dlquantizer

Quantize a deep neural network to 8-bit scaled integer data types

Description

Use the dlquantizer object to reduce the memory requirement of a deep neural network by quantizing weights, biases, and activations to 8-bit scaled integer data types. You can create and verify the behavior of a quantized network for GPU, FPGA, CPU deployment, or explore the quantized network in MATLAB^®.

For CPU and GPU deployment, the software generates code for a convolutional deep neural network by quantizing the weights, biases, and activations of the convolution layers to 8-bit scaled integer data types. The quantization is performed by providing the calibration result file produced by the calibrate function to the codegen (MATLAB Coder) command.

Code generation does not support quantized deep neural networks produced by the quantize function.

This object requires Deep Learning Toolbox Model Quantization Library. To learn about the products required to quantize a deep neural network, see Quantization Workflow Prerequisites.

Creation

Syntax

quantObj = dlquantizer(net)

quantObj = dlquantizer(net,Name,Value)

Description

quantObj = dlquantizer(net) creates a dlquantizer object for the specified deep neural network, net.

quantObj = dlquantizer(net,Name,Value) creates a dlquantizer object for the specified network, with additional options specified by one or more name-value pair arguments.

Input Arguments

net — Pretrained neural network

DAGNetwork object | dlnetwork object | SeriesNetwork object | yolov2ObjectDetector object | ssdObjectDetector object

Pretrained neural network, specified as a DAGNetwork, dlnetwork, SeriesNetwork, yolov2ObjectDetector (Computer Vision Toolbox), or a ssdObjectDetector (Computer Vision Toolbox) object.

Quantization of yolov2ObjectDetector (Computer Vision Toolbox) and ssdObjectDetector (Computer Vision Toolbox) networks requires a GPU Coder™ license.

Properties

NetworkObject — Pretrained neural network

DAGNetwork object | dlnetwork object | SeriesNetwork object | yolov2ObjectDetector object | ssdObjectDetector object

This property is read-only.

Pre-trained neural network, specified as a DAGNetwork, dlnetwork, SeriesNetwork, yolov2ObjectDetector (Computer Vision Toolbox), or a ssdObjectDetector (Computer Vision Toolbox) object.

ExecutionEnvironment — Execution environment

'GPU' (default) | 'FPGA' | 'CPU' | 'MATLAB'

Execution environment for the quantized network, specified as 'GPU', 'FPGA', 'CPU', or 'MATLAB'. How the network is quantized depends on the choice of execution environment.

The 'MATLAB' execution environment indicates a target-agnostic quantization of the neural network will be performed. This option does not require you to have target hardware in order to explore the quantized network in MATLAB.

Example: 'ExecutionEnvironment','FPGA'

Simulation — Validate behavior of network quantized for FPGA environment using simulation

'off' (default) | 'on'

Whether to validate behavior of network quantized for FPGA using simulation, specified as one of these values:

• 'on' — Validate the behavior of the quantized network by simulating the quantized network in MATLAB and comparing the prediction results of the original single-precision network to the simulated prediction results of the quantized network.

• 'off' — Generate code and validate the behavior of the quantized network on the target hardware.

Note

This option is only valid when ExecutionEnvironment is set to 'FPGA'.

Note

Alternatively, you can use the quantize method to create a simulatable quantized network. The simulatable quantized network enables visibility of the quantized layers, weights, and biases of the network, as well as simulatable quantized inference behavior.

Example: 'Simulation', 'on'

Object Functions

calibrate	Simulate and collect ranges of a deep neural network
validate	Quantize and validate a deep neural network
quantize	Quantize deep neural network

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 according to the Train Deep Learning Network to Classify New Images example. 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 squeezenetmerch
net

net = 
  DAGNetwork with properties:

         Layers: [68×1 nnet.cnn.layer.Layer]
    Connections: [75×2 table]
     InputNames: {'data'}
    OutputNames: {'new_classoutput'}

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');
[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.

quantObj = dlquantizer(net);

Define a metric function to use to compare the behavior of the network before and after quantization. This example uses the hComputeModelAccuracy metric function.

function accuracy = hComputeModelAccuracy(predictionScores, net, dataStore)
%% Computes model-level accuracy statistics
    
    % Load ground truth
    tmp = readall(dataStore);
    groundTruth = tmp.response;
    
