maskrcnn example is not training

Hi all,

I'm trying to train a mask r-cnn network on my own data (two classes) and am therefore first trying out the example found on the mathworks website (https://www.mathworks.com/help/deeplearning/ug/instance-segmentation-using-mask-rcnn.html).

The code I'm using (unaltered from the website):

dataFolder = fullfile(tempdir,"coco");
trainedMaskRCNN_url = 'https://www.mathworks.com/supportfiles/vision/data/maskrcnn_object_person_car.mat';
helper.downloadTrainedMaskRCNN(trainedMaskRCNN_url,dataFolder);
pretrained = load(fullfile(dataFolder,'maskrcnn_object_person_car.mat'));
net = pretrained.net;
%%
imTest = imread('visionteam.jpg');
[masks,labels,scores,boxes] = segmentObjects(net,imTest);
overlayedImage = insertObjectMask(imTest,masks);
imshow(overlayedImage)
showShape("rectangle",gather(boxes),"Label",labels,"LineColor",'r')
%%
imageFolder = fullfile(dataFolder,"images");
captionsFolder = fullfile(dataFolder,"annotations");
if ~exist(imageFolder,'dir')
    mkdir(imageFolder)
    mkdir(captionsFolder)
end
annotationFile = fullfile(captionsFolder,"instances_train2014.json");
str = fileread(annotationFile);
trainClassNames = {'person', 'car'};
numClasses = length(trainClassNames);
imageSizeTrain = [800 800 3];
cocoAPIDir = fullfile(dataFolder,"cocoapi-master","MatlabAPI");
addpath(cocoAPIDir);
unpackAnnotationDir = fullfile(dataFolder,"annotations_unpacked","matFiles");
if ~exist(unpackAnnotationDir,'dir')
    mkdir(unpackAnnotationDir)
end
helper.unpackAnnotations(trainClassNames,annotationFile,imageFolder,unpackAnnotationDir);
%%
ds = fileDatastore(unpackAnnotationDir, ...
    'ReadFcn',@(x)helper.cocoAnnotationMATReader(x,imageFolder));
dsTrain = transform(ds,@(x)helper.preprocessData(x,imageSizeTrain));
data = preview(dsTrain)
%%
net = maskrcnn("resnet50-coco",trainClassNames,"InputSize",imageSizeTrain)
params = createMaskRCNNConfig(imageSizeTrain,numClasses,[trainClassNames {'background'}]);
params.ClassAgnosticMasks = false;
params.AnchorBoxes = net.AnchorBoxes;
params.FreezeBackbone = true;
initialLearnRate = 0.0012;
momentum = 0.9;
decay = 0.01;
velocity = [];
maxEpochs = 10;
miniBatchSize = 2;
miniBatchFcn = @(img,boxes,labels,masks) deal(cat(4,img{:}),boxes,labels,masks);
mbqTrain = minibatchqueue(dsTrain,4, ...
    "MiniBatchFormat",["SSCB","","",""], ...
    "MiniBatchSize",miniBatchSize, ...
    "OutputCast",["single","","",""], ...
    "OutputAsDlArray",[true,false,false,false], ...
    "MiniBatchFcn",miniBatchFcn, ...
    "OutputEnvironment",["auto","cpu","cpu","cpu"]);
%%
doTraining = true;
if doTraining
    
    iteration = 1; 
    start = tic;
    
     % Create subplots for the learning rate and mini-batch loss
    fig = figure;
    [lossPlotter, learningratePlotter] = helper.configureTrainingProgressPlotter(fig);
    
    % Initialize verbose output
    helper.initializeVerboseOutput([]);
    
    % Custom training loop
    for epoch = 1:maxEpochs
        reset(mbqTrain)
        shuffle(mbqTrain)
    
        while hasdata(mbqTrain)
            % Get next batch from minibatchqueue
            [X,gtBox,gtClass,gtMask] = next(mbqTrain);
        
            % Evaluate the model gradients and loss using dlfeval
            [gradients,loss,state,learnables] = dlfeval(@networkGradients,X,gtBox,gtClass,gtMask,net,params);
            %dlnet.State = state;
            
            % Compute the learning rate for the current iteration
            learnRate = initialLearnRate/(1 + decay*(epoch-1));
            
            if(~isempty(gradients) && ~isempty(loss))
                [net.AllLearnables,velocity] = sgdmupdate(learnables,gradients,velocity,learnRate,momentum);
            else
                continue;
            end
            
