Weights & Biases¶

We show how LaminDB can be integrated with W&B to track the training process and associate datasets & parameters with models.

# !pip install -q 'lamindb[jupyter,aws]' torch torchvision lightning wandb
!lamin init --storage ./lamin-mlops
!wandb login

import lamindb as ln
import wandb

ln.context.uid = "tULn4Va2yERp0000"
ln.context.track()

Define a model¶

Define a simple autoencoder as an example model using PyTorch Lightning.

from torch import optim, nn, utils
from torchvision.datasets import MNIST
from torchvision.transforms import ToTensor
import lightning


class LitAutoEncoder(lightning.LightningModule):
    def __init__(self, hidden_size, bottleneck_size):
        super().__init__()
        self.encoder = nn.Sequential(
            nn.Linear(28 * 28, hidden_size), 
            nn.ReLU(), 
            nn.Linear(hidden_size, bottleneck_size)
        )
        self.decoder = nn.Sequential(
            nn.Linear(bottleneck_size, hidden_size), 
            nn.ReLU(), 
            nn.Linear(hidden_size, 28 * 28)
        )
        # save hyper-parameters to self.hparams auto-logged by wandb
        self.save_hyperparameters()

    def training_step(self, batch, batch_idx):
        x, y = batch
        x = x.view(x.size(0), -1)
        z = self.encoder(x)
        x_hat = self.decoder(z)
        loss = nn.functional.mse_loss(x_hat, x)
        self.log("train_loss", loss)
        return loss

    def configure_optimizers(self):
        optimizer = optim.Adam(self.parameters(), lr=1e-3)
        return optimizer

Query & download the MNIST dataset¶

We saved the MNIST dataset in curation notebook and it now shows up in the artifact registry:

ln.Artifact.filter(type="dataset").df()

	uid	version	is_latest	description	key	suffix	type	size	hash	n_objects	n_observations	_hash_type	_accessor	visibility	_key_is_virtual	storage_id	transform_id	run_id	created_by_id	updated_at
id
1	egLnkTyXxrvp0f7D0000	None	True	None	testdata/mnist		dataset	54950048	amFx_vXqnUtJr0kmxxWK2Q	4	None	md5-d	None	1	True	1	1	1	1	2024-10-01 18:09:48.027847+00:00

You can also see it on lamin.ai if you connected your instance.

Let’s get the dataset:

artifact = ln.Artifact.get(key="testdata/mnist")
artifact

And download it to a local cache:

path = artifact.cache()
path

Create a pytorch-compatible dataset:

dataset = MNIST(path.as_posix(), transform=ToTensor())
dataset

Monitor training with wandb¶

Train our example model and track the training progress with wandb.

from lightning.pytorch.loggers import WandbLogger

MODEL_CONFIG = {
    "hidden_size": 32,
    "bottleneck_size": 16,
    "batch_size": 32
}

# create the data loader
train_loader = utils.data.DataLoader(dataset, batch_size=MODEL_CONFIG["batch_size"], shuffle=True)

# init model
autoencoder = LitAutoEncoder(MODEL_CONFIG["hidden_size"], MODEL_CONFIG["bottleneck_size"])

# initialize the logger
wandb_logger = WandbLogger(project="lamin")

# add batch size to the wandb config
wandb_logger.experiment.config["batch_size"] = MODEL_CONFIG["batch_size"]

from lightning.pytorch.callbacks import ModelCheckpoint

# store checkpoints to disk and upload to LaminDB after training
checkpoint_callback = ModelCheckpoint(
    dirpath=f"model_checkpoints/{wandb_logger.version}", 
    filename="last_epoch",
    save_top_k=1,
    monitor="train_loss"
)

# train model
trainer = lightning.Trainer(
    accelerator="cpu",
    limit_train_batches=3, 
    max_epochs=2,
    logger=wandb_logger,
    callbacks=[checkpoint_callback]
)
trainer.fit(model=autoencoder, train_dataloaders=train_loader)

Show code cell output Hide code cell output

GPU available: False, used: False

TPU available: False, using: 0 TPU cores

HPU available: False, using: 0 HPUs

  | Name    | Type       | Params | Mode 
-----------------------------------------------
| encoder | Sequential | 25.6 K | train
| decoder | Sequential | 26.4 K | train
-----------------------------------------------
1 K    Trainable params
       Non-trainable params
1 K    Total params
208     Total estimated model params size (MB)
       Modules in train mode
       Modules in eval mode

