v0.9.2: Optimized CUDA version

Like previous release, but CUDA now fast. All MNIST modules implemented.

v0.9.1: Naive CUDA implementation of existing features

First naive CUDA implementation of everything that already was there. Some extra Python functions for better GPU and CPU utilization, such as slicing and permutations. Also added a CUDA MNIST example script and unit tests covering more code and new features.

v0.9.0

First version with training loops.

Features:

• Tensor object
• FeedForward Layer
• Sequential network
• Activation functions: ReLU, LeakyReLU, Sigmoid, Softmax (numerically stable)
• Loss functions: BCE, CrossEntropy, RMSE, BceWithSigmoid, CrossEntropyWithSoftmax
• Autograd solutions for all operations
• Optimizers: SGD, RMSProp
• C++ train loop
• numpy to Tensor and Tensor to numpy conversion
• Compilation into structured Python library

Unit tests:

• Tests for all features in C++ and Python

Examples:

• MNIST dataset learning and verification

v0.1.0

Computational graph on Tensors completed