            % Plot loss/accuracy metric every 10 iterations
            if(mod(iteration,10)==0)
                helper.displayVerboseOutputEveryEpoch(start,learnRate,epoch,iteration,loss);
                D = duration(0,0,toc(start),'Format','hh:mm:ss');
                addpoints(learningratePlotter,iteration,learnRate)
                addpoints(lossPlotter,iteration,double(gather(extractdata(loss))))
                subplot(2,1,2)
                title(strcat("Epoch: ",num2str(epoch),", Elapsed: "+string(D)))
                drawnow
            end
            
            iteration = iteration + 1;    
        end
    
    end
    
    % Save the trained network
    modelDateTime = string(datetime('now','Format',"yyyy-MM-dd-HH-mm-ss"));
    save(strcat("trainedMaskRCNN-",modelDateTime,"-Epoch-",num2str(epoch),".mat"),'net');
    
end

When compiling, I get the following error:

Error using networkGradients
Too many output arguments.

Error in deep.internal.dlfeval (line 17)
[varargout{1:nargout}] = fun(x{:});

Error in dlfeval (line 40)
    [varargout{1:nargout}] =
    deep.internal.dlfeval(fun,varargin{:});

Error in person_car_example (line 80)
            [gradients,loss,state,learnables] =
            dlfeval(@networkGradients,X,gtBox,gtClass,gtMask,net,params);
           

so it seems something goes wrong in the line:

[gradients,loss,state,learnables] = dlfeval(@networkGradients,X,gtBox,gtClass,gtMask,net,params);

with the networkGradients file

I have used the networkGradients file from the example folder and have included it below:

function [gradients, totalLoss, state] = networkGradients(X, gTruthBoxes, gTruthLabels, gTruthMasks, dlnet, params)
% networkGradients - Gradient function to train MaskRCNN using a custom
% training loop.
 
% Copyright 2020 The MathWorks, Inc.


RPNRegDeltas = {'rpnConv1x1BoxDeltas'};regionProposal = {'rpl'};
 
outputNodes = [ RPNRegDeltas, regionProposal, dlnet.OutputNames(:)'];
 
% For training, the first step is to run a forward pass on the MaskRCNN
% network.
% We need the following outputs from the network to compute losses -
% YRPNReg - RPN regression output  [1x1x4xN] {[deltaX deltaY deltaW deltaH]}
% YRPNClass - RPN classification output [1x1x2xN]
% YRCNNReg - RCNN regression output  [1x1x4*(numClasses)xN] {[deltaX deltaY deltaW deltaH]}
% YRCNNClass - RCNN classification output [1x1x(numClasses+1)xN]
% YMask     - RCNN Mask segmentation output [hxwxnumClassesxN]
% proposals - The regionproposals from the RPN. [5xN]
[YRPNRegDeltas, proposal, YRCNNClass, YRCNNReg, YRPNClass, YMask, state] = forward(...
                                            dlnet, X, 'Outputs', outputNodes);

% If there are no proposals don't learn anything
if(isempty(proposal))
   totalLoss = dlarray([]);
   gradients= dlarray([]);
   disp("Empty proposals");
   return;
end

% Proposals are 5XNumProposals (Due to batch restrictions from custom RPL layer)
proposals = gather(extractdata(proposal));

% Convert proposals to numProposals x 5 (as expected by the rest of post processing code)
proposals =proposals';

proposals(:,1:4) = helper.boxUtils.x1y1x2y2ToXYWH(proposals(:,1:4));

numImagesInBatch = size(gTruthBoxes,1);
%Convert numProposalsx5 Proposals to numImagesInBatchx1 (Group by image index)
proposals = helper.groupProposalsByImageIndex(proposals, numImagesInBatch);


% Generate RCNN response targets
%--------------------------------

% Step 1: Match ground truth boxes to proposals                                       
[assignment, positiveIndex, negativeIndex] = helper.bboxMatchAndAssign(...
                                                        proposals, gTruthBoxes,...
                                                        params.PositiveOverlapRange, params.NegativeOverlapRange,...
                                                        false);

