def run_pretrain():
"""pre-train bert_clue"""
parser = argparse_init()
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
device_id=args_opt.device_id)
context.set_context(reserve_class_name_in_scope=False)
is_auto_enable_graph_kernel = _auto_enable_graph_kernel(
args_opt.device_target, args_opt.enable_graph_kernel)
_set_graph_kernel_context(args_opt.device_target,
args_opt.enable_graph_kernel,
is_auto_enable_graph_kernel)
ckpt_save_dir = args_opt.save_checkpoint_path
if args_opt.distribute == "true":
if args_opt.device_target == 'Ascend':
D.init()
device_num = args_opt.device_num
rank = args_opt.device_id % device_num
else:
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(
get_rank()) + '/'
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
device_num=device_num)
_set_bert_all_reduce_split()
else:
rank = 0
device_num = 1
_check_compute_type(args_opt, is_auto_enable_graph_kernel)
if args_opt.accumulation_steps > 1:
logger.info("accumulation steps: {}".format(
args_opt.accumulation_steps))
logger.info("global batch size: {}".format(
cfg.batch_size * args_opt.accumulation_steps))
if args_opt.enable_data_sink == "true":
args_opt.data_sink_steps *= args_opt.accumulation_steps
logger.info("data sink steps: {}".format(args_opt.data_sink_steps))
if args_opt.enable_save_ckpt == "true":
args_opt.save_checkpoint_steps *= args_opt.accumulation_steps
logger.info("save checkpoint steps: {}".format(
args_opt.save_checkpoint_steps))
ds = create_bert_dataset(device_num, rank, args_opt.do_shuffle,
args_opt.data_dir, args_opt.schema_dir)
net_with_loss = BertNetworkWithLoss(bert_net_cfg, True)
new_repeat_count = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.data_sink_steps
if args_opt.train_steps > 0:
train_steps = args_opt.train_steps * args_opt.accumulation_steps
new_repeat_count = min(new_repeat_count,
train_steps // args_opt.data_sink_steps)
else:
args_opt.train_steps = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.accumulation_steps
logger.info("train steps: {}".format(args_opt.train_steps))
optimizer = _get_optimizer(args_opt, net_with_loss)
callback = [
TimeMonitor(args_opt.data_sink_steps),
LossCallBack(ds.get_dataset_size())
]
if args_opt.enable_save_ckpt == "true" and args_opt.device_id % min(
8, device_num) == 0:
config_ck = CheckpointConfig(
save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(
prefix='checkpoint_bert',
directory=None if ckpt_save_dir == "" else ckpt_save_dir,
config=config_ck)
callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path)
load_param_into_net(net_with_loss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
accumulation_steps = args_opt.accumulation_steps
enable_global_norm = cfg.enable_global_norm
if accumulation_steps <= 1:
if cfg.optimizer == 'AdamWeightDecay' and args_opt.device_target == 'GPU':
net_with_grads = BertTrainOneStepWithLossScaleCellForAdam(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell)
else:
net_with_grads = BertTrainOneStepWithLossScaleCell(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell)
else:
allreduce_post = args_opt.distribute == "false" or args_opt.allreduce_post_accumulation == "true"
net_with_accumulation = (
BertTrainAccumulationAllReducePostWithLossScaleCell
if allreduce_post else
BertTrainAccumulationAllReduceEachWithLossScaleCell)
net_with_grads = net_with_accumulation(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell,
accumulation_steps=accumulation_steps,
enable_global_norm=enable_global_norm)
else:
net_with_grads = BertTrainOneStepCell(net_with_loss,
optimizer=optimizer)
model = Model(net_with_grads)
model = ConvertModelUtils().convert_to_thor_model(
model,
network=net_with_grads,
optimizer=optimizer,
frequency=cfg.Thor.frequency)
model.train(new_repeat_count,
ds,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"),
sink_size=args_opt.data_sink_steps)
def run_pretrain(args_opt):
"""pre-train bert"""
global device_id
global device_num
global rank_id
global job_id
args_opt.device_id = device_id
args_opt.device_num = device_num
sync_dataset(args_opt.data_url)
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
device_id=args_opt.device_id)
context.set_context(reserve_class_name_in_scope=False)
context.set_context(variable_memory_max_size="30GB")
ckpt_save_dir = args_opt.save_checkpoint_path
if args_opt.distribute == "true":
if args_opt.device_target == 'Ascend':
D.init('hccl')
device_num = args_opt.device_num
rank = args_opt.device_id % device_num
else:
D.init('nccl')
device_num = D.get_group_size()
rank = D.get_rank()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(
rank) + '/'
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True,
device_num=device_num)
from mindspore.parallel._auto_parallel_context import auto_parallel_context
if bert_net_cfg.num_hidden_layers == 12:
if bert_net_cfg.use_relative_positions:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[29, 58, 87, 116, 145, 174, 203, 217])
else:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[28, 55, 82, 109, 136, 163, 190, 205])
elif bert_net_cfg.num_hidden_layers == 24:
if bert_net_cfg.use_relative_positions:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[30, 90, 150, 210, 270, 330, 390, 421])
else:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[38, 93, 148, 203, 258, 313, 368, 397])
else:
rank = 0
device_num = 1
if args_opt.device_target == 'GPU' and bert_net_cfg.compute_type != mstype.float32:
logger.warning('Gpu only support fp32 temporarily, run with fp32.')
