calibrate
Description
exercises the network and collects the dynamic ranges of the weights and biases in the
convolution and fully connected layers of the network and the dynamic ranges of the
activations in all layers of the network specified by calResults
= calibrate(quantObj
,calData
)dlquantizer
object,
quantObj
, using the data specified by
calData
.
calibrates the network with additional options specified by one or more name-value pair
arguments.calResults
= calibrate(quantObj
,calData
,Name,Value
)
This function requires Deep Learning Toolbox Model Quantization Library. To learn about the products required to quantize a deep neural network, see Quantization Workflow Prerequisites.
Examples
Quantize a Neural Network for GPU Target
This example shows how to quantize learnable parameters in the convolution layers of a neural network for GPU and explore the behavior of the quantized network. In this example, you quantize the squeezenet neural network after retraining the network to classify new images. In this example, the memory required for the network is reduced approximately 75% through quantization while the accuracy of the network is not affected.
Load the pretrained network. net
is the output network of the Train Deep Learning Network to Classify New Images example.
load squeezedlnetmerch
net
net = dlnetwork with properties: Layers: [67×1 nnet.cnn.layer.Layer] Connections: [74×2 table] Learnables: [52×3 table] State: [0×3 table] InputNames: {'data'} OutputNames: {'prob'} Initialized: 1 View summary with summary.
Define calibration and validation data to use for quantization.
The calibration data is used to collect the dynamic ranges of the weights and biases in the convolution and fully connected layers of the network and the dynamic ranges of the activations in all layers of the network. For the best quantization results, the calibration data must be representative of inputs to the network.
The validation data is used to test the network after quantization to understand the effects of the limited range and precision of the quantized convolution layers in the network.
In this example, use the images in the MerchData
data set. Define an augmentedImageDatastore
object to resize the data for the network. Then, split the data into calibration and validation data sets.
unzip('MerchData.zip'); imds = imageDatastore('MerchData', ... 'IncludeSubfolders',true, ... 'LabelSource','foldernames'); classes = categories(imds.Labels); [calData, valData] = splitEachLabel(imds, 0.7, 'randomized'); aug_calData = augmentedImageDatastore([227 227], calData); aug_valData = augmentedImageDatastore([227 227], valData);
Create a dlquantizer
object and specify the network to quantize.
dlquantObj = dlquantizer(net);
Specify the GPU target.
quantOpts = dlquantizationOptions(Target='gpu');
quantOpts.MetricFcn = {@(x)hAccuracy(x,net,aug_valData,classes)}
quantOpts = dlquantizationOptions with properties: Validation Metric Info MetricFcn: {[@(x)hAccuracy(x,net,aug_valData,classes)]} Validation Environment Info Target: 'gpu' Bitstream: ''
Use the calibrate
function to exercise the network with sample inputs and collect range information. The calibrate
function exercises the network and collects the dynamic ranges of the weights and biases in the convolution and fully connected layers of the network and the dynamic ranges of the activations in all layers of the network. The function returns a table. Each row of the table contains range information for a learnable parameter of the optimized network.
calResults = calibrate(dlquantObj, aug_calData)
calResults=120×5 table
Optimized Layer Name Network Layer Name Learnables / Activations MinValue MaxValue
____________________________ ____________________ ________________________ _________ ________
{'conv1_Weights' } {'conv1' } "Weights" -0.91985 0.88489
{'conv1_Bias' } {'conv1' } "Bias" -0.07925 0.26343
{'fire2-squeeze1x1_Weights'} {'fire2-squeeze1x1'} "Weights" -1.38 1.2477
{'fire2-squeeze1x1_Bias' } {'fire2-squeeze1x1'} "Bias" -0.11641 0.24273
{'fire2-expand1x1_Weights' } {'fire2-expand1x1' } "Weights" -0.7406 0.90982
{'fire2-expand1x1_Bias' } {'fire2-expand1x1' } "Bias" -0.060056 0.14602
{'fire2-expand3x3_Weights' } {'fire2-expand3x3' } "Weights" -0.74397 0.66905
{'fire2-expand3x3_Bias' } {'fire2-expand3x3' } "Bias" -0.051778 0.074239
{'fire3-squeeze1x1_Weights'} {'fire3-squeeze1x1'} "Weights" -0.7712 0.68917
{'fire3-squeeze1x1_Bias' } {'fire3-squeeze1x1'} "Bias" -0.10138 0.32675
{'fire3-expand1x1_Weights' } {'fire3-expand1x1' } "Weights" -0.72035 0.9743
{'fire3-expand1x1_Bias' } {'fire3-expand1x1' } "Bias" -0.067029 0.30425
{'fire3-expand3x3_Weights' } {'fire3-expand3x3' } "Weights" -0.61443 0.7741
{'fire3-expand3x3_Bias' } {'fire3-expand3x3' } "Bias" -0.053613 0.10329
{'fire4-squeeze1x1_Weights'} {'fire4-squeeze1x1'} "Weights" -0.7422 1.0877
{'fire4-squeeze1x1_Bias' } {'fire4-squeeze1x1'} "Bias" -0.10885 0.13881
⋮
Use the validate
function to quantize the learnable parameters in the convolution layers of the network and exercise the network. The function uses the metric function defined in the dlquantizationOptions
object to compare the results of the network before and after quantization.
