fid_score.py

"""Calculates the Frechet Inception Distance (FID) to evalulate GANs

The FID metric calculates the distance between two distributions of images.
Typically, we have summary statistics (mean & covariance matrix) of one
of these distributions, while the 2nd distribution is given by a GAN.

When run as a stand-alone program, it compares the distribution of
images that are stored as PNG/JPEG at a specified location with a
distribution given by summary statistics (in pickle format).

The FID is calculated by assuming that X_1 and X_2 are the activations of
the pool_3 layer of the inception net for generated samples and real world
samples respectively.

See --help to see further details.

Code apapted from https://github.com/bioinf-jku/TTUR to use PyTorch instead
of Tensorflow

Copyright 2018 Institute of Bioinformatics, JKU Linz

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
"""

import hashlib
import os
import pathlib
from argparse import ArgumentDefaultsHelpFormatter, ArgumentParser

import numpy as np
import torch
import torchvision.transforms as TF
from PIL import Image
from scipy import linalg
from torch.nn.functional import adaptive_avg_pool2d
from tqdm import tqdm

try:
    from pytorch_fid.inception import InceptionV3
except ImportError:
    raise ImportError("Please install pytorch-fid: pip install pytorch-fid")


parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter)
parser.add_argument("--batch-size", type=int, default=50, help="Batch size to use")
parser.add_argument(
    "--num-workers",
    type=int,
    help=(
        "Number of processes to use for data loading. Defaults to `min(8, num_cpus)`"
    ),
)
parser.add_argument(
    "--device", type=str, default=None, help="Device to use. Like cuda, cuda:0 or cpu"
)
parser.add_argument(
    "--dims",
    type=int,
    default=2048,
    choices=list(InceptionV3.BLOCK_INDEX_BY_DIM),
    help=(
        "Dimensionality of Inception features to use. By default, uses pool3 features"
    ),
)
parser.add_argument(
    "--save-stats",
    action="store_true",
    help=(
        "Generate an npz archive from a directory of samples. "
        "The first path is used as input and the second as output."
    ),
)
parser.add_argument(
    "path",
    type=str,
    nargs=2,
    help=("Paths to the generated images or to .npz statistic files"),
)

IMAGE_EXTENSIONS = {"bmp", "jpg", "jpeg", "pgm", "png", "ppm", "tif", "tiff", "webp"}
TEMP_DIR = pathlib.Path("temp")


class ImagePathDataset(torch.utils.data.Dataset):
    def __init__(self, files, transforms=None):
        self.files = files
        self.transforms = transforms

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

    def __getitem__(self, i):
        path = self.files[i]
        img = Image.open(path).convert("RGB")
        if self.transforms is not None:
            img = self.transforms(img)
        return img


def get_activations(
    files, model, batch_size=50, dims=2048, device="cpu", num_workers=1
):
    """Calculates the activations of the pool_3 layer for all images.

    Params:
    -- files       : List of image files paths
    -- model       : Instance of inception model
    -- batch_size  : Batch size of images for the model to process at once.
                     Make sure that the number of samples is a multiple of
                     the batch size, otherwise some samples are ignored. This
                     behavior is retained to match the original FID score
                     implementation.
    -- dims        : Dimensionality of features returned by Inception
    -- device      : Device to run calculations
    -- num_workers : Number of parallel dataloader workers

    Returns:
    -- A numpy array of dimension (num images, dims) that contains the
       activations of the given tensor when feeding inception with the
       query tensor.
    """
    model.eval()

    if batch_size > len(files):
        print(
            (
                "Warning: batch size is bigger than the data size. "
                "Setting batch size to data size"
            )
        )
        batch_size = len(files)
    # print(files)
    dataset = ImagePathDataset(files, transforms=TF.ToTensor())
    dataloader = torch.utils.data.DataLoader(
        dataset,
        batch_size=batch_size,
        shuffle=False,
        drop_last=False,
        num_workers=num_workers,
    )

    pred_arr = np.empty((len(files), dims))

    start_idx = 0

    for batch in tqdm(dataloader):
        batch = batch.to(device)

        with torch.no_grad():
            pred = model(batch)[0]

        # If model output is not scalar, apply global spatial average pooling.
        # This happens if you choose a dimensionality not equal 2048.
        if pred.size(2) != 1 or pred.size(3) != 1:
            pred = adaptive_avg_pool2d(pred, output_size=(1, 1))

        pred = pred.squeeze(3).squeeze(2).cpu().numpy()

        pred_arr[start_idx : start_idx + pred.shape[0]] = pred

        start_idx = start_idx + pred.shape[0]

    return pred_arr


def calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6):
    """Numpy implementation of the Frechet Distance.
    The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1)
    and X_2 ~ N(mu_2, C_2) is
            d^2 = ||mu_1 - mu_2||^2 + Tr(C_1 + C_2 - 2*sqrt(C_1*C_2)).