bert_net_cfg.compute_type = mstype.float32
ds, new_repeat_count = create_bert_dataset(args_opt.epoch_size, device_num,
rank, args_opt.do_shuffle,
args_opt.enable_data_sink,
args_opt.data_sink_steps,
args_opt.data_dir,
args_opt.schema_dir)
if args_opt.train_steps > 0:
new_repeat_count = min(
new_repeat_count, args_opt.train_steps // args_opt.data_sink_steps)
netwithloss = BertNetworkWithLoss(bert_net_cfg, True)
if cfg.optimizer == 'Lamb':
optimizer = Lamb(netwithloss.trainable_params(),
decay_steps=ds.get_dataset_size() * new_repeat_count,
start_learning_rate=cfg.Lamb.start_learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
power=cfg.Lamb.power,
warmup_steps=cfg.Lamb.warmup_steps,
weight_decay=cfg.Lamb.weight_decay,
eps=cfg.Lamb.eps)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(netwithloss.trainable_params(),
learning_rate=cfg.Momentum.learning_rate,
momentum=cfg.Momentum.momentum)
elif cfg.optimizer == 'AdamWeightDecayDynamicLR':
optimizer = AdamWeightDecayDynamicLR(
netwithloss.trainable_params(),
decay_steps=ds.get_dataset_size() * new_repeat_count,
learning_rate=cfg.AdamWeightDecayDynamicLR.learning_rate,
end_learning_rate=cfg.AdamWeightDecayDynamicLR.end_learning_rate,
power=cfg.AdamWeightDecayDynamicLR.power,
weight_decay=cfg.AdamWeightDecayDynamicLR.weight_decay,
eps=cfg.AdamWeightDecayDynamicLR.eps,
warmup_steps=cfg.AdamWeightDecayDynamicLR.warmup_steps)
else:
raise ValueError(
"Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecayDynamicLR]"
.format(cfg.optimizer))
callback = [TimeMonitor(ds.get_dataset_size()), LossCallBack()]
print("Enable save checkpoint: ", args_opt.enable_save_ckpt)
print("Rank ID: ", rank_id)
if args_opt.enable_save_ckpt == "true" and rank_id % device_num == 0:
print("Enable save checkpoint")
config_ck = CheckpointConfig(
save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_bert',
directory=ckpt_save_dir,
config=config_ck)
callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path)
load_param_into_net(netwithloss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
netwithgrads = BertTrainOneStepWithLossScaleCell(
netwithloss, optimizer=optimizer, scale_update_cell=update_cell)
else:
netwithgrads = BertTrainOneStepCell(netwithloss, optimizer=optimizer)
model = Model(netwithgrads)
model.train(new_repeat_count,
ds,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"))
def run_pretrain():
"""pre-train bert_clue"""
parser = argparse.ArgumentParser(description='bert pre_training')
parser.add_argument(
'--device_target',
type=str,
default='Ascend',
choices=['Ascend', 'GPU'],
help='device where the code will be implemented. (Default: Ascend)')
parser.add_argument("--distribute",
type=str,
default="false",
help="Run distribute, default is false.")
parser.add_argument("--epoch_size",
type=int,
default="1",
help="Epoch size, default is 1.")
parser.add_argument("--device_id",
type=int,
default=0,
help="Device id, default is 0.")
parser.add_argument("--device_num",
type=int,
default=1,
help="Use device nums, default is 1.")
parser.add_argument("--enable_save_ckpt",
type=str,
default="true",
help="Enable save checkpoint, default is true.")
parser.add_argument("--enable_lossscale",
type=str,
default="true",
help="Use lossscale or not, default is not.")
parser.add_argument("--do_shuffle",
type=str,
default="true",
help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink",
type=str,
default="true",
help="Enable data sink, default is true.")
parser.add_argument("--data_sink_steps",
type=int,
default="1",
help="Sink steps for each epoch, default is 1.")
parser.add_argument("--save_checkpoint_path",
type=str,
default="",
help="Save checkpoint path")
parser.add_argument("--load_checkpoint_path",
type=str,
default="",
help="Load checkpoint file path")
parser.add_argument("--save_checkpoint_steps",
type=int,
default=1000,
help="Save checkpoint steps, "
"default is 1000.")
