Squeezenet model not training in MatlabR2017b.

Hello,

I have generated a Squeezenet basic model(vanilla) using Matlab R2017b. I am having exactly same implementation as it is in the Squeezenet implementation using Caffe.

Below is my Matlab code: I am using image datastore object with 10 classes "indsRand10.mat" which is subset of ImageNet dataset.

%%Squeezenet network
%%--Bhushan Muthiyan

imdbPath = fullfile(pwd, 'indsRand10.mat') ;

if exist(imdbPath, 'file')

      imdb = load(imdbPath) ;

      trainingNumFiles = 768;

      valNumFiles = 64;

      rng(1) % For reproducibility

      [imdb.trainDigitData, imdb.testDigitData] = splitEachLabel(imdb.imdsTrain, ...

          trainingNumFiles,'randomize');

      [imdb.testDigitData, a] = splitEachLabel(imdb.testDigitData, ...

          valNumFiles,'randomize');

end

numImages = numel(imdb.trainDigitData.Files);

idx = randperm(numImages,20);

for i = 1:20

      subplot(4,5,i)

      I = readimage(imdb.trainDigitData, idx(i));

      imshow(I)

end

numClasses = numel(categories(imdb.trainDigitData.Labels));

layers = [
    imageInputLayer([224 224 3],'Name','input')

      convolution2dLayer(7,96,'Padding','same','Stride',2,'Name','conv_1')
      reluLayer('Name','relu_1')
      maxPooling2dLayer(3,'Stride',2,'Name','pool_1')

      %%fire 2 
      convolution2dLayer(1,16,'Padding','same','Stride',1,'Name','conv_2')
      reluLayer('Name','relu_11')
      convolution2dLayer(1,64,'Padding','same','Stride',1,'Name','conv_3')
      reluLayer('Name','relu_2')
      depthConcatenationLayer(2,'Name','concat_1')

      %%fire 3
      convolution2dLayer(1,16,'Padding','same','Stride',1,'Name','conv_6')
      reluLayer('Name','relu_12')
      convolution2dLayer(1,64,'Padding','same','Stride',1,'Name','conv_7')
      reluLayer('Name','relu_3')
      depthConcatenationLayer(2,'Name','concat_3')

      maxPooling2dLayer(3,'Stride',2,'Name','pool_2')
      %%fire 4
      convolution2dLayer(1,32,'Padding','same','Stride',1,'Name','conv_9')
      reluLayer('Name','relu_13')
      convolution2dLayer(1,128,'Padding','same','Stride',1,'Name','conv_10')
      reluLayer('Name','relu_4')
      depthConcatenationLayer(2,'Name','concat_5')

      %%fire 5
      convolution2dLayer(1,32,'Padding','same','Stride',1,'Name','conv_12')
      reluLayer('Name','relu_14')
      convolution2dLayer(1,128,'Padding','same','Stride',1,'Name','conv_13')
      reluLayer('Name','relu_5')
      depthConcatenationLayer(2,'Name','concat_6')

      maxPooling2dLayer(3,'Stride',2,'Name','pool_3')

      %fire 6
      convolution2dLayer(1,48,'Padding','same','Stride',1,'Name','conv_15')
      reluLayer('Name','relu_15')
      convolution2dLayer(1,192,'Padding','same','Stride',1,'Name','conv_16')
      reluLayer('Name','relu_6')
      depthConcatenationLayer(2,'Name','concat_8')

      %%fire 7
      convolution2dLayer(1,48,'Padding','same','Stride',1,'Name','conv_18')
      reluLayer('Name','relu_16')
      convolution2dLayer(1,192,'Padding','same','Stride',1,'Name','conv_19')
      reluLayer('Name','relu_7')
      depthConcatenationLayer(2,'Name','concat_9')

      %fire 8
      convolution2dLayer(1,64,'Padding','same','Stride',1,'Name','conv_21')
      reluLayer('Name','relu_17')
      convolution2dLayer(1,256,'Padding','same','Stride',1,'Name','conv_22')
      reluLayer('Name','relu_8')
      depthConcatenationLayer(2,'Name','concat_11')

      % fire 9
      convolution2dLayer(1,64,'Padding','same','Stride',1,'Name','conv_24')
      reluLayer('Name','relu_18')
      convolution2dLayer(1,256,'Padding','same','Stride',1,'Name','conv_25')
      depthConcatenationLayer(2,'Name','concat_12')

      %reluLayer('Name','relu_9')

      dropoutLayer(0.5,'Name','Drop_1')
      convolution2dLayer(1,numClasses,'Padding','same','Stride',1,'Name','conv_27')
      reluLayer('Name','relu_9')
      averagePooling2dLayer(13,'Stride',1,'Name','avg_pool_4')
      %reluLayer('Name','relu_10')
      softmaxLayer('Name','softmax')
      classificationLayer('Name','classOutput')];

lgraph = layerGraph(layers);

figure

plot(lgraph)

conv_4 = convolution2dLayer(3,64,'Padding',1,'Stride',1,'Name','conv_4');

lgraph = addLayers(lgraph,conv_4);

conv_8 = convolution2dLayer(3,64,'Padding',1,'Stride',1,'Name','conv_8');

lgraph = addLayers(lgraph,conv_8);

