blended-tiling

This module adds support for splitting NCHW tensor inputs like images & activations into overlapping tiles of equal size, and then blending those overlapping tiles together after they have been altered. This module is also fully Autograd & JIT / TorchScript compatible.

This tiling solution is intended for situations where one wishes to render / generate outputs that are larger than what their computing device can support. Tiles can be separately rendered and periodically blended together to maintain tile feature coherence.

Setup:

Installation Requirements

• Python >= 3.6
• PyTorch >= 1.6

Installation via pip:

pip install blended-tiling

Dev / Manual install:

git clone https://github.com/progamergov/blended-tiling.git
cd blended-tiling
pip install -e .

# Notebook installs also require appending to environment variables
# import sys
# sys.path.append('/content/blended-tiling')

Documentation

TilingModule

The base blended tiling module.

blended_tiling.TilingModule(tile_size=(224, 224), tile_overlap=(0.25, 0.25), base_size=(512, 512))

Initialization Variables

• tile_size (int or tuple of int): The size of tiles to use. A single integer to use for both the height and width dimensions, or a list / tuple of dimensions with a shape of: [height, width]. The chosen tile sizes should be less than or equal to the sizes of the full NCHW tensor (base_size).
• tile_overlap (int or tuple of int): The amount of overlap to use when creating tiles. A single integer to use for both the height and width dimensions, or a list / tuple of dimensions with a shape of: [height, width]. The chosen overlap percentages should be in the range [0.0, 0.50] (0% - 50%).
• base_size (int or tuple of int): The size of the NCHW tensor being split into tiles. A single integer to use for both the height and width dimensions, or a list / tuple of dimensions with a shape of: [height, width].

Methods

num_tiles()

• Returns
  • num_tiles (int): The number of tiles that the full image shape is divided into based on specified parameters.

tiling_pattern()

• Returns:
  • pattern (list of int): The number of tiles per column and number of tiles per row, in the format of: [n_tiles_per_column, n_tiles_per_row].

split_into_tiles(x): Splits an NCHW image input into overlapping tiles, and then returns the tiles. The base_size parameter is automatically readjusted to match the input.

• Returns:
  • tiles (torch.Tensor): A set of tiles created from the input image.

get_tile_masks(channels=3, device=torch.device("cpu"), dtype=torch.float): Return a stack of NCHW masks corresponding to the tiles outputted by .split_into_tiles(x).

• Variables:
  • channels (int, optional): The number of channels to use for the masks. Default: 3
  • device (torch.device, optional): The desired device to create the masks on. Default: torch.device("cpu")
  • dtype (torch.dtype, optional): The desired dtype to create the masks with. Default: torch.float
• Returns:
  • masks (torch.Tensor): A set of tile masks stacked across the batch dimension.

rebuild(tiles, border=None, colors=None): Creates and returns the full image from a stack of NCHW tiles stacked across the batch dimension.

• Variables:
  • tiles (torch.Tensor): A set of tiles that may or not be masked, stacked across the batch dimension.
  • border (int, optional): Optionally add a border of a specified size to the edges of tiles in the full image for debugging and explainability. Set to None for no border.
  • colors (list of float, optional): A set of floats to use for the border color, if using borders. Default is set to red unless specified.
• Returns:
  • full_image (torch.Tensor): The full image made up of tiles merged together without any blending.

rebuild_with_masks(tiles, border=None, colors=None): Creates and returns the full image from a stack of NCHW tiles stacked across the batch dimension, using tile blend masks.

• Variables:
  • tiles (torch.Tensor): A set of tiles that may or not be masked, stacked across the batch dimension.
  • border (int, optional): Optionally add a border of a specified size to the edges of tiles in the full image for debugging and explainability. Set to None for no border.
  • colors (list of float, optional): A set of floats to use for the border color, if using borders. Default is set to red unless specified.
• Returns:
  • full_image (torch.Tensor): The full image made up of tiles blended together using masks.

forward(x): Takes a stack of tiles, combines them into the full image with blending masks, then splits the image back into tiles.

• Variables:
  • x (torch.Tensor): A set of tiles to blend the overlapping regions together of.
• Returns:
  • x (torch.Tensor): A set of tiles with overlapping regions blended together.

