Problem in YOLOv2 training

I'm doing a YOLOv2 training and I came across the following error:

Error using

trainYOLOv2ObjectDetector>iParseInputsYolov2

(line 240)

Invalid network.

Error in

trainYOLOv2ObjectDetector

(line 174)

[trainingData, lgraph,

params, options] =

iParseInputsYolov2(...

Error in criando (line 13)

[detector,info] =

trainYOLOv2ObjectDetector(ds,lgraph,options);

Caused by:

Network: The input to

the YOLO v2 transform

layer must have 12

channels to support 2

anchor boxes and 1

classes. The number of

channels must equal

numAnchors * (5 +

numClasses). Update the

training data, the

number of anchor boxes

specified in the

yolov2Transform layer,

or the layers preceding

the transform layer.

My code:

clear, clc, close all;

vagem = load('RotulosVagem.mat');

lgraph = load('lgraph.mat');

lgraph = lgraph.lgraph;

gTruth = vagem.gTruth;

[imds, blds] = objectDetectorTrainingData(gTruth);

ds = combine(imds, blds);

options = trainingOptions('sgdm');

[detector,info] = trainYOLOv2ObjectDetector(ds,lgraph,options);

I will attach the files. Thanks in advance.

Answer 1

[Vivek Akkala]

Hi,

There seems to be a mismatch between expected inputs and actual inputs to the yolov2TransformLayer. Based on the "RotulosVagem.mat" and "lgraph" provided by you, I assume you want to train a YOLO v2 network with 2 anchor boxes for 1 class.

For this, the last convolutional layer before yolov2TransformLayer in the "lgraph" must have 12 output filters but the current network is having 20 filters.

The issue can be resolved by updating the output filters of the last convolutional layer. You can try the following code:

lgraph = lgraph.lgraph;

[imds, blds] = objectDetectorTrainingData(gTruth);

ds = combine(imds, blds);

options = trainingOptions('sgdm');

% % Start of the code to be added %%

numClasses= size(vagem.gTruth.LabelData,2);

numAnchorBoxes = size(lgraph.Layers(end,1).AnchorBoxes,1);

outFilters = (5+numClasses).*numAnchorBoxes;

yolov2ConvLayer = convolution2dLayer(3,outFilters,'Name','yolov2ConvUpdated',...

'Padding', 'same',...

'WeightsInitializer',@(sz)randn(sz)*0.01);

yolov2ConvLayer.Bias = zeros(1,1,outFilters);

lgraph = replaceLayer(lgraph,'yolov2ClassConv',yolov2ConvLayer);

% % End of the code to be added %%

[detector,info] = trainYOLOv2ObjectDetector(ds,lgraph,options);

[Abhilasha]

I am writing this code:

data = load('annotated_img.mat');

trainingData = gTruth1;

dataDir = fullfile("Filtered_Combined_Images");

trainingData.imageFilename = fullfile(trainingData.imageFilename);

rng(0); % Set random seed for reproducibility

shuffledIdx = randperm(height(trainingData));

trainingData = trainingData(shuffledIdx,:);

imds = imageDatastore(trainingData.imageFilename);

blds = boxLabelDatastore(trainingData(:,2:end))

dp = combine(imds, blds);

net = load('lgraph.mat');

lgraph = net.lgraph;

analyzeNetwork(lgraph);

lgraph.Layers

options = trainingOptions('sgdm', ...

'InitialLearnRate', 0.001, ...

'Verbose', true, ...

'MiniBatchSize', 16, ...

'MaxEpochs', 200,...

'Shuffle', 'never', ...

'VerboseFrequency', 20, ...

'CheckpointPath', tempdir);

[detector, info] = trainYOLOv2ObjectDetector(dp, lgraph, options);

and I am getting this error:

Network: The input to the YOLO v2 transform layer must have 168 channels to support 8 anchor boxes and 16 classes.

The number of channels must equal numAnchors * (5 + numClasses). Update the training data, the number of anchor

boxes specified in the yolov2Transform layer, or the layers preceding the transform layer.