                                                    
% Step 2: Calcuate regression targets as (dx, dy, log(dw), log(dh))
regressionTargets = helper.generateRegressionTargets(gTruthBoxes, proposals,...
                                                        assignment, positiveIndex,...
                                                        params.NumClasses);
 
classNames = categories(gTruthLabels{1});

% Step 3: Assign groundtrutrh labels to proposals
classificationTargets = helper.generateClassificationTargets (gTruthLabels, assignment,...
                                             positiveIndex, negativeIndex,...
                                             classNames, params.BackgroundClass);
                                         
% Step 4: Calculate instance weights
instanceWeightsReg = helper.regressionResponseInstanceWeights (classificationTargets, params.BackgroundClass);
 
% Step 5: Generate mask targets
 
% Crop and resize the instances based on proposal bboxes and network output size
maskOutputSize = params.MaskOutputSize;
croppedMasks = helper.cropandResizeMasks (gTruthMasks, gTruthBoxes, maskOutputSize);
 
% Generate mask targets
maskTargets = helper.generateMaskTargets(croppedMasks, assignment, classificationTargets, params);

% Stage 2 (RCNN) Loss
% --------------------

% *Classification loss*
classificationTargets = cat(1, classificationTargets{:})';
% onehotencode labels
classificationTargets = onehotencode(classificationTargets,1);
classificationTargets(isnan(classificationTargets)) = 0;
LossRCNNClass = loss.CrossEntropy(YRCNNClass, classificationTargets);
 
% *Weighted regression loss*
regressionTargets = cat(1,regressionTargets{:});
instanceWeightsReg = cat(1, instanceWeightsReg{:});

LossRCNNReg = loss.smoothL1(YRCNNReg, single(regressionTargets'), single(instanceWeightsReg'));
 
% Mask Loss (Weighted cross entropy)
maskTargets= cat(4,maskTargets{:});
positiveIndex = cat(1,positiveIndex{:});
LossRCNNMask = loss.SpatialCrossEntropy(YMask, single(maskTargets), positiveIndex);
 
% Total Stage 2 loss
 LossRCNN = LossRCNNReg + LossRCNNClass + LossRCNNMask;

 
% Generate RCNN response targets
%--------------------------------
featureSize = size(YRPNRegDeltas);
imageSize = params.ImageSize;
[RPNRegressionTargets, RPNRegWeights, assignedLabelsRPN] = helper.rpnRegressionResponse(featureSize, gTruthBoxes, imageSize, params);

RPNClassificationTargets = onehotencode(assignedLabelsRPN, 3);
RPNClassificationTargets(isnan(RPNClassificationTargets)) = 0;


% Stage 1 (RPN) Loss
% --------------------
   
LossRPNClass = loss.CrossEntropy(YRPNClass, RPNClassificationTargets);
 
LossRPNReg = loss.smoothL1(YRPNRegDeltas, RPNRegressionTargets, RPNRegWeights);
 
LossRPN = LossRPNClass + LossRPNReg;


% Total Loss
%------------
totalLoss = LossRCNN + LossRPN;
 
gradients = dlgradient(totalLoss, dlnet.Learnables);
 
end

It seems the "learnables" ouput is one too many, so I remove it but get the following error:

Unrecognized method, property, or field 'OutputNames' for class 'maskrcnn'.

Error in networkGradients (line 10)
outputNodes = [ RPNRegDeltas, regionProposal, dlnet.OutputNames(:)'];

Error in deep.internal.dlfeval (line 17)
[varargout{1:nargout}] = fun(x{:});

Error in dlfeval (line 40)
    [varargout{1:nargout}] = deep.internal.dlfeval(fun,varargin{:});

Error in person_car_example (line 80)
            [gradients,loss,state] = dlfeval(@networkGradients,X,gtBox,gtClass,gtMask,net,params);

The maskrcnn structure does not have an 'OutputNames' field... It seems likethe entire networkGradients file is incompatible with the example.

Any ideas/help would be very welcome!

