Nexora is a revolutionary library designed specifically for Scratch that transforms neural network development into an intuitive and accessible builder. Now you can create your models without leaving your favorite platform! What does Nexora offer? 1. Architecture building layers: (Dense, Dropout, AdaptNorm, Activation, Embedding, Conv1D, Flatten) 2. Powerful optimizers: (SGD with Momentum, Adam with bias correction) 3. Wide choice of activation functions: (ReLU, Tanh, Sigmoid, SiLU, GELU, ELU, Softmax, Linear) 4. A comprehensive training stack thanks to (Fit, WipeTrainData, and InsertExample) with support for shuffling examples during training and batch size support! 5. Intuitive and simple sequence support with the new Embedding layer! And work efficiently with 2D data thanks to Conv1D. 6. Work with multiple models simultaneously: name them as you create them and recall them intuitively—Nexora hides all the complexity under the hood. Why choose Nexora? - Nexora is the first Scratch library to provide such capabilities for creating neural networks. - With an intuitive interface and a powerful core, creating and training models is accessible even for beginners. - Flexibility: Experiment with different combinations of layers, optimizers, and activation functions to find the ideal architecture for your problem. - Open: Nexora uses an open license (CC-BY-SA 2.0), so you can safely use this framework in your projects or improve it for yourself!
last update: {17.09.2026} Fixed: - Chunk loading edge cases where parameters were not fully loaded. modified: {17.08.2026} Debug: -Debugged the backpropagation logic in the Dense and Activation layers. Layers now store values before applying the activation function! Added: -Loss types: (CrossEntropyLoss, MSELoss) -Full Conv1D with two dimensions and Flatten with clipping -Weight compression via multiplication by rounding and conversion to an integer -Asynchronous loading and the ability to load a fixed chunk in a single cycle instead of the entire model. Update: -Interlayer delay mechanism: A layer now passes its name and its position in the model to the label, for correct delays. -BatchNorm has been renamed to AdaptNorm, which uses previously accumulated batches for training, rather than the current batch. -Documentation elements on model operation. - The number of blocks was reduced by creating a markup language for documentation and migrating the documentation to the new language. modified: {26.11.2025} Updated: -Load and save functions accept two parameters (model and load name) -Nexora can now store more than one model in RAM and switch between them without re-initializing the model or loading it. modified: {26.10.2025} Fixed: - An issue where the Dense and BatchNorm layers did not fully initialize their parameters, leaving most of the parameters empty. - A backpropagation hang when using the Embedding layer. modified: {18.10.2025} Updated: -The documentation now includes descriptions and examples of working with new layers (BatchNorm, Activation)! -The project description has been updated to reflect current information! Added: -Added a new Embedding layer for working with sequences! modified: {07.10.2025} Added: -Added a new layer BatchNorm with moving averages (mean, var) and parameters (gamma, beta)! modified: {28.09.2025} Updated: -Splitting a monolithic kernel into a modular kernel! Fixed: -Fixed errors with the activation layer. -Fixed errors with transferring network settings during model creation. modified: {15.08.2025} Fixed: -Fixed incorrect copying of examples for fit! modified: {10.08.2025} Added: -Added support and processing of batches of different sizes! modified: {09.08.2025} Added: -Added training with fit! Added functions: Fit, WipeTrainData, InsertExample! -Additional configurable delay between layers. -New Activation layer for transforming outputs using activation functions! -New Linear activation function. Important: -The batchsize function does not work yet! modified: {03.08.2025} Fixed: -The delay algorithm has been corrected for correct data transmission. modified: {02.08.2025} Fixed: -Neural network prediction algorithm. -Propagation of error to previous layers. Improved: -Delays for correct list retrieval. -Initialization of weights is close to He initialization.
