Utilities

fvdb.utils.fvdbCudaExtension(name: str, sources: Sequence[str], *args: Any, **kwargs: Any) → Extension

Utility function for creating pytorch extensions that depend on fvdb. You then have access to all fVDB’s internal headers to program with. Example usage:

from fvdb.utils import FVDBExtension

ext = FVDBExtension(
    name='my_extension',
    sources=['my_extension.cpp'],
    extra_compile_args={'cxx': ['-std=c++17']},
    libraries=['mylib'],
)

Parameters:

• name (str) – The name of the extension.

• sources (Sequence[str]) – The list of source files.

• args (list[Any]) – Other arguments to pass to torch.utils.cpp_extension.CppExtension().

• kwargs (dict) – Other keyword arguments to pass to torch.utils.cpp_extension.CppExtension().

Returns:

cpp_extension (setuptools.Extension) – to build a PyTorch C++ extension that depends on fVDB.

fvdb.utils.metrics

fvdb.utils.metrics.psnr(noisy_images: Tensor, ground_truth_images: Tensor, max_value: float = 1.0, reduction: Literal['none', 'mean', 'sum'] = 'mean') → Tensor

Compute the Peak-Signal-to-Noise-Ratio (PSNR) ratio between two batches of images.

Parameters:

• noisy_images (torch.Tensor) – A batch of noisy images of shape (B, C, H, W)

• ground_truth_images (torch.Tensor) – A batch of ground truth images of shape (B, C, H, W)

• max_value (float) – The maximum possible value images computed with this loss can have. Default is 1.0.

• reduction (Literal["none", "mean", "sum"]) – How to reduce over the batch dimension. "sum" and "mean" will add-up and average the losses across the batch respectively. "none" will return each loss as a separate entry in the tensor. Default is "mean".

Returns:

psnr (torch.Tensor) – The PSNR between the two images. If reduction is not “none”, the result will be reduced over the batch dimension (i.e. will be a single scalar), otherwise it will be a tensor of shape (B,).

fvdb.utils.metrics.ssim(img1: Tensor, img2: Tensor, padding: Literal['same', 'valid'] = 'same', train: bool = True) → Tensor

Compute the Structural Similarity Index (SSIM) between two images.

Parameters:

• img1 (torch.Tensor) – A batch of images of shape (B, C, H, W)

• img2 (torch.Tensor) – A batch of images of shape (B, C, H, W)

• padding (str) – The padding to use for the images ("same" or "valid"). Default is "same".

• train (bool) – Whether or not to compute the gradients through the SSIM loss. Default is True.

Returns:

ssim (torch.Tensor) – The average SSIM between each image over the batch.