Thanks,

Sam

[Thomas Cimiega]

Hi Sam, how did you get to run the

helper.unpackAnnotations

function? I downloaded the coco repo as mentioned in the example and try to run it, but I always get

Unrecognized function or variable 'gasonMex'

Also, when I follow the instruction inside of the gason.m file by compiling the mex file with

mex('CXXFLAGS=\$CXXFLAGS -std=c++11 -Wall','-largeArrayDims','private/gasonMex.cpp','../common/gason.cpp','-I../common/','-outdir','private');

or

mex('CXXFLAGS=\$CXXFLAGS -std=c++11 -Wall','-largeArrayDims','private/gasonMex.cpp','../common/gason.cpp','-I../common/','-outdir','private','CXX="/usr/bin/g++"');

I get

Error using mex

g++: error: \-fexceptions: No such file or directory

or

Error using mex

Das System kann den angegebenen Pfad nicht finden.

(Cannot find the provided path)

[tao wang]

@Thomas Cimiega hi Thomas Cimiega. Did you solve that problem, I'm stuck at that step too, it says gasonMex can't recognize it

[Thomas Cimiega]

Unfortunately not. It seems that Matlab is not able to provide a working example... Instead I had to write an own example by using other cnn examples and reading the documentations.

[mohd akmal masud]

Dear @Sam Van Der Jeught. I got error like

Unrecognized function or variable 'gasonMex'.

Error in gason (line 53)
out = gasonMex( 'convert', in );

Error in CocoApi (line 61)
coco.data=gason(fileread(annFile)); end

Error in unpackAnnotations (line 7)
coco = CocoApi(annotationFile);

what should I do?

Answer 1

[Zakariya Abousabie]

hi there does anybody find solution to these problem, becuse im trying to train maskrcnn with my on data of two classes but im stuck too and the matworks example doesn't help

Answer 2

[David Willingham]

Hi Sam,

What version of MATLAB are you running?

David

[Sam Van Der Jeught]

Hi David,

I'm using Matlab 2021b, but I've also tried it with 2020b.

In the meantime I have rewritten the networkGradients.m file as follows and now it works. Hope someone else can benefit from this.

function [gradients, totalLoss, state, learnables] = networkGradients(X, gTruthBoxes, gTruthLabels, gTruthMasks, dlnet, params)
% networkGradients - Gradient function to train MaskRCNN using a custom
% training loop.
 
% Copyright 2020 The MathWorks, Inc.


%RPNRegDeltas = {'rpnConv1x1BoxDeltas'};regionProposal = {'rpl'};
 
%outputNodes = [ RPNRegDeltas, regionProposal, dlnet.OutputNames(:)'];
 
% For training, the first step is to run a forward pass on the MaskRCNN
% network.
% We need the following outputs from the network to compute losses -
% YRPNReg - RPN regression output  [1x1x4xN] {[deltaX deltaY deltaW deltaH]}
% YRPNClass - RPN classification output [1x1x2xN]
% YRCNNReg - RCNN regression output  [1x1x4*(numClasses)xN] {[deltaX deltaY deltaW deltaH]}
% YRCNNClass - RCNN classification output [1x1x(numClasses+1)xN]
% YMask     - RCNN Mask segmentation output [hxwxnumClassesxN]
% % proposals - The regionproposals from the RPN. [5xN]
% [YRPNRegDeltas, proposal, YRCNNClass, YRCNNReg, YRPNClass, YMask, state] = forward(...
%                                             dlnet, X, 'Outputs', outputNodes);

[netOut, state] = forward(dlnet, X);

YRPNClass  = netOut{1};
YRPNRegDeltas = netOut{2};
proposal = netOut{3};
YRCNNClass = netOut{4};
YRCNNReg = netOut{5};
YMask = netOut{6};

%save proposal proposal

% If there are no proposals don't learn anything
if(isempty(proposal))
   totalLoss = dlarray([]);
   gradients= dlarray([]);
   disp("Empty proposals");
   return;
end

% Proposals are 5XNumProposals (Due to batch restrictions from custom RPL layer)
proposals = gather(extractdata(proposal));

% Convert proposals to numProposals x 5 (as expected by the rest of post processing code)
proposals =proposals';

proposals(:,1:4) = helper.boxUtils.x1y1x2y2ToXYWH(proposals(:,1:4));

numImagesInBatch = size(gTruthBoxes,1);
%Convert numProposalsx5 Proposals to numImagesInBatchx1 (Group by image index)
proposals = helper.groupProposalsByImageIndex(proposals, numImagesInBatch);