parser.add_argument("--train_steps",
type=int,
default=-1,
help="Training Steps, default is -1, "
"meaning run all steps according to epoch number.")
parser.add_argument("--save_checkpoint_num",
type=int,
default=1,
help="Save checkpoint numbers, default is 1.")
parser.add_argument("--data_dir",
type=str,
default="",
help="Data path, it is better to use absolute path")
parser.add_argument("--schema_dir",
type=str,
default="",
help="Schema path, it is better to use absolute path")
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
device_id=args_opt.device_id)
context.set_context(reserve_class_name_in_scope=False)
context.set_context(variable_memory_max_size="30GB")
ckpt_save_dir = args_opt.save_checkpoint_path
if args_opt.distribute == "true":
if args_opt.device_target == 'Ascend':
D.init('hccl')
device_num = args_opt.device_num
rank = args_opt.device_id % device_num
else:
D.init('nccl')
device_num = D.get_group_size()
rank = D.get_rank()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(
rank) + '/'
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True,
device_num=device_num)
from mindspore.parallel._auto_parallel_context import auto_parallel_context
if bert_net_cfg.num_hidden_layers == 12:
if bert_net_cfg.use_relative_positions:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[29, 58, 87, 116, 145, 174, 203, 217])
else:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[28, 55, 82, 109, 136, 163, 190, 205])
elif bert_net_cfg.num_hidden_layers == 24:
if bert_net_cfg.use_relative_positions:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[30, 90, 150, 210, 270, 330, 390, 421])
else:
auto_parallel_context().set_all_reduce_fusion_split_indices(
[38, 93, 148, 203, 258, 313, 368, 397])
else:
rank = 0
device_num = 1
if args_opt.device_target == 'GPU' and bert_net_cfg.compute_type != mstype.float32:
logger.warning('Gpu only support fp32 temporarily, run with fp32.')
bert_net_cfg.compute_type = mstype.float32
ds, new_repeat_count = create_bert_dataset(args_opt.epoch_size, device_num,
rank, args_opt.do_shuffle,
args_opt.enable_data_sink,
args_opt.data_sink_steps,
args_opt.data_dir,
args_opt.schema_dir)
data_epoch_size = new_repeat_count // args_opt.epoch_size # Epoch nums in one dataset.
if args_opt.train_steps > 0:
new_repeat_count = min(
new_repeat_count, args_opt.train_steps // args_opt.data_sink_steps)
netwithloss = BertNetworkWithLoss(bert_net_cfg, True)
if cfg.optimizer == 'Lamb':
optimizer = Lamb(netwithloss.trainable_params(),
decay_steps=ds.get_dataset_size() * new_repeat_count,
start_learning_rate=cfg.Lamb.start_learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
power=cfg.Lamb.power,
warmup_steps=cfg.Lamb.warmup_steps,
weight_decay=cfg.Lamb.weight_decay,
eps=cfg.Lamb.eps)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(netwithloss.trainable_params(),
learning_rate=cfg.Momentum.learning_rate,
momentum=cfg.Momentum.momentum)
elif cfg.optimizer == 'AdamWeightDecayDynamicLR':
optimizer = AdamWeightDecayDynamicLR(
netwithloss.trainable_params(),
decay_steps=ds.get_dataset_size() * new_repeat_count,
learning_rate=cfg.AdamWeightDecayDynamicLR.learning_rate,
end_learning_rate=cfg.AdamWeightDecayDynamicLR.end_learning_rate,
power=cfg.AdamWeightDecayDynamicLR.power,
weight_decay=cfg.AdamWeightDecayDynamicLR.weight_decay,
eps=cfg.AdamWeightDecayDynamicLR.eps,
warmup_steps=cfg.AdamWeightDecayDynamicLR.warmup_steps)
else:
raise ValueError(
"Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecayDynamicLR]"
.format(cfg.optimizer))
callback = [
TimeMonitor(ds.get_dataset_size()),
LossCallBack(data_epoch_size)
]
if args_opt.enable_save_ckpt == "true":
config_ck = CheckpointConfig(
save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_bert',
directory=ckpt_save_dir,
config=config_ck)
callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path)
load_param_into_net(netwithloss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
netwithgrads = BertTrainOneStepWithLossScaleCell(
netwithloss, optimizer=optimizer, scale_update_cell=update_cell)
else:
netwithgrads = BertTrainOneStepCell(netwithloss, optimizer=optimizer)
model = Model(netwithgrads)
model.train(new_repeat_count,
ds,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"))
def run_pretrain():
"""pre-train bert_clue"""
parser = argparse.ArgumentParser(description='bert pre_training')
parser.add_argument(
'--device_target',
type=str,
default='Ascend',
choices=['Ascend', 'GPU'],
help='device where the code will be implemented. (Default: Ascend)')
parser.add_argument("--distribute",
type=str,
default="false",
choices=["true", "false"],
help="Run distribute, default is false.")
