fid.py

# code was heavily based on https://github.com/mseitzer/pytorch-fid
# This implementation is licensed under the Apache License 2.0.
# Copyright (c) mseitzer


import os
import fnmatch
import numpy as np
import cv2
import paddle
from PIL import Image
from cv2 import imread
from scipy import linalg
from .inception import InceptionV3
from paddle.utils.download import get_weights_path_from_url
from .builder import METRICS

try:
    from tqdm import tqdm
except:

    def tqdm(x):
        return x


""" based on https://github.com/mit-han-lab/gan-compression/blob/master/metric/fid_score.py
"""
"""
inceptionV3 pretrain model is convert from pytorch, pretrain_model url is https://paddle-gan-models.bj.bcebos.com/params_inceptionV3.tar.gz
"""
INCEPTIONV3_WEIGHT_URL = "https://paddlegan.bj.bcebos.com/InceptionV3.pdparams"


@METRICS.register()
class FID(paddle.metric.Metric):
    def __init__(self,
                 batch_size=1,
                 use_GPU=True,
                 dims=2048,
                 premodel_path=None,
                 model=None):
        self.batch_size = batch_size
        self.use_GPU = use_GPU
        self.dims = dims
        self.premodel_path = premodel_path
        if model is None:
            block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
            model = InceptionV3([block_idx], normalize_input=False)
        if premodel_path is None:
            premodel_path = get_weights_path_from_url(INCEPTIONV3_WEIGHT_URL)
        self.model = model
        param_dict = paddle.load(premodel_path)
        self.model.load_dict(param_dict)
        self.model.eval()
        self.reset()

    def reset(self):
        self.preds = []
        self.gts = []
        self.results = []

    def update(self, preds, gts):
        preds_inception, gts_inception = calculate_inception_val(
            preds, gts, self.batch_size, self.model, self.use_GPU, self.dims)
        self.preds.append(preds_inception)
        self.gts.append(gts_inception)

    def accumulate(self):
        self.preds = np.concatenate(self.preds, axis=0)
        self.gts = np.concatenate(self.gts, axis=0)
        value = calculate_fid_given_img(self.preds, self.gts)
        self.reset()
        return value

    def name(self):
        return 'FID'


def _calculate_frechet_distance(mu1, sigma1, mu2, sigma2, eps=1e-6):
    m1 = np.atleast_1d(mu1)
    m2 = 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

    t = sigma1.dot(sigma2)
    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 _get_activations_from_ims(img, model, batch_size, dims, use_gpu):
    n_batches = (len(img) + batch_size - 1) // batch_size
    n_used_img = len(img)

    pred_arr = np.empty((n_used_img, dims))

    for i in range(n_batches):
        start = i * batch_size
        end = start + batch_size
        if end > len(img):
            end = len(img)
        images = img[start:end]
        if images.shape[1] != 3:
            images = images.transpose((0, 3, 1, 2))

        images = paddle.to_tensor(images)
        pred = model(images)[0][0]
        pred_arr[start:end] = pred.reshape([end - start, -1]).cpu().numpy()
    return pred_arr


def _compute_statistic_of_img(act):
    mu = np.mean(act, axis=0)
    sigma = np.cov(act, rowvar=False)
    return mu, sigma


def calculate_inception_val(img_fake,
                            img_real,
                            batch_size,
                            model,
                            use_gpu=True,
                            dims=2048):
    act_fake = _get_activations_from_ims(img_fake, model, batch_size, dims,
                                         use_gpu)
    act_real = _get_activations_from_ims(img_real, model, batch_size, dims,
                                         use_gpu)
    return act_fake, act_real


def calculate_fid_given_img(act_fake, act_real):

    m1, s1 = _compute_statistic_of_img(act_fake)
    m2, s2 = _compute_statistic_of_img(act_real)
    fid_value = _calculate_frechet_distance(m1, s1, m2, s2)
    return fid_value


def _get_activations(files,
                     model,
                     batch_size,
                     dims,
                     use_gpu,
                     premodel_path,
                     style=None):
    if len(files) % batch_size != 0:
        print(('Warning: number of images is not a multiple of the '
               'batch size. Some samples are going to be ignored.'))
    if batch_size > len(files):
        print(('Warning: batch size is bigger than the datasets size. '
               'Setting batch size to datasets size'))
        batch_size = len(files)

    n_batches = len(files) // batch_size
    n_used_imgs = n_batches * batch_size

    pred_arr = np.empty((n_used_imgs, dims))
    for i in tqdm(range(n_batches)):
        start = i * batch_size
        end = start + batch_size