valResults = validate(dlquantObj, aug_valData, quantOpts)
valResults = struct with fields:
NumSamples: 20
MetricResults: [1×1 struct]
Statistics: [2×2 table]
Examine the validation output to see the performance of the quantized network.
valResults.MetricResults.Result
ans=2×2 table
NetworkImplementation MetricOutput
_____________________ ____________
{'Floating-Point'} 1
{'Quantized' } 1
valResults.Statistics
ans=2×2 table
NetworkImplementation LearnableParameterMemory(bytes)
_____________________ _______________________________
{'Floating-Point'} 2.9003e+06
{'Quantized' } 7.3393e+05
In this example, the memory required for the network was reduced approximately 75% through quantization. The accuracy of the network is not affected.
The weights, biases, and activations of the convolution layers of the network specified in the dlquantizer object now use scaled 8-bit integer data types.
Quantize Network for FPGA Deployment
This example uses:
- Deep Learning HDL ToolboxDeep Learning HDL Toolbox
- Deep Learning HDL Toolbox Support Package for Xilinx FPGA and SoC DevicesDeep Learning HDL Toolbox Support Package for Xilinx FPGA and SoC Devices
- Deep Learning ToolboxDeep Learning Toolbox
- Deep Learning Toolbox Model Quantization LibraryDeep Learning Toolbox Model Quantization Library
- MATLAB Coder Interface for Deep LearningMATLAB Coder Interface for Deep Learning
Reduce the memory footprint of a deep neural network by quantizing the weights, biases, and activations of convolution layers to 8-bit scaled integer data types. This example shows how to use Deep Learning Toolbox Model Quantization Library and Deep Learning HDL Toolbox to deploy the int8
network to a target FPGA board.
For this example, you need:
Deep Learning Toolbox™
Deep Learning HDL Toolbox™
Deep Learning Toolbox Model Quantization Library
Deep Learning HDL Toolbox Support Package for Xilinx® FPGA and SoC Devices
MATLAB® Coder™ Interface for Deep Learning.
Load Pretrained Network
Load the pretrained LogoNet network and analyze the network architecture.
snet = getLogoNetwork; deepNetworkDesigner(snet);
Set random number generator for reproducibility.
rng(0);
Load Data
This example uses the logos_dataset data set. The data set consists of 320 images. Each image is 227-by-227 in size and has three color channels (RGB). Create an augmentedImageDatastore
object for calibration and validation.
curDir = pwd; unzip("logos_dataset.zip"); imageData = imageDatastore(fullfile(curDir,'logos_dataset'),... 'IncludeSubfolders',true,'FileExtensions','.JPG','LabelSource','foldernames'); [calibrationData, validationData] = splitEachLabel(imageData, 0.5,'randomized');
Generate Calibration Result File for the Network
Create a dlquantizer
(Deep Learning HDL Toolbox) object and specify the network to quantize. Specify the execution environment as FPGA.
dlQuantObj = dlquantizer(snet,'ExecutionEnvironment',"FPGA");
Use the calibrate
(Deep Learning HDL Toolbox) function to exercise the network with sample inputs and collect the range information. The calibrate
function collects the dynamic ranges of the weights and biases. The calibrate function returns a table. Each row of the table contains range information for a learnable parameter of the quantized network.