parser.add_argument("--epoch_size",
type=int,
default="1",
help="Epoch size, default is 1.")
parser.add_argument("--device_id",
type=int,
default=0,
help="Device id, default is 0.")
parser.add_argument("--device_num",
type=int,
default=1,
help="Use device nums, default is 1.")
parser.add_argument("--enable_save_ckpt",
type=str,
default="true",
choices=["true", "false"],
help="Enable save checkpoint, default is true.")
parser.add_argument("--enable_lossscale",
type=str,
default="true",
choices=["true", "false"],
help="Use lossscale or not, default is not.")
parser.add_argument("--do_shuffle",
type=str,
default="true",
choices=["true", "false"],
help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink",
type=str,
default="true",
choices=["true", "false"],
help="Enable data sink, default is true.")
parser.add_argument("--data_sink_steps",
type=int,
default="1",
help="Sink steps for each epoch, default is 1.")
parser.add_argument(
"--accumulation_steps",
type=int,
default="1",
help=
"Accumulating gradients N times before weight update, default is 1.")
parser.add_argument("--save_checkpoint_path",
type=str,
default="",
help="Save checkpoint path")
parser.add_argument("--load_checkpoint_path",
type=str,
default="",
help="Load checkpoint file path")
parser.add_argument("--save_checkpoint_steps",
type=int,
default=1000,
help="Save checkpoint steps, "
"default is 1000.")
parser.add_argument("--train_steps",
type=int,
default=-1,
help="Training Steps, default is -1, "
"meaning run all steps according to epoch number.")
parser.add_argument("--save_checkpoint_num",
type=int,
default=1,
help="Save checkpoint numbers, default is 1.")
parser.add_argument("--data_dir",
type=str,
default="",
help="Data path, it is better to use absolute path")
parser.add_argument("--schema_dir",
type=str,
default="",
help="Schema path, it is better to use absolute path")
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
device_id=args_opt.device_id)
context.set_context(reserve_class_name_in_scope=False)
ckpt_save_dir = args_opt.save_checkpoint_path
if args_opt.distribute == "true":
if args_opt.device_target == 'Ascend':
D.init()
device_num = args_opt.device_num
rank = args_opt.device_id % device_num
else:
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(
rank) + '/'
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
device_num=device_num)
if bert_net_cfg.num_hidden_layers == 12:
if bert_net_cfg.use_relative_positions:
context.set_auto_parallel_context(all_reduce_fusion_config=[
29, 58, 87, 116, 145, 174, 203, 217
])
else:
context.set_auto_parallel_context(all_reduce_fusion_config=[
28, 55, 82, 109, 136, 163, 190, 205
])
elif bert_net_cfg.num_hidden_layers == 24:
if bert_net_cfg.use_relative_positions:
context.set_auto_parallel_context(all_reduce_fusion_config=[
30, 90, 150, 210, 270, 330, 390, 421
])
else:
context.set_auto_parallel_context(all_reduce_fusion_config=[
38, 93, 148, 203, 258, 313, 368, 397
])
else:
rank = 0
device_num = 1
if args_opt.device_target == 'GPU' and bert_net_cfg.compute_type != mstype.float32:
logger.warning('Gpu only support fp32 temporarily, run with fp32.')
