f110_scripts.train package

Submodules

f110_scripts.train.eval_nn module

eval_nn.py – Print a summary table of best validation metrics for all trained models.

Reads TFEvents logs from data/models/lightning_logs/ and reports the best (minimum) val/loss (MSE) and val/mae achieved across all epochs for every model version found. The reported values correspond to the checkpoint that ModelCheckpoint would have saved.

Usage (from the repo root with the venv active):

python packages/f110_scripts/src/f110_scripts/train/eval_nn.py

f110_scripts.train.eval_nn.main() → None

Print a summary table of best validation metrics from TensorBoard logs.

f110_scripts.train.train_nn module

Advanced training script for F1Tenth LiDAR DNN models using PyTorch Lightning. Supports YAML configuration, advanced dataloading, and automatic tuning.

class f110_scripts.train.train_nn.LidarDataModule(config: dict[str, Any])

Bases: LightningDataModule

Lightning DataModule managing dataset splitting and data loader creation.

setup(stage: str | None = None) → None

Loads and splits the dataset.

Parameters:

stage – Current stage (fit, test, etc.).

train_dataloader() → DataLoader

Creates the training data loader.

val_dataloader() → DataLoader

Creates the validation data loader.

class f110_scripts.train.train_nn.LidarDataset(data_path: str, target_col: str)

Bases: Dataset

Advanced LiDAR Dataset with memory-resident caching.

x

Normalized LiDAR scans.

Type:

np.ndarray

y

Target values (heading_error, left_wall_dist, or track_width).

Type:

np.ndarray

class f110_scripts.train.train_nn.LidarLightningModule(arch_id: int, **kwargs: Any)

Bases: LightningModule

LightningModule wrapping the F1Tenth LiDAR neural network architectures.

configure_optimizers() → dict[str, Any]

Configures the optimizer and LR scheduler.

forward(*args: Any, **kwargs: Any) → Tensor

Performs a forward pass.

Parameters:

• *args – Should contain the input LiDAR tensor at index 0.

• **kwargs – Additional keyword arguments.

Returns:

Network output tensor.

on_after_backward() → None

Logs gradient norms to monitor training stability.

on_train_batch_end(*args: Any, **kwargs: Any) → None

Logs system metrics such as GPU memory usage.

on_train_epoch_end() → None

Logs weight histograms for architecture debugging.

training_step(*args: Any, **kwargs: Any) → Tensor

Executes a training step.

Parameters:

• *args – Should contain (batch, batch_idx).

• **kwargs – Additional keyword arguments.

Returns:

Calculated loss for the step.

validation_step(*args: Any, **kwargs: Any) → Tensor

Executes a validation step.

Parameters:

• *args – Should contain (batch, batch_idx).

• **kwargs – Additional keyword arguments.

Returns:

Calculated loss for the step.

f110_scripts.train.train_nn.main() → None

Main entry point for the training script.

f110_scripts.train.train_nn.run_single_training(config: dict[str, Any], arch_id: int) → None

Run a single training session for a specific architecture.

Checkpoint behaviour is controlled by two YAML flags:

resume: true/false — whether to load model weights from the last

checkpoint. When false, training always starts from random initialisation and resume_weights_only is irrelevant.

resume_weights_only: true — when resume is also true, load only the model

weights and start the optimizer/scheduler from scratch (useful after a cosine-schedule run where the LR has decayed to ~0). When false (default), full Lightning restore: weights + optimizer + scheduler + epoch counter are all restored.

Decision table ────────────── resume=false, resume_weights_only=any → random weights, fresh optimizer resume=true, resume_weights_only=false → weights from ckpt, optimizer from ckpt resume=true, resume_weights_only=true → weights from ckpt, fresh optimizer

f110_scripts.train.train_rl module

Simple script to train an RL policy for cloud scheduling.

This script uses Stable Baselines3 to learn a binary scheduler policy using the f110_gym:f110-cloud-scheduler-v0 environment. The CLI exposes a few key hyperparameters; the training loop is intentionally minimal.

Usage example:

python scripts/train/train_rl.py         --map data/maps/F1/Oschersleben/Oschersleben_map         --waypoints data/maps/F1/Oschersleben/Oschersleben_centerline.tsv         --cloud-latency 10         --timesteps 1000000         --save-path data/models/cloud_scheduler.zip

The trained policy is saved in the provided --save-path.

f110_scripts.train.train_rl.main() → None

Entry point: parse args, build environment, train, and save the policy.

f110_scripts.train.train_rl.parse_args() → Namespace

Parse command-line arguments for RL training.

Module contents