calibrate(dlQuantObj,calibrationData)
ans=35×5 table
Optimized Layer Name Network Layer Name Learnables / Activations MinValue MaxValue
____________________________ __________________ ________________________ ___________ __________
{'conv_1_Weights' } {'conv_1' } "Weights" -0.048978 0.039352
{'conv_1_Bias' } {'conv_1' } "Bias" 0.99996 1.0028
{'conv_2_Weights' } {'conv_2' } "Weights" -0.055518 0.061901
{'conv_2_Bias' } {'conv_2' } "Bias" -0.00061171 0.00227
{'conv_3_Weights' } {'conv_3' } "Weights" -0.045942 0.046927
{'conv_3_Bias' } {'conv_3' } "Bias" -0.0013998 0.0015218
{'conv_4_Weights' } {'conv_4' } "Weights" -0.045967 0.051
{'conv_4_Bias' } {'conv_4' } "Bias" -0.00164 0.0037892
{'fc_1_Weights' } {'fc_1' } "Weights" -0.051394 0.054344
{'fc_1_Bias' } {'fc_1' } "Bias" -0.00052319 0.00084454
{'fc_2_Weights' } {'fc_2' } "Weights" -0.05016 0.051557
{'fc_2_Bias' } {'fc_2' } "Bias" -0.0017564 0.0018502
{'fc_3_Weights' } {'fc_3' } "Weights" -0.050706 0.04678
{'fc_3_Bias' } {'fc_3' } "Bias" -0.02951 0.024855
{'imageinput' } {'imageinput'} "Activations" 0 255
{'imageinput_normalization'} {'imageinput'} "Activations" -139.34 198.72
⋮
Create Target Object
Create a target object with a custom name for your target device and an interface to connect your target device to the host computer. Interface options are JTAG and Ethernet. Interface options are JTAG and Ethernet. To use JTAG, install Xilinx Vivado® Design Suite 2022.1. To set the Xilinx Vivado toolpath, enter:
hdlsetuptoolpath('ToolName', 'Xilinx Vivado', 'ToolPath', 'C:\Xilinx\Vivado\2022.1\bin\vivado.bat');
To create the target object, enter:
hTarget = dlhdl.Target('Xilinx','Interface','Ethernet','IPAddress','10.10.10.15');
Alternatively, you can also use the JTAG interface.
% hTarget = dlhdl.Target('Xilinx', 'Interface', 'JTAG');
Create dlQuantizationOptions
Object
Create a dlquantizationOptions
object. Specify the target bitstream and target board interface. The default metric function is a Top-1 accuracy metric function.
options_FPGA = dlquantizationOptions('Bitstream','zcu102_int8','Target',hTarget); options_emulation = dlquantizationOptions('Target','host');
To use a custom metric function, specify the metric function in the dlquantizationOptions
object.
options_FPGA = dlquantizationOptions('MetricFcn',{@(x)hComputeAccuracy(x,snet,validationData)},'Bitstream','zcu102_int8','Target',hTarget); options_emulation = dlquantizationOptions('MetricFcn',{@(x)hComputeAccuracy(x,snet,validationData)})
Validate Quantized Network
Use the validate
function to quantize the learnable parameters in the convolution layers of the network. The validate
function simulates the quantized network in MATLAB. The validate
function uses the metric function defined in the dlquantizationOptions
object to compare the results of the single-data-type network object to the results of the quantized network object.
