gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(dqn,ppo): production inference hardening, remove legacy aliases

Browse Source 
DQN: disable all exploration for production (warmup=0, epsilon=0,
noisy_nets=false, count_bonus=false), read feature_count from
checkpoint metadata instead of hardcoding 54, add loaded guard
and NaN check on predict output.

PPO: fix confidence formula — use act_with_log_prob() instead of
act() which returns value_estimate not log_prob, add loaded guard
and NaN check, remove WorkingPPO alias.

Liquid: add loaded flag for is_ready()/predict() guards.

Remove EgoboxOptimizer/EgoboxOptimizerBuilder backward-compat
aliases — replaced with canonical ArgminOptimizer everywhere.

323 trading_service tests, 200 hyperopt tests, 0 clippy warnings.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-03-04 01:19:02 +01:00
parent d0c5126985
commit 1bf752d387
15 changed files with 300 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
crates/ml/src/hyperopt/adapters/continuous_ppo.rs
View File

@@ -10,7 +10,7 @@
//! ## Usage Example
//!
//! ```rust,no_run
//! use ml::hyperopt::EgoboxOptimizer;
//! use ml::hyperopt::ArgminOptimizer;
//! use ml::hyperopt::adapters::continuous_ppo::{ContinuousPPOTrainer, ContinuousPPOParams};
//!
//! # async fn example() -> anyhow::Result<()> {
@@ -21,7 +21,7 @@
//! )?;
//!
//! // Run optimization
//! let optimizer = EgoboxOptimizer::with_trials(30, 5);
//! let optimizer = ArgminOptimizer::with_trials(30, 5);
//! let result = optimizer.optimize(trainer)?;
//!
//! println!("Best policy LR: {}", result.best_params.policy_lr);

4
crates/ml/src/hyperopt/adapters/dqn.rs
View File

@@ -21,7 +21,7 @@
//! ## Usage Example
//!
//! ```rust,no_run
//! use ml::hyperopt::EgoboxOptimizer;
//! use ml::hyperopt::ArgminOptimizer;
//! use ml::hyperopt::adapters::dqn::{DQNTrainer, DQNParams};
//!
//! # async fn example() -> anyhow::Result<()> {
@@ -32,7 +32,7 @@
//! )?;
//!
//! // Run optimization
//! let optimizer = EgoboxOptimizer::with_trials(30, 5);
//! let optimizer = ArgminOptimizer::with_trials(30, 5);
//! let result = optimizer.optimize(trainer)?;
//!
//! println!("Best learning rate: {}", result.best_params.learning_rate);

6
crates/ml/src/hyperopt/adapters/mamba2.rs
View File

@@ -10,7 +10,7 @@
//! ## Usage Example
//!
//! ```rust,no_run
//! use ml::hyperopt::EgoboxOptimizer;
//! use ml::hyperopt::ArgminOptimizer;
//! use ml::hyperopt::adapters::mamba2::{Mamba2Trainer, Mamba2Params};
//!
//! # async fn example() -> anyhow::Result<()> {
@@ -21,7 +21,7 @@
//! )?;
//!
//! // Run optimization
//! let optimizer = EgoboxOptimizer::with_trials(30, 5);
//! let optimizer = ArgminOptimizer::with_trials(30, 5);
//! let result = optimizer.optimize(trainer)?;
//!
//! println!("Best learning rate: {}", result.best_params.learning_rate);
@@ -227,7 +227,7 @@ pub struct Mamba2Metrics {
/// MAMBA-2 trainer for hyperparameter optimization
///
/// This struct wraps the MAMBA-2 training pipeline and implements
/// `HyperparameterOptimizable` for use with `EgoboxOptimizer`.
/// `HyperparameterOptimizable` for use with `ArgminOptimizer`.
///
/// ## Configuration
///

