Releases: Gleethos/neureka
Releases · Gleethos/neureka
Neureka 1.0.0
What's Changed
- Important: Final API changes for the first stable version:
Tsr
is now calledTensor
New features
- The reshape operation is now differential
Also
- Important bug fixes
Full Changelog: v0.21.0...v1.0.0
Neureka 0.21.0
What's Changed
- 12 add transpose and permute API and improve related implementation details by @Gleethos in #13
- Introduce dot product support as its own operation by @Gleethos in #15
- 16 introduce reshaping as a differentiable operation by @Gleethos in #17
- Issue 19 tensor instantiation API refinement for next release 0 21 0 by @Gleethos in #20
Also
- Intensive testing and bug fixing of the autograd system.
- Stable and error free convergence of large NN graphs.
- Tensor ND Layout improvements (column major is now only a property concerning the last 2 dimensions)
Full Changelog: v0.20.1...v0.21.0
This is the last release before version 1.0.0 which will have a stable long term API.
v0.20.1 - Bug Fixes
What's Changed
Full Changelog: v0.20.0...v0.20.1
v 0.20.0 - Bug Fixes and the Introduction of Shape
What's Changed
- Added styling to side nav, refactored css by @mischulz14 in #5
- 6 introduce a shape type to improve the neureka api by @Gleethos in #7
- Issue 8 release v0.20.0 preparations by @Gleethos in #9
New Contributors
- @mischulz14 made their first contribution in #5 (Thank you!)
Full Changelog: v0.19.0...v0.20.0
v0.19.0
Full Changelog: v0.18.0...v0.19.0
v0.18.0 - Min, Max, Sum and General API Refinement
Full Changelog: v0.17.0...v0.18.0
v0.17.0 - Tensor Concatenation
- New
concat
function for mergin tensors alongside a specified axis (+ autograd support). - Introduce full
png
support forFileDevice
. - Improve back-end code quality.
- Important bug fixes!
- Improved documentation.
Full Changelog: v0.16.0...v0.17.0
v0.16.0 - New activation functions
- Countless new activation function types (
SeLU
,SiLU
,GeLU
,Softsign
...). - Improved performance for virtual tensors and activations.
- Improved mutability safety of tensors by making its standard data accessing methods return clones.
- Simplify
Device
interface and introduce optionalAccess
API for accessing tensors stored on devices. - Improve tensor type interoperability for both CPU and GPU (only for storing tensors).
- Countless new activation function types.
- Improve back-end code quality.
- Important bug fixes!
Full Changelog: v0.15.0...v0.16.0
v0.15.0 - Small performance improvements
- Improved performance for conversion between column and row major data layout.
- Make performance critical classes final (
NDIterator
types) - Improve back-end code quality by making
ExecutionCall
s hard to mutate. - Important bug fixes!
Full Changelog: v0.14.0...v0.15.0
v0.14.0 - Both Column & Row Major Support and fast OpenCL Matrix Multiplication
- Support for both column and row major data layout.
- Transposing a tensor converts between the two layouts.
- Possibility to convert between the two type (mutable method in the unsafe API).
- New fast OpenCL matrix multiplication for column j´major matrices.
- More specialized
NDConfiguration
andNDIterator
implementations for better general performance. - Important bug fixes!