import sys sys.path.append("src") import os import math import pandas as pd import zlib import yaml import audioldm_train.utilities.audio as Audio from audioldm_train.utilities.tools import load_json from audioldm_train.dataset_plugin import * import librosa from librosa.filters import mel as librosa_mel_fn import threading import random import lmdb from torch.utils.data import Dataset import torch.nn.functional import torch from pydub import AudioSegment import numpy as np import torchaudio import io import json from .datum_all_pb2 import Datum_all as Datum_lmdb from .datum_mos_pb2 import Datum_mos as Datum_lmdb_mos def dynamic_range_compression_torch(x, C=1, clip_val=1e-5): return torch.log(torch.clamp(x, min=clip_val) * C) def dynamic_range_decompression_torch(x, C=1): return torch.exp(x) / C def spectral_normalize_torch(magnitudes): output = dynamic_range_compression_torch(magnitudes) return output def spectral_de_normalize_torch(magnitudes): output = dynamic_range_decompression_torch(magnitudes) return output class AudioDataset(Dataset): def __init__( self, config, lmdb_path, key_path, mos_path, lock=True ): self.config = config # self.lock = threading.Lock() """ Dataset that manages audio recordings """ self.pad_wav_start_sample = 0 self.trim_wav = False self.build_setting_parameters() self.build_dsp() self.lmdb_path = [_.encode("utf-8") for _ in lmdb_path] self.lmdb_env = [lmdb.open(_, readonly=True, lock=False) for _ in self.lmdb_path] self.mos_txn_env = lmdb.open(mos_path, readonly=True, lock=False) self.key_path = [_.encode("utf-8") for id, _ in enumerate(key_path)] self.keys = [] for _ in range(len(key_path)): with open(self.key_path[_]) as f: for line in f: key = line.strip() self.keys.append((_, key.split()[0].encode('utf-8'))) # only for test !!! # if _ > 20: # break # self.keys : [(id, key), ..., ...] # self.lmdb_env = lmdb.open(self.lmdb_path, readonly=True, lock=False) # self.txn = self.lmdb_env.begin() print(f"Dataset initialize finished, dataset_length : {len(self.keys)}") print(f"Initialize of filter start: ") with open('filter_all.lst', 'r') as f: self.filter = {} for _ in f.readlines(): self.filter[_.strip()] = 1 print(f"Initialize of filter finished") #print(f"Initialize of fusion start: ") #with open('new_file.txt', 'r') as f: # self.refined_caption = {} # for _ in f.readlines(): # try: # a, b = _.strip().split("@") # b = b.strip('"\n') # b = b.replace('\n', ',') # self.refined_caption[a] = b # except: # pass #print(f"Initialize of fusion finished") def __getitem__(self, index): ( # name of file, while we use dir of fine here fname, # wav of sr = 16000 waveform, # mel stft, # log mel log_mel_spec, label_vector, # donot start at the begining random_start, # dict or single string which describes the wav file caption, # mos score for single music clip mos ) = self.feature_extraction(index) data = { "text": [caption], # list ... dict ? "fname": [fname], # list # tensor, [batchsize, 1, samples_num] "waveform": "" if (waveform is None) else waveform.float(), # tensor, [batchsize, t-steps, f-bins] "stft": "" if (stft is None) else stft.float(), # tensor, [batchsize, t-steps, mel-bins] "log_mel_spec": "" if (log_mel_spec is None) else log_mel_spec.float(), "duration": self.duration, "sampling_rate": self.sampling_rate, "random_start_sample_in_original_audio_file": random_start, "label_vector": label_vector, "mos":mos } if data["text"] is None: print("Warning: The model return None on key text", fname) data["text"] = "" return data def __len__(self): return len(self.keys) def feature_extraction(self, index): if index > len(self.keys) - 1: print( "The index of the dataloader is out of range: %s/%s" % (index, len(self.data)) ) index = random.randint(0, len(self.keys) - 1) waveform = np.array([]) tyu = 0 flag = 0 last_index = index while(flag == 0): id_, k = self.keys[index] try: if self.filter[k.decode()] == 1: index = random.randint(0, len(self.keys) - 1) else: flag = 1 except: flag = 1 index = last_index while len(waveform) < 1000: id_, k = self.keys[index] with self.lmdb_env[id_].begin(write=False) as