NuMojo v0.5 for Mojo 24.6

26/01/2025 (v0.5)

⭐️ New

• Add support for complex arrays ComplexNDArray, ComplexSIMD, CDType (PR #165). Add creation routines for complex arrays (PR #195).
• Add TypeCoercion struct that calculates the resultant type based on two initial data types. Apply type coercion to math functions (PR #164, PR #189).
• Add OwnData type as container of data buffer for NDArray and Matrix. The property _buf is changed from UnsafePointer to OwnData. Introduced RefData struct and Bufferable trait. This step prepares for future support of array views and facilitates an easy transition once parameterized traits are integrated into Mojo (PR #175, PR #170, PR #178).
• Add NDIter type as a iterator over the array items according to a certain memory layout. Use NDArray.nditer() to construct the iterator (PR #188).
• Add an additional data type mapping for NDArray.to_numpy, i.e., DType.index is mapped to numpy.intp (PR #157).
• Add method NDArray.resize that reshapes the array in-place (PR #158).
• Add a new property flags for NDArray and Matrix to store memory layout of the array, e.g., c-continuous, f-continuous, own data. The memory layout is determined by both strides and shape of the array (PR #170, PR #178).
• Add functions to_tensor and from_tensor (also an overload of array) that convert between NDArray and MAX Tensor (Issue #183, PR #184).
• Add several functions for linear algebra:
  • Add det function calculating the determinant of array (PR #171).
  • Add Householder-based QR decomposition for matrix (PR #172).

🦋 Changed

• Update several methods of NDArray type:
  • Make load and store methods safer by imposing boundary checks. Rename get as load and rename set as store (PR #166).
  • Update item method to allow negative index and conduct boundary checks on that (PR #167).
  • Update item function so that it allows both setting and getting values, e.g., a.item(1,2) and a.item(1,2) = 10 (PR #176).
  • Refine the __init__ overloads (PR #170).
• Allow getting items from NDArray with a list or array of indices, while the array can be multi-dimensional. Allow getting items from NDArray with a list or array of boolean values, while the array can be both one-dimensional or multi-dimensional (PR #180, PR #182).
• Rename Idx struct as Item.You can get scalar from array using a[item(1,2)] (PR #176).
• Update the constructor of the Item type to allow passing int-like values in (PR #185).
• Integrate mat sub-package into core and routines modules, so that users can use a uniformed way to call functions for both Matrix and NDArray types (PR #177).
• Update the following functions to allow operation by any axis:
  • argsort (PR #157).
  • cumsum and cumprod (PR #160).
  • flip (PR #163).
• Update matmul to enable multiplication between two arrays of any dimensions (PR #159).
• Refine the function reshape so that it is working on any dimensions and is working on both row-major and col-major. This also allows us to change the order with the code A.reshape(A.shape, "F")). Also refine functions flatten, ravel (PR #158).
• Remove size property from NDArrayShape and add size_of_array method to get the size of the corresponding array (PR #181).
• Add PrintOption type to customize string representation of NDArray, NDArrayShape, NDArrayStrides, and Item, e.g., str(), repr(), print(). Allow customized separators, paddings, number of items to display, width of formatting, etc, for NDArray._array_to_string method. Auto-adjust width of formatted values and auto-determine wether scientific notations are needed (PR #185, PR #186, PR #190, PR #191, PR #192).
• Rename the auxiliary function _get_index as _get_offset (PR #173).
• Rename the underlying buffer of Idx type to _buf (PR #173).
• Return a view instead of copy for iterator of NDArray (PR #174).

❌ Removed

• Remove order property and datatype property from NDArray. You can use a.flags and a.dtype instead (PR #170).
• Remove the property coefficient of NDArray type (PR #166).
• Remove offset property from the NDArrayStrides type (PR #173).

🛠️ Fixed

• Removes mut keyword before self from NDArray.__add__.
• The methods get and set of NDArray does not check against big negative index values and this leads to overflows. This is fixed (PR #162).
• Fix __setitem__ method so that it can read in variant of int and slices, e.g., a[0, Slice(2, 4)] = a[3, Slice(0, 2)] (PR #176).

📚 Documentatory and testing

• Add readthedocs pages to the repo under /docs/readthedocs (PR #194).
• Add magic run t, magic run f, and magic run p for the magic CLI. They first clear the terminal before running test,final, and package.
• Allow partial testing via command, e.g., magic run test_creation, to avoid overheat.
• Convert the readme file into pure markdown syntax (PR #187).
  NuMojo v0.4 for Mojo 24.6

---

What's Changed

• [update][routines] Update argsort function to allow sorting by any axis by @forFudan in #157
• [update][manipulation routines] Refine functions reshape, resize, flatten, ravel by @forFudan in #158
• [update][linalg] Update matmul to enable any dimensional arrays (n>0) by @forFudan in #159
• [fix][math routines] Allow cumsum and cumprod by any axis by @forFudan in #160
• [fix][ndarray] Fix overflow due to negative index argument for get and set of NDArray by @forFudan in #162
• [update][manipulation] Allow flip by any axis by @forFudan in #163
• [update][core] Update several methods of NDArray type by @forFudan in #166
• [new][core] Type Coercions in Arithmetic dunder methods by @shivasankarka in #164
• [update][core] Update item method by @forFudan in #167
• [core][ndarray] flags replaces order property. Remove ...

