No Batch

Sometimes it may be preferable to operate without a batch dimension. For example, when creating a Dataset, the downstream Dataloader collates the data into batches. TorchSample supports this usecase by either supplying 0 to functions that expect a batch size, or by using the nobatch subfunction for functions that operate on tensors.

Without using this functionality, a Dataset may look like the following:

out = {}
image = torch.rand(3, 480, 640)
image = image[None]  # Add a singleton batch dimension
out["coords"] = ts.coord.randint(1, 4096, (640, 480))  # (1, 4096, 2)
out["rgb"] = ts.sample(out["coords"], image, mode="nearest")  # (1, 4096, 3)

# Remove the singleton dimensions since Dataloader will do the batching
out["coords"] = out["coords"][0]
out["rgb"] = out["rgb"][0]

Not bad, but its more terse and readable by using the nobatch feature:

out = {}
image = torch.rand(3, 480, 640)
out["coords"] = ts.coord.randint(0, 4096, (640, 480))  # (4096, 2)
out["rgb"] = ts.sample.nobatch(out["coords"], image, mode="nearest")  # (4096, 3)