mxnet example training cannot get expected result after 0.5.0

[GeorgeXia1828]

For bugs or installation issues, please provide the following information.
The more information you provide, the more likely people will be able to help you.

Environment info

Operating System: Linux idc01-rank-gpu-01 2.6.32-431.el6.x86_64 #1 SMP Sun Nov 10 22:19:54 EST 2013 x86_64 x86_64 x86_64 GNU/Linux

Compiler: gcc-4.8.5

Package used (Python/R/Scala/Julia): Python

MXNet version: 0.9.4

Or if installed from source: yes

MXNet commit hash (git rev-parse HEAD):

If you are using python package, please provide

Python version and distribution: 2.7.10

If you are using R package, please provide

R sessionInfo():

Error Message:

Please paste the full error message, including stack trace.

Minimum reproducible example

if you are using your own code, please provide a short script that reproduces the error.

###################################
I am trying the new mxnet to do the inception-bn-full test. I was tried successful in mxnet 0.5.0, but I found still not work in mxnet 0.9.4, even try to set fix_gamma = false.

And I found the train acc is very strange which vary unstable. (when doing test in 0.5.0 it increase very stable).

Also I also doing a simple test in mnist, I turn the mnist data to images, then processing then to rec. using alexnet to training the network. It does not work at all.

So I guess is there something wrong with the code in train_imagenet.py or etc. or sth bug exist.

[piiswrong]

There have been some incompatible changes since 0.5. You can try models trained with the new version at data.mxnet.io

[GeorgeXia1828]

@piiswrong thanks,
Before I report the problem, I have already turn into 0.9.4, and failed to do image classification tasks.
Actually, I have some tasks in image classification the scale is around ImageNet tasks, So how can I try to do the train if not turn back to 0.5.0.

[GeorgeXia1828]

@piiswrong
I have been frustrated here, and I try to detect the error, I just found one significant difference in new version training process, the acc is changing unstable while in old version (0.5.0) the changing is very stable and can finally increase to a good score.
I guess something with cuda matters? or can you give some building suggestion?

[szha]

This issue is closed due to lack of activity in the last 90 days. Feel free to ping me to reopen if this is still an active issue. Thanks!