prediction_emulation = dlQuantObj.validate(validationData,options_emulation)
prediction_emulation = struct with fields:
NumSamples: 160
MetricResults: [1×1 struct]
Statistics: []
For validation on an FPGA, the validate function:
Programs the FPGA board by using the output of the
compile
method and the programming fileDownloads the network weights and biases
Compares the performance of the network before and after quantization
prediction_FPGA = dlQuantObj.validate(validationData,options_FPGA)
### Compiling network for Deep Learning FPGA prototyping ... ### Targeting FPGA bitstream zcu102_int8. ### The network includes the following layers: 1 'imageinput' Image Input 227×227×3 images with 'zerocenter' normalization and 'randfliplr' augmentations (SW Layer) 2 'conv_1' 2-D Convolution 96 5×5×3 convolutions with stride [1 1] and padding [0 0 0 0] (HW Layer) 3 'relu_1' ReLU ReLU (HW Layer) 4 'maxpool_1' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 5 'conv_2' 2-D Convolution 128 3×3×96 convolutions with stride [1 1] and padding [0 0 0 0] (HW Layer) 6 'relu_2' ReLU ReLU (HW Layer) 7 'maxpool_2' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 8 'conv_3' 2-D Convolution 384 3×3×128 convolutions with stride [1 1] and padding [0 0 0 0] (HW Layer) 9 'relu_3' ReLU ReLU (HW Layer) 10 'maxpool_3' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 11 'conv_4' 2-D Convolution 128 3×3×384 convolutions with stride [2 2] and padding [0 0 0 0] (HW Layer) 12 'relu_4' ReLU ReLU (HW Layer) 13 'maxpool_4' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 14 'fc_1' Fully Connected 2048 fully connected layer (HW Layer) 15 'relu_5' ReLU ReLU (HW Layer) 16 'fc_2' Fully Connected 2048 fully connected layer (HW Layer) 17 'relu_6' ReLU ReLU (HW Layer) 18 'fc_3' Fully Connected 32 fully connected layer (HW Layer) 19 'softmax' Softmax softmax (SW Layer) 20 'classoutput' Classification Output crossentropyex with 'adidas' and 31 other classes (SW Layer) ### Notice: The layer 'imageinput' with type 'nnet.cnn.layer.ImageInputLayer' is implemented in software. ### Notice: The layer 'softmax' with type 'nnet.cnn.layer.SoftmaxLayer' is implemented in software. ### Notice: The layer 'classoutput' with type 'nnet.cnn.layer.ClassificationOutputLayer' is implemented in software. ### Compiling layer group: conv_1>>relu_4 ... ### Compiling layer group: conv_1>>relu_4 ... complete. ### Compiling layer group: maxpool_4 ... ### Compiling layer group: maxpool_4 ... complete. ### Compiling layer group: fc_1>>fc_3 ... ### Compiling layer group: fc_1>>fc_3 ... complete. ### Allocating external memory buffers: offset_name offset_address allocated_space _______________________ ______________ ________________ "InputDataOffset" "0x00000000" "11.9 MB" "OutputResultOffset" "0x00be0000" "128.0 kB" "SchedulerDataOffset" "0x00c00000" "128.0 kB" "SystemBufferOffset" "0x00c20000" "9.9 MB" "InstructionDataOffset" "0x01600000" "4.6 MB" "ConvWeightDataOffset" "0x01aa0000" "8.2 MB" "FCWeightDataOffset" "0x022e0000" "10.4 MB" "EndOffset" "0x02d40000" "Total: 45.2 MB" ### Network compilation complete. ### FPGA bitstream programming has been skipped as the same bitstream is already loaded on the target FPGA. ### Deep learning network programming has been skipped as the same network is already loaded on the target FPGA. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Finished writing input activations. ### Running single input activation. ### Notice: The layer 'imageinput' of type 'ImageInputLayer' is split into an image input layer 'imageinput' and an addition layer 'imageinput_norm' for normalization on hardware. ### The network includes the following layers: 1 'imageinput' Image Input 227×227×3 images with 'zerocenter' normalization and 'randfliplr' augmentations (SW Layer) 2 'conv_1' 2-D Convolution 96 5×5×3 convolutions with stride [1 1] and padding [0 0 0 0] (HW Layer) 3 'relu_1' ReLU ReLU (HW Layer) 4 'maxpool_1' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 5 'conv_2' 2-D