Skip to content

Latest commit

 

History

History
53 lines (37 loc) · 2.11 KB

NEWS.md

File metadata and controls

53 lines (37 loc) · 2.11 KB
Raw

TensorFlow v0.10.0 Release Notes

  • Support for Julia 1.0 has been added. Support for all prior versions of Julia (except 0.7, which is functionally identical to 1.0) has been dropped.

TensorFlow v0.7 Release Notes

  • Support for Julia 0.5 has been dropped.
  • Enhanced support for visualization with TensorBoard.
  • Operations defined in C are now accessed by import_op(<op name>) instead of Ops.<op name>.

TensorFlow v0.6 Release Notes

API deprecations

The API has changed to resemble the finalized TensorFlow 1.0 API. See the TensorFlow release notes for a list of the changes.

In particular for TensorFlow.jl:

  • The summary operations, like train.scalary_summary, have moved to summary.scalar.
  • train.SummaryWriter has moved to summary.FileWriter.
  • The reduction_indices keyword argument have changed to axisin every function which used to take a reduction_indices argument.
  • mul has changed to multiply, and likewise for sub and neg.
  • pack is now stack and unpack is now unstack.
  • The *_cross_entropy_with_logits family of functions has changed to only accept keyword arguments instead of positional arguments.
  • The order of arguments to concat has switched (tensors to concatenate come first, followed by the concetenation axis).

Highlights

  • dynamic_rnn has been added
  • while_loop has been added, with a convenient @tf while ... end syntax
  • Support for Docker via official Docker images

Switch to 1-based indexing

Operations which take indices as arguments now expect the arguments to be 1-based instead of 0-based, which was a hold-over of TensorFlow's Python legacy. This affects the following functions:

  • The axis parameter for any operation which takes an axis parameter
  • The begin_ argument of slice

All TensorFlow operations now supported

Every operation defined by TensorFlow is now automatically wrapped in a Julia function available in the Ops module.