Understanding the Output

After training completes, your project directory contains several new files. This page explains what each one is and how to use it.

Directory layout after training

my-project/
├── ode.py
├── config.py
├── train.py
├── data/
├── logs/                          # Training logs
│   ├── tensorboard/               # TensorBoard event files
│   │   └── <experiment>/<run>/
│   └── csv/                       # CSV-format metric logs
│       └── <experiment>/<run>/
└── models/                        # Saved artifacts
    └── <experiment>/<run>/
        ├── model.ckpt             # Full Lightning checkpoint
        └── predictions.pt         # Predicted fields + recovered parameters

The <experiment> name comes from EXPERIMENT_NAME in config.py, and <run> defaults to v0 (incremented automatically on subsequent runs).

Training logs

TensorBoard

What is TensorBoard?

TensorBoard is a dashboard that opens in your browser and shows how training is progressing. You'll see loss curves (lower is better) and parameter recovery metrics updating in real time.

Launch TensorBoard to visualize training curves:

uv run tensorboard --logdir logs/tensorboard

Key metrics logged every epoch:

Metric	Meaning
`loss/total`	Weighted sum of all constraint losses
`loss/residuals`	ODE/PDE residual on collocation points
`loss/ic`	Initial condition error
`loss/data`	Fit to observed data
`val/<param>_mse`	MSE between recovered and ground-truth parameter

The val/* metrics are only logged when a ValidationRegistry is provided (all built-in templates include one).

CSV logs

The same metrics are written as CSV files under logs/csv/. These are useful for post-processing or plotting with your own scripts.

Predictions

The predictions.pt file is a Python dictionary saved with torch.save. Load it with:

import torch

predictions = torch.load("models/<experiment>/<run>/predictions.pt")
batch, preds, trues = predictions

It contains three items:

Item	Type	Contents
`batch`	`tuple[Tensor, Tensor]`	`(x_data, y_data)`, the original training data
`preds`	`dict[str, Tensor]`	Predicted fields and parameters, keyed by name
`trues`	`dict[str, Tensor]` or `None`	Ground-truth values (if validation was configured)

Re-running predictions

To reload a trained checkpoint and produce predictions without retraining:

uv run train.py --predict

This skips training entirely and runs the prediction step on the saved checkpoint.

Checkpoints

The model.ckpt file is a standard PyTorch Lightning checkpoint. It contains the full model state (fields, parameters, optimizer, scheduler) and can be loaded for:

Resuming training: Lightning's Trainer handles this automatically
Transfer learning: load fields from a trained model into a new problem
Inspection: examine recovered parameter values directly

import torch

ckpt = torch.load("models/<experiment>/<run>/model.ckpt")
state_dict = ckpt["state_dict"]

# Recovered scalar parameter value
print(state_dict["problem.params.beta.value"])