Fixed size crops

[markusnagel]

I added a transformer that performs a fixed size crop. It is similar to the already existing RandomFixedSizeCrop transformer. One key difference is that the user can specify the crop with a relative location instead of taking a random crop. This is useful for example for predicting on imagenet where often a center crop of the image is used or a center crop plus the 4 corner crops.

[vdumoulin]

@markusnagel Very sorry for the delay! I'll try to have a look at it by the end of the week.

[bartvm]

Sections in the docstring should have newlines.

[vdumoulin]

There seems to be quite a lot of overlap between FixedSizeCrop and RandomFixedSizeCrop. You could make FixedSizeCrop inherit from RandomFixedSizeCrop and factor cropping into _crop_batch and _crop_example methods which accept off_h and off_w arguments.

[aukejw]

The FixedSizeCrop transformer will afaik not benefit from using window_batch_bchw. The same offset is used for all images in the batch, so you can simply use

return source[:, :, off_h:off_h + windowed_height, off_w:off_w + windowed_width]

rather than first constructing arrays offsets_h and offsets_w of the same length as the batch, then invoking window_batch_bchw.

Therefore, the FixedSizeCrop transformer currently does not call _crop_batch. I haven't tested whether it actually makes a difference in terms of speed, but it seems like overkill to use window_batch_bchw here - though I agree that it would be intuitive to abstract the cropping functionality.