def do_train(dataset=None,
network=None,
load_checkpoint_path="",
save_checkpoint_path="",
epoch_num=1):
""" do train """
if load_checkpoint_path == "":
raise ValueError(
"Pretrain model missed, finetune task must load pretrain model!")
steps_per_epoch = dataset.get_dataset_size()
# optimizer
optimizer = Adam(network.trainable_params(),
learning_rate=optimizer_cfg.learning_rate)
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch,
keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(
prefix="classifier",
directory=None if save_checkpoint_path == "" else save_checkpoint_path,
config=ckpt_config)
param_dict = load_checkpoint(load_checkpoint_path)
load_param_into_net(network, param_dict)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32,
scale_factor=2,
scale_window=1000)
netwithgrads = BertFinetuneCell(network,
optimizer=optimizer,
scale_update_cell=update_cell)
model = Model(netwithgrads)
callbacks = [
TimeMonitor(dataset.get_dataset_size()),
LossCallBack(dataset.get_dataset_size()), ckpoint_cb
]
model.train(epoch_num, dataset, callbacks=callbacks)
def run_predistill():
"""
run predistill
"""
cfg = phase1_cfg
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)
load_teacher_checkpoint_path = args_opt.load_teacher_ckpt_path
load_student_checkpoint_path = args_opt.load_gd_ckpt_path
netwithloss = BertNetworkWithLoss_td(teacher_config=td_teacher_net_cfg, teacher_ckpt=load_teacher_checkpoint_path,
student_config=td_student_net_cfg, student_ckpt=load_student_checkpoint_path,
is_training=True, task_type='classification',
num_labels=args_opt.num_labels, is_predistill=True)
rank = 0
device_num = 1
dataset = create_tinybert_dataset('td', td_teacher_net_cfg.batch_size,
device_num, rank, args_opt.do_shuffle,
args_opt.train_data_dir, args_opt.schema_dir)
dataset_size = dataset.get_dataset_size()
print('td1 dataset size: ', dataset_size)
if args_opt.enable_data_sink == 'true':
repeat_count = args_opt.td_phase1_epoch_size * dataset.get_dataset_size() // args_opt.data_sink_steps
time_monitor_steps = args_opt.data_sink_steps
else:
repeat_count = args_opt.td_phase1_epoch_size
time_monitor_steps = dataset_size
optimizer_cfg = cfg.optimizer_cfg
lr_schedule = BertLearningRate(learning_rate=optimizer_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=optimizer_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(dataset_size / 10),
decay_steps=int(dataset_size * args_opt.td_phase1_epoch_size),
power=optimizer_cfg.AdamWeightDecay.power)
params = netwithloss.trainable_params()
decay_params = list(filter(optimizer_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': optimizer_cfg.AdamWeightDecay.weight_decay},
{'params': other_params, 'weight_decay': 0.0},
{'order_params': params}]
optimizer = AdamWeightDecay(group_params, learning_rate=lr_schedule, eps=optimizer_cfg.AdamWeightDecay.eps)
callback = [TimeMonitor(time_monitor_steps), LossCallBack(), ModelSaveCkpt(netwithloss.bert,
args_opt.save_ckpt_step,
args_opt.max_ckpt_num,
td_phase1_save_ckpt_dir)]
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
netwithgrads = BertEvaluationCell(netwithloss, optimizer=optimizer, scale_update_cell=update_cell)
model = Model(netwithgrads)
model.train(repeat_count, dataset, callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == 'true'),
sink_size=args_opt.data_sink_steps)
def compile_net(net, grad_accumulation_step):
context.set_context(save_graphs=True)
learning_rate = 0.1
momentum = 0.9
epoch_size = 2
dataset = Dataset(_x, _b)
opt = Momentum(net.trainable_params(), learning_rate, momentum)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=65536, scale_factor=2, scale_window=1000)
net_wrap = TrainAccumulateStepsWithLossScaleCell(net, opt, scale_update_cell=update_cell,
accumulation_steps=grad_accumulation_step)
model = Model(net_wrap)
model.train(epoch_size, dataset, dataset_sink_mode=False)
context.reset_auto_parallel_context()
def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1):
""" do train """
if load_checkpoint_path == "":
raise ValueError("Pretrain model missed, finetune task must load pretrain model!")
