Improper initialization of classification layer in rcnn

10 次查看（过去 30 天）
Griffon Thomas
Griffon Thomas 2018-4-23
评论： Ameer Hamza 2018-4-23
Hello, I'm a relative newbie to MATLAB and neural networks, and I'm looking at disease spread and analysis in crop fields. I wanted to make an RCNN to help with this. I have some skeleton code, but I'm getting errors I don't understand and don't have the skill to debug.
Here is the code:
load 'D:\Documents\MATLAB\bridgeLabels.mat', 'gTruth';
%these are the labels I made in the image labeler app
trainingData = objectDetectorTrainingData(gTruth);
%this apparently makes the training data for me
layers = [imageInputLayer([2160 3840 3])
convolution2dLayer([5 5],10)
reluLayer()
fullyConnectedLayer(10)
softmaxLayer()
classificationLayer()];
%I understand what all these things do, kind of.
%I just copied this code from the demonstration in the reference
%I'm getting some error with the classification layer I don't know how to fix
options = trainingOptions('sgdm',...
'LearnRateSchedule','piecewise',...
'LearnRateDropFactor',0.2,...
'LearnRateDropPeriod',5,...
'MaxEpochs',20,...
'MiniBatchSize',64,...
'Plots','training-progress');
%again, most of this makes sense to me
detector = trainRCNNObjectDetector(trainingData, layers, options);
%ok so now the network is made apparently
image = imread('D:\Documents\MATLAB\clubroot_shots\lcbo1.png');
%this is my testing image
wid = 10;
rois = zeros(1, (image.width/wid)*(image.height/wid));
for i=1:image.width/wid
for j=1:image.height/wid
rois(i+j*width) = [1+(i-1)*wid, 1+(j-1)*wid, wid, wid];
end
end
%I believe this code will split up the image into 10x10 regions of interest.
%I wrote this block myself.
classifyRegions(detector, image, rois)
%and here the regions get classified. Semicolon off because i want to see what happens
When I run this code, I get the following errors:
Error using vision.internal.cnn.validation.checkNetworkClassificationLayer (line 9)
The number object classes in the network classification layer must be equal to the number of classes
defined in the input trainingData plus 1 for the "Background" class.
Error in vision.internal.rcnn.parseInputs (line 35)
vision.internal.cnn.validation.checkNetworkClassificationLayer(network, trainingData);
Error in trainRCNNObjectDetector (line 185)
params = vision.internal.rcnn.parseInputs(trainingData, network, options, mfilename, varargin{:});
Error in imagenn (line 20)
detector = trainRCNNObjectDetector(trainingData, layers, options);
Error in run (line 91)
evalin('caller', strcat(script, ';'));
I'm not sure, but I believe all these errors stem from an improperly declared classificationLayer. I have two classes, called 'clubroot' and 'healthy'. I'm not sure how to set up the network so it recognizes these two classes.
If anyone could offer help, I would be eternally grateful. Getting this to work is very important to me.

请先登录，再回答此问题。

采纳的回答

Ameer Hamza
Ameer Hamza 2018-4-23
As given here, you are improperly initializing the fullyConnectedLayer. Instead of using fullyConnectedLayer(10) try something like this.
classes = {'first', 'second'}
outputs = 1+numel(classes); % +1 for background class
layers = [imageInputLayer([2160 3840 3])
convolution2dLayer([5 5],10)
reluLayer()
fullyConnectedLayer(outputs)
softmaxLayer()
classificationLayer()];
  2 个评论
Griffon Thomas
Griffon Thomas 2018-4-23
编辑：Griffon Thomas 2018-4-23

Ack! This has seemed to fix the first problem, but it's run into another one. Here's what it looks like now:

*******************************************************************
Training an R-CNN Object Detector for the following object classes:

* clubroot
* healthy

Step 1 of 3: Extracting region proposals from 2 training images...done.

Step 2 of 3: Training a neural network to classify objects in training data...

Error using trainNetwork (line 140)
No public property 'NumObservations' for class 'RegionReader'.

Error in rcnnObjectDetector.train (line 222)
            detector.Network = trainNetwork(dispatcher, layers, opts);

Error in trainRCNNObjectDetector (line 197)
detector = rcnnObjectDetector.train(trainingData, layers, options, params);

Error in imagenn (line 23)
detector = trainRCNNObjectDetector(trainingData, layers, options);

Error in run (line 91)
evalin('caller', strcat(script, ';'));

Caused by:
    No public property 'NumObservations' for class 'RegionReader'.

Now it seems that there is no public property NumObservations for class RegionReader. However, we're through the first step with no problems! Again, I'm way out of my depth here. The deep learning onramp never prepared me for this!

Ameer Hamza
Ameer Hamza 2018-4-23
I am not very familiar with Computer vision toolbox and also don't have computer vision toolbox installed right now. You might want to start a new question since your comment is not directly related to this question topic and other volunteers will not be able to notice this problem.

请先登录，再进行评论。

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Deep Learning Toolbox 的更多信息

标签

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by