souging/TRELLIS_TextTo3D

1import torch2import torch.nn.functional as F3from torch.autograd import Variable4from math import exp5from lpips import LPIPS6 7 8def smooth_l1_loss(pred, target, beta=1.0):9    diff = torch.abs(pred - target)10    loss = torch.where(diff < beta, 0.5 * diff ** 2 / beta, diff - 0.5 * beta)11    return loss.mean()12 13 14def l1_loss(network_output, gt):15    return torch.abs((network_output - gt)).mean()16 17 18def l2_loss(network_output, gt):19    return ((network_output - gt) ** 2).mean()20 21 22def gaussian(window_size, sigma):23    gauss = torch.Tensor([exp(-(x - window_size // 2) ** 2 / float(2 * sigma ** 2)) for x in range(window_size)])24    return gauss / gauss.sum()25 26 27def create_window(window_size, channel):28    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)29    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)30    window = Variable(_2D_window.expand(channel, 1, window_size, window_size).contiguous())31    return window32 33 34def psnr(img1, img2, max_val=1.0):35    mse = F.mse_loss(img1, img2)36    return 20 * torch.log10(max_val / torch.sqrt(mse))37 38 39def ssim(img1, img2, window_size=11, size_average=True):40    channel = img1.size(-3)41    window = create_window(window_size, channel)42 43    if img1.is_cuda:44        window = window.cuda(img1.get_device())45    window = window.type_as(img1)46 47    return _ssim(img1, img2, window, window_size, channel, size_average)48 49def _ssim(img1, img2, window, window_size, channel, size_average=True):50    mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)51    mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)52 53    mu1_sq = mu1.pow(2)54    mu2_sq = mu2.pow(2)55    mu1_mu2 = mu1 * mu256 57    sigma1_sq = F.conv2d(img1 * img1, window, padding=window_size // 2, groups=channel) - mu1_sq58    sigma2_sq = F.conv2d(img2 * img2, window, padding=window_size // 2, groups=channel) - mu2_sq59    sigma12 = F.conv2d(img1 * img2, window, padding=window_size // 2, groups=channel) - mu1_mu260 61    C1 = 0.01 ** 262    C2 = 0.03 ** 263 64    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))65 66    if size_average:67        return ssim_map.mean()68    else:69        return ssim_map.mean(1).mean(1).mean(1)70 71 72loss_fn_vgg = None73def lpips(img1, img2, value_range=(0, 1)):74    global loss_fn_vgg75    if loss_fn_vgg is None:76        loss_fn_vgg = LPIPS(net='vgg').cuda().eval()77    # normalize to [-1, 1]78    img1 = (img1 - value_range[0]) / (value_range[1] - value_range[0]) * 2 - 179    img2 = (img2 - value_range[0]) / (value_range[1] - value_range[0]) * 2 - 180    return loss_fn_vgg(img1, img2).mean()81 82 83def normal_angle(pred, gt):84    pred = pred * 2.0 - 1.085    gt = gt * 2.0 - 1.086    norms = pred.norm(dim=-1) * gt.norm(dim=-1)87    cos_sim = (pred * gt).sum(-1) / (norms + 1e-9)88    cos_sim = torch.clamp(cos_sim, -1.0, 1.0)89    ang = torch.rad2deg(torch.acos(cos_sim[norms > 1e-9])).mean()90    if ang.isnan():91        return -192    return ang93