steps_per_epoch = dataset.get_dataset_size()
# optimizer
if optimizer_cfg.optimizer == 'AdamWeightDecay':
lr_schedule = BertLearningRate(learning_rate=optimizer_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=optimizer_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=optimizer_cfg.AdamWeightDecay.power)
params = network.trainable_params()
decay_params = list(filter(optimizer_cfg.AdamWeightDecay.decay_filter, params))
other_params = list(filter(lambda x: not optimizer_cfg.AdamWeightDecay.decay_filter(x), params))
group_params = [{'params': decay_params, 'weight_decay': optimizer_cfg.AdamWeightDecay.weight_decay},
{'params': other_params, 'weight_decay': 0.0}]
optimizer = AdamWeightDecay(group_params, lr_schedule, eps=optimizer_cfg.AdamWeightDecay.eps)
elif optimizer_cfg.optimizer == 'Lamb':
lr_schedule = BertLearningRate(learning_rate=optimizer_cfg.Lamb.learning_rate,
end_learning_rate=optimizer_cfg.Lamb.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=optimizer_cfg.Lamb.power)
optimizer = Lamb(network.trainable_params(), learning_rate=lr_schedule)
elif optimizer_cfg.optimizer == 'Momentum':
optimizer = Momentum(network.trainable_params(), learning_rate=optimizer_cfg.Momentum.learning_rate,
momentum=optimizer_cfg.Momentum.momentum)
else:
raise Exception("Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]")
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch, keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(prefix="ner",
directory=None if save_checkpoint_path == "" else save_checkpoint_path,
config=ckpt_config)
param_dict = load_checkpoint(load_checkpoint_path)
load_param_into_net(network, param_dict)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32, scale_factor=2, scale_window=1000)
netwithgrads = BertFinetuneCell(network, optimizer=optimizer, scale_update_cell=update_cell)
model = Model(netwithgrads)
callbacks = [TimeMonitor(dataset.get_dataset_size()), LossCallBack(dataset.get_dataset_size()), ckpoint_cb]
train_begin = time.time()
model.train(epoch_num, dataset, callbacks=callbacks)
train_end = time.time()
print("latency: {:.6f} s".format(train_end - train_begin))
def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path=""):
""" do train """
if load_checkpoint_path == "":
raise ValueError("Pretrain model missed, finetune task must load pretrain model!")
steps_per_epoch = dataset.get_dataset_size()
epoch_num = dataset.get_repeat_count()
# optimizer
if optimizer_cfg.optimizer == 'AdamWeightDecayDynamicLR':
optimizer = AdamWeightDecayDynamicLR(network.trainable_params(),
decay_steps=steps_per_epoch * epoch_num,
learning_rate=optimizer_cfg.AdamWeightDecayDynamicLR.learning_rate,
end_learning_rate=optimizer_cfg.AdamWeightDecayDynamicLR.end_learning_rate,
power=optimizer_cfg.AdamWeightDecayDynamicLR.power,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
weight_decay=optimizer_cfg.AdamWeightDecayDynamicLR.weight_decay,
eps=optimizer_cfg.AdamWeightDecayDynamicLR.eps)
elif optimizer_cfg.optimizer == 'Lamb':
optimizer = Lamb(network.trainable_params(), decay_steps=steps_per_epoch * epoch_num,
start_learning_rate=optimizer_cfg.Lamb.start_learning_rate,
end_learning_rate=optimizer_cfg.Lamb.end_learning_rate,
power=optimizer_cfg.Lamb.power, weight_decay=optimizer_cfg.Lamb.weight_decay,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_filter=optimizer_cfg.Lamb.decay_filter)
elif optimizer_cfg.optimizer == 'Momentum':
optimizer = Momentum(network.trainable_params(), learning_rate=optimizer_cfg.Momentum.learning_rate,
momentum=optimizer_cfg.Momentum.momentum)
else:
raise Exception("Optimizer not supported. support: [AdamWeightDecayDynamicLR, Lamb, Momentum]")
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch, keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(prefix="classifier", directory=save_checkpoint_path, config=ckpt_config)
param_dict = load_checkpoint(load_checkpoint_path)
load_param_into_net(network, param_dict)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32, scale_factor=2, scale_window=1000)
netwithgrads = BertFinetuneCell(network, optimizer=optimizer, scale_update_cell=update_cell)
model = Model(netwithgrads)
callbacks = [TimeMonitor(dataset.get_dataset_size()), LossCallBack(), ckpoint_cb]
model.train(epoch_num, dataset, callbacks=callbacks)
def run_general_distill():
"""
run general distill
"""
parser = argparse.ArgumentParser(description='tinybert general distill')
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="3",
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("--save_ckpt_step",
type=int,
default=100,
help="Enable data sink, default is true.")
parser.add_argument("--max_ckpt_num",
type=int,
default=1,
help="Enable data sink, default is true.")