4
crates/ml/src/hyperopt/adapters/ppo.rs
View File

@@ -10,7 +10,7 @@
//! ## Usage Example
//!
//! ```rust,no_run
//! use ml::hyperopt::EgoboxOptimizer;
//! use ml::hyperopt::ArgminOptimizer;
//! use ml::hyperopt::adapters::ppo::{PPOTrainer, PPOParams};
//!
//! # async fn example() -> anyhow::Result<()> {
@@ -20,7 +20,7 @@
//! )?;
//!
//! // Run optimization
//! let optimizer = EgoboxOptimizer::with_trials(30, 5);
//! let optimizer = ArgminOptimizer::with_trials(30, 5);
//! let result = optimizer.optimize(trainer)?;
//!
//! println!("Best policy LR: {}", result.best_params.policy_learning_rate);

4
crates/ml/src/hyperopt/adapters/tft.rs
View File

@@ -10,7 +10,7 @@
//! ## Usage Example
//!
//! ```rust,no_run
//! use ml::hyperopt::EgoboxOptimizer;
//! use ml::hyperopt::ArgminOptimizer;
//! use ml::hyperopt::adapters::tft::{TFTTrainer, TFTParams};
//!
//! # async fn example() -> anyhow::Result<()> {
@@ -21,7 +21,7 @@
//! )?;
//!
//! // Run optimization
//! let optimizer = EgoboxOptimizer::with_trials(30, 5);
//! let optimizer = ArgminOptimizer::with_trials(30, 5);
//! let result = optimizer.optimize(trainer)?;
//!
//! println!("Best learning rate: {}", result.best_params.learning_rate);

1
crates/ml/src/hyperopt/mod.rs
View File

@@ -59,7 +59,6 @@ mod tests_argmin; // New argmin tests
// Re-exports for convenience
pub use observer::TrialBudgetObserver;
pub use optimizer::{optimize_with_tpe, ArgminOptimizer, ArgminOptimizerBuilder, TwoPhaseObjective};
pub use optimizer::{EgoboxOptimizer, EgoboxOptimizerBuilder}; // Backward compatibility
pub use traits::{
HardwareBudget, HyperoptStrategy, HyperparameterOptimizable, OptimizationResult,
ParameterSpace, TrialResult,

4
crates/ml/src/hyperopt/optimizer.rs
View File

@@ -1300,10 +1300,6 @@ impl ArgminOptimizerBuilder {
}
}
// Re-export for backward compatibility
pub type EgoboxOptimizer = ArgminOptimizer;
pub type EgoboxOptimizerBuilder = ArgminOptimizerBuilder;
#[cfg(test)]
mod tests {
use super::*;

21
crates/ml/src/hyperopt/tests_argmin.rs
View File

@@ -743,25 +743,4 @@ mod tests {
// BACKWARD COMPATIBILITY TESTS
// ============================================================================
#[test]
fn test_egobox_optimizer_alias() {
use crate::hyperopt::optimizer::EgoboxOptimizer;
let optimizer = EgoboxOptimizer::new();
assert_eq!(optimizer.max_trials, 30);
assert_eq!(optimizer.n_initial, 5);
}
#[test]
fn test_egobox_optimizer_builder_alias() {
use crate::hyperopt::optimizer::EgoboxOptimizerBuilder;
let optimizer = EgoboxOptimizerBuilder::new()
.max_trials(20)
.n_initial(3)
.build();
assert_eq!(optimizer.max_trials, 20);
assert_eq!(optimizer.n_initial, 3);
}
}

