Skip to content

visin

visin sends a training run to Visin: its epochs, test results, benchmarks, prediction frames and config. It can also read runs back for analysis, and download the datasets Visin hosts. The hosted Visin is at app.visin.eu.

pip install visin

import visin

with visin.init("unet baseline", project="road-seg") as run:
    for epoch in range(1, 21):
        train_loss, val_loss, val_iou = train_one_epoch()
        run.log_epoch(epoch, train={"loss": train_loss}, val={"loss": val_loss, "mean_iou": val_iou})

Why use it rather than requests

Visin's HTTP API is small, and you could post to it with requests. This package does the same, plus the things a long training job needs:

  • It never stops your training. Reports are sent from a background thread. A failure is logged and counted, never raised into your loop.
  • Retries are safe. Each epoch and test result carries an id of its own, so a retried request is recognised and never stored twice. Only requests that are safe to repeat get repeated.
  • It survives losing the network. When Visin stops answering, reports wait on disk and are sent once it answers again. On a machine with no route to Visin at all, run offline and send everything later with visin sync.
  • Crashed and cancelled runs keep their epochs. A run that crashes, or that the scheduler terminates, still sends what it has and is marked failed.
  • Metrics arrive as they are. NumPy and PyTorch scalars are converted. NaN becomes a gap in the chart instead of an error that loses the whole epoch.
  • The same script runs anywhere. With no server configured, every call does nothing.

Where to next