eval/metrics.py

"""

eval/metrics.py



Unified evaluation script for cloud-removal methods on Sen2_MTC datasets.

Computes PSNR, SSIM, FID and LPIPS – same implementation used in the paper.



Usage

-----

# Evaluate every method on both datasets (prints a summary table):

    python metrics.py



# Evaluate one specific method / dataset:

    python metrics.py --dataset Sen2_MTC_New --method diffcr



# Evaluate an arbitrary pair of directories:

    python metrics.py --gt /path/to/GT --pred /path/to/Out



.. note on reproducibility::



    PSNR, FID and LPIPS are fully reproducible on CPU.

    **SSIM requires a CUDA-enabled PyTorch build** to match the exact paper

    values; the 3-D Gaussian kernel implementation uses `.cuda()` internally

    and its floating-point accumulation order differs on CPU, leading to

    slightly different SSIM numbers.  Install the correct torch wheel for

    your GPU from https://pytorch.org before running a full benchmark.



The script expects the following layout (created by migrate.py):



    visualization/

    ├── data/

    │   ├── Sen2_MTC_New/GT/   ← ground-truth images  ({id}.png)

    │   └── Sen2_MTC_Old/GT/

    └── results/

        ├── Sen2_MTC_New/{method}/   ← prediction images ({id}.png)

        └── Sen2_MTC_Old/{method}/

"""

from __future__ import annotations

import argparse
import os
import subprocess
import sys
from glob import glob

import cv2
import lpips
import numpy as np
import torch
from tqdm import tqdm

# ---------------------------------------------------------------------------
# Paths
# ---------------------------------------------------------------------------
# eval/ lives one level below the project root
ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))

DATASETS: list[str] = ["Sen2_MTC_Old", "Sen2_MTC_New"]

# Order matches the paper table (top → bottom)
METHODS: list[str] = [
    "mcgan",
    "pix2pix",
    "ae",
    "stnet",
    "dsen2cr",
    "stgan",
    "ctgan",
    "crtsnet",
    "pmaa",
    "uncrtaints",
    "ddpmcr",
    "diffcr",
]


# ---------------------------------------------------------------------------
# Image helpers
# ---------------------------------------------------------------------------


def _convert_input_type_range(img: np.ndarray) -> np.ndarray:
    """Convert image to float32 in [0, 1]."""
    img_type = img.dtype
    img = img.astype(np.float32)
    if img_type == np.uint8:
        img /= 255.0
    elif img_type != np.float32:
        raise TypeError(f"Unsupported dtype: {img_type}")
    return img


def _convert_output_type_range(img: np.ndarray, dst_type) -> np.ndarray:
    if dst_type == np.uint8:
        img = img.round()
    else:
        img /= 255.0
    return img.astype(dst_type)


def reorder_image(img: np.ndarray, input_order: str = "HWC") -> np.ndarray:
    if input_order not in ("HWC", "CHW"):
        raise ValueError(f"input_order must be 'HWC' or 'CHW', got '{input_order}'")
    if img.ndim == 2:
        img = img[..., None]
    if input_order == "CHW":
        img = img.transpose(1, 2, 0)
    return img


# ---------------------------------------------------------------------------
# PSNR
# ---------------------------------------------------------------------------


def calculate_psnr(

    img1: np.ndarray,

    img2: np.ndarray,

    crop_border: int,

    input_order: str = "HWC",

) -> float:
    """Peak Signal-to-Noise Ratio.



    Accepts uint8 [0, 255] or float32 [0, 1] images.

    """
    assert img1.shape == img2.shape, f"Shape mismatch: {img1.shape} vs {img2.shape}"
    img1 = reorder_image(img1, input_order).astype(np.float64)
    img2 = reorder_image(img2, input_order).astype(np.float64)

    if crop_border:
        img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...]
        img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...]

    mse = np.mean((img1 - img2) ** 2)
    if mse == 0:
        return float("inf")
    max_val = 1.0 if img1.max() <= 1.0 else 255.0
    return 20.0 * np.log10(max_val / np.sqrt(mse))