bert_net_cfg.compute_type = mstype.float32
if args_opt.accumulation_steps > 1:
logger.info("accumulation steps: {}".format(
args_opt.accumulation_steps))
logger.info("global batch size: {}".format(
bert_net_cfg.batch_size * args_opt.accumulation_steps))
if args_opt.enable_data_sink == "true":
args_opt.data_sink_steps *= args_opt.accumulation_steps
logger.info("data sink steps: {}".format(args_opt.data_sink_steps))
if args_opt.enable_save_ckpt == "true":
args_opt.save_checkpoint_steps *= args_opt.accumulation_steps
logger.info("save checkpoint steps: {}".format(
args_opt.save_checkpoint_steps))
ds = create_bert_dataset(device_num, rank, args_opt.do_shuffle,
args_opt.data_dir, args_opt.schema_dir)
net_with_loss = BertNetworkWithLoss(bert_net_cfg, True)
new_repeat_count = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.data_sink_steps
if args_opt.train_steps > 0:
train_steps = args_opt.train_steps * args_opt.accumulation_steps
new_repeat_count = min(new_repeat_count,
train_steps // args_opt.data_sink_steps)
else:
args_opt.train_steps = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.accumulation_steps
logger.info("train steps: {}".format(args_opt.train_steps))
if cfg.optimizer == 'Lamb':
lr_schedule = BertLearningRate(
learning_rate=cfg.Lamb.learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
warmup_steps=cfg.Lamb.warmup_steps,
decay_steps=args_opt.train_steps,
power=cfg.Lamb.power)
params = net_with_loss.trainable_params()
decay_params = list(filter(cfg.Lamb.decay_filter, params))
other_params = list(
filter(lambda x: not cfg.Lamb.decay_filter(x), params))
group_params = [{
'params': decay_params,
'weight_decay': cfg.Lamb.weight_decay
}, {
'params': other_params
}, {
'order_params': params
}]
optimizer = Lamb(group_params,
learning_rate=lr_schedule,
eps=cfg.Lamb.eps)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(net_with_loss.trainable_params(),
learning_rate=cfg.Momentum.learning_rate,
momentum=cfg.Momentum.momentum)
elif cfg.optimizer == 'AdamWeightDecay':
lr_schedule = BertLearningRate(
learning_rate=cfg.AdamWeightDecay.learning_rate,
end_learning_rate=cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=cfg.AdamWeightDecay.warmup_steps,
decay_steps=args_opt.train_steps,
power=cfg.AdamWeightDecay.power)
params = net_with_loss.trainable_params()
decay_params = list(filter(cfg.AdamWeightDecay.decay_filter, params))
other_params = list(
filter(lambda x: not cfg.AdamWeightDecay.decay_filter(x), params))
group_params = [{
'params': decay_params,
'weight_decay': cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}, {
'order_params': params
}]
optimizer = AdamWeightDecay(group_params,
learning_rate=lr_schedule,
eps=cfg.AdamWeightDecay.eps)
else:
raise ValueError(
"Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecay]"
.format(cfg.optimizer))
callback = [
TimeMonitor(args_opt.data_sink_steps),
LossCallBack(ds.get_dataset_size())
]
if args_opt.enable_save_ckpt == "true" and args_opt.device_id % min(
8, device_num) == 0:
config_ck = CheckpointConfig(
save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(
prefix='checkpoint_bert',
directory=None if ckpt_save_dir == "" else ckpt_save_dir,
config=config_ck)
callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path)
load_param_into_net(net_with_loss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
if args_opt.accumulation_steps <= 1:
net_with_grads = BertTrainOneStepWithLossScaleCell(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell)
else:
accumulation_steps = args_opt.accumulation_steps
net_with_grads = BertTrainAccumulateStepsWithLossScaleCell(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell,
accumulation_steps=accumulation_steps,
enable_global_norm=cfg.enable_global_norm)
else:
net_with_grads = BertTrainOneStepCell(net_with_loss,
optimizer=optimizer)
model = Model(net_with_grads)
model.train(new_repeat_count,
ds,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"),
sink_size=args_opt.data_sink_steps)
def run_pretrain():
"""pre-train bert_clue"""
parser = argparse.ArgumentParser(description='bert pre_training')
parser.add_argument(
'--device_target',
type=str,
default='Ascend',
choices=['Ascend', 'GPU'],
help='device where the code will be implemented. (Default: Ascend)')
parser.add_argument("--distribute",
type=str,
default="false",
choices=["true", "false"],
help="Run distribute, default is false.")
parser.add_argument("--epoch_size",
type=int,
default="1",
help="Epoch size, default is 1.")
parser.add_argument("--device_id",
type=int,
default=0,
help="Device id, default is 0.")
parser.add_argument("--device_num",
type=int,
default=1,
help="Use device nums, default is 1.")
parser.add_argument("--enable_save_ckpt",
type=str,
default="true",
choices=["true", "false"],
help="Enable save checkpoint, default is true.")
parser.add_argument("--enable_lossscale",
type=str,
default="true",
choices=["true", "false"],
help="Use lossscale or not, default is not.")
parser.add_argument("--do_shuffle",
type=str,
default="true",
choices=["true", "false"],
help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink",
type=str,
default="true",
choices=["true", "false"],
help="Enable data sink, default is true.")
parser.add_argument("--data_sink_steps",
type=int,
default="1",
help="Sink steps for each epoch, default is 1.")
parser.add_argument(
"--accumulation_steps",
type=int,
default="1",
help=
"Accumulating gradients N times before weight update, default is 1.")
parser.add_argument(
"--allreduce_post_accumulation",
type=str,
default="true",
choices=["true", "false"],
help=
"Whether to allreduce after accumulation of N steps or after each step, default is true."