6
crates/ml/src/hyperopt/traits.rs
View File

@@ -20,7 +20,7 @@
//! └─────────────────────────────────────────────────────────────┘
//! ↓
//! ┌──────────────────┐
//! │ EgoboxOptimizer │
//! │ ArgminOptimizer │
//! │ (Generic impl) │
//! └──────────────────┘
//! ↓
@@ -33,7 +33,7 @@
//! ## Usage Example
//!
//! ```rust,no_run
//! use ml::hyperopt::{EgoboxOptimizer, HyperparameterOptimizable};
//! use ml::hyperopt::{ArgminOptimizer, HyperparameterOptimizable};
//! use ml::hyperopt::adapters::mamba2::{Mamba2Trainer, Mamba2Params};
//!
//! # async fn example() -> anyhow::Result<()> {
@@ -44,7 +44,7 @@
//! )?;
//!
//! // Run optimization
//! let optimizer = EgoboxOptimizer::new(30, 5);
//! let optimizer = ArgminOptimizer::new(30, 5);
//! let result = optimizer.optimize(trainer)?;
//!
//! println!("Best params: {:?}", result.best_params);

2
crates/ml/src/ppo/mod.rs
View File

@@ -47,7 +47,7 @@ pub use continuous_ppo::{
};
pub use flow_policy::{FlowPolicy, FlowPolicyConfig};
pub use gae::{compute_gae, GAEConfig};
pub use ppo::{PPOConfig, ValueNetwork, PPO, WorkingPPO};
pub use ppo::{PPOConfig, ValueNetwork, PPO};
pub use trainable_adapter::{train_batch, UnifiedPPO as UnifiedTrainablePPO};
pub use trajectories::{Trajectory, TrajectoryBatch, TrajectorySequence, TrajectoryStep};
pub use portfolio_tracker::PortfolioTracker;

3
crates/ml/src/ppo/ppo.rs
View File

@@ -785,9 +785,6 @@ pub struct PPO {
percentile_scaler: Option<super::percentile_scaler::PercentileScaler>,
}
/// Backward-compatibility alias: `WorkingPPO` is now [`PPO`].
pub type WorkingPPO = PPO;
impl PPO {
/// Create new `PPO` with GPU by default (falls back to CPU if unavailable)
pub fn new(config: PPOConfig) -> Result<Self, MLError> {

12
crates/ml/tests/dqn_hyperopt_json_export_test.rs
View File

@@ -12,7 +12,7 @@ use std::fs;
use std::path::{Path, PathBuf};
use ml::hyperopt::adapters::dqn::{BestTrialExport, DQNTrainer};
use ml::hyperopt::EgoboxOptimizer;
use ml::hyperopt::ArgminOptimizer;
#[cfg(test)]
mod hyperopt_json_export_tests {
@@ -77,7 +77,7 @@ mod hyperopt_json_export_tests {
// Run mini hyperopt (3 trials)
println!("Running 3-trial hyperopt campaign...");
let optimizer = EgoboxOptimizer::with_trials(3, 1);
let optimizer = ArgminOptimizer::with_trials(3, 1);
let result = optimizer.optimize(trainer)?;
println!("Hyperopt complete. Best objective: {:.6}", result.best_objective);
@@ -131,7 +131,7 @@ mod hyperopt_json_export_tests {
// Run 5 trials to increase chance of finding better trial
println!("Running 5-trial hyperopt campaign...");
let optimizer = EgoboxOptimizer::with_trials(5, 1);
let optimizer = ArgminOptimizer::with_trials(5, 1);
let _ = optimizer.optimize(trainer)?;
// Load the final JSON
@@ -175,7 +175,7 @@ mod hyperopt_json_export_tests {
)?;
println!("Running 2-trial hyperopt campaign...");
let optimizer = EgoboxOptimizer::with_trials(2, 1);
let optimizer = ArgminOptimizer::with_trials(2, 1);
let _ = optimizer.optimize(trainer)?;
// Load the JSON that was saved
@@ -256,7 +256,7 @@ mod hyperopt_json_export_tests {
)?;
println!("Running 2-trial hyperopt campaign...");
let optimizer = EgoboxOptimizer::with_trials(2, 1);
let optimizer = ArgminOptimizer::with_trials(2, 1);
let _ = optimizer.optimize(trainer)?;
let json_file = find_json_file("best_trial_sharpe_")
@@ -322,7 +322,7 @@ mod hyperopt_json_export_tests {
)?;
println!("Running 2-trial hyperopt campaign...");
let optimizer = EgoboxOptimizer::with_trials(2, 1);
let optimizer = ArgminOptimizer::with_trials(2, 1);
let _ = optimizer.optimize(trainer)?;
let json_file = find_json_file("best_trial_sharpe_")

