vmem / modeling /metrics.py
Jiahua0's picture
Upload folder using huggingface_hub
ff47419 verified
from math import exp
import numpy as np
import torch
import torch.nn.functional as F
from torch.autograd import Variable
from torcheval.metrics import PeakSignalNoiseRatio
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):
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)
def _ssim(img1, img2, window, window_size, channel, size_average=True):
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 size_average:
return ssim_map.mean()
else:
return ssim_map.mean(1).mean(1).mean(1)
def calculate_score(pil_img1, pil_img2, lpips_loss_fn, device="cpu"):
"""
Calculate the PSNR and SSIM between two PIL images.
Args:
pil_img1 (PIL.Image): First image.
pil_img2 (PIL.Image): Second image.
win_size (int, optional): The side length of the sliding window used in SSIM.
Must be odd and <= the smallest spatial dimension.
If not provided, defaults to 7 or the largest odd number
that does not exceed the smallest image dimension.
Returns:
tuple: (psnr_value, ssim_value)
Raises:
ValueError: if the images have different shapes, are too small for SSIM, or unsupported dtype.
"""
# Convert PIL images to numpy arrays
img1 = np.array(pil_img1)
img2 = np.array(pil_img2)
# Ensure images have the same dimensions
if img1.shape != img2.shape:
raise ValueError("Images must have the same dimensions.")
# Squeeze singleton dimensions (e.g., single-channel)
if img1.ndim == 3 and img1.shape[-1] == 1:
img1 = img1.squeeze(axis=-1)
img2 = img2.squeeze(axis=-1)
# Determine data_range based on dtype
if img1.dtype == np.uint8:
data_range = 255.0
elif img1.dtype == np.uint16:
data_range = 65535.0
elif np.issubdtype(img1.dtype, np.floating):
data_range = 1.0
else:
raise ValueError("Unsupported dtype. Use uint8, uint16, or float.")
# # # Compute PSNR
# mse = np.mean((img1.astype(np.float32) - img2.astype(np.float32)) ** 2)
# psnr_value_ = 20 * np.log10(data_range / np.sqrt(mse))
img1_tensor = torch.from_numpy(img1)
img2_tensor = torch.from_numpy(img2)
psnr_metric = PeakSignalNoiseRatio(data_range=data_range)
psnr_metric.update(img2_tensor.to(torch.float32), img1_tensor.to(torch.float32))
psnr_value = psnr_metric.compute()
# psnr_metric.reset()
img1 = img1_tensor.permute(2, 0, 1).unsqueeze(0).to(device)
img2 = img2_tensor.permute(2, 0, 1).unsqueeze(0).to(device)
# resize the images to 64x64
img1 = (img1.to(torch.float32)/255.0) * 2.0 - 1.0
img2 = (img2.to(torch.float32)/255.0) * 2.0 - 1.0
img1 = F.interpolate(img1, size=(64, 64), mode='bilinear', align_corners=False)
img2 = F.interpolate(img2, size=(64, 64), mode='bilinear', align_corners=False)
lpips_value = lpips_loss_fn(img1, img2).item()
ssim_value = ssim(img1, img2).item()
return (psnr_value, lpips_value, ssim_value)