load_raster#

deepinv.utils.load_raster(fname, patch=False, patch_start=(0, 0), transform=None, **kwargs)[source]#

Load a raster image and return patches as tensors using rasterio.

This function allows you to stream patches from large rasters e.g. satellite imagery, SAR etc. and supports all file formats supported by rasterio.

This function requires rasterio, and should not rely on external GDAL dependencies. Install it with pip install rasterio.

Parameters:

  • fname (str) – Path to the raster file, such as .geotiff, .tiff, .cos etc., or buffer.

  • patch (bool, int, tuple[int, int],) – Patch extraction mode. If False (default), return the entire image as a torch.Tensor of shape (C, H, W) where C are bands. If True, yield patches based on the raster’s internal block windows (if no block windows are available, raises error; if any block has a dimension of 1 (strip layout), raise warning). If int or (int, int), yield patches of the manually specified size h, w.

  • patch_start (tuple[int, int]) – h and w indices from which to start taking patches. Defaults to 0,0.

  • transform (Callable, None) – Optional transform applied to each patch.

Returns:

Either (where C is the band dimension) * a full image torch.Tensor of shape (C, H, W), if patch=False, or * an iterator of torch tensors over patches of shape (C, h, w), if patch=True or a size is specified, where h,w is the patch size.

Return type:

Tensor | Iterator[Tensor]

Examples:

>>> from deepinv.utils.io import load_raster, load_url
>>> file = load_url("https://download.osgeo.org/geotiff/samples/spot/chicago/SP27GTIF.TIF")
>>> x = load_raster(file, patch=False) # Load whole image
>>> x.shape
torch.Size([1, 929, 699])
>>> x = load_raster(file, patch=True) # Patch via internal block size
>>> next(x).shape
torch.Size([1, 11, 699])
>>> all_patches = list(x) # Load all patches into memory
>>> len(all_patches)
84
>>> from torch.utils.data import DataLoader
>>> dataloader = DataLoader(all_patches, batch_size=2) # You can use this for training
>>>
>>> x = load_raster(file, patch=128, patch_start=(200, 200)) # Patch via manual size, pick away from origin
>>> next(x).shape
torch.Size([1, 128, 128])

Examples using load_raster:#

Loading scientific images

Loading scientific images