def __init__(self,
n_in,
n_depth,
m_conv=1.0,
dilation_growth_rate=1,
dilation_cycle=None,
zero_out=False,
res_scale=False,
reverse_dilation=False,
checkpoint_res=False):
super().__init__()
def _get_depth(depth):
if dilation_cycle is None:
return depth
else:
return depth % dilation_cycle
blocks = [
ResConv1DBlock(n_in,
int(m_conv * n_in),
dilation=dilation_growth_rate**_get_depth(depth),
zero_out=zero_out,
res_scale=1.0 if not res_scale else 1.0 /
math.sqrt(n_depth)) for depth in range(n_depth)
]
if reverse_dilation:
blocks = blocks[::-1]
self.checkpoint_res = checkpoint_res
if self.checkpoint_res == 1:
if dist.get_rank() == 0:
print("Checkpointing convs")
self.blocks = nn.ModuleList(blocks)
else:
self.model = nn.Sequential(*blocks)
def init_dataset(self, hps):
# Load list of files and starts/durations
files = librosa.util.find_files(f'{hps.audio_files_dir}',
['mp3', 'opus', 'm4a', 'aac', 'wav'])
print_all(f"Found {len(files)} files. Getting durations")
cache = dist.get_rank() % 8 == 0 if dist.is_available() else True
durations = np.array([
get_duration_sec(file, cache=cache) * self.sr for file in files
]) # Could be approximate
self.filter(files, durations)
if self.labels:
self.labeller = Labeller(hps.max_bow_genre_size,
hps.n_tokens,
self.sample_length,
v3=hps.labels_v3)
def load_checkpoint(path):
restore = path
if restore[:5] == 'gs://':
gs_path = restore
local_path = os.path.join(os.path.expanduser("~/.cache"), gs_path[5:])
if dist.get_rank() % 8 == 0:
print("Downloading from gce")
if not os.path.exists(os.path.dirname(local_path)):
os.makedirs(os.path.dirname(local_path))
if not os.path.exists(local_path):
download(gs_path, local_path)
restore = local_path
dist.barrier()
checkpoint = t.load(restore, map_location=t.device('cpu'))
print("Restored from {}".format(restore))
return checkpoint
def _setup_dist_from_mpi(master_addr, backend, port, n_attempts, verbose):
from mpi4py import MPI # This must be imported in order to get e rrors from all ranks to show up
mpi_rank = MPI.COMM_WORLD.Get_rank()
mpi_size = MPI.COMM_WORLD.Get_size()
os.environ["RANK"] = str(mpi_rank)
os.environ["WORLD_SIZE"] = str(mpi_size)
os.environ["MASTER_ADDR"] = master_addr
os.environ["MASTER_PORT"] = str(port)
os.environ["NCCL_LL_THRESHOLD"] = "0"
os.environ["NCCL_NSOCKS_PERTHREAD"] = "2"
os.environ["NCCL_SOCKET_NTHREADS"] = "8"
# Pin this rank to a specific GPU on the node
local_rank = mpi_rank % 8
if torch.cuda.is_available():
torch.cuda.set_device(local_rank)
if verbose:
print(f"Connecting to master_addr: {master_addr}")
# There is a race condition when initializing NCCL with a large number of ranks (e.g 500 ranks)
# We guard against the failure and then retry
for attempt_idx in range(n_attempts):
try:
dist.init_process_group(backend=backend, init_method=f"env://")
assert dist.get_rank() == mpi_rank
use_cuda = torch.cuda.is_available()
print(f'Using cuda {use_cuda}')
local_rank = mpi_rank % 8
device = torch.device("cuda", local_rank) if use_cuda else torch.device("cpu")
torch.cuda.set_device(local_rank)
return mpi_rank, local_rank, device
except RuntimeError as e:
print(f"Caught error during NCCL init (attempt {attempt_idx} of {n_attempts}): {e}")
sleep(1 + (0.01 * mpi_rank)) # Sleep to avoid thundering herd
pass
raise RuntimeError("Failed to initialize NCCL")
def calculate_bandwidth(dataset, hps, duration=600):
hps = DefaultSTFTValues(hps)
n_samples = int(dataset.sr * duration)
l1, total, total_sq, n_seen, idx = 0.0, 0.0, 0.0, 0.0, dist.get_rank()
spec_norm_total, spec_nelem = 0.0, 0.0
while n_seen < n_samples:
x = dataset[idx]
if isinstance(x, (tuple, list)):
x, y = x