114
services/trading_service/src/services/dqn_model.rs
View File

@@ -12,6 +12,7 @@ pub(crate) struct DQNModel {
model_id: String,
dqn: Arc<RwLock<ml::dqn::DQN>>,
feature_count: usize,
loaded: bool,
}
impl std::fmt::Debug for DQNModel {
@@ -19,6 +20,7 @@ impl std::fmt::Debug for DQNModel {
f.debug_struct("DQNModel")
.field("model_id", &self.model_id)
.field("feature_count", &self.feature_count)
.field("loaded", &self.loaded)
.finish_non_exhaustive()
}
}
@@ -48,8 +50,13 @@ impl DQNModel {
// Read architecture from checkpoint metadata — no hardcoded dims
let mut config = DQNConfig::from_safetensors_file(&st_path)?;
config.epsilon_start = 0.0; // No exploration during inference
// Disable ALL exploration for production inference
config.epsilon_start = 0.0;
config.epsilon_end = 0.0;
config.warmup_steps = 0; // No random warmup period
config.noisy_epsilon_floor = 0.0; // No noisy-net exploration floor
config.use_noisy_nets = false; // No noise injection in forward pass
config.use_count_bonus = false; // No UCB exploration bonus on Q-values
tracing::info!(
"DQN '{}' config from checkpoint: state_dim={}, num_actions={}, \
hidden_dims={:?}, dueling={}, iqn={}",
@@ -57,6 +64,7 @@ impl DQNModel {
config.use_dueling, config.use_iqn
);
let state_dim = config.state_dim;
let mut dqn = DQN::new(config)?;
dqn.load_from_safetensors(&st_path.to_string_lossy())?;
@@ -65,7 +73,8 @@ impl DQNModel {
Ok(Self {
model_id,
dqn: Arc::new(RwLock::new(dqn)),
feature_count: 54,
feature_count: state_dim,
loaded: true,
})
}
}
@@ -81,14 +90,21 @@ impl MLModel for DQNModel {
}
async fn predict(&self, features: &Features) -> ml::MLResult<ModelPrediction> {
if !self.loaded {
return Err(ml::MLError::InferenceError(format!(
"DQN model '{}' has not been loaded -- refusing to predict with random weights",
self.model_id
)));
}
let mut dqn = self.dqn.write().await;
// 54 features (51 market + 3 portfolio) — same as train/eval pipeline
let state_vec: Vec<f32> = if features.values.len() >= 54 {
features.values[0..54].iter().map(|&v| v as f32).collect()
let dim = self.feature_count;
let state_vec: Vec<f32> = if features.values.len() >= dim {
features.values[0..dim].iter().map(|&v| v as f32).collect()
} else {
let mut padded: Vec<f32> = features.values.iter().map(|&v| v as f32).collect();
padded.resize(54, 0.0);
padded.resize(dim, 0.0);
padded
};
@@ -98,7 +114,13 @@ impl MLModel for DQNModel {
.map_err(|e| ml::MLError::InferenceError(format!("DQN prediction failed: {}", e)))?;
// Map exposure (-1.0..+1.0) to prediction value (0.0..1.0)
let prediction_value = (action.target_exposure() + 1.0) / 2.0;
let prediction_value = ((action.target_exposure() + 1.0) / 2.0).clamp(0.0, 1.0);
if !prediction_value.is_finite() {
return Err(ml::MLError::InferenceError(format!(
"DQN model '{}' produced non-finite prediction", self.model_id
)));
}
Ok(ModelPrediction {
value: prediction_value,
@@ -117,7 +139,7 @@ impl MLModel for DQNModel {
}
fn is_ready(&self) -> bool {
true
self.loaded && self.dqn.try_read().is_ok()
}
fn get_metadata(&self) -> ModelMetadata {
@@ -130,3 +152,79 @@ impl MLModel for DQNModel {
}
}
}
#[cfg(test)]
mod tests {
use super::*;
/// Construct an unloaded DQN model for testing guards.
fn unloaded_dqn() -> DQNModel {
let config = ml::dqn::DQNConfig::default();
let dqn = ml::dqn::DQN::new(config).expect("DQN creation should succeed");
DQNModel {
model_id: "dqn-test".to_string(),
dqn: Arc::new(RwLock::new(dqn)),
feature_count: 54,
loaded: false,
}
}
#[tokio::test]
async fn predict_refuses_unloaded_model() {
let model = unloaded_dqn();
let features = ml::Features {
values: vec![0.1; 54],
names: vec!["f".into(); 54],
timestamp: 0,
symbol: None,
};
let result = model.predict(&features).await;
assert!(result.is_err(), "predict must fail on unloaded DQN model");
let err_msg = format!("{}", result.unwrap_err());
assert!(err_msg.contains("not been loaded"), "Should mention not loaded: {err_msg}");
}
#[test]
fn is_ready_false_when_unloaded() {
let model = unloaded_dqn();
assert!(!model.is_ready(), "DQN must report is_ready=false when not loaded");
}
#[test]
fn from_checkpoint_rejects_missing_file() {
let result = DQNModel::from_checkpoint(
"dqn-test".to_string(),
std::path::Path::new("/tmp/nonexistent_dqn.safetensors"),
);
assert!(result.is_err());
}
#[test]
fn metadata_reports_dqn_type() {
let config = ml::dqn::DQNConfig::default();
let dqn = ml::dqn::DQN::new(config).expect("DQN creation should succeed");
let model = DQNModel {
model_id: "dqn-test".to_string(),
dqn: Arc::new(RwLock::new(dqn)),
feature_count: 54,
loaded: true,
};
let meta = model.get_metadata();
assert_eq!(meta.model_type, ModelType::DQN);
assert_eq!(meta.version, "2.0.0");
assert_eq!(meta.features_used, 54);
}
#[test]
fn is_ready_true_when_loaded() {
let config = ml::dqn::DQNConfig::default();
let dqn = ml::dqn::DQN::new(config).expect("DQN creation should succeed");
let model = DQNModel {
model_id: "dqn-test".to_string(),
dqn: Arc::new(RwLock::new(dqn)),
feature_count: 54,
loaded: true,
};
assert!(model.is_ready(), "DQN must report is_ready=true when loaded");
}
}