conv_11 = convolution2dLayer(3,128,'Padding',1,'Stride',1,'Name','conv_11');

lgraph = addLayers(lgraph,conv_11);

conv_14 = convolution2dLayer(3,128,'Padding',1,'Stride',1,'Name','conv_14');

lgraph = addLayers(lgraph,conv_14);

conv_17 = convolution2dLayer(3,192,'Padding',1,'Stride',1,'Name','conv_17');

lgraph = addLayers(lgraph,conv_17);

conv_20 = convolution2dLayer(3,192,'Padding',1,'Stride',1,'Name','conv_20');

lgraph = addLayers(lgraph,conv_20);

conv_23 = convolution2dLayer(3,256,'Padding',1,'Stride',1,'Name','conv_23');

lgraph = addLayers(lgraph,conv_23);

conv_26 = convolution2dLayer(3,256,'Padding',1,'Stride',1,'Name','conv_26');

lgraph = addLayers(lgraph,conv_26);

relu_19 = reluLayer('Name','relu_19');

lgraph = addLayers(lgraph,relu_19);

relu_20 = reluLayer('Name','relu_20');

lgraph = addLayers(lgraph,relu_20);

relu_21 = reluLayer('Name','relu_21');

lgraph = addLayers(lgraph,relu_21);

relu_22 = reluLayer('Name','relu_22');

lgraph = addLayers(lgraph,relu_22);

relu_23 = reluLayer('Name','relu_23');

lgraph = addLayers(lgraph,relu_23);

relu_24 = reluLayer('Name','relu_24');

lgraph = addLayers(lgraph,relu_24);

relu_25 = reluLayer('Name','relu_25');

lgraph = addLayers(lgraph,relu_25);

relu_26 = reluLayer('Name','relu_26');

lgraph = addLayers(lgraph,relu_26);

lgraph = connectLayers(lgraph,'relu_11','conv_4');

lgraph = connectLayers(lgraph,'conv_4','relu_19');

lgraph = connectLayers(lgraph,'relu_19','concat_1/in2');

lgraph = connectLayers(lgraph,'relu_12','conv_8');

lgraph = connectLayers(lgraph,'conv_8','relu_20');

lgraph = connectLayers(lgraph,'relu_20','concat_3/in2');

lgraph = connectLayers(lgraph,'relu_13','conv_11');

lgraph = connectLayers(lgraph,'conv_11','relu_21');

lgraph = connectLayers(lgraph,'relu_21','concat_5/in2');

lgraph = connectLayers(lgraph,'relu_14','conv_14');

lgraph = connectLayers(lgraph,'conv_14','relu_22');

lgraph = connectLayers(lgraph,'relu_22','concat_6/in2');

lgraph = connectLayers(lgraph,'relu_15','conv_17');

lgraph = connectLayers(lgraph,'conv_17','relu_23');

lgraph = connectLayers(lgraph,'relu_23','concat_8/in2');

lgraph = connectLayers(lgraph,'relu_16','conv_20');

lgraph = connectLayers(lgraph,'conv_20','relu_24');

lgraph = connectLayers(lgraph,'relu_24','concat_9/in2');

lgraph = connectLayers(lgraph,'relu_17','conv_23');

lgraph = connectLayers(lgraph,'conv_23','relu_25');

lgraph = connectLayers(lgraph,'relu_25','concat_11/in2');

lgraph = connectLayers(lgraph,'relu_18','conv_26');

lgraph = connectLayers(lgraph,'conv_26','relu_26');

lgraph = connectLayers(lgraph,'relu_26','concat_12/in2');

figure

plot(lgraph);

optionsTransfer = trainingOptions('sgdm', ...

      'MaxEpochs',25, ...

      'MiniBatchSize',64,...

      'InitialLearnRate',0.04,...

      'LearnRateDropFactor',0.2,...

      'LearnRateDropPeriod',5,...

      'Plots','training-progress',...

      'ExecutionEnvironment','auto');

netTransfer = trainNetwork(imdb.trainDigitData,lgraph,optionsTransfer);

YPred = classify(netTransfer,imdb.testDigitData);
YTest = imdb.testDigitData.Labels;

accuracy = sum(YPred==YTest)/numel(YTest);

fprintf('accuracy = %f\n',accuracy);

numImages = numel(imdb.testDigitData.Files);

idx = randperm(numImages,20);

for i = 1:20

      subplot(4,5,i)

      I = readimage(imdb.testDigitData, idx(i));

      imshow(I)

end

Can someone let me know the reason behind this.

Enclosed here is the image of Squeezenet vanilla model structure.

Answer 1

[Mickaël Tits]

Hi,

If I understand, you are trying to train your Squeezenet model from scratch, with 768 images ? You need a pretrained model if you want a chance that it works.

You can get here a pretrained SqueezeNet, and use it for transfer learning as you want : https://github.com/titsitits/Squeezenet-Matlab-Keras

Mickaël Tits

[BHUSHAN MUTHIYAN]

Hello Mickaël,

I had a look to the link provided by you.

But, the .json file description says that the Keras model generated above has first convolution layer 3x3x64 whereas the original Keras implementation in Caffe has dimension 7x7x96 filter.

Can you please provide me with the exact implementation of Squeezenet model (.h5 file) which matches with original squeezenet implemetation.

Thanks!!