Plan

[ankane]

Current status: the C API doesn't yet support gradients for eager tensors, so work is on-hold until it's added. Two alternatives approaches are 1. use the graph (TensorFlow 1.0) API or 2. use the C++ API instead.

Plan

• Auto-generate ops - use either TF_GetAllOpList or tensorflow/core/ops/ops.pbtxt
  • Generate Tf::RawOps
  • Generate Tf::Math that uses Tf::RawOps
  • Generate Tf methods that uses Tf::Math
• Numo integration
• Data::Dataset module
• Keras Beginner - keras_beginner branch
  • MNIST load data
  • Compile model
  • Fit model
• Keras Advanced (subclassing) - keras_advanced branch
  • Model module
• Save and load models - saved_model branch
• Switch to the unified C API once released

Resources

• Guide for language bindings

Maybe (later)

• Flatten classes
  • Tf::Keras::Models::Sequential -> Tf::Sequential or Tf::SequentialModel
  • Tf::Data::Dataset -> Tf::Dataset
  • etc

[sp00ck]

In future You planing using tensorflow lite? Meybe with compiling tensorflow lite directly in gem? or only convert https://www.tensorflow.org/lite/microcontrollers/build_convert

[eggie5]

doesn't the TF python API use swig? Maybe that'd be a better way to do the ruby port?

* Guide for language bindings

This link is broken. Would you mind having a look here?

[ankane]

@sp00ck (sorry for the long delay) There aren't any plans to support TensorFlow Lite.

@eggie5 From what I can tell, it uses the C++ API (but not SWIG). This is probably the quickest path forward, but requires a complete rewrite. Someone is welcome to try this.

@Ashvith Let me know if you find the updated link.

@ankane I believe it must be this.

[ankane]

Great, updated.

[sp00ck]

@ankane meybe, I can't check. Not working on my hardware x86 with Fedora

[simonstearn1]

Any plans to support the Tensorboard utils ?