Metrics Evaluation

Metrics are implemented as nn.Module classes under src.metrics. Each metric accepts complex analytic signals with shape (batch, n_rois, T).

• forward(ts_pred, ts_target) returns a differentiable scalar loss.
• evaluate(ts_pred, ts_target) returns one or more logging-ready metric values.

Example: evaluate a simulated batch

from src.metrics import (
    FCCorrelation,
    FCMSE,
    FCD,
    PhFCD,
    Metastability,
    PhaseFC,
    PowerSpectrumDistance,
    TemporalCorrelation,
    AutocorrelationDistance,
)

modules = [
    FCCorrelation(),
    FCMSE(),
    FCD(tr=0.72, fcd_win_sec=30.0, fcd_step_sec=2.0),
    PhFCD(),
    Metastability(),
    PhaseFC(),
    PowerSpectrumDistance(),
    TemporalCorrelation(),
    AutocorrelationDistance(),
]

metrics = {}
for module in modules:
    metrics.update(module.evaluate(sim_ts, real_ts))

print(metrics)

Metric groups

• FC: FCCorrelation, FCMSE
• Dynamics / phase: FCD, PhFCD, Metastability, PhaseFC
• Timeseries: PowerSpectrumDistance, TemporalCorrelation, AutocorrelationDistance
• Auxiliary training losses: L2Timeseries, AmplitudeLoss, OmegaLoss

FC, spectrum, and autocorrelation metrics operate on the real part of the analytic signal. Phase-based metrics derive phases with torch.angle(...).

Composite training loss

src.training.CompositeLoss wires the same modules into a weighted training objective:

from src.training import CompositeLoss

loss_fn = CompositeLoss(
    weights={
        "fc_correlation": 1.0,
        "phfcd": 1.0,
        "metastability": 1.0,
    },
    tr=0.72,
    fcd_win_sec=30.0,
    fcd_step_sec=2.0,
)

total_loss, components = loss_fn(sim_ts, real_ts)
print(total_loss)
print(components)

Loader-level evaluation

For end-to-end model evaluation on a DataLoader, use src.utils.evaluate_model_loader_metrics:

from src.utils import EVAL_METRIC_KEYS, evaluate_model_loader_metrics

metrics = evaluate_model_loader_metrics(model, val_loader, cfg, return_std=True)

for key in EVAL_METRIC_KEYS:
    print(key, metrics.get(key), metrics.get(f"{key}_std"))

The default report keys are:

• fc_correlation
• fc_mse
• temporal_correlation
• power_spectrum_distance
• autocorr_distance
• fcd_ks
• phfcd_ks
• phase_fc_correlation
• metastability_diff