AndroidDL.js v1.2.0 – Introducing RNN, GRU & LSTM Support

AndroidDL.js v1.2.0

This release brings a major upgrade to AndroidDL.js with support for recurrent neural networks and new tensor operations.

🚀 What's New

Recurrent Neural Networks

• Added RNNCell and RNN.
• Added GRUCell and GRU.
• Added LSTMCell and LSTM.
• Added sequence processing and backpropagation through time (BPTT).
• Added support for initial hidden and cell states.

Tensor Operations

• Added concatenate and stack.
• Added squeeze and unsqueeze.
• Improved tensor slicing with multidimensional and negative-step support.
• Improved handling of non-contiguous tensors and strided views.

Autograd & Testing

• Extended gradient propagation through recurrent layers.
• Added numerical gradient checks.
• Expanded test coverage across tensor operations and recurrent networks.

📦 What's Included

• Core Tensor and autograd engine
• Neural network modules
• SGD optimizer
• MSE and Cross Entropy losses
• Categorical distribution
• RNN, GRU and LSTM architectures

🎯 What's Next?

Continuing to build a lightweight, PyTorch-inspired deep learning library in JavaScript — one operation at a time.

Build your own PyTorch. Understand every operation.

Thanks for following the AndroidDL.js journey!

Repository: https://github.com/AkashPrajapathi/androiddl-js

Add GRUCell support

v1.1.1

Added GRUCell support

Add GRUCell support

AndroidDL v1.1.0

🚀 New Features

• Added GRUCell for recurrent neural network architectures.
• Supports configurable input size and hidden size.
• Supports optional biases, enabled by default.
• Implements update, reset, and candidate hidden-state gates.
• Integrates with AndroidDL's existing Module, Linear, and autograd system.

🧪 Testing

• Verified forward-pass output shapes and finite values.
• Verified gradients across all GRU parameter groups.
• Tested numerical gradients using finite differences.
• Trained on a small sequence-to-target task and observed a significant reduction in loss.

Example

import { GRUCell } from "androiddl";

const gru = new GRUCell(3, 4);

const nextHidden = gru.call(input, hidden);

input has shape [batchSize, inputSize], and hidden has shape [batchSize, hiddenSize].

Notes

This release introduces a GRU building block for sequence models and reinforcement-learning experiments.

v1.0.1 tanh backward bug fixed

Bug Fix: Tanh Backward Pass

• Fixed a bug in TanhBackward where onesLike() received the output tensor's shape instead of the tensor itself.
• Corrected the call to onesLike(this.output) to ensure proper tensor initialization during backpropagation.
• Resolves the TypeError: e.shape is not iterable error when computing gradients through Tanh.

This fix restores the expected gradient calculation for the tanh activation function.

v1.0.0 - Initial Release

🚀 Version 1.0.0 - Initial Release

We are excited to announce the initial release of androiddl-js! 🎉

androiddl-js is a lightweight JavaScript library built using ES modules and bundled for seamless usage across both npm/Node.js and direct browser environments via <script> tags.

---

✨ Features

• Dual Format Support: Provides ES Module (.js) and UMD/CommonJS (.umd.cjs) bundles.
• Core Neural Network & Math Engine: Includes modules for core, nn, optim, loss, and distribution.
• Zero Config Setup: Simple API designed for high performance in JS runtime environments.

---

📦 Installation & Usage

1. Via npm / Modern Bundlers

Add the GitHub Packages registry to your .npmrc file:

@AkashPrajapathi:registry=[https://npm.pkg.github.com](https://npm.pkg.github.com)