import warnings
import torch
from torch import Tensor
from torch.nn.functional import grid_sample, conv2d, interpolate, pad as torch_pad
from torch.jit.annotations import BroadcastingList2
from typing import Optional, Tuple, List
def _is_tensor_a_torch_image(x: Tensor) -> bool:
return x.ndim >= 2
def _assert_image_tensor(img):
if not _is_tensor_a_torch_image(img):
raise TypeError("Tensor is not a torch image.")
def _get_image_size(img: Tensor) -> List[int]:
# Returns (w, h) of tensor image
_assert_image_tensor(img)
return [img.shape[-1], img.shape[-2]]
def _get_image_num_channels(img: Tensor) -> int:
if img.ndim == 2:
return 1
elif img.ndim > 2:
return img.shape[-3]
raise TypeError("Input ndim should be 2 or more. Got {}".format(img.ndim))
def _max_value(dtype: torch.dtype) -> float:
# TODO: replace this method with torch.iinfo when it gets torchscript support.
# https://github.com/pytorch/pytorch/issues/41492
a = torch.tensor(2, dtype=dtype)
signed = 1 if torch.tensor(0, dtype=dtype).is_signed() else 0
bits = 1
max_value = torch.tensor(-signed, dtype=torch.long)
while True:
next_value = a.pow(bits - signed).sub(1)
if next_value > max_value:
max_value = next_value
bits *= 2
else:
break
return max_value.item()
def _assert_channels(img: Tensor, permitted: List[int]) -> None:
c = _get_image_num_channels(img)
if c not in permitted:
raise TypeError("Input image tensor permitted channel values are {}, but found {}".format(permitted, c))
def convert_image_dtype(image: torch.Tensor, dtype: torch.dtype = torch.float) -> torch.Tensor:
if image.dtype == dtype:
return image
if image.is_floating_point():
# TODO: replace with dtype.is_floating_point when torchscript supports it
if torch.tensor(0, dtype=dtype).is_floating_point():
return image.to(dtype)
# float to int
if (image.dtype == torch.float32 and dtype in (torch.int32, torch.int64)) or (
image.dtype == torch.float64 and dtype == torch.int64
):
msg = f"The cast from {image.dtype} to {dtype} cannot be performed safely."
raise RuntimeError(msg)
# https://github.com/pytorch/vision/pull/2078#issuecomment-612045321
# For data in the range 0-1, (float * 255).to(uint) is only 255
# when float is exactly 1.0.
# `max + 1 - epsilon` provides more evenly distributed mapping of
# ranges of floats to ints.
eps = 1e-3
max_val = _max_value(dtype)
result = image.mul(max_val + 1.0 - eps)
return result.to(dtype)
else:
input_max = _max_value(image.dtype)
# int to float
# TODO: replace with dtype.is_floating_point when torchscript supports it
if torch.tensor(0, dtype=dtype).is_floating_point():
image = image.to(dtype)
return image / input_max
output_max = _max_value(dtype)
# int to int
if input_max > output_max:
# factor should be forced to int for torch jit script
# otherwise factor is a float and image // factor can produce different results
factor = int((input_max + 1) // (output_max + 1))
image = image // factor
return image.to(dtype)
else:
# factor should be forced to int for torch jit script
# otherwise factor is a float and image * factor can produce different results
factor = int((output_max + 1) // (input_max + 1))
image = image.to(dtype)
return image * factor
def vflip(img: Tensor) -> Tensor:
_assert_image_tensor(img)
return img.flip(-2)
def hflip(img: Tensor) -> Tensor:
_assert_image_tensor(img)
return img.flip(-1)
def crop(img: Tensor, top: int, left: int, height: int, width: int) -> Tensor:
_assert_image_tensor(img)
return img[..., top:top + height, left:left + width]
def rgb_to_grayscale(img: Tensor, num_output_channels: int = 1) -> Tensor:
if img.ndim < 3:
raise TypeError("Input image tensor should have at least 3 dimensions, but found {}".format(img.ndim))
_assert_channels(img, [3])
if num_output_channels not in (1, 3):
raise ValueError('num_output_channels should be either 1 or 3')
r, g, b = img.unbind(dim=-3)
# This implementation closely follows the TF one:
# https://github.com/tensorflow/tensorflow/blob/v2.3.0/tensorflow/python/ops/image_ops_impl.py#L2105-L2138
l_img = (0.2989 * r + 0.587 * g + 0.114 * b).to(img.dtype)
l_img = l_img.unsqueeze(dim=-3)
if num_output_channels == 3:
return l_img.expand(img.shape)
return l_img
def adjust_brightness(img: Tensor, brightness_factor: float) -> Tensor:
if brightness_factor < 0:
raise ValueError('brightness_factor ({}) is not non-negative.'.format(brightness_factor))
_assert_image_tensor(img)
_assert_channels(img, [1, 3])
return _blend(img, torch.zeros_like(img), brightness_factor)
def adjust_contrast(img: Tensor, contrast_factor: float) -> Tensor:
if contrast_factor < 0:
raise ValueError('contrast_factor ({}) is not non-negative.'.format(contrast_factor))
_assert_image_tensor(img)
_assert_channels(img, [3])
dtype = img.dtype if torch.is_floating_point(img) else torch.float32
mean = torch.mean(rgb_to_grayscale(img).to(dtype), dim=(-3, -2, -1), keepdim=True)
return _blend(img, mean, contrast_factor)
def adjust_hue(img: Tensor, hue_factor: float) -> Tensor:
if not (-0.5 <= hue_factor <= 0.5):
raise ValueError('hue_factor ({}) is not in [-0.5, 0.5].'.format(hue_factor))
if not (isinstance(img, torch.Tensor)):
raise TypeError('Input img should be Tensor image')
_assert_image_tensor(img)
_assert_channels(img, [1, 3])
if _get_image_num_channels(img) == 1: # Match PIL behaviour
return img
orig_dtype = img.dtype
if img.dtype == torch.uint8:
img = img.to(dtype=torch.float32) / 255.0
img = _rgb2hsv(img)
h, s, v = img.unbind(dim=-3)
h = (h + hue_factor) % 1.0
img = torch.stack((h, s, v), dim=-3)
img_hue_adj = _hsv2rgb(img)
if orig_dtype == torch.uint8:
img_hue_adj = (img_hue_adj * 255.0).to(dtype=orig_dtype)
return img_hue_adj
def adjust_saturation(img: Tensor, saturation_factor: float) -> Tensor:
if saturation_factor < 0:
raise ValueError('saturation_factor ({}) is not non-negative.'.format(saturation_factor))
_assert_image_tensor(img)
_assert_channels(img, [3])
return _blend(img, rgb_to_grayscale(img), saturation_factor)
def adjust_gamma(img: Tensor, gamma: float, gain: float = 1) -> Tensor:
if not isinstance(img, torch.Tensor):
raise TypeError('Input img should be a Tensor.')
