Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
[RELEASE] Announcing v0.9 Release Candidate 1 #3509
We are announcing v0.9 rc1 https://github.com/dmlc/mxnet/tree/v0.9rc1. This includes backend refactor to use NNVM for operator registration and graph optimization.
List of Changes in v0.9 (excluding those in v0.8)
List of Pending Changes
These are not available in v0.9 rc1 and may or may not be available in final v0.9 release.
List of Changes in v0.8
@tqchen @piiswrong @tornadomeet
the newest TX1's compiler is g++-5.0. mxnet has a lot of problems on it.
Using g++-4.8 is ok for me.
v0.9.0rc1 compiling warnings[CPU mode]:
Maybe this is off the topic.
From above, I see:
"Vastly (10x - 100x) improved CPU speed with MKLDNN optimizations from Intel."
Where does "10x-100x" come from ?
I roughly read the source codes of MKL-DNN before. MKL-DNN trys to employ AVX2 and OpenMP to make things faster, and it will try to generate the operators' codes by using JIT at run-time. By using JIT, it could take some advantages by optimally hand coding some of the operators. It will be faster, but it shouldn't be more than 10x faster.
If it's really that fast on single machine, I am really wondering what is the cause. Or I missed something important while I was reading the MKL-DNN or mxnet source codes.
Failed to run "python train_cifar10_resnet.py --gpus 0"