Streamlining Training Process

The major rewrite is now complete 🎉!

The training process has been simplified and can now be run easily.

🔄 Training the model

A YAML configuration file is used for training parameters. For example, the following configuration is used for 2D Burger’s Equation:

device: "cuda"
learning_rate: 0.01
Nb: 1000
Nf: 5000

model:
  input_dim: 3
  hidden_layers: 4
  hidden_units: 20
  output_dim: 2
  epochs: 6000

You can train an example model using the following parameters by:

uv run main.py [--model pinn|pikan] [--config <config_path>] EXAMPLE

The default value of type and config are set to pinn and <default_example_path> respectively. For example, to run burger2d using PIKAN model:

uv run main.py --model pikan burger2d

🛠️ Refactored Examples

The examples are now simplified in the generation of input points. Instead of using multiple functions with similar parameters between them, a single generate_data() function is used.

✨ New Progress Indicator

The training progress is now shown via progress bar in terminal using the alive-progress library.

Epoch 30, Loss: 0.120250269771 |████████████████████████████████████████| 30/30 [100%] in 4.6s (4.53/s)

You can find our repository under gulfgap/gulfgap!