liquidnn 1.1.0

A PyTorch based module for Liquid Networks

🧠 liquidnn

A PyTorch implementation of Liquid Neural Networks (LNNs) and their hybrids (LTC, CNN, RNN, LSTM variants). This library brings continuous-time dynamics into deep learning architectures using learnable time constants.

✨ Features

🔹 LiquidNeuralNetwork – stacked liquid neurons

🔹 LTCLayer – Liquid Time-Constant layer

🔹 LiquidCNN – CNN + Liquid update

🔹 LiquidRNN – RNN + Liquid update

🔹 LiquidLSTM – LSTM + Liquid update (with bidirectional support)

🔹 Per-neuron learnable τ (time constants) with stability clamping

📦 Installation From PyPI (after publishing): pip install liquidnn

🚀 Quick Start (all models in one script)

Copy-paste this script to test every model in liquidnn:

import torch from liquidnn import ( LiquidNeuralNetwork, LTCLayer, LiquidCNN, LiquidRNN, LiquidLSTM )

1. LiquidNeuralNetwork

x = torch.randn(4, 10, 8) # batch=4, seq_len=10, input_size=8 model = LiquidNeuralNetwork(input_size=8, hidden_size=16, num_layers=2) out = model(x) print("LiquidNeuralNetwork output:", out.shape) # [4, 16]

2. LTCLayer

x = torch.randn(5, 20, 10) # batch=5, seq_len=20, input_size=10 ltc = LTCLayer(input_size=10, hidden_size=32, num_layers=2) out = ltc(x) print("LTCLayer output:", out.shape) # [5, 32]

3. LiquidCNN

x = torch.randn(8, 1, 28, 28) # MNIST-like input model = LiquidCNN(input_channels=1, hidden_size=64, num_layers_liq=2, num_layers_conv=2) out = model(x) print("LiquidCNN output:", out.shape) # [8, 64]

4. LiquidRNN

x = torch.randn(15, 4, 12) # seq_len=15, batch=4, input_size=12 model = LiquidRNN(input_size=12, hidden_size=32, num_layers_liq=2, num_layers_rnn=1) out = model(x) print("LiquidRNN output:", out.shape) # [4, 32]

5. LiquidLSTM

x = torch.randn(3, 12, 10) # batch=3, seq_len=12, input_size=10 model = LiquidLSTM(input_size=10, hidden_size=32, num_layers_liq=2, num_layers_lstm=1, batch_first=True, bidirectional=False) out = model(x) print("LiquidLSTM output:", out.shape) # [3, 32]

📚 Architectures

LiquidNeuralNetwork: Stacked liquid neurons

LTCLayer: Continuous-time RNN update rule

LiquidCNN: Convolutional backbone + LTC dynamics

LiquidRNN: RNN + Liquid refinement

LiquidLSTM: LSTM + Liquid refinement

Install dependencies:

pip install torch

🌟 Contribute

PRs and issues are welcome! If you try new liquidized architectures, feel free to share 🚀