txn: cursor = txn.cursor() try: cursor.set_key(k) datum_tmp = Datum_lmdb() datum_tmp.ParseFromString(cursor.value()) zobj = zlib.decompressobj() # obj for decompressing data streams that won’t fit into memory at once. decompressed_bytes = zobj.decompress(datum_tmp.wav_file) # decompressed_bytes = zlib.decompress(file) waveform = np.frombuffer(decompressed_bytes, dtype=np.float32) except: tyu += 1 pass tyu += 1 last_index = index index = random.randint(0, len(self.keys) - 1) if tyu > 1: print('error') index = last_index flag = 0 val = 623787092.84794 while (flag == 0): id_, k = self.keys[index] with self.mos_txn_env.begin(write=False) as txn: cursor = txn.cursor() try: if cursor.set_key(k): datum_mos = Datum_lmdb_mos() datum_mos.ParseFromString(cursor.value()) mos = datum_mos.mos else: mos = -1.0 except : mos = -1.0 if 'pixa_' in k.decode() or 'ifly_' in k.decode(): mos = 5.0 if np.random.rand() < math.exp(5.0 * mos) / val: flag = 1 last_index = index index = random.randint(0, len(self.keys) - 1) index = last_index caption_original = datum_tmp.caption_original try: caption_generated = datum_tmp.caption_generated[0] except: caption_generated = 'None' assert len(caption_generated) > 1 caption_original = caption_original.lower() caption_generated = caption_generated.lower() caption = 'music' if ("msd_" in k.decode()): caption = caption_generated if caption_original == "none" else caption_original elif ("audioset_" in k.decode()): caption = caption_generated if caption_generated != "none" else caption_original elif ("mtt_" in k.decode()): caption = caption_generated if caption_original == "none" else caption_original elif ("fma_" in k.decode()): caption = caption_generated if caption_generated != "none" else caption_original elif ("pixa_" in k.decode() or "ifly_" in k.decode()): caption = caption_generated if caption_generated != "none" else caption_original else: caption = caption_original prefix = 'medium quality' if ("pixa_" in k.decode() or "ifly_" in k.decode()): if caption == 'none': prefix = 'high quality' caption = '' else: prefix = 'high quality' mos = 5.00 else: mos = float(mos) if mos > 3.55 and mos < 4.05: prefix = "medium quality" elif mos >= 4.05: prefix = "high quality" elif mos <= 3.55: prefix = "low quality" else: print(f'mos score for key : {k.decode()} miss, please check') #if 'low quality' or 'quality is low' in caption: # prefix = 'low quality' caption = prefix + ', ' + caption miu = 3.80 sigma = 0.20 if miu - 2 * sigma <= mos < miu - sigma: vq_mos = 2 elif miu - sigma <= mos < miu + sigma: vq_mos = 3 elif miu + sigma <= mos < miu + 2 * sigma: vq_mos = 4 elif mos >= miu + 2 * sigma: vq_mos = 5 else: vq_mos = 1 """ tags = datum_tmp.tags.decode() caption_writing = datum_tmp.caption_writing.decode() caption_paraphrase = datum_tmp.caption_paraphrase.decode() caption_attribute_prediction = datum_tmp.caption_attribute_prediction.decode() caption_summary = datum_tmp.caption_summary.decode() """ ( log_mel_spec, stft, waveform, random_start, ) = self.read_audio_file(waveform, k.decode()) fname = self.keys[index] # t_step = log_mel_spec.size(0) # waveform = torch.FloatTensor(waveform[..., : int(self.hopsize * t_step)]) waveform = torch.FloatTensor(waveform) label_vector = torch.FloatTensor(np.zeros(0, dtype=np.float32)) # finally: # self.lock.release() # import pdb # pdb.set_trace() return ( fname, waveform, stft, log_mel_spec, label_vector, random_start, caption, vq_mos ) def build_setting_parameters(self): # Read from the json config self.melbins = self.config["preprocessing"]["mel"]["n_mel_channels"] # self.freqm = self.config["preprocessing"]["mel"]["freqm"] # self.timem = self.config["preprocessing"]["mel"]["timem"] self.sampling_rate = self.config["preprocessing"]["audio"]["sampling_rate"] self.hopsize = self.config["preprocessing"]["stft"]["hop_length"] self.duration = self.config["preprocessing"]["audio"]["duration"] self.target_length = int(self.duration * self.sampling_rate / self.hopsize) self.mixup = self.config["augmentation"]["mixup"] # Calculate parameter derivations # self.waveform_sample_length = int(self.target_length * self.hopsize) # if (self.config["balance_sampling_weight"]): # self.samples_weight = np.loadtxt( # self.config["balance_sampling_weight"], delimiter="," # ) # if "train" not in self.split: # self.mixup = 0.0 # # self.freqm = 0 # # self.timem = 0 def build_dsp(self): self.mel_basis = {} self.hann_window = {} self.filter_length = self.config["preprocessing"]["stft"]["filter_length"] self.hop_length = self.config["preprocessing"]["stft"]["hop_length"] self.win_length = self.config["preprocessing"]["stft"]["win_length"] self.n_mel = self.config["preprocessing"]["mel"]["n_mel_channels"] self.sampling_rate = self.config["preprocessing"]["audio"]["sampling_rate"] self.mel_fmin = self.config["preprocessing"]["mel"]["mel_fmin"] self.mel_fmax = self.config["preprocessing"]["mel"]["mel_fmax"] self.STFT = Audio.stft.TacotronSTFT( self.config["preprocessing"]["stft"]["filter_length"], self.config["preprocessing"]["stft"]["hop_length"], self.config["preprocessing"]["stft"]["win_length"], self.config["preprocessing"]["mel"]["n_mel_channels"], self.config["preprocessing"]["audio"]["sampling_rate"], self.config["preprocessing"]["mel"]["mel_fmin"], self.config["preprocessing"]["mel"]["mel_fmax"], ) def resample(self, waveform, sr): waveform = torchaudio.functional.resample(waveform, sr, self.sampling_rate) # waveform = librosa.resample(waveform, sr, self.sampling_rate) return waveform # if sr == 16000: # return waveform # if sr == 32000 and self.sampling_rate == 16000: # waveform = waveform[::2] # return waveform # if sr == 48000 and self.sampling_rate == 16000: # waveform = waveform[::3] # return waveform # else: # raise ValueError( # "We currently only support 16k audio generation. You need to resample you audio file to 16k, 32k, or 48k: %s, %s" # % (sr, self.sampling_rate) # ) def normalize_wav(self, waveform): waveform = waveform - np.mean(waveform) waveform = waveform / (np.max(np.abs(waveform)) + 1e-8) return waveform * 0.5 # Manually limit the maximum amplitude into 0.5 def random_segment_wav(self, waveform, target_length): waveform = torch.tensor(waveform) waveform = waveform.unsqueeze(0) waveform_length = waveform.shape[-1] # assert waveform_length > 100, "Waveform is too short, %s" % waveform_length if waveform_length < 100: waveform = torch.nn.functional.pad(waveform, (0, target_length - waveform_length)) # Too short if (waveform_length - target_length) <= 0: return waveform, 0 for i in range(10): random_start = int(self.random_uniform(0, waveform_length - target_length)) if torch.max( torch.abs(waveform[:, random_start : random_start + target_length]) > 1e-4 ): break return waveform[:, random_start : random_start + target_length], random_start def pad_wav(self, waveform, target_length): # print(waveform) # import pdb # pdb.set_trace() waveform_length = waveform.shape[-1] # assert waveform_length > 100, "Waveform is too short, %s" % waveform_length if waveform_length < 100: waveform = torch.nn.functional.pad(waveform, (0, target_length - waveform_length)) if waveform_length == target_length: return waveform # Pad temp_wav = np.zeros((1, target_length), dtype=np.float32) if self.pad_wav_start_sample is None: rand_start = int(self.random_uniform(0, target_length - waveform_length)) else: rand_start = 0 temp_wav[:, rand_start : rand_start + waveform_length] = waveform return temp_wav def trim_wav(self, waveform): if np.max(np.abs(waveform)) < 0.0001: return waveform def detect_leading_silence(waveform, threshold=0.0001): chunk_size = 1000 waveform_length = waveform.shape[0] start = 0 while start + chunk_size < waveform_length: if np.max(np.abs(waveform[start : start + chunk_size])) < threshold: start += chunk_size else: break return start def detect_ending_silence(waveform, threshold=0.0001): chunk_size = 1000 waveform_length = waveform.shape[0] start = waveform_length while start - chunk_size > 0: if np.max(np.abs(waveform[start - chunk_size : start])) < threshold: start -= chunk_size else: break if start == waveform_length: return start else: return start + chunk_size start = detect_leading_silence(waveform) end = detect_ending_silence(waveform) return waveform[start:end] def read_wav_file(self, file, k): #zobj = zlib.decompressobj() # obj for decompressing data streams that won’t fit into memory at once. #decompressed_bytes = zobj.decompress(file) # decompressed_bytes = zlib.decompress(file) #waveform = np.frombuffer(decompressed_bytes, dtype=np.float32) waveform = file # # waveform, sr = librosa.load(filename, sr=None, mono=True) # 4 times slower # if "msd" in k or "fma" in k: # try: # waveform = torch.tensor([(np.array(file.get_array_of_samples(array_type_override='i')) / 2147483648)], dtype=torch.float32) # except: # waveform = torch.tensor([(np.array(file.get_array_of_samples(array_type_override='h')) / 32768)], dtype=torch.float32) # else: # waveform = torch.tensor([(np.array(file.get_array_of_samples(array_type_override='h')) / 32768)], dtype=torch.float32) # # else: # # raise AttributeError # waveform = torch.tensor([(np.array(file.get_array_of_samples(array_type_override='h')) / 32768)], dtype=torch.float32) # import pdb # pdb.set_trace() sr = 16000 waveform, random_start = self.random_segment_wav( waveform, target_length=int(sr * self.duration) ) waveform = self.resample(waveform, sr) # random_start = int(random_start * (self.sampling_rate / sr)) waveform = waveform.numpy()[0, ...] waveform = self.normalize_wav(waveform) if self.trim_wav: waveform = self.trim_wav(waveform) waveform = waveform[None, ...] waveform = self.pad_wav( waveform, target_length=int(self.sampling_rate * self.duration) ) return waveform, random_start def mix_two_waveforms(self, waveform1, waveform2): mix_lambda = np.random.beta(5, 5) mix_waveform = mix_lambda * waveform1 + (1 - mix_lambda) * waveform2 return self.normalize_wav(mix_waveform), mix_lambda def read_audio_file(self, file, k): # target_length = int(self.sampling_rate * self.duration) # import pdb # pdb.set_trace() # print(type(file)) waveform, random_start = self.read_wav_file(file, k) # log_mel_spec, stft = self.wav_feature_extraction_torchaudio(waveform) # this line is faster, but this implementation is not aligned with HiFi-GAN log_mel_spec, stft = self.wav_feature_extraction(waveform) return log_mel_spec, stft, waveform, random_start def mel_spectrogram_train(self, y): if torch.min(y) < -1.0: print("train min value is ", torch.min(y)) if torch.max(y) > 1.0: print("train max value is ", torch.max(y)) # import pdb # pdb.set_trace() if self.mel_fmax not in self.mel_basis: # import pdb # pdb.set_trace() mel = librosa_mel_fn( sr=self.sampling_rate, n_fft=self.filter_length, n_mels=self.n_mel, fmin=self.mel_fmin, fmax=self.mel_fmax, ) self.mel_basis[str(self.mel_fmax) + "_" + str(y.device)] = ( torch.from_numpy(mel).float().to(y.device) ) self.hann_window[str(y.device)] = torch.hann_window(self.win_length).to( y.device ) y = torch.nn.functional.pad( y.unsqueeze(1), ( int((self.filter_length - self.hop_length) / 2), int((self.filter_length - self.hop_length) / 2), ), mode="reflect", ) y = y.squeeze(1) # import pdb # pdb.set_trace() stft_spec = torch.stft( y, self.filter_length, hop_length=self.hop_length, win_length=self.win_length, window=self.hann_window[str(y.device)], center=False, pad_mode="reflect", normalized=False, onesided=True, return_complex=True, ) stft_spec = torch.abs(stft_spec) mel = spectral_normalize_torch( torch.matmul( self.mel_basis[str(self.mel_fmax) + "_" + str(y.device)], stft_spec ) ) return mel[0], stft_spec[0] # This one is significantly slower than "wav_feature_extraction_torchaudio" if num_worker > 1 def wav_feature_extraction(self, waveform): waveform = waveform[0, ...] waveform = torch.FloatTensor(waveform) # log_mel_spec, stft, energy = Audio.tools.get_mel_from_wav(waveform, self.STFT)[0] log_mel_spec, stft = self.mel_spectrogram_train(waveform.unsqueeze(0)) log_mel_spec = torch.FloatTensor(log_mel_spec.T) stft = torch.FloatTensor(stft.T) log_mel_spec, stft = self.pad_spec(log_mel_spec), self.pad_spec(stft) return log_mel_spec, stft def pad_spec(self, log_mel_spec): n_frames = log_mel_spec.shape[0] p = self.target_length - n_frames # cut and pad if p > 0: m = torch.nn.ZeroPad2d((0, 0, 0, p)) log_mel_spec = m(log_mel_spec) elif p < 0: log_mel_spec = log_mel_spec[0 : self.target_length, :] if log_mel_spec.size(-1) % 2 != 0: log_mel_spec = log_mel_spec[..., :-1] return log_mel_spec def _read_datum_caption(self, datum): caption_keys = [x for x in datum.keys() if ("caption" in x)] random_index = torch.randint(0, len(caption_keys), (1,))[0].item() return datum[caption_keys[random_index]] def _is_contain_caption(self, datum): caption_keys = [x for x in datum.keys() if ("caption" in x)] return len(caption_keys) > 0 def label_indices_to_text(self, datum, label_indices): if self._is_contain_caption(datum): return self._read_datum_caption(datum) elif "label" in datum.keys(): name_indices = torch.where(label_indices > 0.1)[0] # description_header = "This audio contains the sound of " description_header = "" labels = "" for id, each in enumerate(name_indices): if id == len(name_indices) - 1: labels += "%s." % self.num2label[int(each)] else: labels += "%s, " % self.num2label[int(each)] return description_header + labels else: return "" # TODO, if both label and caption are not provided, return empty string def random_uniform(self, start, end): val = torch.rand(1).item() return start + (end - start) * val def frequency_masking(self, log_mel_spec, freqm): bs, freq, tsteps = log_mel_spec.size() mask_len = int(self.random_uniform(freqm // 8, freqm)) mask_start = int(self.random_uniform(start=0, end=freq - mask_len)) log_mel_spec[:, mask_start : mask_start + mask_len, :] *= 0.0 return log_mel_spec def time_masking(self, log_mel_spec, timem): bs, freq, tsteps = log_mel_spec.size() mask_len = int(self.random_uniform(timem // 8, timem)) mask_start = int(self.random_uniform(start=0, end=tsteps - mask_len)) log_mel_spec[:, :, mask_start : mask_start + mask_len] *= 0.0 return log_mel_spec class AudioDataset_infer(Dataset): def __init__( self, config, caption_list, lock=True ): self.config = config # self.lock = threading.Lock() """ Dataset that manage caption writings """ self.captions = [] with open(caption_list, 'r') as f: for _ ,line in enumerate(f): key = line.strip() self.captions.append(key.split()[0]) self.duration = self.duration = self.config["preprocessing"]["audio"]["duration"] self.sampling_rate = self.config["variables"]["sampling_rate"] self.target_length = int(self.sampling_rate * self.duration) self.waveform = torch.zeros((1, self.target_length)) def __getitem__(self, index): fname = [f"sample_{index}"] data = { "text": [self.captions[index]], # list ... dict ? "fname": fname, # list # tensor, [batchsize, 1, samples_num] "waveform": "", # tensor, [batchsize, t-steps, f-bins] "stft": "", # tensor, [batchsize, t-steps, mel-bins] "log_mel_spec": "", "duration": self.duration, "sampling_rate": self.sampling_rate, "random_start_sample_in_original_audio_file": 0, "label_vector": torch.FloatTensor(np.zeros(0, dtype=np.float32)), "mos":mos } if data["text"] is None: print("Warning: The model return None on key text", fname) data["text"] = "" return data def __len__(self): return len(self.captions) if __name__ == "__main__": import torch from tqdm import tqdm from pytorch_lightning import seed_everything from torch.utils.data import DataLoader seed_everything(0) def write_json(my_dict, fname): # print("Save json file at "+fname) json_str = json.dumps(my_dict) with open(fname, "w") as json_file: json_file.write(json_str) def load_json(fname): with open(fname, "r") as f: data = json.load(f) return data config = yaml.load( open( "/mnt/bn/lqhaoheliu/project/audio_generation_diffusion/config/vae_48k_256/ds_8_kl_1.0_ch_16.yaml", "r", ), Loader=yaml.FullLoader, ) add_ons = config["data"]["dataloader_add_ons"] # load_json(data) dataset = AudioDataset( config=config, split="train", waveform_only=False, add_ons=add_ons ) loader = DataLoader(dataset, batch_size=1, num_workers=0, shuffle=True) # for cnt, each in tqdm(enumerate(loader)): # print(each["waveform"].size(), each["log_mel_spec"].size()) # print(each['freq_energy_percentile']) # import ipdb # ipdb.set_trace() # pass