)
parser.add_argument("--save_checkpoint_path",
type=str,
default="",
help="Save checkpoint path")
parser.add_argument("--load_checkpoint_path",
type=str,
default="",
help="Load checkpoint file path")
parser.add_argument("--save_checkpoint_steps",
type=int,
default=1000,
help="Save checkpoint steps, "
"default is 1000.")
parser.add_argument("--train_steps",
type=int,
default=-1,
help="Training Steps, default is -1, "
"meaning run all steps according to epoch number.")
parser.add_argument("--save_checkpoint_num",
type=int,
default=1,
help="Save checkpoint numbers, default is 1.")
parser.add_argument("--data_dir",
type=str,
default="",
help="Data path, it is better to use absolute path")
parser.add_argument("--schema_dir",
type=str,
default="",
help="Schema path, it is better to use absolute path")
parser.add_argument("--enable_graph_kernel",
type=str,
default="auto",
choices=["auto", "true", "false"],
help="Accelerate by graph kernel, default is auto.")
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
device_id=args_opt.device_id)
context.set_context(reserve_class_name_in_scope=False)
ckpt_save_dir = args_opt.save_checkpoint_path
if args_opt.distribute == "true":
if args_opt.device_target == 'Ascend':
D.init()
device_num = args_opt.device_num
rank = args_opt.device_id % device_num
else:
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(
get_rank()) + '/'
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
device_num=device_num)
if args_opt.device_target == 'Ascend':
_set_bert_all_reduce_split()
else:
rank = 0
device_num = 1
is_auto_enable_graph_kernel = _auto_enable_graph_kernel(
args_opt.device_target, args_opt.enable_graph_kernel)
if args_opt.enable_graph_kernel == "true" or is_auto_enable_graph_kernel:
context.set_context(enable_graph_kernel=True)
if args_opt.device_target == 'GPU' and bert_net_cfg.compute_type != mstype.float32 and \
not is_auto_enable_graph_kernel:
logger.warning('Gpu only support fp32 temporarily, run with fp32.')
bert_net_cfg.compute_type = mstype.float32
if args_opt.accumulation_steps > 1:
logger.info("accumulation steps: {}".format(
args_opt.accumulation_steps))
logger.info("global batch size: {}".format(
cfg.batch_size * args_opt.accumulation_steps))
if args_opt.enable_data_sink == "true":
args_opt.data_sink_steps *= args_opt.accumulation_steps
logger.info("data sink steps: {}".format(args_opt.data_sink_steps))
if args_opt.enable_save_ckpt == "true":
args_opt.save_checkpoint_steps *= args_opt.accumulation_steps
logger.info("save checkpoint steps: {}".format(
args_opt.save_checkpoint_steps))
ds = create_bert_dataset(device_num, rank, args_opt.do_shuffle,
args_opt.data_dir, args_opt.schema_dir)
net_with_loss = BertNetworkWithLoss(bert_net_cfg, True)
new_repeat_count = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.data_sink_steps
if args_opt.train_steps > 0:
train_steps = args_opt.train_steps * args_opt.accumulation_steps
new_repeat_count = min(new_repeat_count,
train_steps // args_opt.data_sink_steps)
else:
args_opt.train_steps = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.accumulation_steps
logger.info("train steps: {}".format(args_opt.train_steps))
optimizer = _get_optimizer(args_opt, net_with_loss)
callback = [
TimeMonitor(args_opt.data_sink_steps),
LossCallBack(ds.get_dataset_size())
]
if args_opt.enable_save_ckpt == "true" and args_opt.device_id % min(
8, device_num) == 0:
config_ck = CheckpointConfig(
save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(
prefix='checkpoint_bert',
directory=None if ckpt_save_dir == "" else ckpt_save_dir,
config=config_ck)
callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path)
load_param_into_net(net_with_loss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
accumulation_steps = args_opt.accumulation_steps
enable_global_norm = cfg.enable_global_norm
if accumulation_steps <= 1:
if cfg.optimizer == 'AdamWeightDecay' and args_opt.device_target == 'GPU':
net_with_grads = BertTrainOneStepWithLossScaleCellForAdam(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell)
else:
net_with_grads = BertTrainOneStepWithLossScaleCell(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell)
else:
allreduce_post = args_opt.distribute == "false" or args_opt.allreduce_post_accumulation == "true"
net_with_accumulation = (
BertTrainAccumulationAllReducePostWithLossScaleCell
if allreduce_post else
BertTrainAccumulationAllReduceEachWithLossScaleCell)
net_with_grads = net_with_accumulation(
net_with_loss,
optimizer=optimizer,
scale_update_cell=update_cell,
accumulation_steps=accumulation_steps,
enable_global_norm=enable_global_norm)
else:
net_with_grads = BertTrainOneStepCell(net_with_loss,
optimizer=optimizer)
model = Model(net_with_grads)
model.train(new_repeat_count,
ds,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"),
sink_size=args_opt.data_sink_steps)
def run_pretrain():
"""pre-train bert_clue"""
parser = argparse.ArgumentParser(description='bert pre_training')
parser.add_argument(
'--device_target',
type=str,
default='Ascend',
choices=['Ascend', 'GPU'],
help='device where the code will be implemented. (Default: Ascend)')
parser.add_argument("--distribute",
type=str,
default="false",
help="Run distribute, default is false.")
