python - how loss and metric are calculated in cntk -
i new cntk , python. have created python program based on trainresnet_cifar10.py train 4736 of (64x64x3) images , test 2180 images 4 classes. after train 160 epochs, got loss = 0.663 , metric = 29.37%. finished evaluation metric = 18.94%. when evaluate train model based on cntklibrarycsevalexamples.cs test 2180 images, 2180 classified 1 class (second class). questions are:
- i assume loss calculated cross_entropy_with_softmax(z, label_var) , metric using classification_error(z, label_var). correct , how determined?
- what mean of metric = 29.37% , evaluation metric = 18.94%? train , test images, respectively?
- what cause totally wrong evaluate results?
any appreciated.
(1) yes.
(2) 29.37% mean 29.37% of classification correct. evaluation on test data, assuming reading both training , test data.
(3) make sure input same format, mean normalize or subtract mean in python, if need same in c#. can run eval first using python , see result get?
Comments
Post a Comment