NuMojo v0.4 for Mojo 24.6

This is a list of RELEASED changes for the NuMojo Package.

22/12/2024 (v0.4)

⭐️ New

• Implement a static-sized 2D Array type (Matrix type) in numojo.mat sub-package. Matrix is a special case of NDArray but has been optimized since the number of dimensions is known at the compile time. It is useful when users only want to work with 2-dimensional arrays. The indexing and slicing is also more consistent with numpy (PR #138, PR #141, PR #141). It provides:
  • Matrix type (2D array) with basic dunder methods and core methods.
  • Function to construct Matrix from other data objects, e.g., List, NDArray, String, and numpy array.
  • Arithmetic functions for item-wise calculation and broadcasting.
  • Linear algebra: matrix mutiplication, decomposition, inverse of matrix, solve of linear system, Ordinary Least Square, etc.
  • Auxilliary types, e.g., _MatrixIter.
• Implement more the array creation routines from numpy and expand the NuMojo array creation functionality (PR #137).
• Add convolve2d and in the science.signal module (PR #135).
• Add more detailed error messages for NDArray type (PR #140).

🦋 Changed

• Adapt the code to the latest update of Mojo to V24.6 (PR #148).
  • Slice.step is now returning Optional[int]. Thus, it is fixed by using Slice.step.else_to(1).
  • Formattable is now renamed to Writable (same applies to write_to and string.write).
  • width is now inferred from the SIMD's width. So this parameter must be removed when we call UnSafePointer's load and store methods. Due to this, the function fill_pointer, which fill in a width of array with a scale, no longer work. It is replaced by copying with loop.
  • Lifetime is renamed as Origin (same applies to the function __origin_of).
  • inout is renamed as mut.
• Rename the data buffer from data to _buf (PR #136, PR #137)
• To make matmul flexible for different shapes of input arrays (PR #137).
• Change the way to get the shape of the array: array.shape returns the shape of array as NDArrayShape (PR #137).
• The array creation functions are unified in such a way (PR #139).
  • NDAarray.__init__() reads in shape information and initializes an empty ndarray.
  • All other creation routines are implemented by the functions in the array_creation_routine module. For example, to create an array with filled value, the function full should be used. To create an array from a list, the function array should be used.
• Re-organize the functions and modules by topic, so that it is more consistent with numpy (Issue 144, PR #146).
• Rename some attributes of NDArray and make size an attribute instead of a method (PR #145).
• Completely remove import * in init files to fix namespace leak (PR #151).
• Update function sum (PR #149).
• transpose now allows arbitrary dimensions and permutations of axes (PR #152).
• Change buffer type of NDArrayShape and NDArrayStrides to Int (PR #153).
• sort now allows sorting by any axis for high dimensional arrays (PR #154).

❌ Removed

• Temporarily removed negative indexing support in slices since it causes error. Will add just feature in later updates (PR #133).
• Remove inout before self for NDArray.__getitem__ (PR #137).

🛠️ Fixed

• Fixed and rewrote the adjust_slice function that was causing errors (PR #133).
• There is an error in fromstring that the negative signs are not read. It is fixed. Now a valid numeric should start with a digit, a dot, or a hyhen (PR #134).

📚 Documentary and testing

• Added test for slicing in getters. Complete tests for setters will be added later. This is due to python interop limitation (PR #133).
• Move documents from root docs. Delete unused files (PR #147).
• Update the readme.md and features.md to reflect current progress (PR #150).

NuMojo v0.3.1 for Mojo 24.5

19/10/2024 (v0.3.1)

🛠️ Fixed

• Fix the matmul function which returns wrong results for small matrices (PR #129).
• Correct the getter and setter functions, specifically the _adjust_slice (PR #130).

NuMojo v0.3 for Mojo 24.5

NuMojo released changelog

This is a list of RELEASED changes for the NuMojo Package.

