Skip to content

Releases: dmlc/keras

Alpha release for Keras with MXNet backend.

07 Jul 18:52
@piiswrong piiswrong
alpha
9a6e242
Compare
Choose a tag to compare
Alpha release for Keras with MXNet backend. Pre-release
Pre-release

Alpha

Release Notes: Keras 1.2.2 with MXNet Backend 

Highlights

  1. Adding Apache MXNet backend for Keras 1.2.2.
  2. Easy to use multi-gpu training with MXNet backend.
  3. High-performance model training in Keras with MXNet backend.

Unsupported Functionalities

  1. Sparse operations are not supported - ctc_label_dense_to_sparse, ctc_batch_cost, ctc_decode.
  2. Few backend utility functions - stack, gradients, switch.
  3. separable_conv2d - 2D convolution with separable filters are not supported
  4. random_binomial is not supported.
  5. Higher order utility functions - map, foldl, foldr.
  6. Cannot use Numpy’s COOrdinate representation of the array: No sparse support.
  7. Eigenvalue regularizer is not supported.
  8. With MXNet backend, Model concatenation is not supported.

Known Issues

  1. Batch Normalization does not support mode 0 or 2 with MXNet backend.
  2. Optimizer states are not preserved when saving the model, with MXNet backend.
  3. If the batch size is too small and you are using multiple GPUs for training, MXNet backend fails to split the data gracefully and errors out.
  4. SoftmaxCrossEntropy only accepts 1D labels.
  5. MXNet Categorical_crossentropy doesn't support from_logits.
  6. Few failing unit tests and cause for failure - #65
  7. Deconvolution operator with large target shape fails with MXNet backend.

Miscellaneous

  1. Pip installation for dmlc/keras with MXNet backend. Will be added in Beta release.
  2. Latest MXNet v0.10 has few missing functionalities to work with Keras. You need to build from latest source. 3. Refer installation guide in the next section below. Pip wheel will be made available for Beta release.
  3. Performance improvements are not focussed in this release. MXNet’s full power may not be harnessed. 
  4. Training a model with other backends (ex: TF) and loading it with MXNet backend is not tested. This functionality is not supported.
  5. Needs more code documentation for MXNet specific code.
  6. Needs code style fix and clean up.

Contributions

Thanks to all contributors who made significant contributions to get MXNet backend for Keras - @howard0su, @yajiedesign, @piiswrong, @wayao

Other Resources and Getting Started

  1. Getting Started and Installation - https://github.com/dmlc/keras/wiki/Installation
  2. How to use Multi-GPU training - https://github.com/dmlc/keras/wiki/Using-Keras-with-MXNet-in-Multi-GPU-mode
  3. Benchmarking performance of Keras 1.2.2 with MXNet backend - https://github.com/sandeep-krishnamurthy/keras-mxnet-benchmarks
  4. End to end examples and test script - https://github.com/sandeep-krishnamurthy/keras-mxnet-tests
Assets 2
Loading