geowatch.tasks.fusion.utils module

geowatch.tasks.fusion.utils.millify(n)

geowatch.tasks.fusion.utils.load_model_from_package(package_path, with_header=False)

Loads a kitware-flavor torch package (requires a package_header exists)

Notes

• I don’t like that we need to know module_name and arch_name a-priori given a path to a package, I just want to be able to construct the model instance. The package header solves this.

geowatch.tasks.fusion.utils.load_model_header(package_path)

Only grabs header info from a packaged model.

class geowatch.tasks.fusion.utils.Lambda(lambda_)

Bases: Module

forward(x)

class geowatch.tasks.fusion.utils.DimensionDropout(dim, n_keep)

Bases: Module

forward(x)

class geowatch.tasks.fusion.utils.SinePositionalEncoding(dest_dim, dim_to_encode, size=4)

Bases: Module

Parameters:

• dest_dim (int) – feature dimension to concat to

• dim_to_encode (int) – dimension encoding is supposed to represent

• size (int) – number of different encodings for the dim_to_encode

Example

>>> from geowatch.tasks.fusion.utils import *  # NOQA
>>> dest_dim = 3
>>> dim_to_encode = 2
>>> size = 8
>>> self = SinePositionalEncoding(dest_dim, dim_to_encode, size=size)
>>> x = torch.rand(3, 5, 7, 11, 13)
>>> y = self(x)

forward(x)

geowatch.tasks.fusion.utils.model_json(model, max_depth=inf, depth=0)

import torchvision model = torchvision.models.resnet50() info = model_json(model, max_depth=1) print(ub.urepr(info, sort=0, nl=-1))