Convolution 128 3×3×96 convolutions with stride [1 1] and padding [0 0 0 0] (HW Layer) 6 'relu_2' ReLU ReLU (HW Layer) 7 'maxpool_2' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 8 'conv_3' 2-D Convolution 384 3×3×128 convolutions with stride [1 1] and padding [0 0 0 0] (HW Layer) 9 'relu_3' ReLU ReLU (HW Layer) 10 'maxpool_3' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 11 'conv_4' 2-D Convolution 128 3×3×384 convolutions with stride [2 2] and padding [0 0 0 0] (HW Layer) 12 'relu_4' ReLU ReLU (HW Layer) 13 'maxpool_4' 2-D Max Pooling 3×3 max pooling with stride [2 2] and padding [0 0 0 0] (HW Layer) 14 'fc_1' Fully Connected 2048 fully connected layer (HW Layer) 15 'relu_5' ReLU ReLU (HW Layer) 16 'fc_2' Fully Connected 2048 fully connected layer (HW Layer) 17 'relu_6' ReLU ReLU (HW Layer) 18 'fc_3' Fully Connected 32 fully connected layer (HW Layer) 19 'softmax' Softmax softmax (SW Layer) 20 'classoutput' Classification Output crossentropyex with 'adidas' and 31 other classes (SW Layer) ### Notice: The layer 'softmax' with type 'nnet.cnn.layer.SoftmaxLayer' is implemented in software. ### Notice: The layer 'classoutput' with type 'nnet.cnn.layer.ClassificationOutputLayer' is implemented in software. Deep Learning Processor Estimator Performance Results LastFrameLatency(cycles) LastFrameLatency(seconds) FramesNum Total Latency Frames/s ------------- ------------- --------- --------- --------- Network 39136574 0.17789 1 39136574 5.6 imageinput_norm 216472 0.00098 conv_1 6832680 0.03106 maxpool_1 3705912 0.01685 conv_2 10454501 0.04752 maxpool_2 1173810 0.00534 conv_3 9364533 0.04257 maxpool_3 1229970 0.00559 conv_4 1759348 0.00800 maxpool_4 24450 0.00011 fc_1 2651288 0.01205 fc_2 1696632 0.00771 fc_3 26978 0.00012 * The clock frequency of the DL processor is: 220MHz ### Finished writing input activations. ### Running single input activation.
prediction_FPGA = struct with fields:
NumSamples: 160
MetricResults: [1×1 struct]
Statistics: [2×7 table]
View Performance of Quantized Neural Network
Display the accuracy of the quantized network.
prediction_emulation.MetricResults.Result
ans=2×2 table
NetworkImplementation MetricOutput
_____________________ ____________
{'Floating-Point'} 0.9875
{'Quantized' } 0.9875
prediction_FPGA.MetricResults.Result
ans=2×2 table
NetworkImplementation MetricOutput
_____________________ ____________
{'Floating-Point'} 0.9875
{'Quantized' } 0.9875
Display the performance of the quantized network in frames per second.
prediction_FPGA.Statistics
ans=2×7 table
NetworkImplementation FramesPerSecond Number of Threads (Convolution) Number of Threads (Fully Connected) LUT Utilization (%) BlockRAM Utilization (%) DSP Utilization (%)
_____________________ _______________ _______________________________ ___________________________________ ___________________ ________________________ ___________________
{'Floating-Point'} 5.6213 16 4 93.198 63.925 15.595
{'Quantized' } 19.433 64 16 62.31 50.11 32.103
Quantize a Neural Network for CPU Target
This example uses:
- Deep Learning ToolboxDeep Learning Toolbox
- Deep Learning Toolbox Model Quantization LibraryDeep Learning Toolbox Model Quantization Library
- MATLAB CoderMATLAB Coder
- MATLAB Support Package for Raspberry Pi HardwareMATLAB Support Package for Raspberry Pi Hardware
- Embedded CoderEmbedded Coder
- MATLAB Coder Interface for Deep LearningMATLAB Coder Interface for Deep Learning
This example shows how to quantize and validate a neural network for a CPU target. This workflow is similar to other execution environments, but before validating you must establish a raspi
connection and specify it as target using dlquantizationOptions
.
First, load your network. This example uses the pretrained network squeezenet
.
load squeezedlnetmerch
net
net = dlnetwork with properties: Layers: [67×1 nnet.cnn.layer.Layer] Connections: [74×2 table] Learnables: [52×3 table] State: [0×3 table] InputNames: {'data'} OutputNames: {'prob'} Initialized: 1 View summary with summary.
