Compose

class Compose(cfg=None)

Bases: object

Build a transform pipeline from config dicts and run it in order.

Constructed from a list of config dictionaries (the transform=[...] list used throughout pimm). At init each entry is materialized through the TRANSFORMS registry via TRANSFORMS.build – so every step’s type must name a registered transform – and build failures are re-raised with the offending step index and type. Calling the instance threads a single data_dict through every transform in sequence, where a failure is re-raised with the failing step’s index, class name, and the keys available at that point. This is the pipeline runner itself, not a registered transform; do not put Compose in a config list – pass the raw list of dicts and let datasets build it internally.

Parameters:

cfg (Sequence[dict], optional) – ordered list of transform config dicts, each carrying a type plus that transform’s keyword arguments. Defaults to an empty list (an identity pipeline).

Note

Most transforms mutate and return the same data_dict, but a few (e.g. GridSample in mode="test") return a different object such as a list of fragment dicts; the runner simply forwards whatever each step returns to the next.

Example

>>> import numpy as np
>>> from pimm.datasets.transform.base import Compose
>>> pipeline = Compose([
...     dict(type="NormalizeCoord", center=[500., 0., 0.], scale=500.),
...     dict(type="ToTensor"),
...     dict(type="Collect", keys=("coord",), feat_keys=("coord",)),
... ])
>>> data = {"coord": np.array([[0., 0., 0.], [1000., 0., 0.]], dtype="f4")}
>>> out = pipeline(data)  # normalize -> tensor -> collect, run in order
>>> sorted(out)
['coord', 'feat', 'offset']
>>> out["coord"]  # NormalizeCoord mapped [0,1000] -> [-1,1]; ToTensor made it a tensor
tensor([[-1.,  0.,  0.],
        [ 1.,  0.,  0.]])
>>> out["offset"]
tensor([2])