    % Compare with predicted label with actual ground truth 
    predictionError = {};
    for idx=1:numel(groundTruth)
        [~, idy] = max(predictionScores(idx,:)); 
        yActual = net.Layers(end).Classes(idy);
        predictionError{end+1} = (yActual == groundTruth(idx)); %#ok
    end
    
    % Sum all prediction errors.
    predictionError = [predictionError{:}];
    accuracy = sum(predictionError)/numel(predictionError);
end

Specify the metric function in a dlquantizationOptions object.

quantOpts = dlquantizationOptions('MetricFcn',{@(x)hComputeModelAccuracy(x, net, aug_valData)});

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(quantObj, aug_calData)

calResults=121×5 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(quantObj, 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 a Neural Network for FPGA Target

This example shows how to quantize learnable parameters in the convolution layers of a neural network and explore the behavior of the quantized network. In this example, you quantize the logo recognition network (LogoNet). Quantization helps reduce the memory requirement of a deep neural network by quantizing weights, biases and activations of network layers to 8-bit scaled integer data types. Use MATLAB® to retrieve the prediction results from the target device.

This example uses the products listed under FPGA in Quantization Workflow Prerequisites.

Create a file in your current working directory called getLogoNetwork.m. Enter these lines into the file:

function net = getLogoNetwork()
    data = getLogoData();
    net  = data.convnet;
end

function data = getLogoData()
    if ~isfile('LogoNet.mat')
        url = 'https://www.mathworks.com/supportfiles/gpucoder/cnn_models/logo_detection/LogoNet.mat';
        websave('LogoNet.mat',url);
    end
    data = load('LogoNet.mat');
end

Load the pretrained network.

snet = getLogoNetwork();

snet = 

  SeriesNetwork with properties:

         Layers: [22×1 nnet.cnn.layer.Layer]
     InputNames: {'imageinput'}
    OutputNames: {'classoutput'}

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.

This example uses the images in the logos_dataset data set. Define an imageDatastore, then split the data into calibration and validation data sets.

curDir = pwd;
newDir = fullfile(matlabroot,'examples','deeplearning_shared','data','logos_dataset.zip');
copyfile(newDir,curDir);
unzip('logos_dataset.zip');
imageData = imageDatastore(fullfile(curDir,'logos_dataset'),...
 'IncludeSubfolders',true,'FileExtensions','.JPG','LabelSource','foldernames');
[calibrationData,validationData] = splitEachLabel(imageData,0.5,'randomized');

Create a dlquantizer object and specify the network to quantize. Set the execution environment for the quantized network to FPGA.

dlQuantObj = dlquantizer(snet,'ExecutionEnvironment','FPGA');

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.

 dlQuantObj.calibrate(calibrationData)

ans = 
        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 a target object with a custom name for your target device and an interface to connect your target device to the host computer.

hTarget = dlhdl.Target('Intel','Interface','JTAG');

Define a metric function to use to compare the behavior of the network before and after quantization. Save this function in a local file.

function accuracy = hComputeModelAccuracy(predictionScores,net,dataStore)
%% hComputeModelAccuracy test helper function computes model level accuracy statistics

% Copyright 2020 The MathWorks, Inc.
    
    % Load ground truth 
    groundTruth = dataStore.Labels;
    
    % Compare predicted label with ground truth 
    predictionError = {};
    for idx=1:numel(groundTruth)
        [~, idy] = max(predictionScores(idx,:)); 
        yActual = net.Layers(end).Classes(idy);
        predictionError{end+1} = (yActual == groundTruth(idx)); %#ok
    end
    
    % Sum all prediction errors.
    predictionError = [predictionError{:}];
    accuracy = sum(predictionError)/numel(predictionError);
end

Specify the metric function and FPGA execution environment options in a dlquantizationOptions object.

options = dlquantizationOptions('MetricFcn', ...
    {@(x)hComputeModelAccuracy(x,snet,validationData)},'Bitstream','arria10soc_int8',...
'Target',hTarget);

Compile and deploy the quantized network. Use the validate function to quantize the learnable parameters in the convolution layers of the network and exercise the network. This function uses the output of the compile function to program the FPGA board by using the programming file. It also downloads the network weights and biases. The deploy function checks for the Intel^® Quartus^® tool and the supported tool version. It then programs the FPGA device using the sof file, displays progress messages, and the time it takes to deploy the network. The validate function uses the metric function defined in the dlquantizationOptions object to compare the results of the network before and after quantization.

