Skip to first unread message
Buggiant
Aug 8, 2017, 2:27:40 AM8/8/17
to Caffe Users
I was training an AlexNet-like network on Cifar-10.
I got accuracy around 84% using network described in attached file cifar_Alex_Bench.prototxt, with each weight initialized with gaussion and different std each layer.
Then I tried Xavier filler instead of gaussion, as described in attached file cifar_Alex_xavier.prototxt.
And the accuracy went down to 0.1, which means the Xavier filler is not helping at all.
Why would this happen? I thought Xavier should be useful in practise.
I got accuracy around 84% using network described in attached file cifar_Alex_Bench.prototxt, with each weight initialized with gaussion and different std each layer.
Then I tried Xavier filler instead of gaussion, as described in attached file cifar_Alex_xavier.prototxt.
And the accuracy went down to 0.1, which means the Xavier filler is not helping at all.
Why would this happen? I thought Xavier should be useful in practise.
Vidhya Nagarajan
Aug 11, 2017, 2:09:10 AM8/11/17
to Caffe Users
I too faced the same. How filler type influencing the accuracy?
Reply all
Reply to author
Forward
0 new messages