search

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB
Other
1.4k stars 753 forks source link

train the model from scratch #908

Open qiulesun opened 7 years ago

qiulesun commented 7 years ago

I want to train the vgg-vd-16 model from scratch, which parameter initialization method should I choose, gaussian or xavier or xavierimproved?

ankush-me commented 7 years ago

If you use batch-normalization, then training is not too sensitive to the initialization; i.e., with batch-normalization, any / all of the above methods should work.

qiulesun commented 7 years ago

Thank you for your answer.I use the batch-normalization,but the top1err decreased only from 0.99 to 0.91 after 5 epochs, and the convergence rate was slow.Is this normal?