% Generate RCNN response targets
%--------------------------------

% Step 1: Match ground truth boxes to proposals                                       
[assignment, positiveIndex, negativeIndex] = helper.bboxMatchAndAssign(...
                                                        proposals, gTruthBoxes,...
                                                        params.PositiveOverlapRange, params.NegativeOverlapRange,...
                                                        false);
%save assignment assignment
                                                    
% Step 2: Calcuate regression targets as (dx, dy, log(dw), log(dh))
regressionTargets = helper.generateRegressionTargets(gTruthBoxes, proposals,...
                                                        assignment, positiveIndex,...
                                                        params.NumClasses);
 
% classNames = categories(gTruthLabels{1}); % !!! TROUBLE, change as below
classNames = params.ClassNames;

% Step 3: Assign groundtrutrh labels to proposals
classificationTargets = helper.generateClassificationTargets (gTruthLabels, assignment,...
                                             positiveIndex, negativeIndex,...
                                             classNames, params.BackgroundClass);

% Step 4: Calculate instance weights
instanceWeightsReg = helper.regressionResponseInstanceWeights (classificationTargets, params.BackgroundClass);
 
% Step 5: Generate mask targets
 
% Crop and resize the instances based on proposal bboxes and network output size
maskOutputSize = params.MaskOutputSize;
croppedMasks = helper.cropandResizeMasks (gTruthMasks, gTruthBoxes, maskOutputSize);
 
% Generate mask targets
maskTargets = helper.generateMaskTargets(croppedMasks, assignment, classificationTargets, params);

% Stage 2 (RCNN) Loss
% --------------------

% *Classification loss*

classificationTargets = cat(1, classificationTargets{:})';
summary(classificationTargets)

% onehotencode labels                       
classificationTargets = onehotencode(classificationTargets,1);
classificationTargets(isnan(classificationTargets)) = 0;
classificationTargets = reshape(classificationTargets ,1, 1, size(YRCNNClass,3),[]);

 
LossRCNNClass = loss.CrossEntropy(YRCNNClass, classificationTargets);
 
% *Weighted regression loss*

regressionTargets = cat(1,regressionTargets{:})';
instanceWeightsReg = cat(1, instanceWeightsReg{:})';

regressionTargets = reshape(regressionTargets, 1, 1, size(YRCNNReg,3),[]);
instanceWeightsReg = reshape(instanceWeightsReg, 1, 1, size(YRCNNReg,3),[]);
LossRCNNReg = loss.smoothL1(YRCNNReg, single(regressionTargets), single(instanceWeightsReg));
 
% Mask Loss (Weighted cross entropy)
maskTargets= cat(4,maskTargets{:});
positiveIndex = cat(1,positiveIndex{:});
LossRCNNMask = loss.SpatialCrossEntropy(YMask, single(maskTargets), positiveIndex);
 
% Total Stage 2 loss
 LossRCNN = LossRCNNReg + LossRCNNClass + LossRCNNMask;

 
% Generate RCNN response targets
%--------------------------------
featureSize = size(YRPNRegDeltas);
imageSize = params.ImageSize;
[RPNRegressionTargets, RPNRegWeights, assignedLabelsRPN] = helper.rpnRegressionResponse(featureSize, gTruthBoxes, imageSize, params);

RPNClassificationTargets = onehotencode(assignedLabelsRPN, 3);
% detetron2 uses ony foreground class for classification.
RPNClassificationTargets(:,:,2,:) = [];
RPNClassificationTargets  = reshape(RPNClassificationTargets, featureSize(1), featureSize(2), [], numImagesInBatch );
RPNClassificationTargets(isnan(RPNClassificationTargets)) = 0;


% Stage 1 (RPN) Loss
% --------------------
YRPNClass = softmax(YRPNClass);
LossRPNClass = loss.CrossEntropy(YRPNClass, RPNClassificationTargets);
 
LossRPNReg = loss.smoothL1(YRPNRegDeltas, RPNRegressionTargets, RPNRegWeights);
 
LossRPN = LossRPNClass + LossRPNReg;


% Total Loss
%------------
totalLoss = LossRCNN + LossRPN;

if(params.FreezeBackbone)
    learnables = rmfield(dlnet.AllLearnables, {'FeatureExtractionNet','PostPoolFeatureExtractionNet'});
else
    learnables = dlnet.AllLearnables;
end

gradients = dlgradient(totalLoss, learnables);
 
end