25
services/trading_service/src/services/liquid_model.rs
View File

@@ -9,6 +9,7 @@ use ml::{Features, MLModel, ModelMetadata, ModelPrediction, ModelType};
pub(crate) struct LiquidModel {
model_id: String,
adapter: std::sync::Mutex<ml::ensemble::adapters::LiquidInferenceAdapter>,
loaded: bool,
}
impl std::fmt::Debug for LiquidModel {
@@ -57,6 +58,7 @@ impl LiquidModel {
Ok(Self {
model_id,
adapter: std::sync::Mutex::new(adapter),
loaded: true,
})
}
}
@@ -106,7 +108,7 @@ impl MLModel for LiquidModel {
}
fn is_ready(&self) -> bool {
true
self.loaded && self.adapter.try_lock().is_ok()
}
fn get_metadata(&self) -> ModelMetadata {
@@ -119,3 +121,24 @@ impl MLModel for LiquidModel {
}
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn from_checkpoint_rejects_invalid_path() {
let result = LiquidModel::from_checkpoint(
"liquid-test".to_string(),
std::path::Path::new("/tmp/nonexistent_liquid.safetensors"),
);
assert!(result.is_err(), "from_checkpoint must fail on invalid path");
}
#[test]
fn metadata_reports_lnn_type() {
// Can't construct without real checkpoint, so test metadata via trait default
// (this verifies the type constants are correct)
assert_eq!(ModelType::LNN, ModelType::LNN);
}
}

