daimond 0.1.0

checksums.yaml

@@ -0,0 +1,7 @@
+---
+SHA256:
+  metadata.gz: 5f8e7eb4a7668542baca1504c843eb802996481641c22eda1633c24395591f30
+  data.tar.gz: 812a7410ef056cdb8b9a47cab26fa6786191b4cf9715f59b53f1be6f9e6a2c77
+SHA512:
+  metadata.gz: b116749b121c4d779e0884fa512cd8f7d1933c51102ec88f631a3c0422a01c5818abca9093e4c3acea28e9f8ba1f8693609bcde07afddc2b28e58c256e88ba39
+  data.tar.gz: 679a9249cf1171ed4798d5aef7413d595c9910c835dc0ea5eaad715a8307caa4193cdf9fcdb574e2fbb86bb20120c5ecf20aa18582de66ea5b755435ac59bf7e

data/CONTRIBUTIONG.md

@@ -0,0 +1,160 @@
+# Contributing to dAImond 💎
+
+First off, thank you for considering contributing to dAImond! This is a passion project to bring deep learning to Ruby, and every contribution helps.
+
+## Where to Start?
+
+- **⭐ Star the repo** - It helps others discover the project
+- **🐛 Report bugs** - Found an issue? Open an issue with details
+- **📖 Improve docs** - Fix typos, improve examples, add translations
+- **✨ Add features** - New layers, optimizers, or examples
+- **🧪 Add tests** - We need them desperately!
+
+## Development Setup
+
+```bash
+git clone https://github.com/yourusername/daimond.git
+cd daimond
+bundle install
+ruby examples/mnist.rb  # Verify it works
+```
+
+## How to Contribute
+
+**Reporting Bugs**
+
+Use GitHub Issues and include:
+Ruby version (ruby -v)
+Error message with full backtrace
+Minimal code to reproduce
+Expected vs actual behavior
+Suggesting Features
+
+Open an issue with [Feature Request] in title. Explain:
+What and why?
+API design (how should it look?)
+Are you willing to implement it?
+Pull Request Process
+
+Fork the repo and create your branch: git checkout -b feature/amazing-feature
+Make changes following our style guide below
+Test manually with MNIST at minimum: ruby examples/mnist.rb should still achieve 97%+
+Update docs (this README, code comments, examples)
+Commit with clear messages:
+feat: Add BatchNorm layer
+fix: Correct ReLU backward pass
+docs: Update Japanese README
+Push to your fork and open a Pull Request
+Wait for review (usually within 24-48 hours)
+Code Style Guide
+
+**Ruby Style**
+
+Follow Standard Ruby style guide
+Max 100 characters per line
+Prefer do...end for multiline blocks
+Always use frozen_string_literal: true
+
+**For New Layers (NN)**
+If adding new layers (e.g., Conv2D, LSTM):
+```ruby
+module Daimond
+  module NN
+    class Conv2d < Module
+      def initialize(in_channels, out_channels, kernel_size)
+        super()
+        # Initialize parameters as Tensors
+        # Add to @parameters array
+      end
+      
+      def forward(x)
+        # Return Tensor
+        # Store _backward lambda for autograd
+      end
+    end
+  end
+end
+```
+
+**Must include:**
+- super() call in initialize
+- Add parameters to @parameters
+- Return Tensor from forward
+- Implement _backward lambda with correct gradients
+- Update lib/daimond.rb exports
+
+**For Optimizers:**
+```ruby
+module Daimond
+  module Optim
+    class Adam < SGD  # Or Module
+      def initialize(parameters, lr: 0.001, beta1: 0.9, beta2: 0.999)
+        super()
+        @parameters = parameters
+        @lr = lr
+        # Initialize moments
+      end
+      
+      def step
+        # Update parameters using gradients
+      end
+      
+      def zero_grad
+        @parameters.each { |p| p.grad = Numo::DFloat.zeros(*p.shape) }
+      end
+    end
+  end
+end
+```
+
+## Testing (Important!)
+**1. Verify backward pass manually for new operations:**
+```ruby
+# Create simple test case
+x = Daimond::Tensor.new([[1.0, 2.0]])
+y = x.relu
+y.backward!
+
+# Check gradient is correct manually
+puts x.grad.inspect
+```
+**2. Verify MNIST still trains to 95%+ accuracy**
+**3. Add example script in examples/ if adding major features:**
+- examples/conv2d_demo.rb
+- examples/adam_comparison.rb
+
+**Documentation**
+- Code comments: Explain complex math in backward passes
+- README updates: If adding public API
+- Type hints: Optional, but helpful (e.g., # @param [Tensor] input)
+
+## Performance Guidelines
+dAImond is for education and prototyping, but we shouldn't be painfully slow:
+- Prefer Numo operations over Ruby loops for math
+- Minimize object allocation in hot loops (forward/backward)
+- Profile before optimizing: Use ruby -rprofile if needed
+
+Example:
+```ruby
+# ❌ Bad: Ruby loop
+result = []
+1000.times { |i| result << data[i] * 2 }
+
+# ✅ Good: Numo vectorized
+result = data * 2
+```
+
+## Recognition
+- Contributors will be:
+- Added to CONTRIBUTORS.md file
+- Mentioned in release notes
+- Forever appreciated! 🙏
+
+## Questions
+- Open a GitHub Discussion
+- Or email: fyodor@hudzone.com
+
+## Code of Conduct
+Be nice
+
+**Thank you for helping make dAImond shine! Glory Ruby.** 💎✨

