augmentation

Enhancement

A discussion in #614 revealed a good place for improvement - we should ensure that input image is continuous upon start of the augmentation pipeline. This could be implemented by adding image = np.ascontiguousarray(image) to image and mask targets.

A proposed place to add this call - somewhere at the beginning of A.Compose.__call__.

This transform takes a fraction of the end or the start of the audio and treats that part as padding. We can implement several modes:

• constant (zero)
• edge - pads with the edge values of array
• wrap
• reflect

The offset can be randomized, as long as the output has the specified length

Like https://github.com/Spijkervet/torchaudio-augmentations/blob/d044f9d020e12032ab9280acf5f34a337e72d212/torchaudio_augmentations/augmentations/random_resized_crop.py#L5

The idea is that one can have a chain of transforms, and some of them change the input length, but the final length should be fixed. That is where

Using the same concept used in keras segmentation example add pytorch segmentation example