prediction = dlQuantObj.validate(validationData,options);

           offset_name          offset_address     allocated_space 
    _______________________    ______________    _________________

    "InputDataOffset"           "0x00000000"     "48.0 MB"        
    "OutputResultOffset"        "0x03000000"     "4.0 MB"         
    "SystemBufferOffset"        "0x03400000"     "60.0 MB"        
    "InstructionDataOffset"     "0x07000000"     "8.0 MB"         
    "ConvWeightDataOffset"      "0x07800000"     "8.0 MB"         
    "FCWeightDataOffset"        "0x08000000"     "12.0 MB"        
    "EndOffset"                 "0x08c00000"     "Total: 140.0 MB"

### Programming FPGA Bitstream using JTAG...
### Programming the FPGA bitstream has been completed successfully.
### Loading weights to Conv Processor.
### Conv Weights loaded. Current time is 16-Jul-2020 12:45:10
### Loading weights to FC Processor.
### FC Weights loaded. Current time is 16-Jul-2020 12:45:26
### Finished writing input activations.
### Running single input activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13570959                  0.09047                      30          380609145             11.8
    conv_module           12667786                  0.08445 
        conv_1             3938907                  0.02626 
        maxpool_1          1544560                  0.01030 
        conv_2             2910954                  0.01941 
        maxpool_2           577524                  0.00385 
        conv_3             2552707                  0.01702 
        maxpool_3           676542                  0.00451 
        conv_4              455434                  0.00304 
        maxpool_4            11251                  0.00008 
    fc_module               903173                  0.00602 
        fc_1                536164                  0.00357 
        fc_2                342643                  0.00228 
        fc_3                 24364                  0.00016 
 * The clock frequency of the DL processor is: 150MHz


### Finished writing input activations.
### Running single input activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13570364                  0.09047                      30          380612682             11.8
    conv_module           12667103                  0.08445 
        conv_1             3939296                  0.02626 
        maxpool_1          1544371                  0.01030 
        conv_2             2910747                  0.01940 
        maxpool_2           577654                  0.00385 
        conv_3             2551829                  0.01701 
        maxpool_3           676548                  0.00451 
        conv_4              455396                  0.00304 
        maxpool_4            11355                  0.00008 
    fc_module               903261                  0.00602 
        fc_1                536206                  0.00357 
        fc_2                342688                  0.00228 
        fc_3                 24365                  0.00016 
 * The clock frequency of the DL processor is: 150MHz


### Finished writing input activations.
### Running single input activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13571561                  0.09048                      30          380608338             11.8
    conv_module           12668340                  0.08446 
        conv_1             3939070                  0.02626 
        maxpool_1          1545327                  0.01030 
        conv_2             2911061                  0.01941 
        maxpool_2           577557                  0.00385 
        conv_3             2552082                  0.01701 
        maxpool_3           676506                  0.00451 
        conv_4              455582                  0.00304 
        maxpool_4            11248                  0.00007 
    fc_module               903221                  0.00602 
        fc_1                536167                  0.00357 
        fc_2                342643                  0.00228 
        fc_3                 24409                  0.00016 
 * The clock frequency of the DL processor is: 150MHz


### Finished writing input activations.
### Running single input activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13569862                  0.09047                      30          380613327             11.8
    conv_module           12666756                  0.08445 
        conv_1             3939212                  0.02626 
        maxpool_1          1543267                  0.01029 
        conv_2             2911184                  0.01941 
        maxpool_2           577275                  0.00385 
        conv_3             2552868                  0.01702 
        maxpool_3           676438                  0.00451 
        conv_4              455353                  0.00304 
        maxpool_4            11252                  0.00008 
    fc_module               903106                  0.00602 
        fc_1                536050                  0.00357 
        fc_2                342645                  0.00228 
        fc_3                 24409                  0.00016 
 * The clock frequency of the DL processor is: 150MHz


