neuroforge-dl 0.2.0

A full-featured deep learning framework and production AI platform built from first principles in Python & NumPy.

NeuroForge 🧠⚡

A full-featured deep learning framework and production AI platform — built entirely from first principles in Python/NumPy.

---

What is NeuroForge?

NeuroForge is a complete, production-ready deep learning stack implemented from scratch — no PyTorch, no TensorFlow, no Keras. Every component, from the tensor engine to the FastAPI model-serving layer, is hand-crafted:

Layer	What's included
🔢 Tensor Engine	NumPy-backed N-D arrays, dynamic autograd computation graph, full reverse-mode AD
🧱 Neural Modules	Linear, Conv1D, Conv2D, ConvTranspose2D, MaxPool, AvgPool, RNN, LSTM, GRU, Attention, Transformer, Embedding, LayerNorm, BatchNorm, Dropout
🏗️ Model Zoo	MLP, ConvNet, ResNet (ResNet18/34), U-Net, BERT, VAE, GPT, VisionTransformer, VisionLanguageModel
📉 Losses	CrossEntropy, BCE, BCEWithLogits, MSE, VAE Loss (BCE + KL)
⚙️ Optimizers & Schedulers	SGD (+ momentum), Adam, AdamW, StepLR, CosineAnnealingLR, WarmupCosineScheduler, WarmupLinearScheduler, ExponentialLR
🛠️ Training & Utilities	Trainer, Gradient Clipping (norm & value), Numerical Gradient Checker, Weight Initialization (Xavier, Kaiming, Orthogonal), Callbacks
📊 Experiments & Visuals	ExperimentTracker, Plot Training History, Gradient Norm Inspection
📦 Registry & Serving	ModelRegistry (versioned save/load) + FastAPI REST server (/health, /models, /predict, /generate, /metrics)

---

Architecture

┌───────────────────────────────────────────────────────────────────┐
│                          NeuroForge Stack                          │
├───────────────────────────────────────────────────────────────────┤
│  neuroforge.core      │  Tensor + Autograd Engine                  │
│  neuroforge.nn        │  Module, Parameter, 16 Layer Types        │
│  neuroforge.models    │  MLP, ConvNet, ResNet, UNet, BERT, VAE...  │
│  neuroforge.losses    │  CrossEntropy, BCE, MSE, VAE Loss          │
│  neuroforge.optim     │  SGD, Adam, AdamW, Warmup Schedulers       │
│  neuroforge.data      │  TensorDataset, DataLoader                 │
│  neuroforge.training  │  Trainer, Callbacks, Metrics               │
│  neuroforge.utils     │  Init, Grad Clipping, Grad Check, Plots    │
│  neuroforge.experiments│  ExperimentTracker                        │
│  neuroforge.registry  │  ModelRegistry (versioned persistence)     │
│  neuroforge.inference │  InferenceEngine                           │
│  neuroforge.serving   │  FastAPI production REST server            │
└───────────────────────────────────────────────────────────────────┘

---

Quick Start

Installation

cd NeuroForge
pip install -e .

Train a ConvNet in 5 lines

import numpy as np
from neuroforge import ConvNet, CrossEntropyLoss, Adam, TensorDataset, DataLoader, Trainer

X = np.random.randn(200, 1, 28, 28).astype("float32")
y = np.random.randint(0, 10, size=(200,)).astype("int64")

model    = ConvNet(in_channels=1, num_classes=10, channels=[16, 32])
trainer  = Trainer(model, CrossEntropyLoss(), Adam(model.parameters(), lr=0.005))
history  = trainer.fit(DataLoader(TensorDataset(X, y), batch_size=32), epochs=5)

Advanced Models: ResNet18, U-Net, BERT & VAE

import neuroforge as nf

# ResNet-18
resnet = nf.ResNet18(in_channels=3, num_classes=10)
nf.apply_init(resnet, init_fn=nf.kaiming_normal_)

# U-Net Image Segmentation
unet = nf.UNet(in_channels=1, out_channels=1, features=[32, 64])

# BERT Masked Language Model
bert = nf.BERT(vocab_size=1000, d_model=128, nhead=4, num_layers=4)

# VAE with Reparameterization
vae = nf.VAE(input_dim=784, hidden_dim=256, latent_dim=32)

---

Examples

python examples/01_mnist_cnn.py            # ConvNet image classification
python examples/02_text_gpt.py             # GPT language model + text generation
python examples/03_multimodal_demo.py      # Vision-Language multimodal model
python examples/04_advanced_models_demo.py # ResNet, U-Net, BERT, VAE, Grad Clipping & Warmup

---

CLI Training & Serving

# CLI Training
python scripts/train.py --config configs/default_train.yaml

# REST Serving
python scripts/serve.py --host 0.0.0.0 --port 8000

---

Testing

python -m pytest tests/ -v

36 passed in 1.45s  ✅

Test Suite	Tests	Status
tests/unit/test_tensor.py	5	✅
tests/unit/test_autograd.py	3	✅
tests/unit/test_layers.py	5	✅
tests/unit/test_advanced_layers.py	2	✅
tests/unit/test_models.py	5	✅
tests/unit/test_advanced_models.py	4	✅
tests/unit/test_optimizers.py	3	✅
tests/unit/test_utils_and_schedulers.py	4	✅
tests/unit/test_serving.py	4	✅
tests/integration/test_pipeline.py	1	✅

---

License

MIT License. See LICENSE.