samples = x.astype(np.float64)
stft = librosa.core.stft(np.mean(samples, axis=1),
hps.n_fft,
hop_length=hps.hop_length,
win_length=hps.window_size)
spec = np.absolute(stft)
spec_norm_total += np.linalg.norm(spec)
spec_nelem += 1
n_seen += int(np.prod(samples.shape))
l1 += np.sum(np.abs(samples))
total += np.sum(samples)
total_sq += np.sum(samples**2)
idx += max(16, dist.get_world_size())
if dist.is_available():
from jukebox.utils.dist_utils import allreduce
n_seen = allreduce(n_seen)
total = allreduce(total)
total_sq = allreduce(total_sq)
l1 = allreduce(l1)
spec_nelem = allreduce(spec_nelem)
spec_norm_total = allreduce(spec_norm_total)
mean = total / n_seen
bandwidth = dict(l2=total_sq / n_seen - mean**2,
l1=l1 / n_seen,
spec=spec_norm_total / spec_nelem)
print_once(bandwidth)
return bandwidth
def sample(self,
n_samples,
z=None,
z_conds=None,
y=None,
fp16=False,
temp=1.0,
top_k=0,
top_p=0.0,
chunk_size=None,
sample_tokens=None):
N = n_samples
if z is not None:
assert z.shape[
0] == N, f"Expected shape ({N},**), got shape {z.shape}"
if y is not None:
assert y.shape[
0] == N, f"Expected shape ({N},**), got shape {y.shape}"
if z_conds is not None:
for z_cond in z_conds:
assert z_cond.shape[
0] == N, f"Expected shape ({N},**), got shape {z_cond.shape}"
no_past_context = (z is None or z.shape[1] == 0)
if dist.get_rank() == 0:
name = {True: 'Ancestral', False: 'Primed'}[no_past_context]
print(
f"{name} sampling {n_samples} samples with temp={temp}, top_k={top_k}, top_p={top_p}"
)
with t.no_grad():
# Currently x_cond only uses immediately above layer
x_cond, y_cond, prime = self.get_cond(z_conds, y)
if self.single_enc_dec:
# assert chunk_size % self.prime_loss_dims == 0. TODO: Check if needed
if no_past_context:
z, x_cond = self.prior_preprocess([prime], [None, x_cond])
else:
z, x_cond = self.prior_preprocess([prime, z],
[None, x_cond])
if sample_tokens is not None:
sample_tokens += self.n_tokens
z = self.prior.primed_sample(n_samples,
z,
x_cond,
y_cond,
fp16=fp16,
temp=temp,
top_k=top_k,
top_p=top_p,
chunk_size=chunk_size,
sample_tokens=sample_tokens)
z = self.prior_postprocess(z)
else:
encoder_kv = self.get_encoder_kv(prime, fp16=fp16, sample=True)
if no_past_context:
z = self.prior.sample(n_samples,
x_cond,
y_cond,
encoder_kv,
fp16=fp16,
temp=temp,
top_k=top_k,
top_p=top_p,
sample_tokens=sample_tokens)
else:
z = self.prior.primed_sample(n_samples,
z,
x_cond,
y_cond,
encoder_kv,
fp16=fp16,
temp=temp,
top_k=top_k,
top_p=top_p,
chunk_size=chunk_size,
sample_tokens=sample_tokens)
if sample_tokens is None:
assert_shape(z, (N, *self.z_shape))
return z
def __init__(self,
z_shapes,
l_bins,
encoder,
decoder,
level,
downs_t,
strides_t,
labels,
prior_kwargs,
x_cond_kwargs,
y_cond_kwargs,
prime_kwargs,
copy_input,
labels_v3=False,
merged_decoder=False,
single_enc_dec=False):
super().__init__()
self.use_tokens = prime_kwargs.pop('use_tokens')
self.n_tokens = prime_kwargs.pop('n_tokens')
self.prime_loss_fraction = prime_kwargs.pop('prime_loss_fraction')
self.copy_input = copy_input
if self.copy_input:
prime_kwargs['bins'] = l_bins
self.z_shapes = z_shapes
self.levels = len(self.z_shapes)
self.z_shape = self.z_shapes[level]
self.level = level
assert level < self.levels, f"Total levels {self.levels}, got level {level}"
self.l_bins = l_bins
# Passing functions instead of the vqvae module to avoid getting params
self.encoder = encoder
self.decoder = decoder
# X conditioning
self.x_cond = (level != (self.levels - 1))
self.cond_level = level + 1
# Y conditioning
self.y_cond = labels
self.single_enc_dec = single_enc_dec
# X conditioning
if self.x_cond:
self.conditioner_blocks = nn.ModuleList()
