[GitHub] thomasmooon commented on issue #7475: Paradox VRAM demand as a function of batch size: Low batch size, high VRAM demand

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

thomasmooon commented on issue #7475: Paradox VRAM demand as a function of batch size: Low batch size, high VRAM demand
URL: https://github.com/apache/incubator-mxnet/issues/7475#issuecomment-323508007
 
 
   Yes regarding cuDNN.
   And no regarding convolutions.
   Im not processing image Data.
   
   It is a  stacked denoising autoencoder as described in the example simply
   build as follows:
   - Input Layer, corrupted with dropout.
   - 3 fully connected hidden Layers
   - Output layer same dimension as Input
   
   
   
   
   Am 18.08.2017 22:07 schrieb "Przemyslaw Tredak" <no...@github.com>:
   
   I assume you run with cuDNN. Do you have convolutions in your network? If
   so, the most probable explanation is that for smaller batch sizes a
   different algorithm is chosen to perform convolution (since it performs
   faster) with different requirements on additional memory used.
   
   ?
   You are receiving this because you authored the thread.
   Reply to this email directly, view it on GitHub
   <https://github.com/apache/incubator-mxnet/issues/7475#issuecomment-323449186>,
   or mute the thread
   <https://github.com/notifications/unsubscribe-auth/Ab38wQdUoWzO_MMB7U8KMboJeYShLVHaks5sZe8XgaJpZM4O3d_M>
   .
   
 
----------------------------------------------------------------
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