data/README.ja.md

@@ -0,0 +1,115 @@
+# dAImond 💎
+
+PyTorchにインスパイアされた、Rubyのためのディープラーニングフレームワーク。Rubyコミュニティへの愛を込めて、ゼロから作成しました。
+
+[![Ruby](https://img.shields.io/badge/ruby-%23CC342D.svg?style=for-the-badge&logo=ruby&logoColor=white)](https://www.ruby-lang.org/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+
+> **なぜRuby？** 理由はわからない、ただこの言語が好きだから。dAImondはMLに幸せを取り戻します。
+
+## 主な機能
+
+- 🔥 **自動微分** - 計算グラフによる完全なautogradエンジン
+- 🧠 **ニューラルネットワーク** - 線形レイヤー、活性化関数（ReLU、Sigmoid、Softmax、Tanh）
+- 📊 **オプティマイザー** - Momentum付きSGD、学習率スケジューリング
+- 🎯 **損失関数** - MSE、CrossEntropy
+- 💾 **モデルの保存** - Marshalによる保存/読み込み
+- 📈 **データローダー** - バッチ処理、シャッフル、MNIST対応
+- ⚡ **高速バックエンド** - Numo::NArrayによるベクトル化演算（C速度）
+- 🎨 **美しいAPI** - 慣用的なRuby DSL、メソッドチェーン
+
+## インストール
+
+Gemfileに追加：
+
+```ruby
+gem 'daimond'
+```
+
+または手動でインストール：
+```ruby
+gem install daimond
+```
+
+**依存関係:** Ruby 2.7+, numo-narray
+クイックスタート
+```ruby
+require 'daimond'
+
+# モデルの定義
+class NeuralNet < Daimond::NN::Module
+  attr_reader :fc1, :fc2
+  
+  def initialize
+    super()
+    @fc1 = Daimond::NN::Linear.new(784, 128)
+    @fc2 = Daimond::NN::Linear.new(128, 10)
+    @parameters = @fc1.parameters + @fc2.parameters
+  end
+  
+  def forward(x)
+    x = @fc1.forward(x).relu
+    @fc2.forward(x).softmax
+  end
+end
+
+# トレーニングループ
+model = NeuralNet.new
+optimizer = Daimond::Optim::SGD.new(model.parameters, lr: 0.1, momentum: 0.9)
+criterion = Daimond::Loss::CrossEntropyLoss.new
+
+# 順伝播 → 逆伝播 → 更新
+loss = criterion.call(model.forward(input), target)
+optimizer.zero_grad
+loss.backward!
+optimizer.step
+```
+
+## MNISTの例（97%の精度！）
+
+60,000枚の手書き数字で分類器をトレーニング：
+```ruby
+ruby examples/mnist.rb
+```
+
+結果：
+```text
+Epoch 1/5: Loss = 0.2898, Accuracy = 91.35%
+Epoch 2/5: Loss = 0.1638, Accuracy = 95.31%
+Epoch 3/5: Loss = 0.1389, Accuracy = 96.2%
+Epoch 4/5: Loss = 0.1195, Accuracy = 96.68%
+Epoch 5/5: Loss = 0.1083, Accuracy = 97.12%
+```
+
+モデルの保存：
+```ruby
+model.save('models/mnist_model.bin')
+```
+読み込みと予測：
+```ruby
+model = NeuralNet.new
+model.load('models/mnist_model.bin')
+prediction = model.forward(test_image)
+```
+
+## パフォーマンス
+純粋なRubyはPyTorch/CUDAより遅いですが、dAImondはプロトタイピングや小〜中規模のデータセットに対して合理的な速度を実現しています：
+MNIST（60k画像）：現代のCPUで1エポックあたり約2〜3分
+教育、研究、100万パラメーター未満のモデルに最適
+
+## ロードマップ
+- [x] コアautogradエンジン
+- [x] 線形レイヤーと活性化関数
+- [x] MNIST 97%精度
+- [x] モデルのシリアライズ
+- [ ] 畳み込みレイヤー（Conv2D）
+- [ ] Batch NormalizationとDropout
+- [ ] Adam/RMSpropオプティマイザー
+- [ ] GPUサポート（OpenCL/CUDA経由FFI）
+- [ ] ONNXエクスポート/インポート
+
+## コントリビューション
+どんなコントリビューターも歓迎します！CONTRIBUTING.mdをご覧ください。
+
+## ライセンス
+MIT License - LICENSEファイルを参照。

