load_raw#

deepinv.utils.load_raw(fname, dtype=torch.float32, **kwargs)[source]#

Loads any RAW image file as a torch tensor along with a metadata dictionary.

Reads any proprietary raw files supported by LibRaw <https://www.libraw.org> using rawpy (ie .arw, .cr2/.cr3, .nef, .raf, .orf, .rw2,``.pef``, .srw,``.dng``…). Returns the sensor image as a tensor of shape (1, 1, H, W) along with a dict of ISP metadata.

Note

Only \(2\times 2\) Bayer sensors are supported for now; non-Bayer CFAs (e.g. X-Trans) will raise an error.

Warning

Requires rawpy to be installed. Install it with pip install rawpy.

Parameters:

fname (str, pathlib.Path) – path to the raw file.
dtype (torch.dtype) – dtype of the returned tensor.

Returns:

tuple (y, meta) where y is a torch.Tensor of shape (1, 1, H, W) and meta is a dict of ISP metadata with keys cfa_pattern, cfa_colors, color_description, num_colors, black_level_per_channel, camera_white_level_per_channel, camera_whitebalance, daylight_whitebalance, color_matrix, rgb_xyz_matrix, tone_curve, raw_shape, visible_shape, crop_margins, pixel_aspect, flip and raw_type.

Return type:

tuple[Tensor, dict]