Baunsgaard commented on PR #1940: URL: https://github.com/apache/systemds/pull/1940#issuecomment-1796421365
> I don't know how to do the configuration. This is SystemDS out of the box. okay, we will figure it out. https://apache.github.io/systemds/site/run contains some information about MKL in the bottom. But the docs for it is a bit... outdated. > I didn't compare it to pyTorch but I suspect PyTorch is much faster with my current SystemDS configuration. But with my current configuration and the implementation of layers in SystemDS, there is not much to be optimized or runtime gained. I am already using the builtin conv2d layer and the convolutional layers need to happen the way they do in my implementation. Last time i ran our convolution layers yes, pyTorch was (much) faster, we have not worked on that operation for a while. > Feel free to run this forward pass yourself: [test_resnet_big.log](https://github.com/apache/systemds/files/13272287/test_resnet_big.log) ahh, i see. If you are profiling with this, then we cannot trust the results as much as we want. I suggest adding a for loop around the call to the basic_block_forward, to allow the JIT compiler to help you. it should improve the speed of each call. (Java is sometimes very bad if you do not call your things multiple times) -- 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. To unsubscribe, e-mail: dev-unsubscr...@systemds.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org
