evaluations/hed_ssim.py

import matplotlib.pyplot as plt
from tqdm import tqdm
from transformers import DPTImageProcessor, DPTForDepthEstimation
from PIL import Image
import os
import torch
import numpy as np
from torch.utils.data import DataLoader, Dataset
import sys
current_directory = os.getcwd()
sys.path.append(current_directory)
from autoregressive.test.metric import RMSE, SSIM
import torch.nn.functional as F
from condition.hed import HEDdetector
from torchmetrics.image import MultiScaleStructuralSimilarityIndexMeasure
# Define a dataset class for loading image and label pairs
class ImageDataset(Dataset):
    def __init__(self, img_dir, label_dir):
        self.img_dir = img_dir
        self.label_dir = label_dir
        self.images = os.listdir(img_dir)

    def __len__(self):
        return len(self.images)

    def __getitem__(self, idx):
        img_path = os.path.join(self.img_dir, self.images[idx])
        label_path = os.path.join(self.label_dir, self.images[idx])
        image = np.array(Image.open(img_path).convert("RGB"))
        label = np.array(Image.open(label_path))
        return torch.from_numpy(image), torch.from_numpy(label).permute(2, 0, 1)

model = HEDdetector().cuda().eval()

# Define the dataset and data loader
img_dir = 'sample/multigen/hed/visualization'
label_dir = 'sample/multigen/hed/annotations'
dataset = ImageDataset(img_dir, label_dir)
data_loader = DataLoader(dataset, batch_size=16, shuffle=False, num_workers=4)

model.eval()
ssim = MultiScaleStructuralSimilarityIndexMeasure(data_range=1.0).cuda()
ssim_score = []
with torch.no_grad():
    for images, labels in tqdm(data_loader):
        images = images.permute(0,3,1,2).cuda()
        outputs = model(images)
        predicted_hed = outputs.unsqueeze(1)
        labels = labels[:, 0:1, :, :].cuda()
        ssim_score.append(ssim((predicted_hed/255.0).clip(0,1), (labels/255.0).clip(0,1)))

print(f'ssim: {torch.stack(ssim_score).mean()}')