# --------------------------------------------------------------- # Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # This file has been modified from a file in the pytorch_fid library # which was released under the Apache License v2.0 License. # # Source: # https://github.com/mseitzer/pytorch-fid/blob/master/pytorch_fid/fid_score.py # # The license for the original version of this file can be # found in this directory (LICENSE_pytorch_fid). The modifications # to this file are subject to the same Apache License. # --------------------------------------------------------------- """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 os import pathlib from argparse import ArgumentParser, ArgumentDefaultsHelpFormatter from multiprocessing import cpu_count import numpy as np import torch from torch.utils.data import DataLoader import torchvision.transforms as TF from scipy import linalg from torch.nn.functional import adaptive_avg_pool2d from PIL import Image from thirdparty.fid.inception import InceptionV3 parser = ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter) parser.add_argument('--batch-size', type=int, default=50, help='Batch size to use') 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('path', type=str, nargs=2, help=('Paths to the generated images or ' 'to .npz statistic files')) class ImagesPathDataset(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', max_samples=None): """Calculates the activations of the pool_3 layer for all images. Params: -- files : List of image files paths or pytorch data loader -- 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 -- max_samples : Setting this value will stop activation when max_samples is reached 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() model.to(device) if isinstance(files, list): 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) ds = ImagesPathDataset(files, transforms=TF.ToTensor()) dl = DataLoader(ds, batch_size=batch_size, drop_last=False, num_workers=cpu_count()) else: dl = files pred_arr = [] total_processed = 0 #print('Starting to sample.') for batch in dl: # ignore labels if isinstance(batch, list): batch = batch[0] batch = batch.to(device) if batch.shape[1] == 1: # if image is gray scale batch = batch.repeat(1, 3, 1, 1) 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.append(pred) total_processed += pred.shape[0] if max_samples is not None and total_processed > max_samples: #print('Max Samples Reached.') break pred_arr = np.concatenate(pred_arr, axis=0) if max_samples is not None: pred_arr = pred_arr[:max_samples] 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', max_samples=None): """Calculation of the statistics used by the FID. Params: -- files : List of image files paths or pytorch data loader -- 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 -- max_samples : Setting this value will stop activation when max_samples is reached 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, max_samples) mu = np.mean(act, axis=0) sigma = np.cov(act, rowvar=False) return mu, sigma, act def _compute_statistics_of_path(path, model, batch_size, dims, device): if path.endswith('.npz'): f = np.load(path) m, s = f['mu'][:], f['sigma'][:] f.close() else: path = pathlib.Path(path) files = list(path.glob('*.jpg')) + list(path.glob('*.png')) m, s, a = calculate_activation_statistics(files, model, batch_size, dims, device) return m, s, a def compute_statistics_of_generator(data_loader, model, batch_size, dims, device, max_samples=None): m, s, a = calculate_activation_statistics(data_loader, model, batch_size, dims, device, max_samples) return m, s, a def save_statistics(path, m, s, a): assert path.endswith('.npz') #np.savez(path, mu=m, sigma=s, act=a) np.savez(path, mu=m, sigma=s, act=a) def load_statistics(path): assert path.endswith('.npz') f = np.load(path) m, s, a = f['mu'][:], f['sigma'][:], f['act'][:] f.close() return m, s, a def calculate_fid_given_paths(paths, batch_size, device, dims): """Calculates the FID of two paths""" for p in paths: if not os.path.exists(p): raise RuntimeError('Invalid path: %s' % p) 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) m2, s2 = _compute_statistics_of_path(paths[1], model, batch_size, dims, device) fid_value = calculate_frechet_distance(m1, s1, m2, s2) 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) fid_value = calculate_fid_given_paths(args.path, args.batch_size, device, args.dims) print('FID: ', fid_value) if __name__ == '__main__': main()