conditioner_block = lambda _level: Conditioner(
input_shape=z_shapes[_level],
bins=l_bins,
down_t=downs_t[_level],
stride_t=strides_t[_level],
**x_cond_kwargs)
if dist.get_rank() == 0: print(f"Conditioning on 1 above level(s)")
self.conditioner_blocks.append(conditioner_block(self.cond_level))
# Y conditioning
if self.y_cond:
self.n_time = self.z_shape[
0] # Assuming STFT=TF order and raw=T1 order, so T is first dim
self.y_emb = LabelConditioner(n_time=self.n_time,
include_time_signal=not self.x_cond,
**y_cond_kwargs)
# Lyric conditioning
if single_enc_dec:
# Single encoder-decoder transformer
self.prior_shapes = [(self.n_tokens, ),
prior_kwargs.pop('input_shape')]
self.prior_bins = [prime_kwargs['bins'], prior_kwargs.pop('bins')]
self.prior_dims = [np.prod(shape) for shape in self.prior_shapes]
self.prior_bins_shift = np.cumsum([0, *self.prior_bins])[:-1]
self.prior_width = prior_kwargs['width']
print_once(
f'Creating cond. autoregress with prior bins {self.prior_bins}, '
)
print_once(f'dims {self.prior_dims}, ')
print_once(f'shift {self.prior_bins_shift}')
print_once(f'input shape {sum(self.prior_dims)}')
print_once(f'input bins {sum(self.prior_bins)}')
print_once(f'Self copy is {self.copy_input}')
self.prime_loss_dims, self.gen_loss_dims = self.prior_dims[
0], self.prior_dims[1]
self.total_loss_dims = self.prime_loss_dims + self.gen_loss_dims
self.prior = ConditionalAutoregressive2D(
input_shape=(sum(self.prior_dims), ),
bins=sum(self.prior_bins),
x_cond=(self.x_cond or self.y_cond),
y_cond=True,
prime_len=self.prime_loss_dims,
**prior_kwargs)
else:
# Separate encoder-decoder transformer
if self.n_tokens != 0 and self.use_tokens:
from app.jukebox.transformer.ops import Conv1D
prime_input_shape = (self.n_tokens, )
self.prime_loss_dims = np.prod(prime_input_shape)
self.prime_acts_width, self.prime_state_width = prime_kwargs[
'width'], prior_kwargs['width']
self.prime_prior = ConditionalAutoregressive2D(
input_shape=prime_input_shape,
x_cond=False,
y_cond=False,
only_encode=True,
**prime_kwargs)
self.prime_state_proj = Conv1D(
self.prime_acts_width,
self.prime_state_width,
init_scale=prime_kwargs['init_scale'])
self.prime_state_ln = LayerNorm(self.prime_state_width)
self.prime_bins = prime_kwargs['bins']
self.prime_x_out = nn.Linear(self.prime_state_width,
self.prime_bins,
bias=False)
nn.init.normal_(self.prime_x_out.weight,
std=0.02 * prior_kwargs['init_scale'])
else:
self.prime_loss_dims = 0
self.gen_loss_dims = np.prod(self.z_shape)
self.total_loss_dims = self.prime_loss_dims + self.gen_loss_dims
self.prior = ConditionalAutoregressive2D(
x_cond=(self.x_cond or self.y_cond),
y_cond=self.y_cond,
encoder_dims=self.prime_loss_dims,
merged_decoder=merged_decoder,
**prior_kwargs)
self.n_ctx = self.gen_loss_dims
self.downsamples = calculate_strides(strides_t, downs_t)
self.cond_downsample = self.downsamples[
level + 1] if level != self.levels - 1 else None
self.raw_to_tokens = np.prod(self.downsamples[:level + 1])
self.sample_length = self.n_ctx * self.raw_to_tokens
if labels:
self.labels_v3 = labels_v3
self.labeller = Labeller(self.y_emb.max_bow_genre_size,
self.n_tokens,
self.sample_length,
v3=self.labels_v3)
else:
self.labeller = EmptyLabeller()
print(
f"Level:{level}, Cond downsample:{self.cond_downsample}, Raw to tokens:{self.raw_to_tokens}, Sample length:{self.sample_length}"
)
def get_range(x):
if dist.get_rank() == 0:
return def_tqdm(x)
else:
return x
def print_all(msg):
if (not dist.is_available()):
print(msg)
elif dist.get_rank()%8==0:
print(f'{dist.get_rank()//8}: {msg}')
def print_once(msg):
if (not dist.is_available()) or dist.get_rank()==0:
print(msg)