patchEmbeddingLayer
Description
A patch embedding layer maps patches of pixels to vectors. Use this layer in vision transformer neural networks to encode information about patches in images.
Creation
Syntax
Description
creates a patch embedding layer and sets the layer
= patchEmbeddingLayer(patchSize
,outputSize
)PatchSize
and
OutputSize
properties.
This feature requires a Deep Learning Toolbox™ license.
sets additional properties using one or more name-value arguments.layer
= patchEmbeddingLayer(patchSize
,outputSize
,Name=Value
)
Properties
Patch Embedding
PatchSize
— Size of patches to split input images into
positive integer | row vector of positive integers
This property is read-only.
Size of patches to split input images into, specified as a positive integer or row vector of positive integers.
If PatchSize
is a vector, then each element of
PatchSize
is the size of the patch in the corresponding spatial
dimension of the input. If PatchSize
is a scalar, then the layer
uses the same value for all spatial dimensions of the input.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
SpatialFlattenMode
— Mode for flattening output of convolution operation
"column-major"
(default) | "row-major"
Mode for flattening the output of the convolution operation, specified as
"column-major"
or "row-major"
.
If SpatialFlattenMode
is "column-major"
,
then the flatten operation outputs the data in its column-major representation. For
example, consider the input:
A = [ 1 2 3 4 5 6 7 8 9];
AFlat = [1 4 7 2 5 8 3 6 9];
If SpatialFlattenMode
is "row-major"
, then
the flatten operation outputs the data in its row-major representation. For example,
consider the input:
A = [ 1 2 3 4 5 6 7 8 9];
AFlat = [1 2 3 4 5 6 7 8 9];
Set this option when creating or importing models that require this representation.
OutputSize
— Size of output vectors
positive integer
This property is read-only.
Size of output vectors, specified as a positive integer.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
InputSize
— Number of input channels
"auto"
(default) | positive integer
This property is read-only.
Number of input channels, specified as one of these values:
"auto"
— Automatically determine the number of input channels at training time.Positive integer — Configure the layer for the specified number of input channels.
InputSize
and the number of channels in the layer input data must match. For example, if the input is an RGB image, thenInputSize
must be 3.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
| char
| string
Parameters and Initialization
WeightsInitializer
— Function to initialize weights
"glorot"
(default) | "he"
| "narrow-normal"
| "zeros"
| "ones"
| function handle
Function to initialize the weights, specified as one of these values:
"glorot"
— Initialize the weights with the Glorot initializer [1] [2] (also known as Xavier initializer). The Glorot initializer independently samples from a uniform distribution with zero mean and a variance of2/(numIn + numOut)
, wherenumIn
andnumOut
are the values of theInputSize
andOutputSize
properties, respectively."he"
– Initialize the weights with the He initializer [3]. The He initializer samples from a normal distribution with zero mean and a variance of2/numIn
, wherenumIn
is the value of theInputSize
property."narrow-normal"
— Initialize the weights by independently sampling from a normal distribution with zero mean and a standard deviation of 0.01."zeros"
— Initialize the weights with zeros."ones"
— Initialize the weights with ones.Function handle — Initialize the weights with a custom function. If you specify a function handle, then the function must have the form
weights = func(sz)
, wheresz
is the size of the weights.
The layer initializes the weights only when the Weights
property is empty.
Data Types: char
| string
| function_handle
BiasInitializer
— Function to initialize biases
"zeros"
(default) | "narrow-normal"
| "ones"
| function handle
Function to initialize the biases, specified as one of these values:
"zeros"
— Initialize the biases with zeros."ones"
— Initialize the biases with ones."narrow-normal"
— Initialize the biases by independently sampling from a normal distribution with a mean of zero and a standard deviation of 0.01.Function handle — Initialize the biases with a custom function. If you specify a function handle, then the function must have the form
bias = func(sz)
, wheresz
is the size of the biases.
The layer initializes the biases only when the Bias
property is
empty.
The PatchEmbeddingLayer
object stores this property as a character vector or a
function handle.
Data Types: char
| string
| function_handle
Weights
— Learnable weights
[]
(default) | numeric array
Learnable weights.