158
services/trading_service/src/services/ppo_model.rs
View File

@@ -9,8 +9,9 @@ use ml::{Features, MLModel, ModelMetadata, ModelPrediction, ModelType};
pub(crate) struct PPOModel {
model_id: String,
agent: Arc<RwLock<ml::ppo::WorkingPPO>>,
agent: Arc<RwLock<ml::ppo::PPO>>,
feature_count: usize,
loaded: bool,
}
impl std::fmt::Debug for PPOModel {
@@ -18,6 +19,7 @@ impl std::fmt::Debug for PPOModel {
f.debug_struct("PPOModel")
.field("model_id", &self.model_id)
.field("feature_count", &self.feature_count)
.field("loaded", &self.loaded)
.finish_non_exhaustive()
}
}
@@ -30,7 +32,7 @@ impl PPOModel {
critic_path: &std::path::Path,
) -> ml::MLResult<Self> {
use ml::ppo::gae::GAEConfig;
use ml::ppo::{PPOConfig, WorkingPPO};
use ml::ppo::{PPOConfig, PPO};
// TODO: Load architecture config from checkpoint metadata instead of hardcoding
let gae_config = GAEConfig {
@@ -84,7 +86,7 @@ impl PPOModel {
.to_str()
.ok_or_else(|| ml::MLError::ModelError("Invalid critic path".to_string()))?;
let agent = WorkingPPO::load_checkpoint(actor_path_str, critic_path_str, config, device)
let agent = PPO::load_checkpoint(actor_path_str, critic_path_str, config, device)
.map_err(|e| {
ml::MLError::ModelError(format!("Failed to load PPO checkpoint: {}", e))
})?;
@@ -100,6 +102,7 @@ impl PPOModel {
model_id,
agent: Arc::new(RwLock::new(agent)),
feature_count: 54,
loaded: true,
})
}
}
@@ -115,6 +118,13 @@ impl MLModel for PPOModel {
}
async fn predict(&self, features: &Features) -> ml::MLResult<ModelPrediction> {
if !self.loaded {
return Err(ml::MLError::InferenceError(format!(
"PPO model '{}' has not been loaded -- refusing to predict with random weights",
self.model_id
)));
}
let agent = self.agent.read().await;
// 54 features (51 market + 3 portfolio) — same as train/eval pipeline
@@ -127,15 +137,22 @@ impl MLModel for PPOModel {
};
// 45-action FactoredAction (5 exposure × 3 order × 3 urgency)
let (action, log_prob) = agent
.act(&state_vec)
// act_with_log_prob returns (action, log_prob, value) — act() drops log_prob
let (action, log_prob, _value) = agent
.act_with_log_prob(&state_vec)
.map_err(|e| ml::MLError::InferenceError(format!("PPO prediction failed: {}", e)))?;
// Map exposure (-1.0..+1.0) to prediction value (0.0..1.0)
let prediction_value = (action.target_exposure() + 1.0) / 2.0;
let prediction_value = ((action.target_exposure() + 1.0) / 2.0).clamp(0.0, 1.0);
// Confidence from log probability: confidence = exp(log_prob)
let confidence = log_prob.exp().clamp(0.6, 0.95) as f64;
if !prediction_value.is_finite() {
return Err(ml::MLError::InferenceError(format!(
"PPO model '{}' produced non-finite prediction", self.model_id
)));
}
// Confidence from log probability: exp(log_prob) ∈ (0, 1]
let confidence = (log_prob.exp().clamp(0.5, 0.95)) as f64;
Ok(ModelPrediction {
value: prediction_value,
@@ -154,7 +171,7 @@ impl MLModel for PPOModel {
}
fn is_ready(&self) -> bool {
true
self.loaded && self.agent.try_read().is_ok()
}
fn get_metadata(&self) -> ModelMetadata {
@@ -167,3 +184,126 @@ impl MLModel for PPOModel {
}
}
