def get_optimizer(model, hps):
# Optimizer
betas = (hps.beta1, hps.beta2)
if hps.fp16_opt:
opt = FP16FusedAdam(model.parameters(),
lr=hps.lr,
weight_decay=hps.weight_decay,
betas=betas,
eps=hps.eps)
else:
opt = FusedAdam(model.parameters(),
lr=hps.lr,
weight_decay=hps.weight_decay,
betas=betas,
eps=hps.eps)
# lr scheduler
shd = get_lr_scheduler(opt, hps)
restore_path = hps.restore_prior if hps.prior else hps.restore_vqvae
restore_opt(opt, shd, restore_path)
# fp16 dynamic loss scaler
scalar = None
if hps.fp16:
rank = dist.get_rank()
local_rank = rank % 8
scalar = LossScalar(hps.fp16_loss_scale,
scale_factor=2**(1. / hps.fp16_scale_window))
if local_rank == 0: print(scalar.__dict__)
zero_grad(model)
return opt, shd, scalar
def get_ddp(model, hps):
rank = dist.get_rank()
local_rank = rank % 8
ddp = DistributedDataParallel(model,
device_ids=[local_rank],
output_device=local_rank,
broadcast_buffers=False,
bucket_cap_mb=hps.bucket)
return ddp
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 lr_lambda(step):
if hps.lr_use_linear_decay:
lr_scale = hps.lr_scale * min(1.0, step / hps.lr_warmup)
decay = max(
0.0, 1.0 -
max(0.0, step - hps.lr_start_linear_decay) / hps.lr_decay)
if decay == 0.0:
if dist.get_rank() == 0:
print("Reached end of training")
return lr_scale * decay
else:
return hps.lr_scale * (hps.lr_gamma**(step // hps.lr_decay)) * min(
1.0, step / hps.lr_warmup)
def get_ema(model, hps):
mu = hps.mu or (1. - (hps.bs * hps.ngpus / 8.) / 1000)
ema = None
if hps.ema and hps.train:
if hps.cpu_ema:
if dist.get_rank() == 0:
print("Using CPU EMA")
ema = CPUEMA(model.parameters(), mu=mu, freq=hps.cpu_ema_freq)
elif hps.ema_fused:
ema = FusedEMA(model.parameters(), mu=mu)
else:
ema = EMA(model.parameters(), mu=mu)
return ema
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 __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 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
from 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 print_once(msg):
if (not dist.is_available()) or dist.get_rank() == 0:
print(msg)
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 train(model, orig_model, opt, shd, scalar, ema, logger, metrics,
data_processor, hps):
model.train()
orig_model.train()
if hps.prior:
_print_keys = dict(l="loss",
bpd="bpd",
gn="gn",
g_l="gen_loss",
p_l="prime_loss")
else:
_print_keys = dict(l="loss",
sl="spectral_loss",
rl="recons_loss",
e="entropy",
u="usage",
uc="used_curr",
gn="gn",
pn="pn",
dk="dk")
for i, x in logger.get_range(data_processor.train_loader):
if isinstance(x, (tuple, list)):
x, y = x
else:
y = None
x = x.to('cuda', non_blocking=True)
if y is not None:
y = y.to('cuda', non_blocking=True)
x_in = x = audio_preprocess(x, hps)
log_input_output = (logger.iters % hps.save_iters == 0)
if hps.prior:
forw_kwargs = dict(y=y, fp16=hps.fp16, decode=log_input_output)
else:
forw_kwargs = dict(loss_fn=hps.loss_fn, hps=hps)
# Forward
x_out, loss, _metrics = model(x, **forw_kwargs)
# Backward
loss, scale, grad_norm, overflow_loss, overflow_grad = backward(
loss=loss,
params=list(model.parameters()),
scalar=scalar,
fp16=hps.fp16,
logger=logger)
# Skip step if overflow
grad_norm = allreduce(grad_norm, op=dist.ReduceOp.MAX)
if overflow_loss or overflow_grad or grad_norm > hps.ignore_grad_norm > 0:
zero_grad(orig_model)
continue
# Step opt. Divide by scale to include clipping and fp16 scaling
logger.step()
opt.step(scale=clipped_grad_scale(grad_norm, hps.clip, scale))
zero_grad(orig_model)
lr = hps.lr if shd is None else shd.get_lr()[0]
if shd is not None: shd.step()
if ema is not None: ema.step()
next_lr = hps.lr if shd is None else shd.get_lr()[0]
finished_training = (next_lr == 0.0)
# Logging
for key, val in _metrics.items():
_metrics[key] = val.item()
_metrics["loss"] = loss = loss.item(
) * hps.iters_before_update # Make sure to call to free graph
_metrics["gn"] = grad_norm
_metrics["lr"] = lr
_metrics["lg_loss_scale"] = np.log2(scale)
# Average and log
for key, val in _metrics.items():
_metrics[key] = metrics.update(key, val, x.shape[0])
if logger.iters % hps.log_steps == 0:
logger.add_scalar(key, _metrics[key])
# Save checkpoint
with t.no_grad():
if hps.save and (logger.iters % hps.save_iters == 1
or finished_training):
if ema is not None: ema.swap()
orig_model.eval()
name = 'latest' if hps.prior else f'step_{logger.iters}'
if dist.get_rank() % 8 == 0:
save_checkpoint(logger, name, orig_model, opt,
dict(step=logger.iters), hps)
orig_model.train()
if ema is not None: ema.swap()
# Sample
with t.no_grad():
if (logger.iters % 12000) in list(
range(1,
1 + hps.iters_before_update)) or finished_training:
if hps.prior:
sample_prior(orig_model, ema, logger, x_in, y, hps)
# Input/Output
with t.no_grad():
if log_input_output:
log_inputs(orig_model, logger, x_in, y, x_out, hps)
logger.set_postfix(**{
print_key: _metrics[key]
for print_key, key in _print_keys.items()
})
if finished_training:
dist.barrier()
exit()
logger.close_range()
return {key: metrics.avg(key) for key in _metrics.keys()}
def get_range(x):
if dist.get_rank() == 0:
return def_tqdm(x)
else:
return x