### Finished writing input activations.
### Running single input activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13570823                  0.09047                      30          380619836             11.8
    conv_module           12667607                  0.08445 
        conv_1             3939074                  0.02626 
        maxpool_1          1544519                  0.01030 
        conv_2             2910636                  0.01940 
        maxpool_2           577769                  0.00385 
        conv_3             2551800                  0.01701 
        maxpool_3           676795                  0.00451 
        conv_4              455859                  0.00304 
        maxpool_4            11248                  0.00007 
    fc_module               903216                  0.00602 
        fc_1                536165                  0.00357 
        fc_2                342643                  0.00228 
        fc_3                 24406                  0.00016 
 * The clock frequency of the DL processor is: 150MHz


          offset_name          offset_address     allocated_space 
    _______________________    ______________    _________________

    "InputDataOffset"           "0x00000000"     "48.0 MB"        
    "OutputResultOffset"        "0x03000000"     "4.0 MB"         
    "SystemBufferOffset"        "0x03400000"     "60.0 MB"        
    "InstructionDataOffset"     "0x07000000"     "8.0 MB"         
    "ConvWeightDataOffset"      "0x07800000"     "8.0 MB"         
    "FCWeightDataOffset"        "0x08000000"     "12.0 MB"        
    "EndOffset"                 "0x08c00000"     "Total: 140.0 MB"

### 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 activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13572329                  0.09048                      10          127265075             11.8
    conv_module           12669135                  0.08446 
        conv_1             3939559                  0.02626 
        maxpool_1          1545378                  0.01030 
        conv_2             2911243                  0.01941 
        maxpool_2           577422                  0.00385 
        conv_3             2552064                  0.01701 
        maxpool_3           676678                  0.00451 
        conv_4              455657                  0.00304 
        maxpool_4            11227                  0.00007 
    fc_module               903194                  0.00602 
        fc_1                536140                  0.00357 
        fc_2                342688                  0.00228 
        fc_3                 24364                  0.00016 
 * The clock frequency of the DL processor is: 150MHz


### Finished writing input activations.
### Running single input activations.


              Deep Learning Processor Profiler Performance Results

                   LastLayerLatency(cycles)   LastLayerLatency(seconds)       FramesNum      Total Latency     Frames/s
                         -------------             -------------              ---------        ---------       ---------
Network                   13572527                  0.09048                      10          127266427             11.8
    conv_module           12669266                  0.08446 
        conv_1             3939776                  0.02627 
        maxpool_1          1545632                  0.01030 
        conv_2             2911169                  0.01941 
        maxpool_2           577592                  0.00385 
        conv_3             2551613                  0.01701 
        maxpool_3           676811                  0.00451 
        conv_4              455418                  0.00304 
        maxpool_4            11348                  0.00008 
    fc_module               903261                  0.00602 
        fc_1                536205                  0.00357 
        fc_2                342689                  0.00228 
        fc_3                 24365                  0.00016 
 * The clock frequency of the DL processor is: 150MHz

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.

Examine the MetricResults.Result field of the validation output to see the performance of the quantized network.

validateOut = prediction.MetricResults.Result

ans = 
    NetworkImplementation    MetricOutput
    _____________________    ____________

     {'Floating-Point'}         0.9875   
     {'Quantized'     }         0.9875   

Examine the QuantizedNetworkFPS field of the validation output to see the frames per second performance of the quantized network.

prediction.QuantizedNetworkFPS

ans = 11.8126

Import a dlquantizer Object into the Deep Network Quantizer App

This example shows you how to import a dlquantizer object from the base workspace into the Deep Network Quantizer app. This allows you to begin quantization of a deep neural network using the command line or the app, and resume your work later in the app.

Open the Deep Network Quantizer app.

deepNetworkQuantizer

In the app, click New and select Import dlquantizer object.

In the dialog, select the dlquantizer object to import from the base workspace. For this example, use quantObj that you create in the above example Quantize a Neural Network for GPU Target.

The app imports any data contained in the dlquantizer object that was collected at the command line. This data can include the network to quantize, calibration data, validation data, and calibration statistics.

The app displays a table containing the calibration data contained in the imported dlquantizer object, quantObj. To the right of the table, the app displays histograms of the dynamic ranges of the parameters. The gray regions of the histograms indicate data that cannot be represented by the quantized representation. For more information on how to interpret these histograms, see Quantization of Deep Neural Networks.

