def test_assert_different_length_batch_generation():
# prepare batch
batch = 4
length = 32
x = np.random.randint(0, 256, size=(batch, 1))
h = np.random.randn(batch, 28, length)
length_list = sorted(
list(np.random.randint(length // 2, length - 1, batch)))
with torch.no_grad():
net = WaveNet(256, 28, 4, 4, 10, 3, 2)
net.apply(initialize)
net.eval()
# sample-by-sample generation
gen1_list = []
for x_, h_, length in zip(x, h, length_list):
batch_x = torch.from_numpy(np.expand_dims(x_, 0)).long()
batch_h = torch.from_numpy(np.expand_dims(h_, 0)).float()
gen1 = net.fast_generate(batch_x, batch_h, length, 1, "argmax")
gen1_list += [gen1]
# batch generation
batch_x = torch.from_numpy(x).long()
batch_h = torch.from_numpy(h).float()
gen2_list = net.batch_fast_generate(batch_x, batch_h, length_list, 1,
"argmax")
# assertion
for gen1, gen2 in zip(gen1_list, gen2_list):
np.testing.assert_array_equal(gen1, gen2)
def test_generate():
batch = 2
x = np.random.randint(0, 256, size=(batch, 1))
h = np.random.randn(batch, 28, 32)
length = h.shape[-1] - 1
with torch.no_grad():
net = WaveNet(256, 28, 16, 32, 10, 3, 2)
net.apply(initialize)
net.eval()
for x_, h_ in zip(x, h):
batch_x = torch.from_numpy(np.expand_dims(x_, 0)).long()
batch_h = torch.from_numpy(np.expand_dims(h_, 0)).float()
net.generate(batch_x, batch_h, length, 1, "sampling")
net.fast_generate(batch_x, batch_h, length, 1, "sampling")
batch_x = torch.from_numpy(x).long()
batch_h = torch.from_numpy(h).float()
net.batch_fast_generate(batch_x, batch_h, [length] * batch, 1, "sampling")
def gpu_decode(feat_list, gpu):
# set default gpu and do not track gradient
torch.cuda.set_device(gpu)
torch.set_grad_enabled(False)
# define model and load parameters
if config.use_upsampling_layer:
upsampling_factor = config.upsampling_factor
else:
upsampling_factor = 0
model = WaveNet(n_quantize=config.n_quantize,
n_aux=config.n_aux,
n_resch=config.n_resch,
n_skipch=config.n_skipch,
dilation_depth=config.dilation_depth,
dilation_repeat=config.dilation_repeat,
kernel_size=config.kernel_size,
upsampling_factor=upsampling_factor)
model.load_state_dict(
torch.load(args.checkpoint,
map_location=lambda storage, loc: storage)["model"])
model.eval()
model.cuda()
# define generator
generator = decode_generator(
feat_list,
batch_size=args.batch_size,
feature_type=config.feature_type,
wav_transform=wav_transform,
feat_transform=feat_transform,
upsampling_factor=config.upsampling_factor,
use_upsampling_layer=config.use_upsampling_layer,
use_speaker_code=config.use_speaker_code)
# decode
if args.batch_size > 1:
for feat_ids, (batch_x, batch_h, n_samples_list) in generator:
logging.info("decoding start")
samples_list = model.batch_fast_generate(
batch_x, batch_h, n_samples_list, args.intervals)
for feat_id, samples in zip(feat_ids, samples_list):
wav = decode_mu_law(samples, config.n_quantize)
sf.write(args.outdir + "/" + feat_id + ".wav", wav,
args.fs, "PCM_16")
logging.info("wrote %s.wav in %s." %
(feat_id, args.outdir))
else:
for feat_id, (x, h, n_samples) in generator:
logging.info("decoding %s (length = %d)" %
(feat_id, n_samples))
samples = model.fast_generate(x, h, n_samples, args.intervals)
wav = decode_mu_law(samples, config.n_quantize)
sf.write(args.outdir + "/" + feat_id + ".wav", wav, args.fs,
"PCM_16")
logging.info("wrote %s.wav in %s." % (feat_id, args.outdir))
def gpu_decode(feat_list, gpu):
with torch.cuda.device(gpu):
# define model and load parameters
model = WaveNet(n_quantize=config.n_quantize,
n_aux=config.n_aux,
n_resch=config.n_resch,
n_skipch=config.n_skipch,
dilation_depth=config.dilation_depth,
dilation_repeat=config.dilation_repeat,
kernel_size=config.kernel_size,
upsampling_factor=config.upsampling_factor)
model.load_state_dict(
torch.load(args.checkpoint,
map_location=lambda storage, loc: storage.cuda(gpu))
["model"])
model.eval()
model.cuda()
torch.backends.cudnn.benchmark = True
# define generator
generator = decode_generator(
feat_list,
batch_size=args.batch_size,
wav_transform=wav_transform,
feat_transform=feat_transform,
use_speaker_code=config.use_speaker_code,
upsampling_factor=config.upsampling_factor)
# decode
if args.batch_size > 1:
for feat_ids, (batch_x, batch_h, n_samples_list) in generator:
logging.info("decoding start")
samples_list = model.batch_fast_generate(
batch_x, batch_h, n_samples_list, args.intervals)
for feat_id, samples in zip(feat_ids, samples_list):
wav = decode_mu_law(samples, config.n_quantize)
sf.write(args.outdir + "/" + feat_id + ".wav", wav,
args.fs, "PCM_16")
logging.info("wrote %s.wav in %s." %
(feat_id, args.outdir))
else:
for feat_id, (x, h, n_samples) in generator:
logging.info("decoding %s (length = %d)" %