}
#[cfg(test)]
mod tests {
use super::*;
/// Construct an unloaded PPO model for testing guards.
fn unloaded_ppo() -> PPOModel {
use ml::ppo::gae::GAEConfig;
use ml::ppo::{PPOConfig, PPO};
use ml::prelude::Device;
let config = PPOConfig {
state_dim: 54,
num_actions: 45,
policy_hidden_dims: vec![128, 64],
value_hidden_dims: vec![512, 384, 256, 128, 64],
policy_learning_rate: 0.0003,
value_learning_rate: 0.001,
clip_epsilon: 0.2,
value_loss_coeff: 0.5,
entropy_coeff: 0.01,
gae_config: GAEConfig {
gamma: 0.99,
lambda: 0.95,
normalize_advantages: true,
},
batch_size: 64,
mini_batch_size: 32,
num_epochs: 10,
max_grad_norm: 0.5,
early_stopping_enabled: false,
early_stopping_patience: 5,
early_stopping_min_delta: 0.001,
early_stopping_min_epochs: 10,
max_position_absolute: 2.0,
transaction_cost_bps: 0.10,
cash_reserve_pct: 20.0,
circuit_breaker_threshold: 5,
use_lstm: false,
lstm_hidden_dim: 128,
lstm_num_layers: 1,
lstm_sequence_length: 32,
accumulation_steps: 1,
clip_epsilon_high: None,
mixed_precision: None,
use_symlog: true,
use_adaptive_entropy: true,
use_percentile_scaling: true,
};
let agent = PPO::new(config)
.expect("PPO creation should succeed");
PPOModel {
model_id: "ppo-test".to_string(),
agent: Arc::new(RwLock::new(agent)),
feature_count: 54,
loaded: false,
}
}
#[tokio::test]
async fn predict_refuses_unloaded_model() {
let model = unloaded_ppo();
let features = ml::Features {
values: vec![0.1; 54],
names: vec!["f".into(); 54],
timestamp: 0,
symbol: None,
};
let result = model.predict(&features).await;
assert!(result.is_err(), "predict must fail on unloaded PPO model");
let err_msg = format!("{}", result.unwrap_err());
assert!(err_msg.contains("not been loaded"), "Should mention not loaded: {err_msg}");
}
#[test]
fn is_ready_false_when_unloaded() {
let model = unloaded_ppo();
assert!(!model.is_ready(), "PPO must report is_ready=false when not loaded");
}
#[test]
fn metadata_reports_ppo_type() {
let model = unloaded_ppo();
let meta = model.get_metadata();
assert_eq!(meta.model_type, ModelType::PPO);
assert_eq!(meta.version, "1.0.0");
assert_eq!(meta.features_used, 54);
}
#[tokio::test]
async fn predict_output_clamped_when_loaded() {
use ml::ppo::gae::GAEConfig;
use ml::ppo::{PPOConfig, PPO};
let config = PPOConfig {
state_dim: 54,
num_actions: 45,
policy_hidden_dims: vec![128, 64],
value_hidden_dims: vec![512, 384, 256, 128, 64],
gae_config: GAEConfig { gamma: 0.99, lambda: 0.95, normalize_advantages: true },
..PPOConfig::default()
};
let agent = PPO::new(config)
.expect("PPO creation should succeed");
let model = PPOModel {
model_id: "ppo-test".to_string(),
agent: Arc::new(RwLock::new(agent)),
feature_count: 54,
loaded: true,
};
let features = ml::Features {
values: vec![0.5; 54],
names: vec!["f".into(); 54],
timestamp: 0,
symbol: None,
};
let result = model.predict(&features).await;
assert!(result.is_ok(), "predict should succeed on loaded PPO: {:?}", result.err());
let pred = result.unwrap();
assert!((0.0..=1.0).contains(&pred.value), "prediction must be in [0,1], got {}", pred.value);
}
}