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

refactor: remove batch_size CLI override from train_baseline_rl

Browse Source 
All VRAM-derived parameters (batch_size, gpu_n_episodes, buffer_size)
are auto-scaled — CLI overrides bypass this and cause inconsistent
behavior between local testing and production.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-04-02 14:33:00 +02:00
parent 5546a45bc3
commit 1321e11fb2
1 changed files with 5 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
crates/ml/examples/train_baseline_rl.rs
View File

@@ -133,9 +133,7 @@ struct Args {
#[arg(long, default_value_t = 50)]
epochs: usize,
/// Training batch size
#[arg(long, default_value_t = 0)]
batch_size: usize,
// batch_size always auto-scaled from VRAM (no CLI override).
/// Path to directory containing .dbn.zst files (env: FOXHUNT_DATA_DIR)
#[arg(long, env = "FOXHUNT_DATA_DIR")]
@@ -513,7 +511,7 @@ fn train_dqn_fold(
let mut hyperparams = DQNHyperparameters {
learning_rate: hp_f64(hp, "learning_rate").unwrap_or(args.learning_rate),
batch_size: hp_usize(hp, "batch_size").unwrap_or(args.batch_size),
batch_size: hp_usize(hp, "batch_size").unwrap_or(0), // 0 = auto from VRAM
gamma: hp_f64(hp, "gamma").unwrap_or(0.95),
epsilon_start,
epsilon_end: hp_f64(hp, "epsilon_end").unwrap_or(0.01),
@@ -576,7 +574,7 @@ fn train_dqn_fold(
hyperparams.epochs = args.epochs;
// Only override batch_size from CLI/hyperopt if explicitly non-zero.
// batch_size=0 is the auto-compute sentinel — let the constructor handle it.
let hp_batch = hp_usize(hp, "batch_size").unwrap_or(args.batch_size);
let hp_batch = hp_usize(hp, "batch_size").unwrap_or(0_usize /* auto from VRAM */);
if hp_batch > 0 {
hyperparams.batch_size = hp_batch;
}
@@ -704,7 +702,7 @@ fn train_ppo_fold(
learning_rate: hp_f64(hp, "learning_rate").unwrap_or(args.learning_rate),
actor_learning_rate: Some(hp_f64(hp, "policy_learning_rate").unwrap_or(args.learning_rate)),
critic_learning_rate: Some(hp_f64(hp, "value_learning_rate").unwrap_or(args.learning_rate * 3.0)),
batch_size: hp_usize(hp, "batch_size").unwrap_or(args.batch_size.max(64)),
batch_size: hp_usize(hp, "batch_size").unwrap_or(0_usize /* auto from VRAM */.max(64)),
gamma: hp_f64(hp, "gamma").unwrap_or(0.99),
clip_epsilon: hp_f64(hp, "clip_epsilon").unwrap_or(0.2) as f32,
vf_coef: hp_f64(hp, "value_loss_coeff").unwrap_or(0.5) as f32,
@@ -796,7 +794,7 @@ fn run_training(args: &Args) -> Result<Vec<RlTrainingResult>> {
info!(" Model(s): {}", args.model);
info!(" Symbol: {}", args.symbol);
info!(" Epochs: {}", args.epochs);
info!(" Batch size: {}", args.batch_size);
info!(" Batch size: {}", 0_usize /* auto from VRAM */);
info!(" Data dir: {}", args.data_dir.display());
info!(" Output dir: {}", args.output_dir.display());
info!(" Feature dim: {}", args.feature_dim);