littlelearn 1.0.4

machine learning ecosystem

🌱 LittleLearn – Touch the Big World with Little Steps

update Version (1.0.4) date : (29-December-2025):

- change numeric backend be jax.numpy() 
- fix memories leak problem 
- fix LSTM failure training bug
- fix Attention failure training bug 
- add Tensor class mechanism 
- Gradient Reflectot Being autodiff non data can use by general case 
- add Node Mechanism
- add general Tensor ops
- replacing AutoBuildModel and AutoTransformers with instant model in Model
- Tokenizer bug fixed 
- add DiagonalSSM layers with Gating mechanism 

warning : on this update we remove so many feature because paradims changed.

LittleLearn is an experimental and original machine learning framework built from scratch — inspired by the simplicity of Keras and the flexibility of PyTorch, yet designed with its own architecture, philosophy, and gradient engine.

🧠 What Makes LittleLearn Different?

• 🔧 Not a wrapper – LittleLearn is not built on top of TensorFlow, PyTorch, or other major ML libraries.

• 💡 Fully original layers, modules, and autodiff engine (GradientReflector).

• 🧩 Customizable down to the node level: build models from high-level APIs or go low-level for complete control.

• 🛠️ Features unique like:

• Node-level gradient clipping

• Inline graph tracing

• Custom attention mechanisms (e.g., Multi-Head Attention from scratch)

• 🤯 Designed for both research experimentation and deep learning education.

⚙️ Core Philosophy

Touch the Big World with Little Steps. Whether you want rapid prototyping or total model control — LittleLearn gives you both.

📦 Ecosystem Features

• ✅ Deep learning modules: Dense, LSTM, attention mechanisms, and more

• 🤖 instant model by Model Module

• 🔄 Custom training loops with full backend access

• 🧠 All powered by the GradientReflector engine — providing automatic differentiation with transparency and tweakability

🔧 Installation

    pip install littlelearn

🚀 Quick Example :

    import littlelearn as ll 
    import littlelearn.DeepLearning as dl 

    model = dl.layers.Sequential([
        dl.layers.Linear(20,32),
        dl.activations.Relu(),
        dl.layers.Linear(32,1)
    ]) 
    model.train()
    x_train,y_train= datasets()
    optimizer = dl.optimizers.Adam(model.parameter())

    for epoch in range(100) :
        y_pred = model(x_train)
        loss = dl.loss.mse_loss(y_train,y_pred)
        loss.backwardpass()
        optimizer.step()
        loss.reset_grad()
        print(loss.tensor)
    
    model.inference()
    model.save("model.npz")

📌 Disclaimer While inspired by well-known frameworks, LittleLearn is built entirely from scratch with its own mechanics. It is suitable for:

• 🔬 Experimental research

• 🏗️ Framework building

• 📚 Educational purposes

• 🔧 Custom low-level operations

suport this project : https://ko-fi.com/alpin92578

👤 Author Candra Alpin Gunawan 📧 [e-mail] 🌐 GitHub https://github.com/Airinchan818/LittleLearn

youtube : https://youtube.com/@hinamatsuriairin4596?si=KrBtOhXoVYnbBlpY