    Stable version by Dougal J. Sutherland.

    Params:
    -- mu1   : Numpy array containing the activations of a layer of the
               inception net (like returned by the function 'get_predictions')
               for generated samples.
    -- mu2   : The sample mean over activations, precalculated on an
               representative data set.
    -- sigma1: The covariance matrix over activations for generated samples.
    -- sigma2: The covariance matrix over activations, precalculated on an
               representative data set.

    Returns:
    --   : The Frechet Distance.
    """

    mu1 = np.atleast_1d(mu1)
    mu2 = np.atleast_1d(mu2)

    sigma1 = np.atleast_2d(sigma1)
    sigma2 = np.atleast_2d(sigma2)

    assert mu1.shape == mu2.shape, (
        "Training and test mean vectors have different lengths"
    )
    assert sigma1.shape == sigma2.shape, (
        "Training and test covariances have different dimensions"
    )

    diff = mu1 - mu2

    # Product might be almost singular
    covmean, _ = linalg.sqrtm(sigma1.dot(sigma2), disp=False)
    if not np.isfinite(covmean).all():
        msg = (
            "fid calculation produces singular product; "
            "adding %s to diagonal of cov estimates"
        ) % eps
        print(msg)
        offset = np.eye(sigma1.shape[0]) * eps
        covmean = linalg.sqrtm((sigma1 + offset).dot(sigma2 + offset))

    # Numerical error might give slight imaginary component
    if np.iscomplexobj(covmean):
        if not np.allclose(np.diagonal(covmean).imag, 0, atol=1e-3):
            m = np.max(np.abs(covmean.imag))
            raise ValueError("Imaginary component {}".format(m))
        covmean = covmean.real

    tr_covmean = np.trace(covmean)

    return diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean


def calculate_activation_statistics(
    files, model, batch_size=50, dims=2048, device="cpu", num_workers=1
):
    """Calculation of the statistics used by the FID.
    Params:
    -- files       : List of image files paths
    -- model       : Instance of inception model
    -- batch_size  : The images numpy array is split into batches with
                     batch size batch_size. A reasonable batch size
                     depends on the hardware.
    -- dims        : Dimensionality of features returned by Inception
    -- device      : Device to run calculations
    -- num_workers : Number of parallel dataloader workers

    Returns:
    -- mu    : The mean over samples of the activations of the pool_3 layer of
               the inception model.
    -- sigma : The covariance matrix of the activations of the pool_3 layer of
               the inception model.
    """
    act = get_activations(files, model, batch_size, dims, device, num_workers)
    mu = np.mean(act, axis=0)
    sigma = np.cov(act, rowvar=False)
    return mu, sigma


def compute_statistics_of_path(path, model, batch_size, dims, device, num_workers=1):
    # if path.endswith('.npz'):
    #     with np.load(path) as f:
    #         m, s = f['mu'][:], f['sigma'][:]
    # else:
    m, s = calculate_activation_statistics(
        path, model, batch_size, dims, device, num_workers
    )
    # else:
    #     path = pathlib.Path(path)
    #     files = sorted([file for ext in IMAGE_EXTENSIONS
    #                    for file in path.glob('*.{}'.format(ext))])
    #     m, s = calculate_activation_statistics(files, model, batch_size,
    #                                            dims, device, num_workers)

    return m, s


def _fid_cache_paths():
    tmp_dir = TEMP_DIR
    tmp_dir.mkdir(exist_ok=True)
    stats_path = tmp_dir / "fid_stats.npz"
    hash_path = tmp_dir / "fid_stats.hash"
    return stats_path, hash_path


def _load_fid_stats(stats_path):
    arr = np.load(stats_path)
    return arr["mu"], arr["sigma"]


def _save_fid_stats(stats_path, mu, sigma):
    np.savez_compressed(stats_path, mu=mu, sigma=sigma)


def calculate_fid_given_paths(paths, batch_size, device, dims, num_workers=1):
    """Calculates the FID of two paths, with caching for ground truth stats if the second path is a directory of images."""
    import pathlib

    if isinstance(paths[1], (str, pathlib.Path)) and pathlib.Path(paths[1]).is_dir():
        # Get all PNGs in the directory
        gt_images = list(pathlib.Path(paths[1]).glob("*.png"))
        stats_path, hash_path = _fid_cache_paths()
        if stats_path.exists():
            print(f"Using cached FID stats from {stats_path}")
            print("WARNING: Cache may be stale if ground truth images have changed.")
            m1, s1 = _load_fid_stats(stats_path)
        else:
            print("Computing FID stats for ground truth images...")
            block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
            model = InceptionV3([block_idx]).to(device)
            m1, s1 = calculate_activation_statistics(
                gt_images, model, batch_size, dims, device, num_workers
            )
            _save_fid_stats(stats_path, m1, s1)
        # m2, s2 for denoised images
        block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
        model = InceptionV3([block_idx]).to(device)
        m2, s2 = calculate_activation_statistics(
            paths[0], model, batch_size, dims, device, num_workers
        )
        fid_value = calculate_frechet_distance(m1, s1, m2, s2)
        return fid_value