data/README.md

@@ -0,0 +1,115 @@
+# dAImond 💎
+
+Deep Learning framework for Ruby, inspired by PyTorch. Written from scratch with love for the Ruby community.
+
+[![Ruby](https://img.shields.io/badge/ruby-%23CC342D.svg?style=for-the-badge&logo=ruby&logoColor=white)](https://www.ruby-lang.org/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+
+> **Why Ruby?** IDK, i just love this lang. dAImond brings back the happiness to ML.
+
+## Features
+
+- 🔥 **Automatic Differentiation** - Full autograd engine with computational graphs
+- 🧠 **Neural Networks** - Linear layers, activations (ReLU, Sigmoid, Softmax, Tanh)
+- 📊 **Optimizers** - SGD with momentum, learning rate scheduling
+- 🎯 **Loss Functions** - MSE, CrossEntropy
+- 💾 **Model Serialization** - Save/load trained models with Marshal
+- 📈 **Data Loaders** - Batch processing, shuffling, MNIST support
+- ⚡ **Fast Backend** - Numo::NArray for vectorized operations (C-speed)
+- 🎨 **Beautiful API** - Idiomatic Ruby DSL, chainable methods
+
+## Installation
+
+Add this line to your Gemfile:
+
+```ruby
+gem 'daimond'
+```
+
+Or install manually:
+```ruby
+gem install daimond
+```
+
+**Dependencies:** Ruby 2.7+, numo-narray
+
+## Quick Start
+```ruby
+require 'daimond'
+
+# Define your model
+class NeuralNet < Daimond::NN::Module
+  attr_reader :fc1, :fc2
+  
+  def initialize
+    super()
+    @fc1 = Daimond::NN::Linear.new(784, 128)
+    @fc2 = Daimond::NN::Linear.new(128, 10)
+    @parameters = @fc1.parameters + @fc2.parameters
+  end
+  
+  def forward(x)
+    x = @fc1.forward(x).relu
+    @fc2.forward(x).softmax
+  end
+end
+
+# Training loop
+model = NeuralNet.new
+optimizer = Daimond::Optim::SGD.new(model.parameters, lr: 0.1, momentum: 0.9)
+criterion = Daimond::Loss::CrossEntropyLoss.new
+
+# Forward → Backward → Update
+loss = criterion.call(model.forward(input), target)
+optimizer.zero_grad
+loss.backward!
+optimizer.step
+```
+
+## MNIST Example (97% Accuracy!)
+**Train a classifier on 60,000 handwritten digits:**
+```ruby
+ruby examples/mnist.rb
+```
+**Results:**
+```text
+Epoch 1/5: Loss = 0.2898, Accuracy = 91.35%
+Epoch 2/5: Loss = 0.1638, Accuracy = 95.31%
+Epoch 3/5: Loss = 0.1389, Accuracy = 96.2%
+Epoch 4/5: Loss = 0.1195, Accuracy = 96.68%
+Epoch 5/5: Loss = 0.1083, Accuracy = 97.12%
+```
+
+**Save your model:**
+```ruby
+model.save('models/mnist_model.bin')
+```
+
+**Load and predict:**
+```ruby
+model = NeuralNet.new
+model.load('models/mnist_model.bin')
+prediction = model.forward(test_image)
+```
+
+## Performance
+While pure Ruby is slower than PyTorch/CUDA, dAImond achieves reasonable speeds for prototyping and small-to-medium datasets:
+MNIST (60k images): ~2-3 minutes per epoch on modern CPU
+Perfect for education, research, and production models < 1M parameters
+
+## Roadmap
+- [x] Core autograd engine
+- [x] Linear layers & activations
+- [x] MNIST 97% accuracy
+- [x] Model serialization
+- [ ] Convolutional layers (Conv2D)
+- [ ] Batch Normalization & Dropout
+- [ ] Adam/RMSprop optimizers
+- [ ] GPU support (OpenCL/CUDA via FFI)
+- [ ] ONNX export/import
+
+## Contributing
+I'll be happy to see any contributors! Please read CONTRIBUTING.md for details.
+
+## License
+MIT License - see LICENSE file.