If PatchSize
is a positive integer, then
Weights
is an PatchSize
-by-...-by-PatchSize
-by-InputSize
-by-OutputSize
numeric array
or []
, where the number of dimensions of size
PatchSize
is the number of spatial dimensions of the
input.
If PatchSize
is a vector, then Weights
is an
PatchSize(1)
-by-...-by-PatchSize(K)
-by-InputSize
-by-OutputSize
numeric array
or []
, where K
is the number of spatial
dimensions of the input.
The layer weights are learnable parameters. You can specify the initial value of the weights
directly using the Weights
property of the layer. When
you train a network, if the Weights
property of the layer
is nonempty, then the trainnet
(Deep Learning Toolbox)
function uses the Weights
property as the initial value.
If the Weights
property is empty, then the software uses
the initializer specified by the WeightsInitializer
property of the layer.
Data Types: single
| double
Bias
— Layer biases
[]
(default) | column vector
Layer biases, specified as a numeric column vector of length
OutputSize
or []
.
The layer biases are learnable parameters. When you train a neural network, if Bias
is nonempty, then the trainnet
(Deep Learning Toolbox)
function uses the Bias
property as the initial value. If
Bias
is empty, then software uses the initializer
specified by BiasInitializer
.
Data Types: single
| double
Learning Rate and Regularization
WeightLearnRateFactor
— Learning rate factor for weights
1
(default) | nonnegative scalar
Learning rate factor for the weights, specified as a nonnegative scalar.
The software multiplies this factor by the global learning rate to determine the learning rate for the weights in this layer. For example, if WeightLearnRateFactor
is 2
, then the learning rate for the weights in this layer is twice the current global learning rate. The software determines the global learning rate based on the settings you specify using the trainingOptions
(Deep Learning Toolbox) function.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
BiasLearnRateFactor
— Learning rate factor for biases
1
(default) | nonnegative scalar
Learning rate factor for the biases, specified as a nonnegative scalar.
The software multiplies this factor by the global learning rate to determine the learning rate for the biases in this layer. For example, if BiasLearnRateFactor
is 2
, then the learning rate for the biases in the layer is twice the current global learning rate. The software determines the global learning rate based on the settings you specify using the trainingOptions
(Deep Learning Toolbox) function.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
WeightL2Factor
— L2 regularization factor for
weights
1 (default) | nonnegative scalar
L2 regularization factor for the weights, specified as a nonnegative scalar.
The software multiplies this factor by the global L2 regularization factor to determine the L2 regularization for the weights in this layer. For example, if WeightL2Factor
is 2
, then the L2 regularization for the weights in this layer is twice the global L2 regularization factor. You can specify the global L2 regularization factor using the trainingOptions
(Deep Learning Toolbox) function.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
BiasL2Factor
— L2 regularization factor for biases
0
(default) | nonnegative scalar
L2 regularization factor for the biases, specified as a nonnegative scalar.
The software multiplies this factor by the global L2 regularization factor to determine the L2 regularization for the biases in this layer. For example, if BiasL2Factor
is 2
, then the L2 regularization for the biases in this layer is twice the global L2 regularization factor. The software determines the global L2 regularization factor based on the settings you specify using the trainingOptions
(Deep Learning Toolbox) function.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
Layer
Name
— Layer name
""
(default) | character vector | string scalar
Layer name, specified as a character vector or string scalar.
For Layer
array input, the trainnet
(Deep Learning Toolbox) and
dlnetwork
(Deep Learning Toolbox) functions automatically assign
names to layers with the name ""
.
The PatchEmbeddingLayer
object stores this property as a character vector.
Data Types: char
| string
NumInputs
— Number of inputs
1
(default)
This property is read-only.
Number of inputs to the layer, returned as 1
. This layer accepts a
single input only.
Data Types: double
InputNames
— Input names
{'in'}
(default)
This property is read-only.
Input names, returned as {'in'}
. This layer accepts a single input
only.
Data Types: cell
NumOutputs
— Number of outputs
1
(default)
This property is read-only.
Number of outputs from the layer, returned as 1
. This layer has a
single output only.
Data Types: double
OutputNames
— Output names
{'out'}
(default)
This property is read-only.