14/10/2024 (v0.3)

⭐️ New

• Add support for magic system and MAX 24.5 (PR #91 #109 by @shivasankarka).
• Add some basic functions, e.g., diagflat, tri, trace, T (PR #91 by @shivasankarka).
• Add a constructor which reads arrays from numpy arrays (PR #91 by @shivasankarka).
• Add functions solve and inv for solving linear algebra AX = Y for X, finding inverse of a matrix, and conducting LU decomposition (PR #101 #104 #105 by @forFudan).
• Add itemset method to fill a scalar into an NDArray (PR #102 by @forFudan).
• Add Idx struct to represent the index of an NDArray (PR #118 by @shivasankarka).
• Add NDArray initialization from numpy arrays (PR #118 by @shivasankarka).
• Created a new io module with some basic functions, e.g., format_float_scientific, loadtxt, savetxt (PR #118 by @shivasankarka).

🦋 Changed

• Make some methods, e.g., sort, flatten, inplace (Issue #87 by @mmenendezg, PR #91 by @shivasankarka).
• Modify initialization of NDArray (PR #97 by @MadAlex1997)
• Added Formattable trait and fixed the print function (PR #108 by @shivasankarka)
• Refine the NDArray initializers and array creation routines (Discussion #90, Issue #110).
  • Remove random argument from NDArray constructors. Make random initializer a standalone functions (Issue #96 by @MadAlex1997, PR #97 by @MadAlex1997, PR #98 by @shivasankarka).
  • Remove initializers from String. Make fromstring a standalone function (#113 by @forFudan).
  • Add several array overloads for initializing NDArray (PR #118 by @shivasankarka).
• Modify the behavior of __get__ and __set__. Passing in a sequence of Int or Slice returns an NDArray. Passing in an Idx returns a scalar. Allow users to set a multiple items in one NDArray with another NDArray, using __set__ (PR #118 by @shivasankarka, Discussion #70).

❌ Removed

• Removed all instances of getters and setters with List[Int], VariadicList[Int] (PR #118 by @shivasankarka).

🛠️ Fixed

• Fix the issues in parallelization (due to Max 24.5) for some linear algebra functions, e.g, matmul, solve, inv, etc (PR #115 #117 by @forFudan).

📚 Documentary and testing

• Add workflows with unit tests and linting (PR #95 by @sandstromviktor).
• Add multi-lingual support (Chinese, Japanese) for the readme file (PR #99 #120 by @forFudan, PR #100 by @shivasankarka).
• Update the test flow file to accommodate MAX 24.5 and the magic system (PR #116 by @forFudan).

17/08/2024 (V0.2)

⭐️ New

Array operations:

• Introduced diagflat() method for creating diagonal arrays
• Implemented most basic array creation routines
• Implemented initializer from string numojo.NDArray("[[1,2,3],[4,5,6]]")
• Enhanced NDArray constructor methods
• Added boolean masking for NDArrays
• Introduced new mathematical methods: floordiv, mod, and more
• __getitem__ and __setitem__ now fully compatible with NumPy behavior

Others:

• Introduced Rust-like data type aliases (e.g., DType.float64 → f64)
• Implemented function overloading for core mathematical operations (add, sub, etc.)
• Added a new differentiation module

🦋 Changed

• Improved slicing functionality to match NumPy behavior

❌ Removed

• Removed in_dtype and out_dtype parameters for simplified type handling

📚 Documentatory and testing

Documentation updates:

• Expanded and updated README
• Improved docstrings for functions
• Added style guide and examples

Testing updates:

• Introduced new test files that are compatible with mojo test.

Contributors

• @shivasankarka
• @forFudan
• @MadAlex1997
• @mmenendezg
• @sandstromviktor

NuMojo V0.2

NuMojo V0.2 Release Notes

Major Changes

• Removed in_dtype and out_dtype parameters for simplified type handling
• Introduced Rust-like data type aliases (e.g., DType.float64 → f64)
• Implemented function overloading for core mathematical operations (add, sub, etc.)
• Added a new differentiation module

New Features

Array Operations

• Introduced diagflat() method for creating diagonal arrays
• Implemented most basic array creation routines
• Implemented initializer from string numojo.NDArray("[[1,2,3],[4,5,6]]")
• Enhanced NDArray constructor methods
• Improved slicing functionality to match NumPy behavior
• Added boolean masking for NDArrays
• Introduced new mathematical methods: floordiv, mod, and more
• __getitem__ and __setitem__ now fully compatible with NumPy behavior

Documentation Updates

• Expanded and updated README
• Improved docstrings for functions
• Added style guide and examples

Testing

• Introduced new test files that are compatible with mojo test.

Compatibility

• Works with Mojo 24.4
• No external dependencies required

Notes

• Binary distributions are not included due to current limitations in Mojo's packaging system

Contributors

• @MadAlex1997
• @shivasankarka
• @forFudan

New Contributors

• @mmenendezg made their first contribution in #88

Full Changelog: v0.1...v0.2

NuMojo v0.1 for Mojo 24.4

The first tracked version of NuMojo, is compatible with Mojo 24.4. Binaries are not included (Mojo doesn't have a packaging system yet), but no dependencies other than Mojo compiler version 24.4.