[GitHub] xingzhaoo commented on issue #5460: very low training accuracies with alexnet or vgg but good with resnet

[gi...@git.apache.org]

xingzhaoo commented on issue #5460: very low training accuracies with alexnet or vgg but good with resnet
URL: https://github.com/apache/incubator-mxnet/issues/5460#issuecomment-319545407
 
 
   I have the same problem for alexnet. I am using CPU only and running alexnet on asp over 2 workers.
   I check the alexnet.py under the symbols directory. The code appears that the original alexnet with 2 columns(in 2 GPUs) are integrated into one. Is that the reason which causes this underfitting  problem? 
   ```
   INFO:root:Epoch[7] Batch [3120]	Speed: 5.91 samples/sec	accuracy=0.078125	top_k_accuracy_5=0.025000
   INFO:root:Epoch[8] Batch [2000]	Speed: 6.60 samples/sec	accuracy=0.062500	top_k_accuracy_5=0.015625
   INFO:root:Epoch[7] Batch [3140]	Speed: 6.07 samples/sec	accuracy=0.056250	top_k_accuracy_5=0.025000
   INFO:root:Epoch[8] Batch [2020]	Speed: 6.55 samples/sec	accuracy=0.078125	top_k_accuracy_5=0.031250
   INFO:root:Epoch[7] Batch [3160]	Speed: 6.24 samples/sec	accuracy=0.062500	top_k_accuracy_5=0.006250
   ```
 
----------------------------------------------------------------
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:
users@infra.apache.org


With regards,
Apache Git Services