I have usecase where dlib based models need to used to perform face detection and face alignment. Is there a way to use these models to perform face detection and alignment as part of pytorch transforms?
In the common use case you would apply any transformation in the Dataset.__getitem__ method on each sample. Assuming your dlib transformation works on a single numpy array (or any other object you could transform into a PyTorch tensor) you could just apply it in the __getitem__, too.
Additionally, you could also write a custom transform class which expects to initialize internal states in the __init__ and apply the transformation in its __call__ method. Take a look at e.g torchvision.transforms.ToTensor for an example. This would allow you to add this custom transformation into e.g. torchvision.transforms.Compose, but make sure the expected dtype is used.