Then define your calibration and validation data, calDS
and valDS
respectively.
unzip('MerchData.zip'); imds = imageDatastore('MerchData', ... 'IncludeSubfolders',true, ... 'LabelSource','foldernames'); classes = categories(imds.Labels); [calData, valData] = splitEachLabel(imds, 0.7, 'randomized'); aug_calData = augmentedImageDatastore([227 227],calData); aug_valData = augmentedImageDatastore([227 227],valData);
Create the dlquantizer
object and specify a CPU execution environment.
dq = dlquantizer(net,'ExecutionEnvironment','CPU')
dq = dlquantizer with properties: NetworkObject: [1×1 dlnetwork] ExecutionEnvironment: 'CPU'
Calibrate the network.
calResults = calibrate(dq,aug_calData,'UseGPU','off')
calResults=120×5 table
Optimized Layer Name Network Layer Name Learnables / Activations MinValue MaxValue
__________________________ ____________________ ________________________ _________ ________
"conv1_Weights" {'conv1' } "Weights" -0.91985 0.88489
"conv1_Bias" {'conv1' } "Bias" -0.07925 0.26343
"fire2-squeeze1x1_Weights" {'fire2-squeeze1x1'} "Weights" -1.38 1.2477
"fire2-squeeze1x1_Bias" {'fire2-squeeze1x1'} "Bias" -0.11641 0.24273
"fire2-expand1x1_Weights" {'fire2-expand1x1' } "Weights" -0.7406 0.90982
"fire2-expand1x1_Bias" {'fire2-expand1x1' } "Bias" -0.060056 0.14602
"fire2-expand3x3_Weights" {'fire2-expand3x3' } "Weights" -0.74397 0.66905
"fire2-expand3x3_Bias" {'fire2-expand3x3' } "Bias" -0.051778 0.074239
"fire3-squeeze1x1_Weights" {'fire3-squeeze1x1'} "Weights" -0.7712 0.68917
"fire3-squeeze1x1_Bias" {'fire3-squeeze1x1'} "Bias" -0.10138 0.32675
"fire3-expand1x1_Weights" {'fire3-expand1x1' } "Weights" -0.72035 0.9743
"fire3-expand1x1_Bias" {'fire3-expand1x1' } "Bias" -0.067029 0.30425
"fire3-expand3x3_Weights" {'fire3-expand3x3' } "Weights" -0.61443 0.7741
"fire3-expand3x3_Bias" {'fire3-expand3x3' } "Bias" -0.053613 0.10329
"fire4-squeeze1x1_Weights" {'fire4-squeeze1x1'} "Weights" -0.7422 1.0877
"fire4-squeeze1x1_Bias" {'fire4-squeeze1x1'} "Bias" -0.10885 0.13881
⋮
Use the MATLAB Support Package for Raspberry Pi Hardware function, raspi
, to create a connection to the Raspberry Pi. In the following code, replace:
raspiname
with the name or address of your Raspberry Piusername
with your user namepassword
with your password
% r = raspi('raspiname','username','password')
For example,
r = raspi('gpucoder-raspberrypi-8','pi','matlab')
r = raspi with properties: DeviceAddress: 'gpucoder-raspberrypi-8' Port: 18734 BoardName: 'Raspberry Pi 3 Model B+' AvailableLEDs: {'led0'} AvailableDigitalPins: [2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27] AvailableSPIChannels: {} AvailableI2CBuses: {} AvailableWebcams: {} I2CBusSpeed: AvailableCANInterfaces: {} Supported peripherals
Specify raspi
object as the target for the quantized network.
opts = dlquantizationOptions('Target',r);
opts.MetricFcn = {@(x)hAccuracy(x,net,aug_valData,classes)}
opts = dlquantizationOptions with properties: Validation Metric Info MetricFcn: {[@(x)hAccuracy(x,net,aug_valData,classes)]} Validation Environment Info Target: [1×1 raspi] Bitstream: ''
Validate the quantized network with the validate
function.
valResults = validate(dq,aug_valData,opts)
### Starting application: 'codegen/lib/validate_predict_int8/pil/validate_predict_int8.elf' To terminate execution: clear validate_predict_int8_pil ### Launching application validate_predict_int8.elf... ### Host application produced the following standard output (stdout) and standard error (stderr) messages:
valResults = struct with fields:
NumSamples: 20
MetricResults: [1×1 struct]
Statistics: []
Examine the validation output to see the performance of the quantized network.
valResults.MetricResults.Result
ans=2×2 table
NetworkImplementation MetricOutput
_____________________ ____________
{'Floating-Point'} 1
{'Quantized' } 1
Quantize YOLO v3 Object Detector
This example shows how to quantize a yolov3ObjectDetector
(Computer Vision Toolbox) object using preprocessed calibration and validation data.