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("--save_ckpt_path",
type=str,
default="",
help="Save checkpoint path")
parser.add_argument("--load_teacher_ckpt_path",
type=str,
default="",
help="Load checkpoint file path")
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(
"--dataset_type",
type=str,
default="tfrecord",
help="dataset type tfrecord/mindrecord, default is tfrecord")
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")
save_ckpt_dir = os.path.join(
args_opt.save_ckpt_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
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()
save_ckpt_dir = save_ckpt_dir + '_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)
else:
rank = 0
device_num = 1
if not os.path.exists(save_ckpt_dir):
os.makedirs(save_ckpt_dir)
enable_loss_scale = True
if args_opt.device_target == "GPU":
if bert_student_net_cfg.compute_type != mstype.float32:
logger.warning(
'Compute about the student only support float32 temporarily, run with float32.'
)
bert_student_net_cfg.compute_type = mstype.float32
# Backward of the network are calculated using fp32,
# and the loss scale is not necessary
enable_loss_scale = False
netwithloss = BertNetworkWithLoss_gd(
teacher_config=bert_teacher_net_cfg,
teacher_ckpt=args_opt.load_teacher_ckpt_path,
student_config=bert_student_net_cfg,
is_training=True,
use_one_hot_embeddings=False)
if args_opt.dataset_type == "tfrecord":
dataset_type = DataType.TFRECORD
elif args_opt.dataset_type == "mindrecord":
dataset_type = DataType.MINDRECORD
else:
raise Exception("dataset format is not supported yet")
dataset = create_tinybert_dataset('gd',
common_cfg.batch_size,
device_num,
rank,
args_opt.do_shuffle,
args_opt.data_dir,
args_opt.schema_dir,
data_type=dataset_type)
dataset_size = dataset.get_dataset_size()
print('dataset size: ', dataset_size)
print("dataset repeatcount: ", dataset.get_repeat_count())
if args_opt.enable_data_sink == "true":
repeat_count = args_opt.epoch_size * dataset_size // args_opt.data_sink_steps
time_monitor_steps = args_opt.data_sink_steps
else:
repeat_count = args_opt.epoch_size
time_monitor_steps = dataset_size
lr_schedule = BertLearningRate(
learning_rate=common_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=common_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(dataset_size * args_opt.epoch_size / 10),
decay_steps=int(dataset_size * args_opt.epoch_size),
power=common_cfg.AdamWeightDecay.power)
params = netwithloss.trainable_params()
decay_params = list(filter(common_cfg.AdamWeightDecay.decay_filter,
params))
other_params = list(
filter(lambda x: not common_cfg.AdamWeightDecay.decay_filter(x),
params))
group_params = [{
'params': decay_params,
'weight_decay': common_cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}, {
'order_params': params
}]
optimizer = AdamWeightDecay(group_params,
learning_rate=lr_schedule,
eps=common_cfg.AdamWeightDecay.eps)
callback = [
TimeMonitor(time_monitor_steps),
LossCallBack(),
ModelSaveCkpt(netwithloss.bert, args_opt.save_ckpt_step,
args_opt.max_ckpt_num, save_ckpt_dir)
]
if enable_loss_scale:
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=common_cfg.loss_scale_value,
scale_factor=common_cfg.scale_factor,
scale_window=common_cfg.scale_window)
netwithgrads = BertTrainWithLossScaleCell(
netwithloss, optimizer=optimizer, scale_update_cell=update_cell)
else:
netwithgrads = BertTrainCell(netwithloss, optimizer=optimizer)
model = Model(netwithgrads)
model.train(repeat_count,
dataset,
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=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 do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1,resume=False):
"""
Do train
Args:
dataset: the train dataset.
network: the network with loss
load_checkpoint_path: the file path which saved pretrain model checkpoint.