Supported Tensor Types

The TilingModule class has been tested with and is confirmed to work with the following PyTorch Tensor Data types / dtypes: torch.float32 / torch.float, torch.float64 / torch.double, torch.float16 / torch.half, & torch.bfloat16.

Usage

The TilingModule class is pretty easy to use.

from blended_tiling import TilingModule


full_size = [512, 512]
tile_size = [224, 224]
tile_overlap = [0.25, 0.25]  # 25% overlap on both H & W

tiling_module = TilingModule(
    tile_size=tile_size,
    tile_overlap=tile_overlap,
    base_size=full_size,
)

# Shape of tiles expected in forward pass
input_shape = [tiling_module.num_tiles(), 3] + tile_size

# Tiles are blended together and then split apart by default
blended_tiles = tiling_module(torch.ones(input_shape))

Tiles can be created and then merged back into the original tensor like this:

full_tensor = torch.ones(1, 3, 512, 512)

tiles = tiling_module.split_into_tiles(full_tensor)

full_tensor = tiling_module.rebuild_with_masks(tiles)

The tile boundaries can be viewed on the full tensor like this:

tiles = torch.ones(9, 3, 224, 224)
full_tensor = tiling_module.rebuild_with_masks(tiles, border=2)

And the number of tiles and tiling pattern can be obtained like this:

num_tiles = tiling_module.num_tiles()

tiling_pattern = tiling_module.tiling_pattern()
print("{}x{}".format(tiling_pattern[0], tiling_pattern[1]))

Custom Classes

It's also easy to modify the forward function of the tiling module:

from typing import Union, List, Tuple
from blended_tiling import TilingModule

class CustomTilingModule(TilingModule):
    def __init__(
        self,
        tile_size: Union[int, List[int], Tuple[int, int]] = [224, 224],
        tile_overlap: Union[float, List[float], Tuple[float, float]] = [0.25, 0.25],
        base_size: Union[int, List[int], Tuple[int, int]] = [512, 512],
    ) -> None:
        TilingModule.__init__(self, tile_size, tile_overlap, base_size)
        self.custom_module = torch.nn.Identity()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.rebuild_with_masks(x)
        x = self.custom_module(x) + 4.0
        return self._get_tiles_and_coords(x)[0]

Examples

To demonstrate the tile blending abilities of the TilingModule class, an example has been created below.

First we'll create a set of tiles & give them all unique colors for this example:

# Setup TilingModule instance
full_size = [768, 1014]
tile_size = [256, 448]
tile_overlap = [0.25, 0.25]
tiling_module = TilingModule(
    tile_size=tile_size,
    tile_overlap=tile_overlap,
    base_size=full_size,
)

# Create unique colors for tiles
tile_colors = [
    [0.5334, 0.0, 0.8459],
    [0.0, 1.0, 0.0],
    [0.0, 0.7071, 0.7071],
    [0.7071, 0.7071, 0.0],
    [1.0, 0.0, 0.0],
    [0.8459, 0.0, 0.5334],
    [0.7071, 0.0, 0.7071],
    [0.0, 0.8459, 0.5334],
    [0.5334, 0.8459, 0.0],
    [0.0, 0.5334, 0.8459],
    [0.0, 0.0, 1.0],
    [0.8459, 0.5334, 0.0],
]
tile_colors = torch.as_tensor(tile_colors).view(12, 3, 1, 1)

# Create tiles
tiles = torch.ones([tiling_module.num_tiles(), 3] + tile_size)

# Color tiles
tiles = tiles * tile_colors

Next we apply the blend masks to the tiles:

tiles = tiles * tiling_module.get_tile_masks()

We can now combine the masked tiles into the full image:

# Build full tiled image
output = tiling_module.rebuild(tiles)

We can also view the tile boundaries like so:

# Build full tiled image
output = tiling_module.rebuild(tiles, border=2, colors=[0,0,0])

We can view an animation of the tiles being added like this:

from torchvision.transforms import ToPILImage

tile_steps = [
    tiling_module.rebuild(tiles[: i + 1]) for i in range(tiles.shape[0])
]
tile_frames = [
    ToPILImage()(x[0])
    for x in [torch.zeros_like(tile_steps[0])] + tile_steps + [tile_steps[-1]]
]
tile_frames[0].save(
    "tiles.gif",
    format="GIF",
    append_images=tile_frames[1:],
    save_all=True,
    duration=700,
    loop=0,
)