First, download a pretrained YOLO v3 object detector.
detector = downloadPretrainedNetwork();
This example uses a small labeled data set that contains one or two labeled instances of a vehicle. Many of these images come from the Caltech Cars 1999 and 2001 data sets, created by Pietro Perona and used with permission.
Unzip the vehicle images and load the vehicle ground truth data.
unzip vehicleDatasetImages.zip data = load('vehicleDatasetGroundTruth.mat'); vehicleDataset = data.vehicleDataset;
Add the full path to the local vehicle data folder.
vehicleDataset.imageFilename = fullfile(pwd, vehicleDataset.imageFilename);
Create an imageDatastore
for loading the images and a boxLabelDatastore
(Computer Vision Toolbox) for the ground truth bounding boxes.
imds = imageDatastore(vehicleDataset.imageFilename); blds = boxLabelDatastore(vehicleDataset(:,2));
Use the combine
function to combine both the datastores into a CombinedDatastore
.
combinedDS = combine(imds, blds);
Split the data into calibration and validation data.
calData = combinedDS.subset(1:32); valData = combinedDS.subset(33:64);
Use the preprocess
(Computer Vision Toolbox) method of yolov3ObjectDetector
(Computer Vision Toolbox) object with transform
function to prepare the data for calibration and validation.
The transform
function returns a TransformedDatastore
object.
processedCalData = transform(calData, @(data)preprocess(detector,data)); processedValData = transform(valData, @(data)preprocess(detector,data));
Create the dlquantizer
object. When you use the MATLAB execution environment, quantization is performed using the fi
fixed-point data type which requires a Fixed-Point Designer™ license.
dq = dlquantizer(detector, 'ExecutionEnvironment', 'MATLAB');
Calibrate the network.
calResults = calibrate(dq, processedCalData,'UseGPU','off')
calResults=135×5 table
Optimized Layer Name Network Layer Name Learnables / Activations MinValue MaxValue
____________________________ ____________________ ________________________ _________ ________
{'conv1_Weights' } {'conv1' } "Weights" -0.92189 0.85687
{'conv1_Bias' } {'conv1' } "Bias" -0.096271 0.26628
{'fire2-squeeze1x1_Weights'} {'fire2-squeeze1x1'} "Weights" -1.3751 1.2444
{'fire2-squeeze1x1_Bias' } {'fire2-squeeze1x1'} "Bias" -0.12068 0.23104
{'fire2-expand1x1_Weights' } {'fire2-expand1x1' } "Weights" -0.75275 0.91615
{'fire2-expand1x1_Bias' } {'fire2-expand1x1' } "Bias" -0.059252 0.14035
{'fire2-expand3x3_Weights' } {'fire2-expand3x3' } "Weights" -0.75271 0.6774
{'fire2-expand3x3_Bias' } {'fire2-expand3x3' } "Bias" -0.062214 0.088242
{'fire3-squeeze1x1_Weights'} {'fire3-squeeze1x1'} "Weights" -0.7586 0.68772
{'fire3-squeeze1x1_Bias' } {'fire3-squeeze1x1'} "Bias" -0.10206 0.31645
{'fire3-expand1x1_Weights' } {'fire3-expand1x1' } "Weights" -0.71566 0.97678
{'fire3-expand1x1_Bias' } {'fire3-expand1x1' } "Bias" -0.069313 0.32881
{'fire3-expand3x3_Weights' } {'fire3-expand3x3' } "Weights" -0.60079 0.77642
{'fire3-expand3x3_Bias' } {'fire3-expand3x3' } "Bias" -0.058045 0.11229
{'fire4-squeeze1x1_Weights'} {'fire4-squeeze1x1'} "Weights" -0.738 1.0805
{'fire4-squeeze1x1_Bias' } {'fire4-squeeze1x1'} "Bias" -0.11189 0.13698
⋮
Validate the quantized network with the validate
function.
valResults = validate(dq, processedValData)
valResults = struct with fields:
NumSamples: 32
MetricResults: [1×1 struct]
Statistics: []
function detector = downloadPretrainedNetwork() pretrainedURL = 'https://ssd.mathworks.com/supportfiles/vision/data/yolov3SqueezeNetVehicleExample_21aSPKG.zip'; websave('yolov3SqueezeNetVehicleExample_21aSPKG.zip', pretrainedURL); unzip('yolov3SqueezeNetVehicleExample_21aSPKG.zip'); pretrained = load("yolov3SqueezeNetVehicleExample_21aSPKG.mat"); detector = pretrained.detector; end
Input Arguments
quantObj
— Network to quantize
dlquantizer
object
Network to quantize, specified as a dlquantizer
object.