parser.add_argument("--epoch_size",
type=int,
default="1",
help="Epoch size, default is 1.")
parser.add_argument("--device_id",
type=int,
default=4,
help="Device id, default is 0.")
parser.add_argument("--device_num",
type=int,
default=1,
help="Use device nums, default is 1.")
parser.add_argument("--enable_save_ckpt",
type=str,
default="true",
help="Enable save checkpoint, default is true.")
parser.add_argument("--enable_lossscale",
type=str,
default="false",
help="Use lossscale or not, default is not.")
parser.add_argument("--do_shuffle",
type=str,
default="false",
help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink",
type=str,
default="true",
help="Enable data sink, default is true.")
parser.add_argument("--data_sink_steps",
type=int,
default="100",
help="Sink steps for each epoch, default is 1.")
parser.add_argument("--save_checkpoint_path",
type=str,
default="",
help="Save checkpoint path")
parser.add_argument("--load_checkpoint_path",
type=str,
default="",
help="Load checkpoint file path")
parser.add_argument("--save_checkpoint_steps",
type=int,
default=1000,
help="Save checkpoint steps, "
"default is 1000.")
parser.add_argument("--train_steps",
type=int,
default=-1,
help="Training Steps, default is -1, "
"meaning run all steps according to epoch number.")
parser.add_argument("--save_checkpoint_num",
type=int,
default=1,
help="Save checkpoint numbers, default is 1.")
parser.add_argument("--data_dir",
type=str,
default="",
help="Data path, it is better to use absolute path")
parser.add_argument("--schema_dir",
type=str,
default="",
help="Schema path, it is better to use absolute path")
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
device_id=args_opt.device_id,
save_graphs=False)
context.set_context(reserve_class_name_in_scope=False)
context.set_context(variable_memory_max_size="30GB")
context.set_context(max_call_depth=3000)
ckpt_save_dir = args_opt.save_checkpoint_path
if args_opt.distribute == "true":
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
ckpt_save_dir = args_opt.save_checkpoint_path + 'ckpt_' + str(
rank) + '/'
context.reset_auto_parallel_context()
_set_bert_all_reduce_split()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
gradients_mean=True,
device_num=device_num)
else:
rank = 0
device_num = 1
if args_opt.device_target == 'GPU' and bert_net_cfg.compute_type != mstype.float32:
logger.warning('Gpu only support fp32 temporarily, run with fp32.')
bert_net_cfg.compute_type = mstype.float32
ds = create_bert_dataset(device_num, rank, args_opt.do_shuffle,
args_opt.data_dir, args_opt.schema_dir)
net_with_loss = BertNetworkWithLoss(bert_net_cfg, True)
new_repeat_count = args_opt.epoch_size * ds.get_dataset_size(
) // args_opt.data_sink_steps
if args_opt.train_steps > 0:
new_repeat_count = min(
new_repeat_count, args_opt.train_steps // args_opt.data_sink_steps)
else:
args_opt.train_steps = args_opt.epoch_size * ds.get_dataset_size()
logger.info("train steps: {}".format(args_opt.train_steps))
optimizer = _get_optimizer(args_opt, net_with_loss)
callback = [TimeMonitor(args_opt.data_sink_steps), LossCallBack()]
if args_opt.enable_save_ckpt == "true" and rank == 0:
config_ck = CheckpointConfig(
save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_bert',
directory=ckpt_save_dir,
config=config_ck)
callback.append(ckpoint_cb)
if args_opt.load_checkpoint_path:
param_dict = load_checkpoint(args_opt.load_checkpoint_path)
load_param_into_net(net_with_loss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
net_with_grads = BertTrainOneStepWithLossScaleCell(
net_with_loss, optimizer=optimizer, scale_update_cell=update_cell)
else:
net_with_grads = BertTrainOneStepCell(net_with_loss,
optimizer=optimizer)
model = Model(net_with_grads, frequency=cfg.Thor.frequency)
model.train(new_repeat_count,
ds,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"),
sink_size=args_opt.data_sink_steps)
def run_pretrain():
"""pre-train bert_clue"""
parser = argparse.ArgumentParser(description='bert pre_training')
parser.add_argument("--distribute", type=str, default="false", help="Run distribute, default is false.")
parser.add_argument("--epoch_size", type=int, default="1", help="Epoch size, default is 1.")
parser.add_argument("--device_id", type=int, default=0, help="Device id, default is 0.")
parser.add_argument("--device_num", type=int, default=1, help="Use device nums, default is 1.")
