custom-rnn 0.1.0

A NumPy-based neural network library featuring RNN, GRU cells, CTC loss/decoding, greedy and beam search decoders — built from scratch without PyTorch.

custom-rnn

A lightweight, NumPy-only neural network library implementing recurrent architectures and CTC (Connectionist Temporal Classification) from scratch.

Features

• RNN Cell — vanilla recurrent neural network cell with forward and backward (BPTT) support
• GRU Cell — gated recurrent unit with reset, update, and new gates
• RNN Phoneme Classifier — multi-layer RNN for sequence classification
• Character Predictor — GRU-based character-level prediction with inference utilities
• CTC Loss — full CTC forward-backward algorithm with posterior probability computation
• CTC Decoding — greedy search and beam search decoders
• Neural Network Primitives — Linear layer, Sigmoid, Tanh, Softmax Cross-Entropy loss

Installation

pip install custom-rnn

Quick Start

import numpy as np
from mytorch import RNNCell, GRUCell
from CTC import CTC, CTCLoss, GreedySearchDecoder, BeamSearchDecoder

# Create an RNN cell
rnn = RNNCell(input_size=10, hidden_size=20)
x = np.random.randn(1, 10)
h = np.zeros((1, 20))
h_next = rnn(x, h)

# Create a GRU cell
gru = GRUCell(input_size=10, hidden_size=20)
x = np.random.randn(10)
h = np.zeros(20)
h_next = gru(x, h)

# CTC greedy decoding
symbols = ["a", "b", "c"]
decoder = GreedySearchDecoder(symbols)

# CTC beam search decoding
beam_decoder = BeamSearchDecoder(symbols, beam_width=3)

Requirements

• Python >= 3.8
• NumPy