        # same as stargan-v2 official implementation: resize to 256 first, then resize to 299
        if style == 'stargan':
            img_list = []
            for f in files[start:end]:
                im = Image.open(str(f)).convert('RGB')
                if im.size[0] != 299:
                    im = im.resize((256, 256), 2)
                    im = im.resize((299, 299), 2)

                img_list.append(np.array(im).astype('float32'))

            images = np.array(img_list)
        else:
            images = np.array(
                [imread(str(f)).astype(np.float32) for f in files[start:end]])

        if len(images.shape) != 4:
            images = imread(str(files[start]))
            images = cv2.cvtColor(images, cv2.COLOR_BGR2GRAY)
            images = np.array([images.astype(np.float32)])

        images = images.transpose((0, 3, 1, 2))
        images /= 255

        # imagenet normalization
        if style == 'stargan':
            mean = np.array([0.485, 0.456, 0.406]).astype('float32')
            std = np.array([0.229, 0.224, 0.225]).astype('float32')
            images[:] = (images[:] - mean[:, None, None]) / std[:, None, None]

        if style == 'stargan':
            pred_arr[start:end] = inception_infer(images, premodel_path)
        else:
            with paddle.guard():
                images = paddle.to_tensor(images)
                param_dict, _ = paddle.load(premodel_path)
                model.set_dict(param_dict)
                model.eval()

                pred = model(images)[0][0].numpy()

                pred_arr[start:end] = pred.reshape(end - start, -1)

    return pred_arr


def inception_infer(x, model_path):
    exe = paddle.static.Executor()
    [inference_program, feed_target_names,
     fetch_targets] = paddle.static.load_inference_model(model_path, exe)
    results = exe.run(inference_program,
                      feed={feed_target_names[0]: x},
                      fetch_list=fetch_targets)
    return results[0]


def _calculate_activation_statistics(files,
                                     model,
                                     premodel_path,
                                     batch_size=50,
                                     dims=2048,
                                     use_gpu=False,
                                     style=None):
    act = _get_activations(files, model, batch_size, dims, use_gpu,
                           premodel_path, style)
    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,
                                use_gpu,
                                premodel_path,
                                style=None):
    if path.endswith('.npz'):
        f = np.load(path)
        m, s = f['mu'][:], f['sigma'][:]
        f.close()
    else:
        files = []
        for root, dirnames, filenames in os.walk(path):
            for filename in fnmatch.filter(
                    filenames, '*.jpg') or fnmatch.filter(filenames, '*.png'):
                files.append(os.path.join(root, filename))
        m, s = _calculate_activation_statistics(files, model, premodel_path,
                                                batch_size, dims, use_gpu,
                                                style)
    return m, s


def calculate_fid_given_paths(paths,
                              premodel_path,
                              batch_size,
                              use_gpu,
                              dims,
                              model=None,
                              style=None):
    assert os.path.exists(
        premodel_path
    ), 'pretrain_model path {} is not exists! Please download it first'.format(
        premodel_path)
    for p in paths:
        if not os.path.exists(p):
            raise RuntimeError('Invalid path: %s' % p)

    if model is None and style != 'stargan':
        with paddle.guard():
            block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
            model = InceptionV3([block_idx], class_dim=1008)

    m1, s1 = _compute_statistics_of_path(paths[0], model, batch_size, dims,
                                         use_gpu, premodel_path, style)
    m2, s2 = _compute_statistics_of_path(paths[1], model, batch_size, dims,
                                         use_gpu, premodel_path, style)

    fid_value = _calculate_frechet_distance(m1, s1, m2, s2)
    return fid_value