""" Calculate Frechet Audio Distance betweeen two audio directories. Frechet distance implementation adapted from: https://github.com/mseitzer/pytorch-fid VGGish adapted from: https://github.com/harritaylor/torchvggish """ import os import numpy as np from glob import glob import torch from torch import nn from scipy import linalg from tqdm import tqdm import soundfile as sf import resampy from multiprocessing.dummy import Pool as ThreadPool from src.torchvggish.torchvggish.vggish import VGGishlocal SAMPLE_RATE = 16000 # resample audio file to SAMPLE_RATE. since uses the pretrained vggish model which takes wav_data as input def load_audio_task(fname):# load wav file and resample to SAMPLE_RATE wav_data, sr = sf.read(fname, dtype='int16') assert wav_data.dtype == np.int16, 'Bad sample type: %r' % wav_data.dtype wav_data = wav_data / 32768.0 # Convert to [-1.0, +1.0] # Convert to mono if len(wav_data.shape) > 1: wav_data = np.mean(wav_data, axis=1) if sr != SAMPLE_RATE: wav_data = resampy.resample(wav_data, sr, SAMPLE_RATE) return wav_data, SAMPLE_RATE # use pretrained torchvggish as embedding extractor, and calculate the statistic of wav file class FrechetAudioDistance: def __init__(self, use_pca=False, use_activation=False, verbose=False, audio_load_worker=8): # self.__get_model(use_pca=use_pca, use_activation=use_activation) self.__get_local_model(local_path='src/torchvggish/docs',use_pca=use_pca, use_activation=use_activation) self.verbose = verbose self.audio_load_worker = audio_load_worker def __get_model(self, use_pca=False, use_activation=False): """ Params: -- x : Either (i) a string which is the directory of a set of audio files, or (ii) a np.ndarray of shape (num_samples, sample_length) """ self.model = torch.hub.load('harritaylor/torchvggish', 'vggish') if not use_pca: self.model.postprocess = False if not use_activation: self.model.embeddings = nn.Sequential(*list(self.model.embeddings.children())[:-1]) self.model.eval() def __get_local_model(self,local_path,use_pca=False, use_activation=False): self.model = VGGishlocal(local_path) if not use_pca: self.model.postprocess = False if not use_activation: self.model.embeddings = nn.Sequential(*list(self.model.embeddings.children())[:-1]) self.model.eval() def get_embeddings(self, x, sr=16000): """ Get embeddings using VGGish model. Params: -- x : Either (i) a string which is the directory of a set of audio files, or (ii) a list of np.ndarray audio samples -- sr : Sampling rate, if x is a list of audio samples. Default value is 16000. """ embd_lst = [] if isinstance(x, list):# np.ndarray try: for audio, sr in tqdm(x, disable=(not self.verbose)): embd = self.model.forward(audio, sr) if self.model.device == torch.device('cuda'): embd = embd.cpu() embd = embd.detach().numpy() embd_lst.append(embd) except Exception as e: print("[Frechet Audio Distance] get_embeddings throw an exception: {}".format(str(e))) elif isinstance(x, str): try: for fname in tqdm(os.listdir(x), disable=(not self.verbose)): embd = self.model.forward(os.path.join(x, fname)) if self.model.device == torch.device('cuda'): embd = embd.cpu() embd = embd.detach().numpy() embd_lst.append(embd) except Exception as e: print("[Frechet Audio Distance] get_embeddings throw an exception: {}".format(str(e))) else: raise AttributeError # print("embd_lst_len",len(embd_lst)) return np.concatenate(embd_lst, axis=0) def calculate_embd_statistics(self, embd_lst): if isinstance(embd_lst, list): embd_lst = np.array(embd_lst) mu = np.mean(embd_lst, axis=0) sigma = np.cov(embd_lst, rowvar=False) return mu, sigma def calculate_frechet_distance(self, mu1, sigma1, mu2, sigma2, eps=1e-6): """ Adapted from: https://github.com/mseitzer/pytorch-fid/blob/master/src/pytorch_fid/fid_score.py 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. """ # print(f"mu1.shape:{mu1.shape},mu2.shape:{sigma1.shape}") mu1 = np.atleast_1d(mu1) # shape(128,) mu2 = np.atleast_1d(mu2) sigma1 = np.atleast_2d(sigma1)# shape(128,128) 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) print(f"diff^2:{diff.dot(diff)}, sigma1:{np.trace(sigma1)},sigma2:{np.trace(sigma2)},2 * tr_covmean{2 * tr_covmean}") return (diff.dot(diff) + np.trace(sigma1) + np.trace(sigma2) - 2 * tr_covmean) def load_audio_files(self, dir):# load_audio_task会resample task_results = [] all_wav_files = glob(os.path.join(dir,"*.wav")) pool = ThreadPool(self.audio_load_worker) pbar = tqdm(total=len(all_wav_files), disable=(not self.verbose)) def update(*a): pbar.update() if self.verbose: print("[Frechet Audio Distance] Loading audio from {}...".format(dir)) for fname in all_wav_files: res = pool.apply_async(load_audio_task, args=(fname,), callback=update)# load_audio_task会resample task_results.append(res) pool.close() pool.join() return [k.get() for k in task_results] # get return value,each is (wav_data, sample_rate) def score(self, background_dir, eval_dir, store_embds=False): try: audio_background = self.load_audio_files(background_dir) audio_eval = self.load_audio_files(eval_dir) print("audios len",len(audio_background),len(audio_eval)) embds_background = self.get_embeddings(audio_background) # (N,128) embds_eval = self.get_embeddings(audio_eval) # (M,128) # print(embds_background.shape,embds_eval.shape) if store_embds: np.save("embds_background.npy", embds_background) np.save("embds_eval.npy", embds_eval) if len(embds_background) == 0: print("[Frechet Audio Distance] background set dir is empty, exitting...") return -1 if len(embds_eval) == 0: print("[Frechet Audio Distance] eval set dir is empty, exitting...") return -1 mu_background, sigma_background = self.calculate_embd_statistics(embds_background) mu_eval, sigma_eval = self.calculate_embd_statistics(embds_eval) fad_score = self.calculate_frechet_distance( mu_background, sigma_background, mu_eval, sigma_eval ) return fad_score except Exception as e: print("[Frechet Audio Distance] exception thrown, {}".format(str(e))) return -1