(feat_id, n_samples))
samples = model.fast_generate(x, h, n_samples,
args.intervals)
wav = decode_mu_law(samples, config.n_quantize)
sf.write(args.outdir + "/" + feat_id + ".wav", wav,
args.fs, "PCM_16")
logging.info("wrote %s.wav in %s." %
(feat_id, args.outdir))
def test_assert_fast_generation():
# get batch
batch = 2
x = np.random.randint(0, 256, size=(batch, 1))
h = np.random.randn(batch, 28, 32)
length = h.shape[-1] - 1
with torch.no_grad():
# --------------------------------------------------------
# define model without upsampling and with kernel size = 2
# --------------------------------------------------------
net = WaveNet(256, 28, 4, 4, 10, 3, 2)
net.apply(initialize)
net.eval()
# sample-by-sample generation
gen1_list = []
gen2_list = []
for x_, h_ in zip(x, h):
batch_x = torch.from_numpy(np.expand_dims(x_, 0)).long()
batch_h = torch.from_numpy(np.expand_dims(h_, 0)).float()
gen1 = net.generate(batch_x, batch_h, length, 1, "argmax")
gen2 = net.fast_generate(batch_x, batch_h, length, 1, "argmax")
np.testing.assert_array_equal(gen1, gen2)
gen1_list += [gen1]
gen2_list += [gen2]
gen1 = np.stack(gen1_list)
gen2 = np.stack(gen2_list)
np.testing.assert_array_equal(gen1, gen2)
# batch generation
batch_x = torch.from_numpy(x).long()
batch_h = torch.from_numpy(h).float()
gen3_list = net.batch_fast_generate(batch_x, batch_h, [length] * batch,
1, "argmax")
gen3 = np.stack(gen3_list)
np.testing.assert_array_equal(gen3, gen2)
# --------------------------------------------------------
# define model without upsampling and with kernel size = 3
# --------------------------------------------------------
net = WaveNet(256, 28, 4, 4, 10, 3, 3)
net.apply(initialize)
net.eval()
# sample-by-sample generation
gen1_list = []
gen2_list = []
for x_, h_ in zip(x, h):
batch_x = torch.from_numpy(np.expand_dims(x_, 0)).long()
batch_h = torch.from_numpy(np.expand_dims(h_, 0)).float()
gen1 = net.generate(batch_x, batch_h, length, 1, "argmax")
gen2 = net.fast_generate(batch_x, batch_h, length, 1, "argmax")
np.testing.assert_array_equal(gen1, gen2)
gen1_list += [gen1]
gen2_list += [gen2]
gen1 = np.stack(gen1_list)
gen2 = np.stack(gen2_list)
np.testing.assert_array_equal(gen1, gen2)
# batch generation
batch_x = torch.from_numpy(x).long()
batch_h = torch.from_numpy(h).float()
gen3_list = net.batch_fast_generate(batch_x, batch_h, [length] * batch,
1, "argmax")
gen3 = np.stack(gen3_list)
np.testing.assert_array_equal(gen3, gen2)
# get batch
batch = 2
upsampling_factor = 10
x = np.random.randint(0, 256, size=(batch, 1))
h = np.random.randn(batch, 28, 3)
length = h.shape[-1] * upsampling_factor - 1
# -----------------------------------------------------
# define model with upsampling and with kernel size = 2
# -----------------------------------------------------
net = WaveNet(256, 28, 4, 4, 10, 3, 2, upsampling_factor)
net.apply(initialize)
net.eval()
# sample-by-sample generation
gen1_list = []
gen2_list = []
for x_, h_ in zip(x, h):
batch_x = torch.from_numpy(np.expand_dims(x_, 0)).long()
batch_h = torch.from_numpy(np.expand_dims(h_, 0)).float()
gen1 = net.generate(batch_x, batch_h, length, 1, "argmax")
gen2 = net.fast_generate(batch_x, batch_h, length, 1, "argmax")
np.testing.assert_array_equal(gen1, gen2)
gen1_list += [gen1]
gen2_list += [gen2]
gen1 = np.stack(gen1_list)
gen2 = np.stack(gen2_list)
np.testing.assert_array_equal(gen1, gen2)
# batch generation
batch_x = torch.from_numpy(x).long()
batch_h = torch.from_numpy(h).float()
gen3_list = net.batch_fast_generate(batch_x, batch_h, [length] * batch,
1, "argmax")
gen3 = np.stack(gen3_list)
np.testing.assert_array_equal(gen3, gen2)
# -----------------------------------------------------
# define model with upsampling and with kernel size = 3
# -----------------------------------------------------
net = WaveNet(256, 28, 4, 4, 10, 3, 2, upsampling_factor)
net.apply(initialize)
net.eval()
# sample-by-sample generation
gen1_list = []
gen2_list = []
for x_, h_ in zip(x, h):
batch_x = torch.from_numpy(np.expand_dims(x_, 0)).long()
batch_h = torch.from_numpy(np.expand_dims(h_, 0)).float()
gen1 = net.generate(batch_x, batch_h, length, 1, "argmax")
gen2 = net.fast_generate(batch_x, batch_h, length, 1, "argmax")
np.testing.assert_array_equal(gen1, gen2)
gen1_list += [gen1]
gen2_list += [gen2]
gen1 = np.stack(gen1_list)
gen2 = np.stack(gen2_list)
np.testing.assert_array_equal(gen1, gen2)
# batch generation
batch_x = torch.from_numpy(x).long()
batch_h = torch.from_numpy(h).float()
gen3_list = net.batch_fast_generate(batch_x, batch_h, [length] * batch,
1, "argmax")
gen3 = np.stack(gen3_list)
np.testing.assert_array_equal(gen3, gen2)