parser.add_argument("--enable_save_ckpt", type=str, default="true", help="Enable save checkpoint, default is true.")
parser.add_argument("--enable_lossscale", type=str, default="true", help="Use lossscale or not, default is not.")
parser.add_argument("--do_shuffle", type=str, default="true", help="Enable shuffle for dataset, default is true.")
parser.add_argument("--enable_data_sink", type=str, default="true", help="Enable data sink, default is true.")
parser.add_argument("--data_sink_steps", type=int, default="1", help="Sink steps for each epoch, default is 1.")
parser.add_argument("--checkpoint_path", type=str, default="", help="Checkpoint file path")
parser.add_argument("--save_checkpoint_steps", type=int, default=1000, help="Save checkpoint steps, "
"default is 1000.")
parser.add_argument("--save_checkpoint_num", type=int, default=1, help="Save checkpoint numbers, default is 1.")
parser.add_argument("--data_dir", type=str, default="", help="Data path, it is better to use absolute path")
parser.add_argument("--schema_dir", type=str, default="", help="Schema path, it is better to use absolute path")
args_opt = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=args_opt.device_id)
context.set_context(reserve_class_name_in_scope=False)
if args_opt.distribute == "true":
device_num = args_opt.device_num
context.reset_auto_parallel_context()
context.set_auto_parallel_context(parallel_mode=ParallelMode.DATA_PARALLEL, mirror_mean=True,
device_num=device_num)
from mindspore.parallel._auto_parallel_context import auto_parallel_context
if bert_net_cfg.num_hidden_layers == 12:
auto_parallel_context().set_all_reduce_fusion_split_indices([28, 55, 82, 109, 136, 163, 190, 205])
elif bert_net_cfg.num_hidden_layers == 24:
auto_parallel_context().set_all_reduce_fusion_split_indices([38, 93, 148, 203, 258, 313, 368, 397])
D.init()
rank = args_opt.device_id % device_num
else:
rank = 0
device_num = 1
ds, new_repeat_count = create_bert_dataset(args_opt.epoch_size, device_num, rank, args_opt.do_shuffle,
args_opt.enable_data_sink, args_opt.data_sink_steps,
args_opt.data_dir, args_opt.schema_dir)
netwithloss = BertNetworkWithLoss(bert_net_cfg, True)
if cfg.optimizer == 'Lamb':
optimizer = Lamb(netwithloss.trainable_params(), decay_steps=ds.get_dataset_size() * ds.get_repeat_count(),
start_learning_rate=cfg.Lamb.start_learning_rate, end_learning_rate=cfg.Lamb.end_learning_rate,
power=cfg.Lamb.power, warmup_steps=cfg.Lamb.warmup_steps, weight_decay=cfg.Lamb.weight_decay,
eps=cfg.Lamb.eps)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(netwithloss.trainable_params(), learning_rate=cfg.Momentum.learning_rate,
momentum=cfg.Momentum.momentum)
elif cfg.optimizer == 'AdamWeightDecayDynamicLR':
optimizer = AdamWeightDecayDynamicLR(netwithloss.trainable_params(),
decay_steps=ds.get_dataset_size() * ds.get_repeat_count(),
learning_rate=cfg.AdamWeightDecayDynamicLR.learning_rate,
end_learning_rate=cfg.AdamWeightDecayDynamicLR.end_learning_rate,
power=cfg.AdamWeightDecayDynamicLR.power,
weight_decay=cfg.AdamWeightDecayDynamicLR.weight_decay,
eps=cfg.AdamWeightDecayDynamicLR.eps,
warmup_steps=cfg.AdamWeightDecayDynamicLR.warmup_steps)
else:
raise ValueError("Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecayDynamicLR]".
format(cfg.optimizer))
callback = [TimeMonitor(ds.get_dataset_size()), LossCallBack()]
if args_opt.enable_save_ckpt == "true":
config_ck = CheckpointConfig(save_checkpoint_steps=args_opt.save_checkpoint_steps,
keep_checkpoint_max=args_opt.save_checkpoint_num)
ckpoint_cb = ModelCheckpoint(prefix='checkpoint_bert', config=config_ck)
callback.append(ckpoint_cb)
if args_opt.checkpoint_path:
param_dict = load_checkpoint(args_opt.checkpoint_path)
load_param_into_net(netwithloss, param_dict)
if args_opt.enable_lossscale == "true":
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
netwithgrads = BertTrainOneStepWithLossScaleCell(netwithloss, optimizer=optimizer,
scale_update_cell=update_cell)
else:
netwithgrads = BertTrainOneStepCell(netwithloss, optimizer=optimizer)
model = Model(netwithgrads)
model.train(new_repeat_count, ds, callbacks=callback, dataset_sink_mode=(args_opt.enable_data_sink == "true"))
