Issues with Faster R-CNN regarding Input Image Size, Mini-Batch Loss (NaN value) and another error.

Question

Lathi20 2020-11-12

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/645053-issues-with-faster-r-cnn-regarding-input-image-size-mini-batch-loss-nan-value-and-another-error

编辑： ramin takmil 2022-2-22

采纳的回答： Raynier Suresh

在 MATLAB Online 中打开

Hello Everyone,

I am trying to use Faster R-CNN for an object detection task. My code is based on the info in the following link:

https://au.mathworks.com/help/vision/ug/object-detection-using-faster-r-cnn-deep-learning.html

I had several issues as follow:

1) The size of the images that I have is 512x640x1 (jpg format) and the size of the object vary from 3x3x1 to 10x10x1. The input layer in the ResNet-50 is 224x224x3. My question is how to change the input layer size of ResNet-50 to match the size of my input images?

2) What is the reason to get (NaN value) for the (Mini-batch Loss) and (empty column) for (Mini-batch Accuracy) and (Mini-batch RMSE)?

Knowing that I also decreased the 'InitialLearnRate' from 1e-3 to 1e-5 as suggested in one of answers in other questions, but still the issue has not been solved. (Image Attached below)

3) If I rescale the input images to 224x244x3 Using the code below, I got the following error. What is the reason of this error (Image attached below)?

ERROR: Undefined function 'categories' for input arguments of type 'double'.

% First, specify the network input size.
inputSize = [224 224 3];
%inputSize = [512 640 3];
% Estimate anchor boxes based on the size of objects in the training data. 
preprocessedTrainingData = transform(trainingData, @(data)preprocessData(data,inputSize));
numAnchors = 3;
anchorBoxes = estimateAnchorBoxes(preprocessedTrainingData,numAnchors)
% load a pretrained ResNet-50 model.
featureExtractionNetwork = resnet50;
% This feature extraction layer outputs feature maps that are downsampled by a factor of 16. 
featureLayer = 'activation_40_relu';
% analyzeNetwork(resnet50)
% Define the number of classes to detect.
numClasses = width(vehicleDataset)-1;
% Create the Faster R-CNN object detection network.
lgraph = fasterRCNNLayers(inputSize,numClasses,anchorBoxes,featureExtractionNetwork,featureLayer);

Thank you so much for your help in adavnce

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

ramin takmil 2022-2-22

编辑：ramin takmil 2022-2-22

hello I have this problem too

my input size is [512 512 1] but resnet-50 input is [224 224 1] so when I run the code it gives me this error : Input size mismatch. Size of input to this layer [512 512 1] is different from the expected input size [512 512 3].

I have tried to change the input layer using 'replaceLayer' function but it gives me error when I want to build "lgraph"

how did you solve the input size problem???

thanks in advance

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Raynier Suresh 2021-1-20

1
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/645053-issues-with-faster-r-cnn-regarding-input-image-size-mini-batch-loss-nan-value-and-another-error#answer_602139

To change the input layer size, you can replace the layer either in MATLAB Terminal using ‘replaceLayer’ function or in Deep Network Designer App.

Refer the below links for more information:

https://www.mathworks.com/help/deeplearning/ref/replacelayer.html

https://www.mathworks.com/help/deeplearning/ref/deepnetworkdesigner-app.html?searchHighlight=deep%20network%20designer&s_tid=srchtitle

The function categories will output the number of categories in a categorical array¸ I think that the training data might be improper, this could also be the reason for NaN’s in the Mini-Batch Loss.