LavenderDataLoader - PyTorch Integration

Convert your iteration to a PyTorch DataLoader for seamless integration with PyTorch, by calling the .torch() method.

Parameter	Description	Default
prefetch_factor	Number of batches to prefetch.	None
pin_memory	Pin memory for faster GPU transfer.	False
pin_memory_device	Device to pin memory to.	""
in_order	Whether to iterate in order. If False, the order is not guaranteed but can be slightly faster.	True
poll_interval	How often to check for new samples. Smaller values can make the iteration faster but can lead to more cpu usage on the server side.	0.01

dataloader = LavenderDataLoader(
    dataset_id=dataset.id,
    shardsets=[shardset.id],
    batch_size=10,
    shuffle=True,
).torch(
    prefetch_factor=4,
    pin_memory=True,
    pin_memory_device="cuda:0",
    in_order=True,
)

for batch in dataloader:
    # Use in PyTorch training loop
    outputs = model(batch)
    loss = criterion(outputs, batch["labels"])
    # ...