def process_components(
model: Model,
criterion: Criterion = None,
optimizer: Optimizer = None,
scheduler: Scheduler = None,
distributed_params: Dict = None,
device: Device = None,
) -> Tuple[Model, Criterion, Optimizer, Scheduler, Device]:
"""
Returns the processed model, criterion, optimizer, scheduler and device
Args:
model (Model): torch model
criterion (Criterion): criterion function
optimizer (Optimizer): optimizer
scheduler (Scheduler): scheduler
distributed_params (dict, optional): dict with the parameters
for distributed and FP16 methond
device (Device, optional): device
"""
distributed_params = distributed_params or {}
distributed_params = copy.deepcopy(distributed_params)
if device is None:
device = utils.get_device()
model: Model = utils.maybe_recursive_call(model, "to", device=device)
if utils.is_wrapped_with_ddp(model):
pass
elif len(distributed_params) > 0:
assert isinstance(model, nn.Module)
distributed_rank = distributed_params.pop("rank", -1)
syncbn = distributed_params.pop("syncbn", False)
if distributed_rank > -1:
torch.cuda.set_device(distributed_rank)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
if "opt_level" in distributed_params:
utils.assert_fp16_available()
from apex import amp
amp_result = amp.initialize(model, optimizer, **distributed_params)
if optimizer is not None:
model, optimizer = amp_result
else:
model = amp_result
if distributed_rank > -1:
from apex.parallel import DistributedDataParallel
model = DistributedDataParallel(model)
if syncbn:
from apex.parallel import convert_syncbn_model
model = convert_syncbn_model(model)
if distributed_rank <= -1 and torch.cuda.device_count() > 1:
model = torch.nn.DataParallel(model)
elif torch.cuda.device_count() > 1:
if isinstance(model, nn.Module):
model = torch.nn.DataParallel(model)
elif isinstance(model, dict):
model = {k: torch.nn.DataParallel(v) for k, v in model.items()}
model: Model = utils.maybe_recursive_call(model, "to", device=device)
return model, criterion, optimizer, scheduler, device
def process_components(
model: Model,
criterion: Criterion = None,
optimizer: Optimizer = None,
scheduler: Scheduler = None,
distributed_params: Dict = None,
device: Device = None,
) -> Tuple[Model, Criterion, Optimizer, Scheduler, Device]:
"""
Returns the processed model, criterion, optimizer, scheduler and device
Args:
model (Model): torch model
criterion (Criterion): criterion function
optimizer (Optimizer): optimizer
scheduler (Scheduler): scheduler
distributed_params (dict, optional): dict with the parameters
for distributed and FP16 methond
device (Device, optional): device
"""
distributed_params = distributed_params or {}
distributed_params = copy.deepcopy(distributed_params)
distributed_params.update(get_distributed_params())
if device is None:
device = utils.get_device()
model: Model = utils.maybe_recursive_call(model, "to", device=device)
if utils.is_wrapped_with_ddp(model):
pass
elif get_rank() >= 0:
assert isinstance(model, nn.Module)
local_rank = distributed_params.pop("local_rank", 0)
device = f"cuda:{local_rank}"
model = utils.maybe_recursive_call(model, "to", device=device)
syncbn = distributed_params.pop("syncbn", False)
use_apex = distributed_params.pop("apex", True) and is_apex_available()
if use_apex:
import apex
amp_params = get_default_params(apex.amp.initialize,
["models", "optimizers"])
amp_params["opt_level"] = "O0"
for dp in distributed_params:
if dp in amp_params:
amp_params[dp] = distributed_params[dp]
amp_result = apex.amp.initialize(model, optimizer, **amp_params)
if optimizer is not None:
model, optimizer = amp_result
else:
model = amp_result
model = apex.parallel.DistributedDataParallel(model)
if syncbn:
model = apex.parallel.convert_syncbn_model(model)
else:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[local_rank], output_device=local_rank)
elif torch.cuda.device_count() > 1:
if isinstance(model, nn.Module):
model = torch.nn.DataParallel(model)
elif isinstance(model, dict):
model = {k: torch.nn.DataParallel(v) for k, v in model.items()}
model: Model = utils.maybe_recursive_call(model, "to", device=device)
return model, criterion, optimizer, scheduler, device
def process_components(
model: Model,
criterion: Criterion = None,
optimizer: Optimizer = None,
scheduler: Scheduler = None,
distributed_params: Dict = None,
device: Device = None,
) -> Tuple[Model, Criterion, Optimizer, Scheduler, Device]:
"""
Returns the processed model, criterion, optimizer, scheduler and device.
Args:
model (Model): torch model
criterion (Criterion): criterion function
optimizer (Optimizer): optimizer
scheduler (Scheduler): scheduler
distributed_params (dict, optional): dict with the parameters
for distributed and FP16 method
device (Device, optional): device
"""
distributed_params = distributed_params or {}
distributed_params = copy.deepcopy(distributed_params)
distributed_params.update(get_distributed_params())
if device is None:
device = get_device()
is_apex_available = (distributed_params.pop("apex", True)
and check_apex_available())
model: Model = maybe_recursive_call(model, "to", device=device)
if check_ddp_wrapped(model):
pass
# distributed data parallel run (ddp) (with apex support)
elif get_rank() >= 0:
assert isinstance(
model,
nn.Module), "Distributed training is not available for KV model"
local_rank = distributed_params.pop("local_rank", 0) or 0
device = f"cuda:{local_rank}"
model = maybe_recursive_call(model, "to", device=device)
syncbn = distributed_params.pop("syncbn", False)
if is_apex_available:
import apex
model, optimizer = initialize_apex(model, optimizer,
**distributed_params)
model = apex.parallel.DistributedDataParallel(model)
if syncbn:
model = apex.parallel.convert_syncbn_model(model)
else:
model = nn.parallel.DistributedDataParallel(
model,
device_ids=[local_rank],
output_device=local_rank,
find_unused_parameters=True)
# data parallel run (dp) (with apex support)
else:
# apex issue https://github.com/deepset-ai/FARM/issues/210
use_apex = (is_apex_available and torch.cuda.device_count() == 1) or (
is_apex_available and torch.cuda.device_count() > 1
and distributed_params.get("opt_level", "O0") == "O1")
if use_apex:
assert isinstance(
model,
nn.Module), "Apex training is not available for KV model"
model, optimizer = initialize_apex(model, optimizer,
**distributed_params)
if torch.cuda.device_count() > 1:
if isinstance(model, nn.Module):
model = nn.DataParallel(model)
elif isinstance(model, dict):
model = {k: nn.DataParallel(v) for k, v in model.items()}
else:
raise NotImplementedError()
model: Model = maybe_recursive_call(model, "to", device=device)
return model, criterion, optimizer, scheduler, device
