InstantMLDocs

Integrations overview

Track training runs from PyTorch Lightning, Hugging Face Transformers, Keras, and Weights & Biases with one-line InstantML adapters.

Open .md

Pick the adapter for the framework you already train with and wire it in with one line. Every adapter calls the same public run.log(...) methods you would call by hand, so you can always drop down to direct logging for custom loops.

Install the adapters

Install the SDK with the frameworks extra:

bash
python -m pip install "instantml[frameworks]"

Pick a framework adapter

PyTorch LightningPass `im.LightningLogger(...)` to `Trainer(logger=...)`.Hugging Face TransformersAdd `im.TransformersCallback()` to the `Trainer`.KerasPass `im.InstantMLKerasCallback(...)` to `model.fit(callbacks=...)`.Weights & BiasesSwap the import, shadow, or mirror — three migration modes.

Migrate from Weights & Biases

There are three ways to move, depending on how much risk you want to take:

ModeImportSource of truthUse when
Drop-inimport instantml.compat.wandb as wandbInstantMLYou are ready to switch and want a minimal code change.
Shadowim.init(..., shadow_wandb=True)BothYou want InstantML and W&B side by side during evaluation.
Mirrorimport instantml.wandb_mirror as wandbW&BYou want to keep W&B primary and mirror metrics into InstantML.

The Weights & Biases migration guide covers all three with runnable examples.

Import existing history

To bring past experiments along, the CLI imports run history from W&B, Neptune, and MLflow exports (instantml import wandb|neptune|mlflow) and syncs TensorBoard event files (instantml sync tensorboard). See Imports.

Next steps