Hectiq Lab

Training

A complete training script: it reads a dataset, trains in stages, keeps a figure per epoch, and saves the model.

Setup

pip install "hectiqlab[cli]"
hlab login
hlab init hectiq-ai/demo

The script

import hectiqlab as hl


@hl.stage
def prepare(folder):
    ...  # load the data and build the model
    return loader, model


def main():
    folder = hl.datasets.download("interactions").result()
    loader, model = prepare(folder)

    with hl.stage("train"):
        for epoch in hl.each(range(10), "epoch"):
            for batch in hl.each(loader):
                loss = model.step(batch)
            hl.add_metric("loss", step=epoch, value=loss)

            figure = plot_confusion(model)
            figure.savefig("confusion.png")
            hl.artifacts.create("confusion.png", group="confusion", step=epoch)

    model.save("checkpoints/")
    hl.models.create("checkpoints/", name="two-tower")


with hl.run("Two-tower, 10 epochs", category="training", tags=["ENG-42"], config={"lr": 1e-3}):
    main()

In the web app, the run shows the stages prepare, then train with epoch-0 to epoch-9 inside it, each with a bar for its batches. It lists the dataset it read, which download records, the model it made, the loss, and the confusion figure of every epoch.

On this page