is this an overfitting problem

[himora...@gmail.com]

Hi everyone,

I have a situation, don't know whether it is the overfitting or not?

I got 3 sets of data (used data augmentation: rotation and flipping)

  + Training data: 2 millions images

  + Validation data: 52k images (take random from the training data)

  + Testing data

The procedure include 3 steps

  + Step1: train CNN on training data and observe the accuracy on the validation data

  + Step2: use the trained CNN as feature extractor, extract feature on validation data and run SVM on the extracted feature to obtain the final model

  + Step3: test the final model on the testing data

I have run 3 cases

  + Case1: NW1(conv1=48,conv2=128,conv3=128,conv4=128,conv5=128,fc6=4096,fc7=4096,fc8=9699) => accuracy of step 1 is 57%, accuracy of step 3 is x

  + Case2: NW1(conv1=48,conv2=128,conv3=256,conv4=256,conv5=256,fc6=2048,fc7=2048,fc8=9699) => accuracy of step 1 is 61%, accuracy of step 3 is x1 (x1<x)

  + Case3: NW1(conv1=48,conv2=128,conv3=384,conv4=384,conv5=384,fc6=2048,fc7=2048,fc8=9699) => accuracy of step 1 is 64%, accuracy of step 3 is x2 (x2<x)

I have calculate the number of parameter in the case1: 100mil, case2: 70mil, case3: 90mil

because the case2 and case3 have less paramter, higher accuracy in step1, i expect it to have higher accuracy in step 3 but it didn't. Is this the overfitting problem

Hope to hear your suggestion and opinion in my situation

Best regards