/opt/hostedtoolcache/Python/3.10.15/x64/lib/python3.10/site-packages/lightning/pytorch/trainer/connectors/data_connector.py:424: The 'train_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=3` in the `DataLoader` to improve performance.
/opt/hostedtoolcache/Python/3.10.15/x64/lib/python3.10/site-packages/lightning/pytorch/loops/fit_loop.py:298: The number of training batches (3) is smaller than the logging interval Trainer(log_every_n_steps=50). Set a lower value for log_every_n_steps if you want to see logs for the training epoch.

Training: |          | 0/? [00:00<?, ?it/s]

Training:   0%|          | 0/3 [00:00<?, ?it/s]

Epoch 0:   0%|          | 0/3 [00:00<?, ?it/s]

Epoch 0:  33%|███▎      | 1/3 [00:00<00:00, 42.10it/s]

Epoch 0:  33%|███▎      | 1/3 [00:00<00:00, 40.79it/s, v_num=uxj9]

Epoch 0:  67%|██████▋   | 2/3 [00:00<00:00, 62.90it/s, v_num=uxj9]

Epoch 0:  67%|██████▋   | 2/3 [00:00<00:00, 61.59it/s, v_num=uxj9]

Epoch 0: 100%|██████████| 3/3 [00:00<00:00, 75.99it/s, v_num=uxj9]

Epoch 0: 100%|██████████| 3/3 [00:00<00:00, 74.44it/s, v_num=uxj9]

Epoch 0: 100%|██████████| 3/3 [00:00<00:00, 72.30it/s, v_num=uxj9]

Epoch 0:   0%|          | 0/3 [00:00<?, ?it/s, v_num=uxj9]

Epoch 1:   0%|          | 0/3 [00:00<?, ?it/s, v_num=uxj9]

Epoch 1:  33%|███▎      | 1/3 [00:00<00:00, 89.31it/s, v_num=uxj9]

Epoch 1:  33%|███▎      | 1/3 [00:00<00:00, 83.77it/s, v_num=uxj9]

Epoch 1:  67%|██████▋   | 2/3 [00:00<00:00, 105.34it/s, v_num=uxj9]

Epoch 1:  67%|██████▋   | 2/3 [00:00<00:00, 100.88it/s, v_num=uxj9]

Epoch 1: 100%|██████████| 3/3 [00:00<00:00, 111.97it/s, v_num=uxj9]

Epoch 1: 100%|██████████| 3/3 [00:00<00:00, 108.86it/s, v_num=uxj9]

Epoch 1: 100%|██████████| 3/3 [00:00<00:00, 105.42it/s, v_num=uxj9]

`Trainer.fit` stopped: `max_epochs=2` reached.

Epoch 1: 100%|██████████| 3/3 [00:00<00:00, 96.03it/s, v_num=uxj9] 

wandb_logger.experiment.name

wandb_logger.version

wandb.finish()

See the training progress in the wandb UI:

Save model in LaminDB¶

# save checkpoint as a model in LaminDB
artifact = ln.Artifact(
    f"model_checkpoints/{wandb_logger.version}",
    key="testmodels/litautoencoder",  # is automatically versioned
    type="model",
).save()

# create a label with the wandb experiment name
experiment_label = ln.ULabel(
    name=wandb_logger.experiment.name, 
    description="wandb experiment name"
).save()

# annotate the model artifact
artifact.ulabels.add(experiment_label)

# define the associated model hyperparameters in ln.Param
for k, v in MODEL_CONFIG.items():
    ln.Param(name=k, dtype=type(v).__name__).save()
artifact.params.add_values(MODEL_CONFIG)

# describe the artifact
artifact.describe()

See the checkpoints:

If later on, you want to re-use the checkpoint, you can download it like so:

ln.Artifact.get(key='testmodels/litautoencoder').cache()

PosixUPath('/home/runner/work/lamin-mlops/lamin-mlops/docs/lamin-mlops/.lamindb/ZxuYcxrfpwuFuxkJ')

Or on the CLI:

lamin get artifact --key 'testmodels/litautoencoder'

# save notebook
# ln.context.finish()