def __init__(self, config, training=True):
super(AudioDataset, self).__init__()
self.config = config
self.hop_length = config.data_config.hop_length
self.training = training
if self.training:
self.segment_length = config.training_config.segment_length
self.sample_rate = config.data_config.sample_rate
self.filelist_path = config.training_config.train_filelist_path \
if self.training else config.training_config.test_filelist_path
self.audio_paths = parse_filelist(self.filelist_path)
def add_authors(self, dir=None, format=None):
if dir is None:
return None
files = []
if os.path.isdir(dir):
files = utils.parse_filelist(dir, format)
c = self.get_class('人物')
name = self.get_data_property(c, '姓名')
# brief = self.get_data_property(c, '生平简介')
for file in files:
# print(file, os.path.join(dir, file))
with open(os.path.join(dir, file), 'r') as f:
data = json.load(f)
# print(data)
for v in data:
rid = self.new_entity_resource()
self.poetry_graph.add((rid, RDF.type, c))
self.poetry_graph.add((rid, name, Literal(v['诗人姓名'], lang='zh')))
model.load_state_dict(torch.load(args.checkpoint_path)['model'],
strict=False)
# Set noise schedule
noise_schedule = torch.load(args.noise_schedule_path)
n_iter = noise_schedule.shape[-1]
init_fn = lambda **kwargs: noise_schedule
init_kwargs = {'steps': n_iter}
model.set_new_noise_schedule(init_fn, init_kwargs)
# Trying to run inference on GPU
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = model.to(device)
# Inference
filelist = parse_filelist(args.mel_filelist)
rtfs = []
for mel_path in (tqdm(filelist, leave=False)
if args.verbose else filelist):
with torch.no_grad():
mel = torch.load(mel_path).unsqueeze(0).to(device)
start = datetime.now()
outputs = model.forward(mel, store_intermediate_states=False)
end = datetime.now()
outputs = outputs.cpu().squeeze()
baseidx = os.path.basename(
os.path.abspath(mel_path)).split('_')[-1].replace('.pt', '')
save_path = f'{os.path.dirname(os.path.abspath(mel_path))}/predicted_{baseidx}.wav'
torchaudio.save(save_path,
def add_poetry(self, dir=None, format=None):
"""
诗的信息，包括诗名、作者、诗的内容，利用这个可以添加诗句实体，
以及诗和人物的关系、诗和诗句的关系、诗句和人物的关系、诗句和诗句的关系。
:param dir:
:param format:
:return:
"""
if dir is None:
return None
files = []
if os.path.isdir(dir):
files = utils.parse_filelist(dir, format)
# 涉及的类
poetry = self.get_class('诗词')
verse_class = self.get_class('诗句')
person = self.get_class('人物')
# 涉及的属性
title = self.get_data_property(poetry, '题名')
poem_content = self.get_data_property(poetry, '内容')
poem_author = self.get_object_property('作者', poetry, person)
verse_author = self.get_object_property('作者', verse_class, person)
verse_belongto_poem = self.get_object_property('属于', verse_class, poetry)
next_verse = self.get_object_property('下一句', verse_class, verse_class)
prev_verse = self.get_object_property('上一句', verse_class, verse_class)
for file in files:
# print(file, os.path.join(dir, file))
with open(os.path.join(dir, file), 'r') as f:
data = json.load(f)
# print(data)
for v in data:
poem = self.new_entity_resource()
# 添加 诗词 实体
self.poetry_graph.add((poem, RDF.type, poetry))
self.poetry_graph.add((poem, title, Literal(v['诗名'], lang='zh')))
self.poetry_graph.add((poem, poem_content, Literal(v['内容'], lang='zh')))
# 添加 诗词 与 人物 关系
personal = self.get_resource(person, '姓名', v['作者'])
if not personal:
personal = self.new_entity_resource()
self.poetry_graph.add((personal, RDF.type, person))
self.poetry_graph.add((personal, self.get_data_property(person, '姓名'), Literal(v['作者'], lang='zh')))
else:
self.poetry_graph.add((poem, poem_author, personal))
# 添加 诗句 实体以及 诗句与诗词 的关系
prev_v = None
for verse in re.split('，|。', v['内容']):
vres = self.new_entity_resource()
self.poetry_graph.add((vres, RDF.type, verse_class))
self.poetry_graph.add((vres, RDFS.label, Literal(verse, lang='zh')))
self.poetry_graph.add((vres, verse_author, personal))
self.poetry_graph.add((vres, verse_belongto_poem, poem))
if not prev_v:
prev_v = vres
else:
self.poetry_graph.add((prev_v, next_verse, vres))
self.poetry_graph.add((vres, prev_verse, prev_v))
prev_v = vres
with open(args.config) as f:
config = ConfigWrapper(**json.load(f))
mel_fn = MelSpectrogramFixed(sample_rate=config.data_config.sample_rate,
n_fft=config.data_config.n_fft,
win_length=config.data_config.win_length,
hop_length=config.data_config.hop_length,
f_min=config.data_config.f_min,
f_max=config.data_config.f_max,
n_mels=config.data_config.n_mels,
window_fn=torch.hann_window).cuda()
dataset = AudioDataset(config, training=True)
dataset.filelist_path = args.filelist
dataset.audio_paths = parse_filelist(dataset.filelist_path)
loader = DataLoader(dataset, batch_size=48)
nans, infs = [], []
for batch in tqdm(loader, total=int(np.ceil(len(dataset) / 48))):
batch = batch.cuda()
mels = mel_fn(batch)
nan_mask = torch.isnan(mels)
inf_mask = torch.isinf(mels)
nans.append(nan_mask.sum().cpu())
infs.append(inf_mask.sum().cpu())
print(f'Dataset has nans: {any([item != 0 for item in nans])}')