Release 0.1.5

Version 0.1.5

New Features

• Performance: Enjoy enhanced performance on Apple Silicon devices with the new Metal Performance Shaders (MPS) backend support, enabling hardware acceleration for various tensor operations. [#9]
• Buffer Management: Experience improved data handling with the new BufferManager, designed to streamline memory management and enhance performance. [993a075]
• Serialization: Benefit from the new serialization and deserialization support, allowing for easier data handling and storage across different formats. [b726e79]
• Examples: Explore new examples demonstrating serialization features and reorganized device examples for better clarity and usability. [0a26aa8]

Improvements

• Core Functionality: The buffer interface is redesigned to include offset support, providing more flexibility in data management and safer memory access. [c9b8180]
• Tensor Operations: The to_flatten_vec method now supports additional data types (u16, u64, i16), enhancing its versatility for various applications. [cffeaf1]

Others

• Documentation: The feature support tables are updated to replace CPU, CUDA, and MPS columns with a new Notes column for clearer information presentation. [6dc7c23]

Release 0.1.3

Version 0.1.3

New Features

Neural Network

• Added Embedding layer with support for padding and gradient scaling [71915be]

Tensor Operations

• Added tensor reduction operations including norm, variance, and standard deviation [0076a86]
• Implemented stride and offset support for padding operations to handle non-contiguous tensors [d52daa9]

Indexing Operations

• Added index_select and index_add_ functions to enhance tensor indexing capabilities [aaebba8]
• Added bincount operation for efficient counting of occurrences in tensors [222d9e4]

Data Types

• Added support for u16, u64, and i16 data types for more flexibility in tensor operations [7bb020c]

Improvements

• Enhanced performance of the contiguous operation for CUDA tensors [#6, 68b342f]
• Unified type promotion across all tensor operations, enhancing compatibility and reducing errors [316dd26]

Fixes

• Fixed gather operation to ensure the output uses the source tensor's data type [91128cd]
• Fixed alignment issues in CPU kernels to match the CUDA implementation [040e2f1]
• Fixed overflow issues in read/write_scalar operations for u8 and i8 tensor types [#5, fd901bc]

Release 0.1.2

Version 0.1.2

New Features

Mathematical Operations

• Added trigonometric functions (sine, cosine, tangent) [6ae44d3]
• Added logarithmic and exponential functions (ln, log10, log2, exp, exp2, exp10) [bcd79bb]

Neural Network

• Implemented LayerNorm with keepdims option for mean function [d7face4]
• Added Dropout layer and LayerState for independent training mode tracking [4981cac]
• Added CrossEntropyLoss implementation [31aa816]

Tensor Operations

• Added arange and range functions for evenly spaced tensors [8b9ba03]
• Added gather and scatter operations for indexed tensor manipulation [076f2a0]
• Added min and max operations [89fda3b]
• Added recip operation for calculating reciprocals [93dd124]
• Added softmax function [f39cbdf]
• Added broadcast_like function [b2f4d4e]
• Added log function with ln alias [b4259f0]
• Enjoy boolean support for scalar comparison operations, enhancing flexibility in your computations. [316dd26]

Improvements

• Optimized im2col and col2im CUDA kernels [a01cca5]
• Added keep_dims parameter to the sum function [d2a717f]
• Improved naming of loss function classes [5344285]
• Fixed boolean promotion issues and scalar type conversion in arange function [c31db84]

Fixes

• Fixed compilation errors in softmax CUDA kernel [14aea5e]
• Fixed namespace from nn::functional to nn::alias [5d2fbbc]
• Corrected minor typos in the codebase [d290383, aa7041f, 65affd1]

Release 0.1.1

Version 0.1.1

Added Features

Tensor Operations

• Added read_scalar and write_scalar for both CPU and CUDA devices [bd3d4b5]
• Implemented tensor offset support for memory-efficient views [82a2e30]
• Introduced new padding operations [5531e67]
• Added operator overriding capabilities [f1c957b]
• Added metadata support in tensors [030b9f8]

Neural Network Components

• Added LeakyReLU, GELU, and ELU activation functions [06ff9d2]

Enhancements

• Improved performance by replacing RwLock with Arc for buffer management [af8b5d5]
• Updated tensor indexing methods to support offsets [89a382b]

Fixes

• Fixed borrowing issues in inplace operations [af8b5d5]

Release 0.1.0

Version 0.1.0

Added Features

DType Support

• Floating-Point Types: bf16, f16, f32, f64
• Integer Types: u8, u32, i8, i32, i64

Neural Network Components

• Modules: Linear, Bilinear, Conv2d layers (CPU and CUDA)
• Activations: ReLU, Sigmoid, Tanh
• Loss Functions: MSE, MAE, Huber
• Optimizers: SGD, Adam

Tensor Operations

• Basic Operations: element-wise operations with gradient tracking and broadcasting
• Shape Operations: transpose, reshape, and broadcasting utilities
• Reduction Operations: sum, mean
• Scalar Operations: addition, subtraction, multiplication, division
• Comparison Operations: scalar-based comparisons

Enhancements

• Gradient Tracking via automatic differentiation
• Full tensor broadcasting support