2df1ea92e1
BREAKING CHANGES: - Removed orphaned dqn.rs monolithic trainer (4,975 lines) - Removed orphaned dqn_ensemble.rs module (816 lines) - Removed orphaned tft.rs and tft_complete_int8_integration_test.rs - TFT trainer split into modular directory structure DQN Module Refactoring: - Split trainers/dqn.rs into modular structure (config.rs, statistics.rs, trainer.rs) - Fixed hyperopt 39D search space (continuous params only) - Boolean flags (use_dueling, use_double_dqn, use_per, use_noisy_nets) are now FIXED architectural decisions - use_distributional defaults to false (Candle BUG #36 - scatter_add gradient issues) Clean Module Structure: - ml/src/trainers/dqn/ directory with proper mod.rs exports - ml/src/trainers/tft/ directory with config.rs, types.rs, model.rs, trainer.rs, tests.rs - All P0 features validated: TD-error clamping, batch diversity, LR scheduler, priority staleness Documentation: - Added comprehensive docs in docs/codebase-cleanup/ - ADR-001 for DQN refactoring decisions - Rainbow DQN component matrix and quick reference guides Build Status: Compiles with zero errors 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
76 lines
2.3 KiB
Markdown
76 lines
2.3 KiB
Markdown
# DQN Trainer Refactoring Plan
|
|
|
|
## Current State
|
|
- **File**: `ml/src/trainers/dqn.rs`
|
|
- **Lines**: 4,975
|
|
- **Target**: <1,000 lines per module
|
|
|
|
## Module Extraction Strategy
|
|
|
|
### Module 1: `dqn/config.rs` (~800 lines)
|
|
**Lines 48-747**
|
|
- Constants: `EPISODE_LENGTH`, type aliases (`FeatureVector`, `FeatureVector51`)
|
|
- `FeatureStatistics` struct + impl (lines 78-157)
|
|
- `DQNHyperparameters` struct (lines 425-610)
|
|
- `DQNHyperparameters` impl methods (lines 615-747)
|
|
|
|
### Module 2: `dqn/agent_wrapper.rs` (~260 lines)
|
|
**Lines 164-424**
|
|
- `DQNAgentType` enum (wrapper for Standard/RegimeConditional)
|
|
- `QValueStats` struct
|
|
- All `DQNAgentType` impl methods (unified API)
|
|
|
|
### Module 3: `dqn/training_monitor.rs` (~270 lines)
|
|
**Lines 748-1012**
|
|
- `TrainingMonitor` struct
|
|
- All validation methods (rewards, actions, Q-values)
|
|
- Episode tracking
|
|
|
|
### Module 4: `dqn/trainer.rs` (~3900 lines -> split further)
|
|
**Lines 1013-4975**
|
|
This needs further breakdown:
|
|
- 4a. `trainer_core.rs`: Struct definition, constructors, epoch metrics
|
|
- 4b. `training_loop.rs`: Main `train()` and `train_from_parquet()` methods
|
|
- 4c. `data_loading.rs`: OHLCV extraction, feature creation
|
|
- 4d. `checkpointing.rs`: Checkpoint save/load logic
|
|
|
|
### Module 5: `dqn/mod.rs` (~50 lines)
|
|
- Public re-exports maintaining current API
|
|
- Module declarations
|
|
|
|
## Public API Preservation
|
|
|
|
**Current public exports from `trainers/dqn`:**
|
|
```rust
|
|
pub use dqn::{DQNHyperparameters, DQNTrainer};
|
|
```
|
|
|
|
**New structure ensures:**
|
|
```rust
|
|
// dqn/mod.rs re-exports everything
|
|
pub use config::{DQNHyperparameters, FeatureStatistics, EPISODE_LENGTH};
|
|
pub use agent_wrapper::{DQNAgentType, QValueStats};
|
|
pub use training_monitor::TrainingMonitor;
|
|
pub use trainer::DQNTrainer;
|
|
```
|
|
|
|
## Dependencies to Update
|
|
|
|
1. **Tests**: 19 DQN test files reference `trainers::dqn::`
|
|
2. **Hyperopt adapter**: `hyperopt/adapters/dqn.rs` (3,162 lines)
|
|
3. **Trainers mod**: `trainers/mod.rs` re-export
|
|
|
|
## Verification Steps
|
|
|
|
1. Extract modules incrementally
|
|
2. `cargo check --package ml` after each extraction
|
|
3. Update imports in dependent files
|
|
4. Run test suite: `cargo test --package ml --lib`
|
|
5. Verify no breaking changes to public API
|
|
|
|
## Risk Mitigation
|
|
|
|
- Keep original `dqn.rs` as `dqn.rs.backup` until verification complete
|
|
- Incremental extraction with build checks
|
|
- Test after each module extraction
|