# ---------------------------------------------------------------------------
# SSIM – 3-D Gaussian kernel (paper implementation)
# ---------------------------------------------------------------------------


def _generate_3d_gaussian_kernel(device: torch.device) -> torch.nn.Conv3d:
    """Build the 11×11×11 separable Gaussian Conv3d used in the paper."""
    kernel_1d = cv2.getGaussianKernel(11, 1.5)  # (11, 1)
    window_2d = np.outer(kernel_1d, kernel_1d.T)  # (11, 11)
    kernel_3d = np.stack(
        [window_2d * k for k in kernel_1d],
        axis=0,  # (11, 11, 11)
    )
    conv3d = torch.nn.Conv3d(
        1,
        1,
        (11, 11, 11),
        stride=1,
        padding=(5, 5, 5),
        bias=False,
        padding_mode="replicate",
    )
    conv3d.weight.requires_grad_(False)
    conv3d.weight[0, 0] = torch.tensor(kernel_3d)
    return conv3d.to(device)


def _apply_3d_gaussian(img: torch.Tensor, conv3d: torch.nn.Conv3d) -> torch.Tensor:
    return conv3d(img.unsqueeze(0).unsqueeze(0)).squeeze(0).squeeze(0)


def _ssim_3d(

    img1: np.ndarray,

    img2: np.ndarray,

    max_value: float,

    device: torch.device,

) -> float:
    """3-D SSIM over all three channels simultaneously (paper metric)."""
    assert img1.ndim == 3 and img2.ndim == 3
    C1 = (0.01 * max_value) ** 2
    C2 = (0.03 * max_value) ** 2

    kernel = _generate_3d_gaussian_kernel(device)

    t1 = torch.tensor(img1.astype(np.float64)).float().to(device)
    t2 = torch.tensor(img2.astype(np.float64)).float().to(device)

    mu1 = _apply_3d_gaussian(t1, kernel)
    mu2 = _apply_3d_gaussian(t2, kernel)

    mu1_sq = mu1**2
    mu2_sq = mu2**2
    mu1_mu2 = mu1 * mu2

    sigma1_sq = _apply_3d_gaussian(t1**2, kernel) - mu1_sq
    sigma2_sq = _apply_3d_gaussian(t2**2, kernel) - mu2_sq
    sigma12 = _apply_3d_gaussian(t1 * t2, kernel) - mu1_mu2

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


def calculate_ssim(

    img1: np.ndarray,

    img2: np.ndarray,

    crop_border: int,

    input_order: str = "HWC",

    device: torch.device | None = None,

) -> float:
    """Structural Similarity using the 3-D Gaussian kernel (paper implementation).



    Requires CUDA by default; falls back to CPU if no GPU is available.

    """
    if device is None:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    assert img1.shape == img2.shape, f"Shape mismatch: {img1.shape} vs {img2.shape}"

    img1 = reorder_image(img1, input_order).astype(np.float64)
    img2 = reorder_image(img2, input_order).astype(np.float64)

    if crop_border:
        img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...]
        img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...]

    max_val = 1 if img1.max() <= 1.0 else 255

    with torch.no_grad():
        return _ssim_3d(img1, img2, max_val, device)


# ---------------------------------------------------------------------------
# FID
# ---------------------------------------------------------------------------


def calculate_fid(gt_dir: str, pred_dir: str) -> float:
    """Compute FID via the pytorch-fid command-line tool."""
    device = "cuda" if torch.cuda.is_available() else "cpu"
    num_workers = 0 if sys.platform == "win32" else 4

    result = subprocess.run(
        [
            sys.executable,
            "-m",
            "pytorch_fid",
            gt_dir,
            pred_dir,
            "--device",
            device,
            "--batch-size",
            "4",
            "--num-workers",
            str(num_workers),
        ],
        capture_output=True,
        text=True,
    )
    output = result.stdout + result.stderr
    for line in output.splitlines():
        line = line.strip()
        if "fid" in line.lower():
            try:
                return float(line.split()[-1])
            except ValueError:
                pass
    print(f"[WARN] Could not parse FID output:\n{output}", file=sys.stderr)
    return float("nan")


# ---------------------------------------------------------------------------
# LPIPS
# ---------------------------------------------------------------------------


def calculate_lpips(

    gt_dir: str,

    pred_dir: str,

    device: torch.device | None = None,

) -> float:
    """Mean LPIPS (AlexNet backbone) over matched image pairs."""
    if device is None:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    loss_fn = lpips.LPIPS(net="alex", verbose=False).to(device)

    gt_map = _stem_map(gt_dir)
    pred_map = _stem_map(pred_dir)
    keys = sorted(set(gt_map) & set(pred_map))

    if not keys:
        print(
            f"[WARN] No common files between {gt_dir} and {pred_dir}", file=sys.stderr
        )
        return float("nan")

    scores: list[float] = []
    for k in tqdm(keys, desc="LPIPS", leave=False):
        t1 = lpips.im2tensor(lpips.load_image(gt_map[k])).to(device)
        t2 = lpips.im2tensor(lpips.load_image(pred_map[k])).to(device)
        scores.append(loss_fn(t1, t2).item())

    return float(np.mean(scores))


