` (Deep Learning Toolbox)
`nnet.onnx.layer.VerifyBatchSizeLayer` (Deep Learning Toolbox)

Custom Layers

Layer Name CUDA TensorRT and cuDNN

Layer Name	CUDA	TensorRT and cuDNN
`Custom layers` More information Custom layers, with or without learnable parameters, that you define for your problem. To learn how to define custom deep learning layers, see Define Custom Deep Learning Layers (Deep Learning Toolbox) and Define Custom Deep Learning Layer for Code Generation (Deep Learning Toolbox). For an example on how to generate code for a network with custom layers, see Code Generation for Object Detection Using YOLO v3 Deep Learning Network. Code generation Limitations The outputs of the custom layer must be fixed-size arrays. Setting the `MallocMode` property to `"unified"` in the `coder.gpuConfig` requires extra memory copies, which leads to slower performance. It is recommended to use `"discrete"`. For more information on GPU memory allocation, see Discrete and Managed Modes cuDNN targets support both row-major and column-major code generation for custom layers. TensorRT targets support only column-major code generation. Custom layers must contain the `%#codegen` pragma. A sequence network that contains custom layers and LSTM or GRU layers is not supported. You can pass `dlarray` to custom layers if: The custom layer is in `dlnetwork`. The custom layer is in a DAG or series network and either inherits from `nnet.layer.Formattable` or has no backward propagation. For unsupported `dlarray` methods, you must extract the underlying data from the `dlarray`, perform the computations, and reconstruct the data into the `dlarray` object. For example: function Z = predict(layer, X) if coder.target('MATLAB') Z = doPredict(X); else if isdlarray(X) X1 = extractdata(X); Z1 = doPredict(X1); Z = dlarray(Z1); else Z = doPredict(X); end end end

Custom layers

More information

Code generation Limitations

The outputs of the custom layer must be fixed-size arrays.
Setting the MallocMode property to "unified" in the coder.gpuConfig requires extra memory copies, which leads to slower performance. It is recommended to use "discrete". For more information on GPU memory allocation, see Discrete and Managed Modes
cuDNN targets support both row-major and column-major code generation for custom layers. TensorRT targets support only column-major code generation.
Custom layers must contain the %#codegen pragma.
A sequence network that contains custom layers and LSTM or GRU layers is not supported.
You can pass dlarray to custom layers if:
- The custom layer is in dlnetwork.
- The custom layer is in a DAG or series network and either inherits from nnet.layer.Formattable or has no backward propagation.
For unsupported dlarray methods, you must extract the underlying data from the dlarray, perform the computations, and reconstruct the data into the dlarray object. For example:
```
function Z = predict(layer, X)

if coder.target('MATLAB')
   Z = doPredict(X);
else
   if isdlarray(X)
      X1 = extractdata(X);
      Z1 = doPredict(X1);
      Z = dlarray(Z1);
  else
      Z = doPredict(X);
  end
end

end
```

Supported Classes

These classes are supported for code generation by GPU Coder for these target deep learning libraries.

Name	CUDA	TensorRT and cuDNN
`dlnetwork` (Deep Learning Toolbox)
`DAGNetwork` (Deep Learning Toolbox)
`pointPillarsObjectDetector` (Lidar Toolbox)
`SeriesNetwork` (Deep Learning Toolbox)
`ssdObjectDetector` (Computer Vision Toolbox)
`yolov2ObjectDetector` (Computer Vision Toolbox)
`yolov3ObjectDetector` (Computer Vision Toolbox)
`yolov4ObjectDetector` (Computer Vision Toolbox)
`yoloxObjectDetector` (Computer Vision Toolbox)