v2.1.0 Released — LSTM, GRU & More

Deep Learning for

LowMind is a pure-NumPy deep learning framework built from scratch. Get a PyTorch-like API without the multi-GB installation — just NumPy and psutil.

Install LowMind View API Docs

Dependencies

Unit Tests

Examples

Version

quickstart.py

import lowmind as lm
import numpy as np

# Build a neural network
model = lm.Sequential(
    lm.Linear(784, 128),
    lm.ReLU(),
    lm.Dropout(0.3),
    lm.Linear(128, 10),
)

# Create optimizer
optimizer = lm.Adam(
    model.parameters(), lr=1e-3
)

# Train
for epoch in range(20):
    optimizer.zero_grad()
    out = model(X_batch)
    loss = lm.cross_entropy_loss(
        out, y_batch
    )
    loss.backward()
    optimizer.step()

print(f"Loss: {loss.item():.4f}")
              

🍓 Raspberry Pi Ready

✓ Only 2 Dependencies

Quick Install

pip install lowmind

✓ numpy>=1.19.0

✓ psutil>=5.8.0

✓ Python 3.7+

Full Install Guide →

🧠 About

What is LowMind?

A full-featured deep learning framework that runs anywhere — no GPU required, no 10GB installs. Built for the real world of embedded AI.

🔋

Ultra Low Footprint

Only numpy and psutil as dependencies. Runs on Raspberry Pi, Jetson Nano, and any Linux/macOS/Windows machine with minimal RAM.

🔄

PyTorch-Like API

Familiar Module, Sequential, autograd, and training loop patterns. Switch from PyTorch with minimal code changes.

⚡

Production Ready

88+ passing unit tests, full package structure, model save/load with gzip compression, system monitoring, and comprehensive error handling.

🧮

Full Autograd Engine

Reverse-mode automatic differentiation with full broadcasting, tuple-axis support, and gradient clipping built-in.

📊

Complete Toolkit

Layers, activations, losses, optimizers, schedulers, metrics, data utilities, trainer, callbacks — everything in one package.

🍓

Pi-First Design

Memory manager with LRU eviction, system monitor for CPU/RAM/temperature, memory tracing, and inference optimization for constrained devices.

Framework at a Glance

Everything you need to build, train, and deploy deep learning models

Optimizers & Schedulers

Layer & Activation Types

Loss Functions

Passing Unit Tests

📋 Full Feature List

Everything Included

Category	What's Included
Autograd	Reverse-mode automatic differentiation, full broadcasting, tuple-axis support
Layers	Linear, Conv2d, BatchNorm1d/2d, MaxPool2d, AvgPool2d, Flatten, Dropout, Embedding, LSTM, GRU
Activations	ReLU, LeakyReLU, ELU, GELU, Sigmoid, Tanh, Softmax, LogSoftmax
Loss Functions	CrossEntropy, BCE, MSE, MAE, Huber, NLL
Optimizers	SGD (+ Nesterov), Adam, AdamW, RMSprop, AdaGrad — all support parameter groups
LR Schedulers	StepLR, MultiStepLR, ExponentialLR, CosineAnnealingLR, ReduceLROnPlateau, CyclicLR, LinearWarmup
Data	Dataset, TensorDataset, DataLoader, train_test_split, Data Transforms (Compose, Normalize, Flip, Crop…)
Metrics	accuracy, top-k accuracy, precision, recall, F1, confusion matrix, R², MSE, MAE
Trainer	High-level training loop with callbacks, gradient clipping, gradient accumulation, validation
Callbacks	EarlyStopping, ModelCheckpoint, LRSchedulerCallback, History
Models	MicroMLP, MicroCNN, TinyResNet
Weight Init	xavier_uniform/normal, kaiming_uniform/normal, orthogonal, normal, uniform, constant, zeros, ones, eye
Monitoring	SystemMonitor, memory_trace, health_score, ModelProfiler (FLOPs, throughput)
Model I/O	save/load (gzip or plain pickle), state_dict, load_state_dict, optimizer state save/restore

🚀 Get Started

Ready to Build?

Install LowMind in seconds and start building deep learning models that run on any device.

📦 Install Now 💡 See Examples