Quickstart

The examples below use synthetic descriptor values so they run without calling PaDEL. For real QSPR workflows, construct a dataset from SMILES (or use a bundled load_* helper) with the default backend='padel'.

Train a small model

from ecnet import ECNet
from ecnet.datasets import QSPRDatasetFromValues

desc_vals = [
    [0.0, 0.1, 0.2, 0.3],
    [0.1, 0.2, 0.3, 0.4],
    [0.2, 0.3, 0.4, 0.5],
    [0.3, 0.4, 0.5, 0.6],
]
target_vals = [[1.0], [2.0], [3.0], [4.0]]
dataset = QSPRDatasetFromValues(desc_vals, target_vals)

model = ECNet(input_dim=4, output_dim=1, hidden_dim=16, n_hidden=1)
train_loss, valid_loss = model.fit(
    dataset=dataset,
    epochs=10,
    batch_size=2,
    random_state=0,
)

Save and reload

ECNet.save writes an ecnet-state-v1 checkpoint. load_model also reads legacy full-module .pt pickles.

from ecnet.model import load_model

model.save("example_model.pt")
restored = load_model("example_model.pt")

Next steps

• API reference — autodoc for the frozen public surface

• API stability policy — compatibility and versioning policy

• Units and numeric conventions — cloud point (°C), kinematic viscosity (cSt), and scales

• Example notebooks under examples/ on GitHub (PaDEL/Java; run manually — not executed in CI; see CONTRIBUTING.md).