Training and Evaluating the System

train_network(opt, ds_path, cp_path; kws...)

Starts the training process with the given configuration.

Arguments

• opt: Optimiser that is used for training.
• ds_path: Path to the dataset folder.
• cp_path: Path where checkpoints are being saved to.
• kws: Keyword arguments that customize the training process.

Keyword Arguments

• mps = 15: Number of message passing steps.
• layer_size = 128: Latent size of the hidden layers inside MLPs.
• hidden_layers = 2: Number of hidden layers inside MLPs.
• batchsize = 1: Size per batch (not implemented yet).
• epochs = 1: Number of epochs.
• steps = 10e6: Number of training steps.
• checkpoint = 10000: Number of steps after which checkpoints are created.
• norm_steps = 1000: Number of steps before training (accumulate normalization stats).
• max_norm_steps = 10.0f6: Number of steps after which no more normalization stats are collected.
• types_inflow = [4]: Node types whose values are prescribed by the dataset during solver rollouts.
• types_updated = [0, 5]: Array containing node types which are updated after each step.
• types_noisy = [0]: Array containing node types which noise is added to.
• noise_stddevs = [0.0f0]: Array containing the standard deviation of noise that is added to the target features.
• training_strategy = DerivativeTraining(): Methods used for training. See documentation.
• use_cuda = true: Whether a GPU is used for training or not (if available). Currently only CUDA GPUs are supported.
• gpu_device = CUDA.device() when CUDA is available, otherwise nothing: CUDA device used for training. See nvidia-smi for reference.
• cell_idxs = [0]: Reserved cell indices; currently unused by the training loop.
• use_valid = true: Shared checkpoint-selection setting; currently unused by the training loop.
• solver_valid = Tsit5(): Which solver should be used for validation during training.
• solver_valid_dt = nothing: If set, the solver for validation will use fixed timesteps.
• wandb_logger = nothing: If set, a Wandb WandbLogger will be used for logging the training.
• reset_valid = false: If set, the previous minimal validation loss will be overwritten.
• ad = :Zygote: Automatic-differentiation backend. Currently only :Zygote is supported.

Training Strategies

• DerivativeTraining
• SolverTraining
• SolverBatchTraining
• MultipleShooting

See CylinderFlow Example for reference.

Returns

• Trained network as a GraphNetwork struct.
• Minimum of validation loss (for hyperparameter tuning).

eval_network(ds_path, cp_path, out_path, solver = nothing;
             start, stop, dt = nothing, saves, mse_steps, kws...)

Starts the evaluation process with the given configuration.

Arguments

• ds_path: Path to the dataset folder.
• cp_path: Path from which checkpoints are loaded.
• out_path: Path where the results are saved.
• solver = nothing: Solver that is used for evaluating the system.

Keyword Arguments

• start: Start time of the simulation.
• stop: Stop time of the simulation.
• dt = nothing: If provided, changes the solver to use fixed step sizes.
• saves: Time steps where the solution is saved at.
• mse_steps: Time steps where the relative error is printed at.
• kws: Additional keyword arguments that customize evaluation. The network configuration has to be the same as during training.
• mps = 15: Number of message passing steps.
• layer_size = 128: Latent size of the hidden layers inside MLPs.
• hidden_layers = 2: Number of hidden layers inside MLPs.
• types_inflow = [4]: Node types whose values are prescribed by the dataset during the rollout.
• types_updated = [0, 5]: Array containing node types which are updated after each step.
• use_cuda = true: Whether a GPU is used for evaluation (if available). Currently only CUDA GPUs are supported.
• gpu_device = CUDA.device() when CUDA is available, otherwise nothing: CUDA device used for evaluation. See nvidia-smi for reference.
• use_valid = true: Whether the last checkpoint with the minimal validation loss should be used.