Dzmitry Pranchuk

pretrained DPN on pytorch https://github.com/rwightman/pytorch-dpn-pretrained

dpn68 180-180 val_accuracy  val loss  at the step ~40k learning rate was increased in attempting to push out training process from plateau

@asomoza Hi, I don't understand why inpainting region should have the same size as initial image. For example, if I have 2k X 4k image and want to inpaint only...

> For this to work and to match the positions of the original image and the "paste back" generated image, you need to use a mask that is the same...

Ok, we have following signature for the `apply_overlay` method: ``` def apply_overlay( self, mask: PIL.Image.Image, init_image: PIL.Image.Image, image: PIL.Image.Image, crop_coords: Optional[Tuple[int, int, int, int]] = None, ) -> PIL.Image.Image: ```...

When we inpaint with `padding_mask_crop` the generated area is not white area but smth like this:  From the official doc: > padding_mask_crop (int, optional, defaults to None) — The...

Yes, I think it could be a great feature for the inpainting pipeline. However, I believe that `image_processor.apply_overlay` lacks sufficient documentation, making it difficult to understand how it's supposed to...