Emulate Target Agnostic Quantized Network

This example shows how to create a target agnostic, simulatable quantized deep neural network in MATLAB.

Target agnostic quantization allows you to see the effect quantization has on your neural network without target hardware or target-specific quantization schemes. Creating a target agnostic quantized network is useful if you:

• Do not have access to your target hardware.

• Want to preview whether or not your network is suitable for quantization.

• Want to find layers that are sensitive to quantization.

Quantized networks emulate quantized behavior for quantization-compatible layers. Network architecture like layers and connections are the same as the original network, but inference behavior uses limited precision types. Once you have quantized your network, you can use the quantizationDetails function to retrieve details on what was quantized.

Load the pretrained network. net is a SqueezeNet network that has been retrained using transfer learning to classify images in the MerchData data set.

load squeezenetmerch
net

net = 
  DAGNetwork with properties:

         Layers: [68×1 nnet.cnn.layer.Layer]
    Connections: [75×2 table]
     InputNames: {'data'}
    OutputNames: {'new_classoutput'}

You can use the quantizationDetails function to see that the network is not quantized.

qDetailsOriginal = quantizationDetails(net)

qDetailsOriginal = struct with fields:
            IsQuantized: 0
          TargetLibrary: ""
    QuantizedLayerNames: [0×0 string]
    QuantizedLearnables: [0×3 table]

Unzip and load the MerchData images as an image datastore.

unzip('MerchData.zip')
imds = imageDatastore('MerchData', ...
    'IncludeSubfolders',true, ...
    'LabelSource','foldernames');

Define calibration and validation data to use for quantization. The output size of the images are changed for both calibration and validation data according to network requirements.

[calData,valData] = splitEachLabel(imds,0.7,'randomized');
augCalData = augmentedImageDatastore([227 227],calData);
augValData = augmentedImageDatastore([227 227],valData);

Create dlquantizer object and specify the network to quantize. Set the execution environment to MATLAB. How the network is quantized depends on the execution environment. The MATLAB execution environment is agnostic to the target hardware and allows you to prototype quantized behavior.

quantObj = dlquantizer(net,'ExecutionEnvironment','MATLAB');

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(quantObj,augCalData);

Use the quantize method to quantize the network object and return a simulatable quantized network.

qNet = quantize(quantObj)  

qNet = 
Quantized DAGNetwork with properties:

         Layers: [68×1 nnet.cnn.layer.Layer]
    Connections: [75×2 table]
     InputNames: {'data'}
    OutputNames: {'new_classoutput'}

Use the quantizationDetails method to extract quantization details.

You can use the quantizationDetails function to see that the network is now quantized.

qDetailsQuantized = quantizationDetails(qNet)

qDetailsQuantized = struct with fields:
            IsQuantized: 1
          TargetLibrary: "none"
    QuantizedLayerNames: [26×1 string]
    QuantizedLearnables: [52×3 table]

Make predictions using the original, single-precision floating-point network, and the quantized INT8 network.

predOriginal = classify(net,augValData);       % Predictions for the non-quantized network
predQuantized = classify(qNet,augValData);     % Predictions for the quantized network 

Compute the relative accuracy of the quantized network as compared to the original network.

ccrQuantized = mean(predQuantized == valData.Labels)*100

ccrQuantized = 100

ccrOriginal = mean(predOriginal == valData.Labels)*100

ccrOriginal = 100

For this validation data set, the quantized network gives the same predictions as the floating-point network.

Version History

Introduced in R2020a

R2022b: dlnetwork support

dlquantizer now supports dlnetwork objects for quantization using the calibrate and validate functions.

R2022a: Validate the performance of quantized network for CPU target

You can now use the dlquantizer object and the validate function to quantize a network and generate code for CPU targets.

R2022a: Quantize neural networks without a specific target

Specify MATLAB as the ExecutionEnvironment to quantize your neural networks without generating code or committing to a specific target for code deployment. This can be useful if you:

• Do not have access to your target hardware.

• Want to inspect your quantized network without generating code.

Your quantized network implements int8 data instead of single data. It keeps the same layers and connections as the original network, and it has the same inference behavior as it would when running on hardware.

Once you have quantized your network, you can use the quantizationDetails function to inspect your quantized network. Additionally, you also have the option to deploy the code to a GPU target.