C XLA release 0.0.10

C XLA release 0.0.10

C XLA release 0.0.9

C XLA release 0.0.9

C XLA release 0.0.8

C XLA release 0.0.8

C XLA release 0.0.7

C XLA release 0.0.7

C XLA release 0.0.1

C XLA release 0.0.1

Release 0.0.6

This is a small release which fixes an incompatibility with a dependency

Tested on Ubuntu 64-bit with Idris2 nightly build and Nx XLA CPU binaries v0.3.0

API changes

• implement non-finite Bounded for scalar tensors #332
• add argmin, argmax for finding the index of minina, maxima in a vector #329
• remove Hashable implementation for Pair #335

Bug fixes

• fix min and max for NaN inputs #334

Build changes

• enable manual trigger for documentation build #324
• build locally with latest version of hashable #335

Release 0.0.5

Tested on Ubuntu 64-bit with Idris2 nightly build and Nx XLA CPU binaries v0.3.0

API changes

• support execution on GPU #259
• add pure Idris tensor Literal for improved shape inference #183 #260 #261 #263 #304 #305 #306 #308
• add support for unsigned 64-bit XLA type U64 #282
• remove Each and # from element-wise names #232
• hide IO in eval and rename to toArray #241
• hide IO in toString; implement Show for Tensor #242
• add/implement tensor functions
  • support for slicing with dynamic (tensor) indices #303
  • support for slicing multiple axes at once #269
  • unary functions: (inverse-)(hyperbolic-)trigonometric; recip; square #268
  • select for choosing values element-wise from two tensors #233
  • cond for evaluating one of a pair of functions #238
  • transpose for transposing arbitrary axes #299 #307 #316
  • reverse for reversing elements along specified tensor axes #286 #307 #316
  • sort for sorting elements along tensor axes #283
  • (|\) and (\|) for solving linear equations with lower- and upper- triangular matrices #228
  • triangle for fetching the triangular parts of a matrix #229
  • stateless uniform sampling #281 #284 #287 #297
  • stateless univariate normal sampling #292 #297
• extend reduce to operate on multiple axes at once #298 #307 #316
• move runtime shapes from tensor APIs to data constructor #231
• use Int32 rather than Int for S32 #311
• erase proofs where possible #276
• implement Gaussian univariate CDF #239
• remove Error interface #255
• replace Morphism with Reader in Bayesian optimization tooling #319

Bug fixes

• fix reading/writing unsigned int from xla::Literal #279
• ensure we can read/write numeric values at XLA bounds #291

Documentation

• add tutorial on basic dependent types for tensors #301
• move installation instructions to INSTALL.md #248
• provide copy-paste installation instructions #321

Build changes

• adopt Idris2 package manager pack, thus using the latest Idris2 build #321

Release 0.0.4

Tested on Ubuntu 64-bit with Idris2 v0.5.1 and Nx XLA CPU binaries v0.3.0

This release implements (or, in some cases, removes) all functions the Tensor API.

API changes

• introduce Idris counterparts to XLA data types (S32, U32, F64 etc.) for use in Tensor #168. Add custom numerical interfaces to mirror Prelude numerical constraints #174
• support boolean primitive type PRED #135
• use List Nat for tensor shapes #184 #185
• make dtype explicit in Array #156
• make broadcasting explicit #138
• delete the Variable mutable tensor type #186
• delete trace_product #189, cast_dtype, inverse #198, det #209
• make ProbabilisticModel an interface #121
• implement new functions and operators. Most of these APIs have changed during implementation
  • identity for constructing identity matrices #199
  • index for basic tensor indexing #191
  • reshape for reshaping tensors #192
  • expand for adding a dimension to a Tensor #193
  • squeeze for removing dimensions of length one #194
  • slice for slicing along any one axis #195
  • split for splitting a tensor in two along a given axis #196
  • concat for concatenating Tensors #197
  • .T for transposing matrices #208
  • diag for fetching the diagonal of a matrix #211
  • element-wise unary functions: absolute value absEach #144 #163; numeric negation negate #157; natural exponential expEach and logarithm logEach #187; add floor floorEach, ceiling ceilEach, logistic logisticEach, sine sinEach, cosine cosEach, hyperbolic tangent tanhEach, and square root sqrtEach #188; boolean negation notEach #160
  • element-wise binary functions: equality ==# and inequality /=# #136; subtraction - #145; multiplication *# #134; division /# #158; comparison ># <# >=# <=# #143; logical and &&# and or ||# #160; exponential wrt any base ^# #206; minimum minEach and maximum maxEach #179
  • division by a constant / #158
  • multiplication by a constant * #142
  • functor-like map #165, applicative-like map2 #171 #178 and foldable-like reduce #176 along with monoid instances for Tensor #180
  • @@ for matrix and vector multiplication #201
  • trace for summing along a matrix diagonal #202
  • cholesky for lower triangular Cholesky decomposition #203
  • (\\) for solving linear equations wrt triangular matrices #207

Bug fixes

• pdf for Gaussian now uses covariance correctly in exponent #209

Documentation

• expand introduction to Bayesian optimisation in tutorial #172

Build changes

• compile XLA C wrapper into single shared library #215
• update XLA binary to version 0.0.3 (for XLA/TensorFlow version 2.8.0) #216

Release 0.0.3

Tested on Ubuntu 64-bit with Idris2 v0.5.1 and Nx XLA binaries v0.2.0

API enhancements

• implement Gaussian PDF (#42)
• refactor distribution functionality (#86)
• replace Connection with Morphism (#89)
• improve efficiency of GP training with Cholesky factorization (#93)
• allow any data representation in Bayesian optimization (#100)
• erase as many values as possible (#106)
• simplify namespace names (#107)
• add Matern 5/2 kernel (#109)
• simplify Distribution with functional dependencies (#112)
• implement conjugate GP regression (#113)
• introduce Dataset data type (#115)
• add Bayesian optimization loop (#114)
• add XLA backend (#82 #90 #123 #124 #126 #127 #128 #129 #131) with executable implementations for addition (#128) and broadcasting (#130)

Other

• add installation instructions (#133)
• update Idris 2 to v0.5.1 (#78)

Release 0.0.2

API enhancements

• add RBF kernel (#75)
• implement zero mean function (#74)
• constrain element type of tensors to have corresponding types in backend (#47)
• replace Maybe usage with Either and error types (#45)
• add support for squeezing Tensors (#40)
• add numerous tensor operators (#39)
• add mutable (linear) tensor Variable type and in-place ops (#36)
• hide Tensor constructor, enabling arbitrary Tensor representation (#34)
• implement acquisition functions: (negative) lower confidence bound, probability of feasibility, expected improvement (#24)

Bug fixes

• improve type inference for tensor contraction operator @@ (#72)

Other

• convert Bayesian optimization integration test to tutorial (#43) (#51)
• add automatic API reference build and deployment on release (#77)
• update Idris 2 to v0.4.0 (#78)