Train a model

Train on Kaggle's free GPU

After AnalytiqAI trains in your browser, a panel under the results runs LightGBM, XGBoost and CatBoost, NVIDIA's pretrained KumoTabular model, and the AutoML tools FLAML and AutoGluon in a private notebook on your own Kaggle account. They are scored on the same test rows as the browser models, and every result lands in one table.

One-time setup

Everything happens on kaggle.com and on this site. There is nothing to install and no command to run.

  1. Create a free account on kaggle.com.
  2. Verify your phone number under Settings. Kaggle only gives GPU and internet access to verified accounts, and KumoTabular, FLAML and AutoGluon need internet so they can be installed.

Every run

  1. Train a model in the lab as usual.
  2. Under the results, in Go further on Kaggle's free GPU, press Sign in with Kaggle. Kaggle's own sign-in page opens in a small window: sign in, or create a free account there. The window closes by itself. Tick Keep me signed in to skip this next time. You can also paste an API token from Settings, API on kaggle.com under Use an API token instead.
  3. Pick the engines and how long each AutoML tool may search, tick the box confirming the upload, and press Start Kaggle run. Keep the tab open while it runs.

Most of the time goes on waiting for a GPU and installing the AutoML tools. The results appear in the same table as the browser models.

What is uploaded

  • Only the columns you trained on, plus a row number and whether each row is for training or testing.
  • The data travels compressed inside one private script notebook on your account. No dataset is created. By default the notebook is deleted when the run finishes. If the deletion fails, the page tells you.
  • You sign in on kaggle.com itself, so AnalytiqAI never sees your password. The sign-in gives this tab a token that goes straight to Kaggle's API; AnalytiqAI has no server, so no one else sees it. It is kept only in the open tab, or in this browser's storage if you tick Keep me signed in. Disconnect clears it.
  • Until you connect, AnalytiqAI makes no network requests. The request counter under the Train models button counts every Kaggle call once you do.

Scores from a real run

The churn sample on a Tesla T4, all six scored on the same 600 test rows. Bars start at 75%.

EngineAccuracy
FLAML77.3%
XGBoost77.0%
AnalytiqAI blend, in the browser76.8%
LightGBM76.5%
CatBoost76.5%
AutoGluon76.3%

On small, clean tables the browser models are often within a point. The larger libraries help most with messy data, many categories, or many more rows. With 10,000 rows or fewer the GPU barely speeds things up; what you gain is the extra libraries and AutoML.

Kaggle free GPU limit

Kaggle gives about 30 GPU hours a week for free. Each run has a hard time limit of at least 30 minutes, so a stuck run can't use up your week.

When a run fails

Kaggle did not accept this token
Tokens can expire or be revoked. Create a new one under Settings, API on kaggle.com and connect again.
Kaggle refused the notebook
Verify your phone number. GPU and internet stay off until you do.
Kaggle could not be reached
Check your connection. Some privacy extensions and company networks block api.kaggle.com.
FLAML or AutoGluon failed
Usually internet is off, so the install was blocked. The other engines still report.

Back to the lab