source/losses.py

# Code is copied from the gaussian-splatting/utils/loss_utils.py

import torch
import torch.nn.functional as F
from torch.autograd import Variable
from math import exp

def l1_loss(network_output, gt, mean=True):
    return torch.abs((network_output - gt)).mean() if mean else torch.abs((network_output - gt))

def l2_loss(network_output, gt):
    return ((network_output - gt) ** 2).mean()

def gaussian(window_size, sigma):
    gauss = torch.Tensor([exp(-(x - window_size // 2) ** 2 / float(2 * sigma ** 2)) for x in range(window_size)])
    return gauss / gauss.sum()

def create_window(window_size, channel):
    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
    window = Variable(_2D_window.expand(channel, 1, window_size, window_size).contiguous())
    return window

def ssim(img1, img2, window_size=11, size_average=True, mask = None):
    channel = img1.size(-3)
    window = create_window(window_size, channel)

    if img1.is_cuda:
        window = window.cuda(img1.get_device())
    window = window.type_as(img1)

    return _ssim(img1, img2, window, window_size, channel, size_average, mask)

def _ssim(img1, img2, window, window_size, channel, size_average=True, mask = None):
    mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
    mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)

    mu1_sq = mu1.pow(2)
    mu2_sq = mu2.pow(2)
    mu1_mu2 = mu1 * mu2

    sigma1_sq = F.conv2d(img1 * img1, window, padding=window_size // 2, groups=channel) - mu1_sq
    sigma2_sq = F.conv2d(img2 * img2, window, padding=window_size // 2, groups=channel) - mu2_sq
    sigma12 = F.conv2d(img1 * img2, window, padding=window_size // 2, groups=channel) - mu1_mu2

    C1 = 0.01 ** 2
    C2 = 0.03 ** 2

    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))

    if mask is not None:
        ssim_map = ssim_map * mask

    if size_average:
        return ssim_map.mean()
    else:
        return ssim_map.mean(1).mean(1).mean(1)


def mse(img1, img2):
    return (((img1 - img2)) ** 2).view(img1.shape[0], -1).mean(1, keepdim=True)

def psnr(img1, img2):
    """
    Computes the Peak Signal-to-Noise Ratio (PSNR) between two single images. NOT BATCHED!
    Args:
        img1 (torch.Tensor): The first image tensor, with pixel values scaled between 0 and 1.
                             Shape should be (channels, height, width).
        img2 (torch.Tensor): The second image tensor with the same shape as img1, used for comparison.

    Returns:
        torch.Tensor: A scalar tensor containing the PSNR value in decibels (dB).
    """
    mse = (((img1 - img2)) ** 2).view(img1.shape[0], -1).mean(1, keepdim=True)
    return 20 * torch.log10(1.0 / torch.sqrt(mse))


def tv_loss(image):
    """
    Computes the total variation (TV) loss for an image of shape [3, H, W].

    Args:
        image (torch.Tensor): Input image of shape [3, H, W]

    Returns:
        torch.Tensor: Scalar value representing the total variation loss.
    """
    # Ensure the image has the correct dimensions
    assert image.ndim == 3 and image.shape[0] == 3, "Input must be of shape [3, H, W]"

    # Compute the difference between adjacent pixels in the x-direction (width)
    diff_x = torch.abs(image[:, :, 1:] - image[:, :, :-1])

    # Compute the difference between adjacent pixels in the y-direction (height)
    diff_y = torch.abs(image[:, 1:, :] - image[:, :-1, :])

    # Sum the total variation in both directions
    tv_loss_value = torch.mean(diff_x) + torch.mean(diff_y)

    return tv_loss_value