data/README.ru.md

@@ -0,0 +1,116 @@
+# dAImond 💎
+
+Deep Learnin фреймворк для Ruby, вдохновлённый PyTorch.
+
+[![Ruby](https://img.shields.io/badge/ruby-%23CC342D.svg?style=for-the-badge&logo=ruby&logoColor=white)](https://www.ruby-lang.org/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
+
+> **Почему Ruby?** ХЗ, захотелось. dAImond возвращает радость в возюкании с ML, потому что это Ruby.
+
+## Возможности
+
+- 🔥 **Автоматическое дифференцирование** - Полноценный autograd с вычислительными графами
+- 🧠 **Нейронные сети** - Линейные слои, активации (ReLU, Sigmoid, Softmax, Tanh)
+- 📊 **Оптимизаторы** - SGD с моментумом, планирование learning rate
+- 🎯 **Функции потерь** - MSE, CrossEntropy
+- 💾 **Сериализация моделей** - Сохранение/загрузка через Marshal
+- 📈 **Загрузчики данных** - Batch processing, шаффл, поддержка MNIST
+- ⚡ **Быстрый бэкенд** - Numo::NArray для векторизованных операций (скорость C)
+- 🎨 **Красивый API** - Идиоматичный Ruby DSL, чейнящиеся методы
+
+## Установка
+
+Добавьте в Gemfile:
+
+```ruby
+gem 'daimond'
+```
+
+
+Или установите ручками:
+```ruby
+gem install daimond
+```
+
+**Зависимости:** Ruby 2.7+, numo-narray
+
+## Быстрый старт
+```ruby
+require 'daimond'
+
+# Define your model
+class NeuralNet < Daimond::NN::Module
+  attr_reader :fc1, :fc2
+  
+  def initialize
+    super()
+    @fc1 = Daimond::NN::Linear.new(784, 128)
+    @fc2 = Daimond::NN::Linear.new(128, 10)
+    @parameters = @fc1.parameters + @fc2.parameters
+  end
+  
+  def forward(x)
+    x = @fc1.forward(x).relu
+    @fc2.forward(x).softmax
+  end
+end
+
+# Training loop
+model = NeuralNet.new
+optimizer = Daimond::Optim::SGD.new(model.parameters, lr: 0.1, momentum: 0.9)
+criterion = Daimond::Loss::CrossEntropyLoss.new
+
+# Forward → Backward → Update
+loss = criterion.call(model.forward(input), target)
+optimizer.zero_grad
+loss.backward!
+optimizer.step
+```
+
+## Пример MNIST (97% Accuracy!)
+**Обучение классификатора на 60к рукописных цифрах:**
+```ruby
+ruby examples/mnist.rb
+```
+**Результаты:**
+```text
+Epoch 1/5: Loss = 0.2898, Accuracy = 91.35%
+Epoch 2/5: Loss = 0.1638, Accuracy = 95.31%
+Epoch 3/5: Loss = 0.1389, Accuracy = 96.2%
+Epoch 4/5: Loss = 0.1195, Accuracy = 96.68%
+Epoch 5/5: Loss = 0.1083, Accuracy = 97.12%
+```
+
+**Сохранение модели:**
+```ruby
+model.save('models/mnist_model.bin')
+```
+
+**Загрузка и предикт:**
+```ruby
+model = NeuralNet.new
+model.load('models/mnist_model.bin')
+prediction = model.forward(test_image)
+```
+
+## Производительность
+Хотя чистый Ruby медленнее PyTorch/CUDA, dAImond обеспечивает разумную скорость для прототипирования и небольших/средних датасетов:
+MNIST (60k изображений): ~2-3 минуты на эпоху на современном CPU
+Идеально для обучения, исследований и моделей < 1M параметров
+
+## Планы
+- [x] Ядро autograd
+- [x] Линейные слои и активации
+- [x] MNIST 97% точности
+- [x] Сериализация моделей
+- [ ] Свёрточные слои (Conv2D)
+- [ ] Batch Normalization и Dropout
+- [ ] Оптимизаторы Adam/RMSprop
+- [ ] Поддержка GPU (OpenCL/CUDA via FFI)
+- [ ] ONNX экспорт/импорт
+
+## Помощь
+Буду рад любой помощи! Инфа в CONTRIBUTING.md.
+
+## Лицензия
+MIT License