845f8707b9
Per-trade online learning with Elastic Weight Consolidation to prevent catastrophic forgetting. Rolling 10K experience buffer, mini-updates every 100 trades. Safety rails: grad clip 1.0, LR×0.1, auto-rollback at 20% Sharpe degradation, kill switch at Sharpe < -1.0. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
111 lines
4.0 KiB
Rust
111 lines
4.0 KiB
Rust
//! ML Model Trainers with gRPC Integration
|
|
//!
|
|
//! Production-grade trainers for all ML models in the Foxhunt system,
|
|
//! designed to integrate seamlessly with the ML Training Service gRPC interface.
|
|
//!
|
|
//! ## Available Trainers
|
|
//!
|
|
//! - **MAMBA-2**: State space model optimized for 4GB VRAM with gradient checkpointing
|
|
//! - **TFT (Temporal Fusion Transformer)**: Time series forecasting with attention
|
|
//! - **DQN**: Deep Q-Network for reinforcement learning
|
|
//! - **PPO**: Proximal Policy Optimization
|
|
//! - **Liquid Networks**: Continuous-time neural ODEs
|
|
//! - **TLOB**: Temporal Limit Order Book transformer
|
|
//!
|
|
//! ## Features
|
|
//!
|
|
//! - Real-time training progress streaming via gRPC
|
|
//! - GPU acceleration with memory-efficient operations (4GB VRAM optimized)
|
|
//! - Checkpoint management with MinIO/S3 integration
|
|
//! - Comprehensive metrics reporting (loss, perplexity, state statistics)
|
|
//! - Resource usage monitoring (CPU, memory, GPU)
|
|
//! - Early stopping and learning rate scheduling
|
|
//! - Teacher forcing for sequence models
|
|
//!
|
|
//! ## MAMBA-2 Trainer Example
|
|
//!
|
|
//! ```rust,no_run
|
|
//! use ml::trainers::mamba2::{Mamba2Trainer, Mamba2Hyperparameters};
|
|
//! use candle_core::{Device, Tensor, DType};
|
|
//!
|
|
//! #[tokio::main]
|
|
//! async fn main() -> Result<(), Box<dyn std::error::Error>> {
|
|
//! // Configure MAMBA-2 trainer (validated for 4GB VRAM)
|
|
//! let hyperparameters = Mamba2Hyperparameters {
|
|
//! learning_rate: 1e-4,
|
|
//! batch_size: 8, // Conservative for GPU memory
|
|
//! d_model: 256, // Hidden dimension
|
|
//! n_layers: 6, // Number of layers
|
|
//! state_size: 32, // SSM state dimension
|
|
//! dropout: 0.1,
|
|
//! epochs: 100,
|
|
//! seq_len: 128, // Sequence length
|
|
//! grad_clip: 1.0,
|
|
//! weight_decay: 1e-4,
|
|
//! warmup_steps: 1000,
|
|
//! };
|
|
//!
|
|
//! // Create trainer with checkpoint path
|
|
//! let checkpoint_path = Some("s3://foxhunt-ml-models/mamba2/my-job".to_owned());
|
|
//! let mut trainer = Mamba2Trainer::new(hyperparameters, checkpoint_path)?;
|
|
//!
|
|
//! // Create dummy training data (replace with real market data)
|
|
//! let device = Device::cuda_if_available(0)?;
|
|
//! let train_data: Vec<(Tensor, Tensor)> = vec![
|
|
//! (
|
|
//! Tensor::randn(0.0, 1.0, (8, 128, 256), &device)?,
|
|
//! Tensor::randn(0.0, 1.0, (8, 1), &device)?,
|
|
//! )
|
|
//! ];
|
|
//! let val_data = train_data.clone();
|
|
//!
|
|
//! // Train with real-time progress updates
|
|
//! let training_history = trainer.train(&train_data, &val_data).await?;
|
|
//!
|
|
//! println!("Training complete: {} epochs", training_history.len());
|
|
//! println!("Best validation loss: {:.6}", trainer.best_val_loss);
|
|
//!
|
|
//! Ok(())
|
|
//! }
|
|
//! ```
|
|
|
|
pub mod curriculum;
|
|
pub mod dqn;
|
|
pub mod liquid;
|
|
pub mod mamba2;
|
|
pub mod online_learning;
|
|
pub mod ppo;
|
|
pub mod tft;
|
|
pub mod tft_parquet; // Parquet lazy-loading extension for TFT
|
|
pub mod tlob;
|
|
|
|
/// Target network update strategy for DQN training
|
|
#[derive(Debug, Clone, Copy)]
|
|
pub enum TargetUpdateMode {
|
|
/// Polyak averaging (soft target updates): θ_target = (1-τ)θ_target + τθ_online
|
|
/// Rainbow DQN standard with smoother Q-value tracking
|
|
Soft,
|
|
/// Hard target updates: Full copy every N steps (legacy)
|
|
/// Causes sudden Q-value shifts but simpler implementation
|
|
Hard,
|
|
}
|
|
|
|
impl Default for TargetUpdateMode {
|
|
fn default() -> Self {
|
|
TargetUpdateMode::Soft
|
|
}
|
|
}
|
|
|
|
// Re-export commonly used types
|
|
pub use dqn::{DQNHyperparameters, DQNTrainer};
|
|
pub use mamba2::{
|
|
Mamba2Hyperparameters, Mamba2Trainer, ProgressCallback, TrainingMetrics, TrainingProgress,
|
|
};
|
|
pub use ppo::{PpoHyperparameters, PpoTrainer, PpoTrainingMetrics};
|
|
pub use tft::{
|
|
ResourceUsage as TFTResourceUsage, TFTTrainer, TFTTrainerConfig,
|
|
TrainingMetrics as TFTTrainingMetrics, TrainingProgress as TFTTrainingProgress,
|
|
};
|
|
pub use liquid::{LiquidHyperparameters, LiquidTrainer, LiquidTrainingMetrics};
|
|
pub use tlob::{TLOBHyperparameters, TLOBTrainer, TLOBTrainingMetrics};
|