    block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
    model = InceptionV3([block_idx]).to(device)
    m1, s1 = compute_statistics_of_path(
        paths[0], model, batch_size, dims, device, num_workers
    )
    print(paths[1])
    m2, s2 = compute_statistics_of_path(
        paths[1], model, batch_size, dims, device, num_workers
    )
    fid_value = calculate_frechet_distance(m1, s1, m2, s2)
    return fid_value


def save_fid_stats(paths, batch_size, device, dims, num_workers=1):
    """Calculates the FID of two paths"""
    if not os.path.exists(paths[0]):
        raise RuntimeError("Invalid path: %s" % paths[0])

    if os.path.exists(paths[1]):
        raise RuntimeError("Existing output file: %s" % paths[1])

    block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]

    model = InceptionV3([block_idx]).to(device)

    print(f"Saving statistics for {paths[0]}")

    m1, s1 = compute_statistics_of_path(
        paths[0], model, batch_size, dims, device, num_workers
    )

    np.savez_compressed(paths[1], mu=m1, sigma=s1)


def calculate_fid_with_cached_ground_truth(
    denoised_image_dirs,
    ground_truth_image_dirs,
    batch_size=50,
    device="cpu",
    dims=2048,
    num_workers=1,
):
    """
    Calculates the FID between denoised images and ground truth images, using cached stats for ground truth if possible.
    Args:
        denoised_image_dirs: list of denoised image paths
        ground_truth_image_dirs: list of ground truth image paths (or a directory)
        batch_size, device, dims, num_workers: same as calculate_fid_given_paths
    Returns:
        FID value
    """
    # If ground_truth_image_dirs is a directory, get all PNGs
    if isinstance(ground_truth_image_dirs, (str, pathlib.Path)):
        ground_truth_image_dirs = list(
            pathlib.Path(ground_truth_image_dirs).glob("*.png")
        )

    # Compute hash for cache
    def compute_file_hashes(file_list):
        hash_md5 = hashlib.md5()
        for fname in sorted(map(str, file_list)):
            try:
                stat = os.stat(fname)
                hash_md5.update(fname.encode())
                hash_md5.update(str(stat.st_mtime).encode())
            except Exception:
                continue
        return hash_md5.hexdigest()

    tmp_dir = TEMP_DIR
    tmp_dir.mkdir(exist_ok=True)
    stats_path = tmp_dir / "fid_stats.npz"
    hash_path = tmp_dir / "fid_stats.hash"
    # TODO: caching shouldn't be based on ground truth image dirs
    # since we can have multiple reconstructions of same ground truth
    current_hash = compute_file_hashes(ground_truth_image_dirs)
    cache_valid = False
    if stats_path.exists() and hash_path.exists():
        try:
            with open(hash_path, "r") as f:
                cached_hash = f.read().strip()
            if cached_hash == current_hash:
                cache_valid = True
        except Exception:
            pass
    # TODO: need more sophisticated caching for sweeps
    if cache_valid:
        print(f"Using cached FID stats from {stats_path}")
        arr = np.load(stats_path)
        mu, sigma = arr["mu"], arr["sigma"]
    else:
        print("Computing FID stats for ground truth images...")
        block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
        model = InceptionV3([block_idx]).to(device)
        mu, sigma = calculate_activation_statistics(
            ground_truth_image_dirs,
            model,
            batch_size=batch_size,
            dims=dims,
            device=device,
            num_workers=num_workers,
        )
        np.savez_compressed(stats_path, mu=mu, sigma=sigma)
        with open(hash_path, "w") as f:
            f.write(current_hash)
    # Compute stats for denoised images
    block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
    model = InceptionV3([block_idx]).to(device)
    mu2, sigma2 = calculate_activation_statistics(
        denoised_image_dirs,
        model,
        batch_size=batch_size,
        dims=dims,
        device=device,
        num_workers=num_workers,
    )
    fid_value = calculate_frechet_distance(mu, sigma, mu2, sigma2)
    return fid_value


def main():
    args = parser.parse_args()

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

    if args.num_workers is None:
        try:
            num_cpus = len(os.sched_getaffinity(0))
        except AttributeError:
            # os.sched_getaffinity is not available under Windows, use
            # os.cpu_count instead (which may not return the *available* number
            # of CPUs).
            num_cpus = os.cpu_count()

        num_workers = min(num_cpus, 8) if num_cpus is not None else 0
    else:
        num_workers = args.num_workers

    if args.save_stats:
        save_fid_stats(args.path, args.batch_size, device, args.dims, num_workers)
        return

    fid_value = calculate_fid_given_paths(
        args.path, args.batch_size, device, args.dims, num_workers
    )
    print("FID: ", fid_value)


if __name__ == "__main__":
    main()