How to increase network input size for 'Tiny-yolov3-coco'?

Greetings,

I want to use 'tiny-yolov3-coco' for detecting the cells. As shown in the Figure, I used a network input size of 416x416. However, the total loss is still high. Therefore I want to increase the network input size to 832x832. The code that I used as follows:

networkInputSize = [832 832 3];

baseNetwork = 'tiny-yolov3-coco';

detector = yolov3ObjectDetector(baseNetwork);

net = detector.Network;

lgraph = layerGraph(net);

imgLayer = imageInputLayer(networkInputSize,"Name","input","Normalization","none");

lgraph = replaceLayer(lgraph,"input",imgLayer);

newbaseNetwork = assembleNetwork(lgraph);

However, I got the error as follows:

Error in trainyolov3 (line 61)

newbaseNetwork = assembleNetwork(lgraph);

Caused by:

Network: Missing output layer. The network must have at least one output layer.

Layer 'conv2d_10': Unconnected output. Each layer output must be connected to the input of another layer.

Layer 'conv2d_13': Unconnected output. Each layer output must be connected to the input of another layer.

Answer 1

[Khushboo]

Hello,

The network that you are using is a dlnetwork (which does not have output layers) and not a DAG network. The assembleNetwork function returns a DAG network ready for prediction. If using a dlnetwork works for your usecase, you can do the following:

networkInputSize = [832 832 3];
baseNetwork = 'tiny-yolov3-coco';
detector = yolov3ObjectDetector(baseNetwork);
net = detector.Network;
lgraph = layerGraph(net);
imgLayer = imageInputLayer(networkInputSize,"Name","input","Normalization","none");
lgraph = replaceLayer(lgraph,"input",imgLayer);

newbaseNetwork = dlnetwork(lgraph);

Hope this answers your question!

[Fahmi Akmal Dzulkifli]

Thank you Sir for the responses. I just want to ask one more question. When I want to create the yolov3 detector, the error appears as follows:

Error using yolov3ObjectDetector>iValidateYOLOv3Network (line 1234)

Number of filters in the output convolutional layer must be 56 for 8 anchor boxes and 2 classes.

My code for anhor boxes is:

numAnchors = 16;
[anchors, meanIoU] = estimateAnchorBoxes(trainingDataForEstimation, numAnchors);
area = anchors(:, 1).*anchors(:, 2);
[~, idx] = sort(area, 'descend');
anchors = anchors(idx, :);
anchorBoxes = {anchors(1:8,:)
    anchors(9:16,:)
    };
classNames = trainingDataTbl.Properties.VariableNames(2:end);

Can you detect what is the problem with the code?

[Khushboo]

Hello,

I am not able to understand what this code does. But from what I assume, the error is related to the number of dimensions defined for the last convolution layer. As you are changing the input size, I think the rest of the dimensions will have to be changes correspondingly. Kindly make sure you have done that.