chinakook commented on issue #7820: Low accuracy of pvanet
URL: 
https://github.com/apache/incubator-mxnet/issues/7820#issuecomment-328280564
 
 
   1. You can fix the same net weights in both caffe and mxnet, and debug layer 
by layer to see the result is whether the same.
   2. Check mean file and std scale value.
   3. Weights initialized with mx.random.normal in mxnet is worse than Xaviar 
or msra. So weights initialization is very sensitive in mxnet, otherwise you 
will get bad result as my experience.
   4. I can't understand the meaning of the line 'pool5 = 
mx.sym.Pooling(data=bsr, kernel=(1, 1), stride=(1, 1), pad=(0, 0), 
pool_type='max', name='pool5')'. Moreover, pooling is 'valid' in mxnet but 
'full' in caffe.
   
   Anyway, If you want to port caffe to mxnet, fix the same net weights and 
compare the result of inferencing result of both framework is needed. As for 
training, It's complicated because the platform differences.
 
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
[email protected]


With regards,
Apache Git Services

Reply via email to