Output names, returned as {'out'}
. This layer has a single output
only.
Data Types: cell
Examples
Create Patch Embedding Layer
Create a patch embedding layer that embeds patches of size 16 with an output size of 768.
patchSize = 16; embeddingOutputSize = 768; layer = patchEmbeddingLayer(patchSize,embeddingOutputSize)
layer = PatchEmbeddingLayer with properties: Name: '' PatchSize: 16 InputSize: 'auto' OutputSize: 768 SpatialFlattenMode: 'column-major' WeightsInitializer: 'glorot' BiasInitializer: 'zeros' WeightLearnRateFactor: 1 BiasLearnRateFactor: 1 WeightL2Factor: 1 BiasL2Factor: 1 Learnable Parameters Weights: [] Bias: [] State Parameters No properties. Use properties method to see a list of all properties.
Create a dlnetwork
object.
net = dlnetwork;
Specify layers of the network, including a patch embedding layer.
inputSize = [384 384 3]; maxPosition = (inputSize(1)/patchSize)^2 + 1; numHeads = 4; numKeyChannels = 4*embeddingOutputSize; numClasses = 1000; layers = [ imageInputLayer(inputSize) patchEmbeddingLayer(patchSize,embeddingOutputSize,Name="patch-emb") embeddingConcatenationLayer(Name="emb-cat") positionEmbeddingLayer(embeddingOutputSize,maxPosition,Name="pos-emb"); additionLayer(2,Name="add") selfAttentionLayer(numHeads,numKeyChannels,AttentionMask="causal") indexing1dLayer(Name="idx-first") fullyConnectedLayer(numClasses) softmaxLayer]; net = addLayers(net,layers);
Connect the embedding concatenation layer with the "in2" input of the addition layer.
net = connectLayers(net,"emb-cat","add/in2");
View the neural network architecture.
plot(net)
Algorithms
Patch Embedding Layer
A patch embedding layer maps patches of pixels to vectors. You can use this layer in vision transformer neural networks to encode information about patches in images.
The layer uses a convolution operation with the layer weights and biases to extract and project patches from the input. In particular, the layer:
Splits the input into non-overlapping patches.
Flattens the patches.
Projects the flattened patches to the output size.
Flattens the spatial dimensions of the projected output.
Layer Input and Output Formats
Layers in a layer array or layer graph pass data to subsequent layers as formatted dlarray
(Deep Learning Toolbox) objects.
The format of a dlarray
object is a string of characters in which each
character describes the corresponding dimension of the data. The formats consist of one or
more of these characters:
"S"
— Spatial"C"
— Channel"B"
— Batch"T"
— Time"U"
— Unspecified
For example, you can describe 2-D image data that is represented as a 4-D array, where the
first two dimensions correspond to the spatial dimensions of the images, the third
dimension corresponds to the channels of the images, and the fourth dimension
corresponds to the batch dimension, as having the format "SSCB"
(spatial, spatial, channel, batch).
You can interact with these dlarray
objects in automatic differentiation
workflows, such as those for developing a custom layer, using a functionLayer
(Deep Learning Toolbox)
object, or using the forward
(Deep Learning Toolbox) and predict
(Deep Learning Toolbox) functions with
dlnetwork
objects.
This table shows the supported input formats of PatchEmbeddingLayer
objects and the
corresponding output format. If the software passes the output of the layer to a custom
layer that does not inherit from the nnet.layer.Formattable
class, or a
FunctionLayer
object with the Formattable
property
set to 0
(false
), then the layer receives an
unformatted dlarray
object with dimensions ordered according to the formats
in this table. The formats listed here are only a subset. The layer may support additional
formats such as formats with additional "S"
(spatial) or
"U"
(unspecified) dimensions.