_assert_channels(img, [1, 3])
if gamma < 0:
raise ValueError('Gamma should be a non-negative real number')
result = img
dtype = img.dtype
if not torch.is_floating_point(img):
result = convert_image_dtype(result, torch.float32)
result = (gain * result ** gamma).clamp(0, 1)
result = convert_image_dtype(result, dtype)
result = result.to(dtype)
return result
def center_crop(img: Tensor, output_size: BroadcastingList2[int]) -> Tensor:
"""DEPRECATED
"""
warnings.warn(
"This method is deprecated and will be removed in future releases. "
"Please, use ``F.center_crop`` instead."
)
_assert_image_tensor(img)
_, image_width, image_height = img.size()
crop_height, crop_width = output_size
# crop_top = int(round((image_height - crop_height) / 2.))
# Result can be different between python func and scripted func
# Temporary workaround:
crop_top = int((image_height - crop_height + 1) * 0.5)
# crop_left = int(round((image_width - crop_width) / 2.))
# Result can be different between python func and scripted func
# Temporary workaround:
crop_left = int((image_width - crop_width + 1) * 0.5)
return crop(img, crop_top, crop_left, crop_height, crop_width)
def five_crop(img: Tensor, size: BroadcastingList2[int]) -> List[Tensor]:
"""DEPRECATED
"""
warnings.warn(
"This method is deprecated and will be removed in future releases. "
"Please, use ``F.five_crop`` instead."
)
_assert_image_tensor(img)
assert len(size) == 2, "Please provide only two dimensions (h, w) for size."
_, image_width, image_height = img.size()
crop_height, crop_width = size
if crop_width > image_width or crop_height > image_height:
msg = "Requested crop size {} is bigger than input size {}"
raise ValueError(msg.format(size, (image_height, image_width)))
tl = crop(img, 0, 0, crop_width, crop_height)
tr = crop(img, image_width - crop_width, 0, image_width, crop_height)
bl = crop(img, 0, image_height - crop_height, crop_width, image_height)
br = crop(img, image_width - crop_width, image_height - crop_height, image_width, image_height)
center = center_crop(img, (crop_height, crop_width))
return [tl, tr, bl, br, center]
def ten_crop(img: Tensor, size: BroadcastingList2[int], vertical_flip: bool = False) -> List[Tensor]:
"""DEPRECATED
"""
warnings.warn(
"This method is deprecated and will be removed in future releases. "
"Please, use ``F.ten_crop`` instead."
)
_assert_image_tensor(img)
assert len(size) == 2, "Please provide only two dimensions (h, w) for size."
first_five = five_crop(img, size)
if vertical_flip:
img = vflip(img)
else:
img = hflip(img)
second_five = five_crop(img, size)
return first_five + second_five
def _blend(img1: Tensor, img2: Tensor, ratio: float) -> Tensor:
ratio = float(ratio)
bound = 1.0 if img1.is_floating_point() else 255.0
return (ratio * img1 + (1.0 - ratio) * img2).clamp(0, bound).to(img1.dtype)
def _rgb2hsv(img):
r, g, b = img.unbind(dim=-3)
# Implementation is based on https://github.com/python-pillow/Pillow/blob/4174d4267616897df3746d315d5a2d0f82c656ee/
# src/libImaging/Convert.c#L330
maxc = torch.max(img, dim=-3).values
minc = torch.min(img, dim=-3).values
# The algorithm erases S and H channel where `maxc = minc`. This avoids NaN
# from happening in the results, because
# + S channel has division by `maxc`, which is zero only if `maxc = minc`
# + H channel has division by `(maxc - minc)`.
#
# Instead of overwriting NaN afterwards, we just prevent it from occuring so
# we don't need to deal with it in case we save the NaN in a buffer in
# backprop, if it is ever supported, but it doesn't hurt to do so.
eqc = maxc == minc
cr = maxc - minc
# Since `eqc => cr = 0`, replacing denominator with 1 when `eqc` is fine.
ones = torch.ones_like(maxc)
s = cr / torch.where(eqc, ones, maxc)
# Note that `eqc => maxc = minc = r = g = b`. So the following calculation
# of `h` would reduce to `bc - gc + 2 + rc - bc + 4 + rc - bc = 6` so it
# would not matter what values `rc`, `gc`, and `bc` have here, and thus
# replacing denominator with 1 when `eqc` is fine.
cr_divisor = torch.where(eqc, ones, cr)
rc = (maxc - r) / cr_divisor
gc = (maxc - g) / cr_divisor
bc = (maxc - b) / cr_divisor
hr = (maxc == r) * (bc - gc)
Loading ...