save_checkpoint_path: the file path which will save finetune model checkpoint.
epoch_num: the number of epoch
"""
if load_checkpoint_path == "":
raise ValueError("Pretrain model missed, finetune task must load pretrain model!")
steps_per_epoch = dataset.get_dataset_size() # samples / batch_size
#print info
print("="*30,"TRAIN INFO","="*30)
print("optimizer: {}".format(cfg.optimizer))
#Select Optimizer
if cfg.optimizer == 'AdamWeightDecay':
lr_schedule = GPT2LearningRate(learning_rate=cfg.AdamWeightDecay.learning_rate,
end_learning_rate=cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=cfg.AdamWeightDecay.power)
params = network.trainable_params() # return a list of all trainable parmeters of the network
# Use parameter groups and set different values
decay_params = list(filter(cfg.AdamWeightDecay.decay_filter, params)) # without layernorm and bias
other_params = list(filter(lambda x: not cfg.AdamWeightDecay.decay_filter(x), params)) # with layernorm and bias
group_params = [{'params': decay_params, 'weight_decay': cfg.AdamWeightDecay.weight_decay},
{'params': other_params, 'weight_decay': 0.0}]
optimizer = AdamWeightDecay(group_params, lr_schedule, eps=cfg.AdamWeightDecay.eps)
elif cfg.optimizer == 'Lamb':
#print info
print("lr: {}".format(cfg.Lamb.learning_rate))
print("end_learning_rate: {}".format(cfg.Lamb.end_learning_rate))
#print("warmup_steps: {}".format(int(steps_per_epoch * epoch_num * 0.1)))
print("power: {}".format(cfg.Lamb.power))
lr_schedule = GPT2LearningRate(learning_rate=cfg.Lamb.learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=cfg.Lamb.power)
optimizer = Lamb(network.trainable_params(), lr_schedule)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(network.trainable_params(), cfg.Momentum.learning_rate, cfg.Momentum.momentum)
else:
raise Exception("Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]")
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch, keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(prefix="gpt2_summarization",
directory=None if save_checkpoint_path == "" else save_checkpoint_path,
config=ckpt_config)
param_dict = load_checkpoint(load_checkpoint_path)
reorganized_param_dict = dict()
if resume == False :
print("Do not resume.\nRESUME STATE: {}".format(resume))
for netName in param_dict:
reorganized_param_dict['gpt2.gpt2.'+netName] = param_dict[netName]
reorganized_param_dict['gpt2.lm_head.weight'] = param_dict['gpt2_embedding_lookup.embedding_table']
else:
print("Start to resume training.\nRESUME STATE: {}".format(resume))
reorganized_param_dict = param_dict
load_param_into_net(network, reorganized_param_dict)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32, scale_factor=2, scale_window=1000)
netwithgrads = GPT2FinetuneCell(network, optimizer=optimizer, scale_update_cell=update_cell)
netwithgrads.set_train(True)
loss_cb = LossMonitor(per_print_times=1)
model = Model(netwithgrads)
callbacks = [TimeMonitor(dataset.get_dataset_size()), loss_cb, ckpoint_cb]
print("============== Starting Training For Summrization Task ==============")
model.train(epoch_num, dataset, callbacks=callbacks, dataset_sink_mode=False)
print("============== Summrization Training Success ==============")
def train():
'''
finetune function
'''
# BertCLS train for classification
# BertNER train for sequence labeling
if cfg.task == 'NER':
tag_to_index = None
if cfg.use_crf:
tag_to_index = json.loads(open(cfg.label2id_file).read())
print(tag_to_index)
max_val = len(tag_to_index)
tag_to_index["<START>"] = max_val
tag_to_index["<STOP>"] = max_val + 1
number_labels = len(tag_to_index)
else:
number_labels = cfg.num_labels
netwithloss = BertNER(bert_net_cfg,
cfg.batch_size,
True,
num_labels=number_labels,
use_crf=cfg.use_crf,
tag_to_index=tag_to_index,
dropout_prob=0.1)
elif cfg.task == 'Classification':
netwithloss = BertCLS(bert_net_cfg,
True,
num_labels=cfg.num_labels,
dropout_prob=0.1,
assessment_method=cfg.assessment_method)
else:
raise Exception("task error, NER or Classification is supported.")