Input Format | Output Format |
---|---|
"SCB" (spatial, channel, batch) | "SCB" (spatial, channel, batch) |
"SSCB" (spatial, spatial, channel, batch) | "SCB" (spatial, channel, batch) |
"SSSCB" (spatial, spatial, spatial, channel,
batch) | "SCB" (spatial, channel, batch) |
"SCBT" (spatial, channel, batch, time) | "SCBT" (spatial, channel, batch, time) |
"SSCBT" (spatial, spatial, channel, batch, time) | "SCBT" (spatial, channel, batch, time) |
"SSSCBT" (spatial, spatial, spatial, channel, batch,
time) | "SCBT" (spatial, channel, batch, time) |
"SC" (spatial, channel) | "SC" (spatial, channel) |
"SSC" (spatial, spatial, channel) | "SC" (spatial, channel) |
"SSSC" (spatial, spatial, spatial, channel) | "SC" (spatial, channel) |
In dlnetwork
objects, PatchEmbeddingLayer
objects also support
these input and output format combinations.
Input Format | Output Format |
---|---|
"SCT" (spatial, channel, time) | "SCT" (spatial, channel, time) |
"SSCT" (spatial, spatial, channel, time) | "SCT" (spatial, channel, time) |
"SSSCT" (spatial, spatial, spatial, channel, time) | "SCT" (spatial, channel, time) |
References
[1] Dosovitskiy, Alexey, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani et al. "An Image is Worth 16x16 words: Transformers for Image Recognition at Scale." Preprint, submitted June 3, 2021. https://doi.org/10.48550/arXiv.2010.11929.
[2] Glorot, Xavier, and Yoshua Bengio. "Understanding the Difficulty of Training Deep Feedforward Neural Networks." In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, 249–356. Sardinia, Italy: AISTATS, 2010. https://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf
[3] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification." In 2015 IEEE International Conference on Computer Vision (ICCV), 1026–34. Santiago, Chile: IEEE, 2015. https://doi.org/10.1109/ICCV.2015.123
Extended Capabilities
C/C++ Code Generation
Generate C and C++ code using MATLAB® Coder™.
Usage notes and limitations:
Code generation supports only 1-D and 2-D spatial data. 3-D spatial or more than 3-D spatial data format such as "SSS" or "SSSS" is not supported.
You can generate generic C/C++ code that does not depend on third-party libraries and deploy the generated code to hardware platforms.
GPU Code Generation
Generate CUDA® code for NVIDIA® GPUs using GPU Coder™.
Usage notes and limitations:
Code generation supports only 1-D and 2-D spatial data. 3-D spatial or more than 3-D spatial data format such as "SSS" or "SSSS" is not supported.
You can generate CUDA code that is independent of deep learning libraries and deploy the generated code to platforms that use NVIDIA® GPU processors.
Version History
Introduced in R2023bR2024a: Specify spatial flattening mode of patch embedding layers
Specify the mode for flattening output of the convolution operation using the SpatialFlattenMode
option. Set this option when creating or importing models
that require this representation.
See Also
visionTransformer
| selfAttentionLayer
(Deep Learning Toolbox) | embeddingConcatenationLayer
(Deep Learning Toolbox) | indexing1dLayer
(Deep Learning Toolbox) | trainnet
(Deep Learning Toolbox) | trainingOptions
(Deep Learning Toolbox) | dlnetwork
(Deep Learning Toolbox)
Topics
- Train Vision Transformer Network for Image Classification
- Deep Learning in MATLAB (Deep Learning Toolbox)
- List of Deep Learning Layers (Deep Learning Toolbox)
- Deep Learning Tips and Tricks (Deep Learning Toolbox)
- Data Sets for Deep Learning (Deep Learning Toolbox)
MATLAB Command
You clicked a link that corresponds to this MATLAB command:
Run the command by entering it in the MATLAB Command Window. Web browsers do not support MATLAB commands.
Select a Web Site
Choose a web site to get translated content where available and see local events and offers. Based on your location, we recommend that you select: .
You can also select a web site from the following list:
How to Get Best Site Performance
Select the China site (in Chinese or English) for best site performance. Other MathWorks country sites are not optimized for visits from your location.
Americas
- América Latina (Español)
- Canada (English)
- United States (English)
Europe
- Belgium (English)
- Denmark (English)
- Deutschland (Deutsch)
- España (Español)
- Finland (English)
- France (Français)
- Ireland (English)
- Italia (Italiano)
- Luxembourg (English)
- Netherlands (English)
- Norway (English)
- Österreich (Deutsch)
- Portugal (English)
- Sweden (English)
- Switzerland
- United Kingdom (English)