kitepy - Dead Simple Deep Learning 🚀

Build and train deep learning models with one line of code.

from kitepy import CNN
CNN("resnet18").train("cifar10")

That's it. No boilerplate. No configuration hell. Just magic. ✨

---

Features

• 🎯 One-Line Training: CNN("resnet18").train("cifar10")
• 🔧 Zero Configuration: Smart defaults that just work
• 🎨 Full Customization: Override anything when needed
• 🌐 Multi-Modal: CNNs & Transformers (LLMs, VLMs coming soon)
• ⚡ Production Ready: Built on PyTorch and timm (HuggingFace coming soon)
• 🚄 Fast: Automatic mixed precision (multi-GPU & distributed coming soon)

---

Installation

# Minimal install (CNN support)
pip install kitepy

# With LLM support
pip install kitepy[llm]

# Full install (all features)
pip install kitepy[all]

# Development install
git clone https://github.com/ExtraKaizen/kitepy
cd kitepy
pip install -e .

---

Quick Start

Example 1: Simplest Possible

from kitepy import CNN

# Train ResNet-18 on CIFAR-10
CNN("resnet18").train("cifar10")

Example 2: Custom Settings

# Override any parameter
CNN("resnet50").train(
    data="cifar100",
    epochs=50,
    batch_size=128,
    lr=0.001,
)

Example 3: Save and Load

# Train and save
model = CNN("resnet18")
model.train("cifar10", epochs=10)
model.save("my_model.pt")

# Load and use
loaded = CNN.load("my_model.pt")
predictions = loaded.predict(my_images)

Example 4: Custom PyTorch Model

import torch.nn as nn

# Use your own model
my_model = nn.Sequential(
    nn.Conv2d(3, 64, 3),
    nn.ReLU(),
    # ... your architecture
)

CNN(my_model).train("my_dataset")

---

Supported Models

CNNs (Vision)

• ResNet: resnet18, resnet34, resnet50, resnet101, resnet152
• EfficientNet: efficientnet_b0 through efficientnet_b7
• MobileNet: mobilenet_v2, mobilenet_v3_small, mobilenet_v3_large
• VGG: vgg11, vgg13, vgg16, vgg19
• DenseNet: densenet121, densenet169, densenet201
• And 800+ more from timm!

Vision Transformers

• ViT: vit_tiny, vit_small, vit_base, vit_large, vit_huge
• DeiT: deit_tiny, deit_small, deit_base
• Swin: swin_tiny, swin_small, swin_base, swin_large

Coming Soon (Phase 2+)

• LLMs: GPT, LLaMA, Mistral (Phase 3)
• VLMs: CLIP, BLIP, Flamingo (Phase 4)
• Audio: Whisper, AST (Phase 5)
• Diffusion: Stable Diffusion (Phase 5)

---

Supported Datasets

Built-in Vision Datasets

• cifar10, cifar100
• mnist, fashion_mnist
• imagenet (requires manual download)

Custom Datasets

# From path (ImageFolder structure)
CNN("resnet18").train("/path/to/my/dataset")

# From PyTorch DataLoader
from torch.utils.data import DataLoader
my_loader = DataLoader(...)
CNN("resnet18").train(my_loader)

---

Advanced Usage

Configuration System

from kitepy import CNN, CNNConfig, TrainingConfig

# Method 1: Config objects
model_config = CNNConfig(
    depth=50,
    num_classes=100,
    dropout=0.5
)

train_config = TrainingConfig(
    epochs=100,
    batch_size=256,
    lr=0.01,
    optimizer="adamw",
    scheduler="cosine",
)

model = CNN("resnet50", config=model_config)
model.train("cifar100", **train_config.to_dict())

YAML Configuration

# config.yaml
model:
  arch: resnet
  depth: 50
  num_classes: 100

training:
  epochs: 100
  batch_size: 256
  lr: 0.01
  optimizer: adamw

model = CNN(config="config.yaml")
model.train("cifar100")

Model Inspection

model = CNN("resnet18")

# Show architecture and config
model.describe()

# Show parameter count
model.summary()

# Show all config options
model.explain_config()

# Get underlying PyTorch model
pytorch_model = model.unwrap()

List Available Models

from kitepy import list_models, list_datasets

# List all CNN models
models = list_models("cnn")
print(models)  # ['resnet18', 'resnet50', ...]

# List all datasets
datasets = list_datasets()
print(datasets)  # ['cifar10', 'imagenet', ...]

Register Custom Presets

from kitepy import register_preset

register_preset("cnn", "my_custom_resnet", {
    "arch": "resnet",
    "depth": 101,
    "width_multiplier": 2.0,
})

# Now use it
CNN("my_custom_resnet").train("imagenet")

---

API Reference

CNN

CNN(
    model: str | nn.Module,           # Model name or custom PyTorch model
    config: dict | Config | str,      # Config dict, object, or YAML path
    **kwargs                          # Override any config value
)

Methods:

• .train(data, epochs, **kwargs) - Train the model
• .evaluate(data) - Evaluate on test data
• .predict(input) - Run inference
• .save(path) - Save checkpoint
• .load(path) - Load checkpoint (classmethod)
• .describe() - Print model info
• .summary() - Print parameter count
• .explain_config() - Show config options
• .unwrap() - Get PyTorch model

---

Architecture

kitepy/
├── api/           # Public APIs (CNN, LLM, VLM, etc.)
├── core/          # Core engines (training, data, optimization)
├── config/        # Configuration system
├── models/        # Architecture templates
├── data/          # Data loading
├── wrappers/      # timm, HuggingFace wrappers
└── utils/         # Utilities

Key Principle: We wrap existing ecosystems (PyTorch, timm, HuggingFace) instead of reimplementing.

---

Development Roadmap

• Phase 1: CNN Module ✅
• Phase 2: Vision Transformers ✅
• Phase 3: LLM Module (Coming Soon)
• Phase 4: VLM Module (Planned)
• Phase 5: Audio & Diffusion (Planned)

---

Requirements

• Python >= 3.8
• PyTorch >= 2.0.0
• timm >= 0.9.0 (for vision models)
• transformers >= 4.30.0 (for LLMs, optional)

---

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

---

License

MIT License - see LICENSE for details.

---

Citation

If you use kitepy in your research, please cite:

@software{kitepy2026,
  title={kitepy: Dead Simple Deep Learning},
  author={ExtraKaizen},
  year={2026},
  url={https://github.com/ExtraKaizen/kitepy}
}

---

Acknowledgments

Built on the shoulders of giants:

• PyTorch
• timm
• HuggingFace Transformers
• PyTorch Lightning

---

Made with ❤️ for the AI community