calData
— Data to use for calibration of quantized network
imageDatastore
object | augmentedImageDatastore
object | pixelLabelImageDatastore
object | CombinedDatastore
object | TransformedDatastore
object | arrayDatastore
object | dlarray
object | numeric array
Data to use for calibration of quantized network, specified as one of these values:
imageDatastore
objectaugmentedImageDatastore
objectpixelLabelImageDatastore
(Computer Vision Toolbox) objectCombinedDatastore
objectTransformedDatastore
objectarrayDatastore
objectdlarray
objectnumeric array
When using a dlarray
object, format the dlarray
to
assure the data has the appropriate shape. Unformatted dlarray
objects
are treated in the same way as numeric arrays.
You must preprocess the data used for calibration of a yolov3ObjectDetector
(Computer Vision Toolbox) object using the preprocess
(Computer Vision Toolbox)
function. For an example of using preprocessed data for calibration of a
yolov3ObjectDetector
, see Quantize YOLO v3 Object Detector.
For more information on supported datastores, see Prepare Data for Quantizing Networks.
Data Types: single
| double
| int8
| int16
| int32
| uint8
| uint16
| uint32
Name-Value Arguments
Specify optional pairs of arguments as
Name1=Value1,...,NameN=ValueN
, where Name
is
the argument name and Value
is the corresponding value.
Name-value arguments must appear after other arguments, but the order of the
pairs does not matter.
Before R2021a, use commas to separate each name and value, and enclose
Name
in quotes.
Example: calResults =
calibrate(quantObj,calData,'UseGPU','on')
MiniBatchSize
— Size of mini-batches
32
(default) | positive integer
Size of the mini-batches to use for calibration, specified as a positive integer. Larger mini-batch sizes require more memory, but can lead to faster calibration.
Data Types: single
| double
| int8
| int16
| int32
| int64
| uint8
| uint16
| uint32
| uint64
UseGPU
— Whether to use host GPU for calibration
'auto' (default) | 'on'
| 'off
Whether to use host GPU for calibration, specified as one of the following:
'auto'
— Use host GPU for calibration if one is available. Otherwise, use host CPU for calibration.'on'
— Use host GPU for calibration.'off'
— Use host CPU for calibration.
Data Types: char
Output Arguments
calResults
— Dynamic ranges of network
table
Dynamic ranges of layers of the network, returned as a table. Each row in the table displays the minimum and maximum values of a learnable parameter of a convolution layer of the optimized network. The software uses these minimum and maximum values to determine the scaling for the data type of the quantized parameter.
Version History
Introduced in R2020aR2022b: Calibrate on host GPU or host CPU
You can now choose whether to calibrate your network using the host GPU or host CPU. By
default, the calibrate
function and the Deep Network
Quantizer app will calibrate on the host GPU if one is available.
In previous versions, it was required that the execution environment was the same as the instrumentation environment used for the calibration step of quantization.
R2022b: Specify mini-batch size to use for calibration
Use MiniBatchSize
to specify the size of mini-batches to use for
calibration.
R2021a: ARM Cortex-A calibration support
The Deep Learning Toolbox™ Model Quantization Library now supports calibration of a network for quantization and deployment on ARM® Cortex®-A microcontrollers.
See Also
Apps
Functions
validate
|dlquantizer
|dlquantizationOptions
|quantize
|quantizationDetails
|estimateNetworkMetrics
Topics
- Quantization Workflow Prerequisites
- Prepare Data for Quantizing Networks
- Quantization of Deep Neural Networks
- Quantize Residual Network Trained for Image Classification and Generate CUDA Code
- Quantize Network for FPGA Deployment (Deep Learning HDL Toolbox)
- Generate int8 Code for Deep Learning Networks (MATLAB Coder)
Comando MATLAB
Hai fatto clic su un collegamento che corrisponde a questo comando MATLAB:
Esegui il comando inserendolo nella finestra di comando MATLAB. I browser web non supportano i comandi MATLAB.
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)