def identify_genre():
global music_genre
spectro = []
y = load_audio(audio_file)[0]
signals = split_audio(y)
spec_array = to_melspec(signals)
spectro.extend(spec_array)
spectro = np.array(spectro)
spectro = np.squeeze(np.stack((spectro, ) * 3, -1))
predictions = np.array(model.predict(spectro))
preds = np.argmax(predictions, axis=1)
print('-' * 40)
music_genre = genres[np.bincount(preds).argmax()]
print(music_genre)
def main(args):
CWD = Path(hydra.utils.get_original_cwd())
os.chdir(CWD)
# Load model args
trained_dirpath = Path(args.trained_dirpath)
run_args = torch.load(trained_dirpath / 'args.pth')
# define args from trained model
sr = run_args.sr
num_scales = run_args.num_scales
scale_factor = run_args.scale_factor
max_value = run_args.max_val
max_value_f0 = run_args.max_val_f0
cond_freq = run_args.cond_freq
# Convert filepaths
input_dirpath = Path(args.input_dirpath)
input_files = input_dirpath.glob('*.wav')
output_dirpath = trained_dirpath.joinpath(args.exp_name)
try:
output_dirpath.mkdir()
except FileExistsError:
print('Directory already exists')
# Pytorch device
device = torch.device("cuda")
#load input file
base_audio = BaseAudio(args.crepe_path, device, args.unvoiced_flag)
srs = create_srs(sr, num_scales, scale_factor)
samplers = create_samplers(srs, device=device)
if args.norm_loudness_flag:
norm_dicts = load_norm_dicts(trained_dirpath / 'loudness.json')
else:
norm_dicts = None
octave_shifts = [2**x for x in args.octaves]
# Load Trained models
Gs = load_trained_pyramid(trained_dirpath, network_params=run_args.generator_params, device=device, srs=srs)
filepath = next(iter(input_files))
octave = next(iter(octave_shifts))
real_audio = load_audio(filepath, sr, max_value)
loudness_hop = 8 * sr // cond_freq
real_audio = real_audio[:len(real_audio) // loudness_hop * loudness_hop]
loudness_list = calc_loudness_list(audio=real_audio, srs=srs, device=device,
sr_in=sr, norm_dicts=norm_dicts)
real_audio = base_audio.forward(real_audio, sr, max_value_f0, numpy_flag=True, octave=octave)
real_audio_orig = real_audio[None, None, ...].to(device)
# resample input to the wanted scale
real_audio = resample_torch(real_audio_orig, sr, srs[0], max_val=max_value_f0)
BUFFER_SIZES = [256, 512, 1024, 2048, 4096, 8192, 16384, 32768]
num_iters = 100
times = []
for bs in BUFFER_SIZES:
this_ra = real_audio[..., :bs // 8]
this_ll = [l[..., :bs // (64000 // l.shape[-1])] for l in loudness_list]
with torch.no_grad():
for i in trange(num_iters):
start_time = time.time()
audio_outputs = f0_transfer(this_ra, this_ll, Gs, samplers, max_val=max_value, save_all = False)
time_elapsed = time.time() - start_time
times.append(
["htp", "gpu", bs, time_elapsed]
)
df = pd.DataFrame(times)
df.to_csv("htp_gpu_rtf.csv")
def main(args):
CWD = Path(hydra.utils.get_original_cwd())
os.chdir(CWD)
# Load model args
trained_dirpath = Path(args.trained_dirpath)
run_args = torch.load(trained_dirpath / 'args.pth')
# define args from trained model
sr = run_args.sr
num_scales = run_args.num_scales
scale_factor = run_args.scale_factor
max_value = run_args.max_val
max_value_f0 = run_args.max_val_f0
cond_freq = run_args.cond_freq
# Convert filepaths
input_dirpath = Path(args.input_dirpath)
input_files = input_dirpath.glob('*.wav')
output_dirpath = trained_dirpath.joinpath(args.exp_name)
try:
output_dirpath.mkdir()
except FileExistsError:
print('Directory already exists')
# Pytorch device
device = torch.device("cuda")
#load input file
base_audio = BaseAudio(args.crepe_path, device, args.unvoiced_flag)
