InstantMLDocs

PyTorch Lightning

Log and compare PyTorch Lightning training runs in InstantML with a one-line logger that drops into the Lightning Trainer.

Open .md

Pass im.LightningLogger to the Lightning Trainer and everything you self.log(...) lands in InstantML. The logger implements the Lightning Logger interface and is rank-zero safe: in distributed training only the main process logs.

Install

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

Authenticate with `instantml login` or INSTANTML_API_KEY if you haven't already.

Wire the logger into the Trainer

This complete example trains a tiny regressor on synthetic data and logs train/loss every step:

python
import instantml as im
import lightning as L
import torch
from torch import nn
from torch.utils.data import DataLoader, TensorDataset


class TinyRegressor(L.LightningModule):
    def __init__(self, lr=1e-3):
        super().__init__()
        self.save_hyperparameters()
        self.layer = nn.Linear(8, 1)

    def training_step(self, batch, batch_idx):
        x, y = batch
        loss = nn.functional.mse_loss(self.layer(x), y)
        self.log("train/loss", loss)
        return loss

    def configure_optimizers(self):
        return torch.optim.Adam(self.parameters(), lr=self.hparams.lr)


dataset = TensorDataset(torch.randn(256, 8), torch.randn(256, 1))

logger = im.LightningLogger(
    project="lightning-demo",
    name="tiny-regressor",
    tags=["lightning"],
)

trainer = L.Trainer(max_epochs=5, logger=logger, log_every_n_steps=1)
trainer.fit(TinyRegressor(), DataLoader(dataset, batch_size=32))

Open the lightning-demo project at instantml.ai to watch the train/loss chart live. The run is finished automatically when Lightning calls finalize(...) at the end of training.

LightningLogger accepts the same keyword arguments as im.init(...) (name, config, tags, notes, api_key, ...) and forwards them when it creates the run on first use.

What gets logged

  • Metrics — everything you pass to self.log(...) inside your LightningModule is forwarded through log_metrics(...) at the current step.
  • Hyperparameters — log_hyperparams(...) records them as run config, so self.save_hyperparameters() in your module captures your config automatically.
  • Media — Lightning's log_image, log_audio, and log_video map to InstantML rich objects.

Reuse an existing run

Pass an already-created run when you want the logger to share a run with other code instead of creating its own:

python
run = im.init(project="lightning-demo", name="tiny-regressor")
logger = im.LightningLogger(run=run)
trainer = L.Trainer(logger=logger)

Next steps