chrishkchris commented on issue #646: SINGA-487 Test case for distributed module
URL: https://github.com/apache/singa/pull/646#issuecomment-607628829
 
 
   > > > @XJDKC I also added `self.fused_all_reduce([p.data], send=False)` in 
the function `backward_and_partial_update`
   > > > to adopt the change last time in the `fused_all_reduce` function
   > > 
   > > 
   > > You can pass the whole plist to the fused_all_reduce function, no need 
to pass tensors one by one.
   > 
   > done, it becomes
   > 
   > ```
   >                         self.fused_all_reduce(plist, send=False)
   >                         self.fused_all_reduce(plist)
   > ```
   
   
   
   
   The asychronous training is also okay now:
   ```
   root@71ac539cda77:~/dcsysh/singa/examples/autograd# python3 
cifar10_multiprocess.py
   Starting Epoch 0:
   Training loss = 2910.528564, training accuracy = 0.335027
   Evaluation accuracy = 0.438902, Elapsed Time = 148.023807s
   Starting Epoch 1:
   Training loss = 2098.108887, training accuracy = 0.512984
   Evaluation accuracy = 0.558393, Elapsed Time = 147.847368s
   Starting Epoch 2:
   Training loss = 1696.253052, training accuracy = 0.612596
   Evaluation accuracy = 0.634615, Elapsed Time = 145.889916s
   Starting Epoch 3:
   Training loss = 1424.486328, training accuracy = 0.678877
   Evaluation accuracy = 0.725461, Elapsed Time = 144.518893s
   Starting Epoch 4:
   Training loss = 1202.804688, training accuracy = 0.732714
   Evaluation accuracy = 0.731070, Elapsed Time = 145.818738s
   
   ```

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to 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