Hi Anton, Thanks for driving the patch release. Besides the MKL improvements, I suggest we include two changes for *performance improvement* for NLP tasks below:
CUDNN support for LSTM with projection & clipping: - https://github.com/apache/incubator-mxnet/pull/13056 - It is used in state of the art language models such as BIG-LSTM [1] and Elmo (ACL 2018 best paper) [2] sample_like operators: - https://github.com/apache/incubator-mxnet/pull/13034 - Many models require candidate sampling (e.g. word2vec [3], fasttext [4]) for training. The sample_like operator enables drawing random samples without shape information, therefore the candidate sampling blocks can now be hybridized and be accelerated a lot. If there is no concern I will open two PRs for the above two changes to 1.3.x branch. Thanks! Best, Haibin [1] https://arxiv.org/pdf/1602.02410.pdf [2] https://arxiv.org/pdf/1802.05365.pdf [3] https://papers.nips.cc/paper/5021-distributed-representations-of-words-and-phrases-and-their-compositionality.pdf [4] https://arxiv.org/pdf/1607.01759.pdf On Mon, Nov 5, 2018 at 11:11 AM Anton Chernov <mecher...@gmail.com> wrote: > It seems that there is a problem porting following changes to the v1.3.x > release branch: > > Implement mkldnn convolution fusion and quantization > https://github.com/apache/incubator-mxnet/pull/12530 > > MKL-DNN Quantization Examples and README > https://github.com/apache/incubator-mxnet/pull/12808 > > The bases are different. > > I would need help from authors of these changes to make a backport PR. > > @ZhennanQin, @xinyu-intel would you be able to assist me and create the > corresponding PR's? > > Without proper history and domain knowledge I would not be able to create > them by my own in reasonable amount of time, I'm afraid. > > Best regards, > Anton > > пн, 5 нояб. 2018 г. в 19:45, Anton Chernov <mecher...@gmail.com>: > > > > > As part of: > > > > Implement mkldnn convolution fusion and quantization > > https://github.com/apache/incubator-mxnet/pull/12530 > > > > I propose to add the examples and documentation PR as well: > > > > MKL-DNN Quantization Examples and README > > https://github.com/apache/incubator-mxnet/pull/12808 > > > > > > Best regards, > > Anton > > > > пн, 5 нояб. 2018 г. в 19:02, Anton Chernov <mecher...@gmail.com>: > > > >> Dear MXNet community, > >> > >> I will be the release manager for the upcoming 1.3.1 patch release. > >> Naveen will be co-managing the release and providing help from the > >> committers side. > >> > >> The following dates have been set: > >> > >> Code Freeze: 31st October 2018 > >> Release published: 13th November 2018 > >> > >> Release notes have been drafted here [1]. > >> > >> > >> * Known issues > >> > >> Update MKL-DNN dependency > >> https://github.com/apache/incubator-mxnet/pull/12953 > >> > >> This PR hasn't been merged even to master yet. Requires additional > >> discussion and merge. > >> > >> distributed kvstore bug in MXNet > >> https://github.com/apache/incubator-mxnet/issues/12713 > >> > >> > When distributed kvstore is used, by default gluon.Trainer doesn't > work > >> with mx.optimizer.LRScheduler if a worker has more than 1 GPU. To be > more > >> specific, the trainer updates once per GPU, the LRScheduler object is > >> shared across GPUs and get a wrong update count. > >> > >> This needs to be fixed. [6] > >> > >> > >> * Changes > >> > >> The following changes will be ported to the release branch, per [2]: > >> > >> Infer dtype in SymbolBlock import from input symbol [3] > >> https://github.com/apache/incubator-mxnet/pull/12412 > >> > >> [MXNET-953] Fix oob memory read > >> https://github.com/apache/incubator-mxnet/pull/12631 > >> > >> [MXNET-969] Fix buffer overflow in RNNOp > >> https://github.com/apache/incubator-mxnet/pull/12603 > >> > >> [MXNET-922] Fix memleak in profiler > >> https://github.com/apache/incubator-mxnet/pull/12499 > >> > >> Implement mkldnn convolution fusion and quantization (MXNet Graph > >> Optimization and Quantization based on subgraph and MKL-DNN proposal > [4]) > >> https://github.com/apache/incubator-mxnet/pull/12530 > >> > >> Following items (test cases) should be already part of 1.3.0: > >> > >> [MXNET-486] Create CPP test for concat MKLDNN operator > >> https://github.com/apache/incubator-mxnet/pull/11371 > >> > >> [MXNET-489] MKLDNN Pool test > >> https://github.com/apache/incubator-mxnet/pull/11608 > >> > >> [MXNET-484] MKLDNN C++ test for LRN operator > >> https://github.com/apache/incubator-mxnet/pull/11831 > >> > >> [MXNET-546] Add unit test for MKLDNNSum > >> https://github.com/apache/incubator-mxnet/pull/11272 > >> > >> [MXNET-498] Test MKLDNN backward operators > >> https://github.com/apache/incubator-mxnet/pull/11232 > >> > >> [MXNET-500] Test cases improvement for MKLDNN on Gluon > >> https://github.com/apache/incubator-mxnet/pull/10921 > >> > >> Set correct update on kvstore flag in dist_device_sync mode (as part of > >> fixing [5]) > >> https://github.com/apache/incubator-mxnet/pull/12786 > >> > >> upgrade mshadow version > >> https://github.com/apache/incubator-mxnet/pull/12692 > >> But another PR will be used instead: > >> update mshadow > >> https://github.com/apache/incubator-mxnet/pull/12674 > >> > >> CudnnFind() usage improvements > >> https://github.com/apache/incubator-mxnet/pull/12804 > >> A critical CUDNN fix that reduces GPU memory consumption and addresses > >> this memory leak issue. This is an important fix to include in 1.3.1 > >> > >> > >> From discussion about gluon toolkits: > >> > >> disable opencv threading for forked process > >> https://github.com/apache/incubator-mxnet/pull/12025 > >> > >> Fix lazy record io when used with dataloader and multi_worker > 0 > >> https://github.com/apache/incubator-mxnet/pull/12554 > >> > >> fix potential floating number overflow, enable float16 > >> https://github.com/apache/incubator-mxnet/pull/12118 > >> > >> > >> > >> * Resolved issues > >> > >> MxNet 1.2.1–module get_outputs() > >> https://discuss.mxnet.io/t/mxnet-1-2-1-module-get-outputs/1882 > >> > >> As far as I can see from the comments the issue has been resolved, no > >> actions need to be taken for this release. [7] is mentioned in this > >> regards, but I don't see any action points here either. > >> > >> > >> I will start with help of Naveen port the mentioned PR's to the 1.3.x > >> branch. > >> > >> > >> Best regards, > >> Anton > >> > >> [1] https://cwiki.apache.org/confluence/x/eZGzBQ > >> [2] > >> > https://cwiki.apache.org/confluence/display/MXNET/Project+Proposals+for+next+MXNet+Release > >> [3] https://github.com/apache/incubator-mxnet/issues/11849 > >> [4] > >> > https://cwiki.apache.org/confluence/display/MXNET/MXNet+Graph+Optimization+and+Quantization+based+on+subgraph+and+MKL-DNN > >> [5] https://github.com/apache/incubator-mxnet/issues/12713 > >> [6] > >> > https://github.com/apache/incubator-mxnet/issues/12713#issuecomment-435773777 > >> [7] https://github.com/apache/incubator-mxnet/pull/11005 > >> > >> >