I want to apply Alexnet to faster RCNN

Hello, I am Jungmoon. I am studying the faster RCNN. And I am trying to apply the Alexnet to faster RCNN. I found the contents of Alexnet's network in matlab' help. Here is alexnet's layer array.

Then, I modified faster RCNN with the contents of Alexnet. There is big problem. I show the problem with a picture.

Pleas help me.

[Jungmoon Kim]

Here is my code.

Please....

Thank you~!

inputLayer = imageInputLayer([227 227 3]);
middleLayers = [
    convolution2dLayer([11 11], 96,'stride',4, 'Padding', 0, 'NumChannels',3 )
    reluLayer()
    crossChannelNormalizationLayer(5)
    maxPooling2dLayer(3,'stride',2, 'Padding',0)
    convolution2dLayer([5 5], 256,'stride',1, 'Padding', 2, 'NumChannels',96)    
    reluLayer()
    crossChannelNormalizationLayer(5)
    maxPooling2dLayer(3,'stride',2, 'Padding',0)
    convolution2dLayer([3 3], 384,'stride',1, 'Padding', 1,'NumChannels',256)
    reluLayer()
    convolution2dLayer([3 3], 384,'stride',1, 'Padding', 1, 'NumChannels',384)
    reluLayer()
    convolution2dLayer([3 3], 256,'stride',1, 'Padding', 1,'NumChannels',384)
    reluLayer()
    maxPooling2dLayer(3,'stride',2, 'Padding',0)
    ];
finalLayers = [
    % Add a fully connected layer with 64 output neurons. The output size
    % of this layer will be an array with a length of 64.
    fullyConnectedLayer(4096)
    reluLayer()
    dropoutLayer(0.5)
    fullyConnectedLayer(4096)
    reluLayer()
    dropoutLayer(0.5)

      fullyConnectedLayer(width(Dataset))

      % Add the softmax loss layer and classification layer. 
      softmaxLayer()
      classificationLayer()
  ];

layers = [
    inputLayer
    middleLayers
    finalLayers
    ]

optionsStage1 = trainingOptions('sgdm', ...
    'MiniBatchSize', 4, ...
    'MaxEpochs', 10, ...
    'InitialLearnRate', 1e-5, ...
    'CheckpointPath', tempdir);

% Options for step 2
optionsStage2 = trainingOptions('sgdm', ...
    'MiniBatchSize', 4, ...
    'MaxEpochs', 10, ...
    'InitialLearnRate', 1e-5, ...
    'CheckpointPath', tempdir);

% Options for step 3.
optionsStage3 = trainingOptions('sgdm', ...
    'MiniBatchSize', 4, ...
    'MaxEpochs', 10, ...
    'InitialLearnRate', 1e-6, ...
    'CheckpointPath', tempdir);

% Options for step 4.
optionsStage4 = trainingOptions('sgdm', ...
    'MiniBatchSize', 4, ...
    'MaxEpochs', 10, ...
    'InitialLearnRate', 1e-6, ...
    'CheckpointPath', tempdir);

options = [
    optionsStage1
    optionsStage2
    optionsStage3
    optionsStage4
    ];

tic
% A trained network is loaded from disk to save time when running the
% example. Set this flag to true to train the network. 
doTrainingAndEval = true %false;

if doTrainingAndEval
    % Set random seed to ensure example training reproducibility.
    rng(0);

      % Train Faster R-CNN detector. Select a BoxPyramidScale of 1.2 to allow
      % for finer resolution for multiscale object detection.
      detector = trainFasterRCNNObjectDetector(trainingData, layers, options, ...
          'NegativeOverlapRange', [0 0.3], ...
          'PositiveOverlapRange', [0.6 1], ...
          'SmallestImageDimension', [400], ...
          'BoxPyramidScale', 1.2);
  else
      %         'NumStrongestRegions', 500, ...
  %         'MinBoxSizes', [21 21], ...
      % Load pretrained detector for the example.
      detector = data.detector;
  end
  toc

[Joss Knight]

AlexNet comes with Matlab as an add-on. Just type net = alexnet at the command line.

Answer 1

[Carl]

Hi Jungmoon. The [48,48] and [128,128] indicate that the 96 channels and 256 filters are split up into two groups. This is part of the implementation of AlexNet. Each set of filters are applied to a different group. As of now, there is no way of replicating this behavior using the "convolution2dLayer" function in the Neural Network Toolbox.

You should be fine specifying 96 and 256 for NumChannels and NumFilters. I believe the original intention of this behavior was just to make it easier to split the computation between GPUs. Functionally, I believe your current setup should behave similarly.

[Jungmoon Kim]

Thank you very much~!

[Elif BAYKAL]

I also met the same problem. Could you share how you solved it, please?

[Carlos]

share the solution, please

Answer 2

[taha khan]

https://www.mathworks.com/matlabcentral/answers/353563-i-want-to-apply-alexnet-to-faster-rcnn#answer_317984 sir i am retrain AlexNet throught this way... retrain Alex Net .....but my training speed is too much slow? sir any one help me how to increase my training speed ?? how to increase my processing power of CPU ???

Answer 3

[taha khan]

after training AlexNet according to this what up next ? https://www.mathworks.com/videos/deep-learning-with-matlab-transfer-learning-in-10-lines-of-matlab-code-1487714838381.html

how to do live streaming with train network ?