srs = create_srs(sr, num_scales, scale_factor)
samplers = create_samplers(srs, device=device)
if args.norm_loudness_flag:
norm_dicts = load_norm_dicts(trained_dirpath / 'loudness.json')
else:
norm_dicts = None
octave_shifts = [2**x for x in args.octaves]
# Load Trained models
Gs = load_trained_pyramid(trained_dirpath, network_params=run_args.generator_params, device=device, srs=srs)
for filepath in tqdm(input_files, desc='Generating audio file'):
for octave in tqdm(octave_shifts, desc='Octave'):
real_audio = load_audio(filepath, sr, max_value)
loudness_hop = 8 * sr // cond_freq
real_audio = real_audio[:len(real_audio) // loudness_hop * loudness_hop]
loudness_list = calc_loudness_list(audio=real_audio, srs=srs, device=device,
sr_in=sr, norm_dicts=norm_dicts)
real_audio = base_audio.forward(real_audio, sr, max_value_f0, numpy_flag=True, octave=octave)
real_audio_orig = real_audio[None, None, ...].to(device)
# resample input to the wanted scale
real_audio = resample_torch(real_audio_orig, sr, srs[0], max_val=max_value_f0)
audio_outputs = f0_transfer(real_audio,loudness_list, Gs, samplers, max_val=max_value, save_all = False)
# add f0 sine input
audio_outputs.append(real_audio_orig)
save_audios(output_dirpath, audio_outputs,f'{filepath.stem}_{octave}', [srs[-1]])
def main(args):
CWD = Path(hydra.utils.get_original_cwd())
os.chdir(CWD)
# Load model args
trained_dirpath = Path(args.trained_dirpath)
run_args = torch.load(trained_dirpath / 'args.pth')
# define args from trained model
sr = run_args.sr
num_scales = run_args.num_scales
scale_factor = run_args.scale_factor
max_value = run_args.max_val
max_value_f0 = run_args.max_val_f0
cond_freq = run_args.cond_freq
# Convert filepaths
input_dirpath = Path(args.input_dirpath)
input_files = input_dirpath.glob('*.wav')
output_dirpath = trained_dirpath.joinpath(args.exp_name)
try:
output_dirpath.mkdir()
except FileExistsError:
print('Directory already exists')
# Pytorch device
device = torch.device("cuda")
#load input file
base_audio = BaseAudio(args.crepe_path, device, args.unvoiced_flag)
srs = create_srs(sr, num_scales, scale_factor)
samplers = create_samplers(srs, device=device)
if args.norm_loudness_flag:
norm_dicts = load_norm_dicts(trained_dirpath / 'loudness.json')
else:
norm_dicts = None
octave_shifts = [2**x for x in args.octaves]
# Load Trained models
Gs = load_trained_pyramid(trained_dirpath, network_params=run_args.generator_params, device=device, srs=srs)
filepath = next(iter(input_files))
octave = next(iter(octave_shifts))
real_audio = load_audio(filepath, sr, max_value)
loudness_hop = 8 * sr // cond_freq
real_audio = real_audio[:len(real_audio) // loudness_hop * loudness_hop]
loudness_list = calc_loudness_list(audio=real_audio, srs=srs, device=device,
sr_in=sr, norm_dicts=norm_dicts)
real_audio = base_audio.forward(real_audio, sr, max_value_f0, numpy_flag=True, octave=octave)
real_audio_orig = real_audio[None, None, ...].to(device)
# resample input to the wanted scale
real_audio = resample_torch(real_audio_orig, sr, srs[0], max_val=max_value_f0)
num_iters = 100
times = []
with torch.no_grad():
for i in trange(num_iters):
start_time = time.time()
audio_outputs = f0_transfer(real_audio,loudness_list, Gs, samplers, max_val=max_value, save_all = False)
time_elapsed = time.time() - start_time
times.append(time_elapsed)
length_in_seconds = 4
rtfs = np.array(times) / length_in_seconds
print("Mean RTF: %.4f" % np.mean(rtfs))
print("90th percentile RTF: %.4f" % np.percentile(rtfs, 90))