# ---------------------------------------------------------------------------
# Helpers
# ---------------------------------------------------------------------------


def _stem_map(directory: str) -> dict[str, str]:
    """Return {stem: full_path} for every .png in *directory*."""
    return {
        os.path.splitext(os.path.basename(f))[0]: f
        for f in glob(os.path.join(directory, "*.png"))
    }


def evaluate_pair(

    gt_dir: str,

    pred_dir: str,

    desc: str = "",

    device: torch.device | None = None,

) -> dict | None:
    """Compute all four metrics for one (GT, prediction) directory pair.



    Returns a dict with keys: n, PSNR, SSIM, FID, LPIPS.

    Returns None if no common files are found.

    """
    if device is None:
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

    gt_map = _stem_map(gt_dir)
    pred_map = _stem_map(pred_dir)
    keys = sorted(set(gt_map) & set(pred_map))

    if not keys:
        print(
            f"[WARN] No common files – GT: {gt_dir!r}  Pred: {pred_dir!r}",
            file=sys.stderr,
        )
        return None

    psnr_list: list[float] = []
    ssim_list: list[float] = []

    for k in tqdm(keys, desc=desc or "PSNR/SSIM", leave=False):
        img_gt = cv2.imread(gt_map[k])
        img_pred = cv2.imread(pred_map[k])
        if img_gt is None or img_pred is None:
            print(f"[WARN] Could not read image for key '{k}'", file=sys.stderr)
            continue
        if img_gt.shape != img_pred.shape:
            print(
                f"[WARN] Shape mismatch for '{k}': {img_gt.shape} vs {img_pred.shape}",
                file=sys.stderr,
            )
            continue
        psnr_list.append(calculate_psnr(img_gt, img_pred, crop_border=0))
        ssim_list.append(calculate_ssim(img_gt, img_pred, crop_border=0, device=device))

    if not psnr_list:
        return None

    fid_score = calculate_fid(gt_dir, pred_dir)
    lpips_score = calculate_lpips(gt_dir, pred_dir, device=device)

    return {
        "n": len(psnr_list),
        "PSNR": float(np.mean(psnr_list)),
        "SSIM": float(np.mean(ssim_list)),
        "FID": fid_score,
        "LPIPS": lpips_score,
    }


# ---------------------------------------------------------------------------
# Pretty table
# ---------------------------------------------------------------------------


def _fmt(v: float | None, width: int, decimals: int) -> str:
    if v is None or (isinstance(v, float) and np.isnan(v)):
        return f"{'N/A':>{width}}"
    return f"{v:>{width}.{decimals}f}"


def print_table(

    all_results: dict[str, dict[str, dict]],

    datasets: list[str],

    methods: list[str],

) -> None:
    col = 38  # width of one dataset block
    sep = "-" * (16 + col * len(datasets))

    # Header
    print("\n" + "=" * len(sep))
    header = f"{'Method':<16}"
    for ds in datasets:
        label = ds.replace("Sen2_MTC_", "Sen2_MTC ")
        header += f"{'| ' + label:<{col}}"
    print(header)

    sub = f"{'':16}"
    for _ in datasets:
        sub += f"| {'PSNR':>7} {'SSIM':>6} {'FID':>9} {'LPIPS':>6} "
    print(sub)
    print(sep)

    for m in methods:
        row = f"{m:<16}"
        for ds in datasets:
            r = all_results.get(ds, {}).get(m)
            if r:
                row += (
                    f"| {_fmt(r['PSNR'], 7, 3)}"
                    f"  {_fmt(r['SSIM'], 6, 3)}"
                    f"  {_fmt(r['FID'], 9, 3)}"
                    f"  {_fmt(r['LPIPS'], 6, 3)} "
                )
            else:
                row += f"|{'SKIP':^{col - 2}} "
        print(row)

    print("=" * len(sep))


# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------


def _parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(
        description="Evaluate cloud-removal metrics (PSNR / SSIM / FID / LPIPS)"
    )
    p.add_argument(
        "--dataset",
        type=str,
        default=None,
        choices=DATASETS,
        help="Evaluate only this dataset (default: both)",
    )
    p.add_argument(
        "--method",
        type=str,
        default=None,
        help="Evaluate only this method (default: all)",
    )
    p.add_argument(
        "--gt",
        type=str,
        default=None,
        help="Ground-truth directory (use together with --pred for a custom pair)",
    )
    p.add_argument(
        "--pred",
        type=str,
        default=None,
        help="Prediction directory",
    )
    p.add_argument(
        "--no-fid",
        action="store_true",
        help="Skip FID computation (much faster, useful for quick checks)",
    )
    p.add_argument(
        "--no-lpips",
        action="store_true",
        help="Skip LPIPS computation",
    )
    return p.parse_args()


def main() -> None:
    args = _parse_args()
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Device: {device}")

    # ---- custom pair -------------------------------------------------------
    if args.gt and args.pred:
        print(f"GT  : {args.gt}")
        print(f"Pred: {args.pred}")
        res = evaluate_pair(args.gt, args.pred, desc="custom", device=device)
        if res:
            print(
                f"\nPSNR  = {res['PSNR']:.3f}\n"
                f"SSIM  = {res['SSIM']:.3f}\n"
                f"FID   = {res['FID']:.3f}\n"
                f"LPIPS = {res['LPIPS']:.3f}\n"
                f"(n = {res['n']})"
            )
        return

    # ---- standard evaluation loop ------------------------------------------
    datasets = [args.dataset] if args.dataset else DATASETS
    methods = [args.method] if args.method else METHODS

    all_results: dict[str, dict[str, dict]] = {ds: {} for ds in datasets}

    for ds in datasets:
        gt_dir = os.path.join(ROOT, "data", ds, "GT")
        if not os.path.isdir(gt_dir):
            print(f"[ERROR] GT directory not found: {gt_dir}", file=sys.stderr)
            continue

        for m in methods:
            pred_dir = os.path.join(ROOT, "results", ds, m)
            if not os.path.isdir(pred_dir):
                print(f"  SKIP  {ds}/{m}  (not found)")
                continue

            print(f"\n[{ds}] [{m}]")

            gt_map = _stem_map(gt_dir)
            pred_map = _stem_map(pred_dir)
            n_common = len(set(gt_map) & set(pred_map))
            print(
                f"  GT: {len(gt_map)} imgs  |  Pred: {len(pred_map)} imgs  |  Common: {n_common}"
            )

            # PSNR / SSIM
            psnr_list: list[float] = []
            ssim_list: list[float] = []
            keys = sorted(set(gt_map) & set(pred_map))

            for k in tqdm(keys, desc="PSNR/SSIM", leave=False):
                ig = cv2.imread(gt_map[k])
                ip = cv2.imread(pred_map[k])
                if ig is None or ip is None or ig.shape != ip.shape:
                    continue
                psnr_list.append(calculate_psnr(ig, ip, crop_border=0))
                ssim_list.append(calculate_ssim(ig, ip, crop_border=0, device=device))

            if not psnr_list:
                print("  [WARN] No valid image pairs found.")
                continue

            psnr_mean = float(np.mean(psnr_list))
            ssim_mean = float(np.mean(ssim_list))
            print(f"  PSNR = {psnr_mean:.3f}  |  SSIM = {ssim_mean:.3f}")

            # FID
            fid_score: float = float("nan")
            if not args.no_fid:
                print("  Computing FID ...", end=" ", flush=True)
                fid_score = calculate_fid(gt_dir, pred_dir)
                print(f"{fid_score:.3f}")

            # LPIPS
            lpips_score: float = float("nan")
            if not args.no_lpips:
                lpips_score = calculate_lpips(gt_dir, pred_dir, device=device)
                print(f"  LPIPS = {lpips_score:.3f}")

            all_results[ds][m] = {
                "n": len(psnr_list),
                "PSNR": psnr_mean,
                "SSIM": ssim_mean,
                "FID": fid_score,
                "LPIPS": lpips_score,
            }

    # ---- summary table -----------------------------------------------------
    if any(all_results[ds] for ds in datasets):
        print_table(all_results, datasets, methods)


if __name__ == "__main__":
    main()