dataset = get_dataset(data_file=cfg.data_file, batch_size=cfg.batch_size)
steps_per_epoch = dataset.get_dataset_size()
print('steps_per_epoch:', steps_per_epoch)
# optimizer
steps_per_epoch = dataset.get_dataset_size()
if cfg.optimizer == 'AdamWeightDecay':
lr_schedule = BertLearningRate(
learning_rate=optimizer_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=optimizer_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(steps_per_epoch * cfg.epoch_num * 0.1),
decay_steps=steps_per_epoch * cfg.epoch_num,
power=optimizer_cfg.AdamWeightDecay.power)
params = netwithloss.trainable_params()
decay_params = list(
filter(optimizer_cfg.AdamWeightDecay.decay_filter, params))
other_params = list(
filter(lambda x: not optimizer_cfg.AdamWeightDecay.decay_filter(x),
params))
group_params = [{
'params':
decay_params,
'weight_decay':
optimizer_cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}]
optimizer = AdamWeightDecay(group_params,
lr_schedule,
eps=optimizer_cfg.AdamWeightDecay.eps)
elif cfg.optimizer == 'Lamb':
lr_schedule = BertLearningRate(
learning_rate=optimizer_cfg.Lamb.learning_rate,
end_learning_rate=optimizer_cfg.Lamb.end_learning_rate,
warmup_steps=int(steps_per_epoch * cfg.epoch_num * 0.1),
decay_steps=steps_per_epoch * cfg.epoch_num,
power=optimizer_cfg.Lamb.power)
optimizer = Lamb(netwithloss.trainable_params(),
learning_rate=lr_schedule)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(
netwithloss.trainable_params(),
learning_rate=optimizer_cfg.Momentum.learning_rate,
momentum=optimizer_cfg.Momentum.momentum)
else:
raise Exception("Optimizer not supported.")
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch,
keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(prefix=cfg.ckpt_prefix,
directory=cfg.ckpt_dir,
config=ckpt_config)
param_dict = load_checkpoint(cfg.pre_training_ckpt)
load_param_into_net(netwithloss, param_dict)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32,
scale_factor=2,
scale_window=1000)
netwithgrads = BertFinetuneCell(netwithloss,
optimizer=optimizer,
scale_update_cell=update_cell)
model = Model(netwithgrads)
callbacks = [
TimeMonitor(dataset.get_dataset_size()),
LossCallBack(dataset.get_dataset_size()), ckpoint_cb
]
model.train(cfg.epoch_num,
dataset,
callbacks=callbacks,
dataset_sink_mode=True)
def run_general_distill():
"""
run general distill
"""
parser = argparse.ArgumentParser(description='tinybert general distill')
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="3",
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("--save_ckpt_step",
type=int,
default=100,
help="Enable data sink, default is true.")
parser.add_argument("--max_ckpt_num",
type=int,
default=1,
help="Enable data sink, default is true.")
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_ckpt_path",
type=str,
default="",
help="Save checkpoint path")
parser.add_argument("--load_teacher_ckpt_path",
type=str,
default="",
help="Load checkpoint file path")
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(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")
save_ckpt_dir = os.path.join(
args_opt.save_ckpt_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
if not os.path.exists(save_ckpt_dir):
os.makedirs(save_ckpt_dir)
if args_opt.distribute == "true":
D.init('hccl')
device_num = args_opt.device_num
rank = args_opt.device_id % device_num
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.DATA_PARALLEL,
mirror_mean=True,
device_num=device_num)
else:
rank = 0
device_num = 1
netwithloss = BertNetworkWithLoss_gd(
teacher_config=bert_teacher_net_cfg,
teacher_ckpt=args_opt.load_teacher_ckpt_path,
student_config=bert_student_net_cfg,
is_training=True,
use_one_hot_embeddings=False)
dataset = create_tinybert_dataset('gd', bert_teacher_net_cfg.batch_size,
device_num, rank, args_opt.do_shuffle,
args_opt.data_dir, args_opt.schema_dir)
dataset_size = dataset.get_dataset_size()
print('dataset size: ', dataset_size)
if args_opt.enable_data_sink == "true":
repeat_count = args_opt.epoch_size * dataset.get_dataset_size(
) // args_opt.data_sink_steps
time_monitor_steps = args_opt.data_sink_steps
else:
repeat_count = args_opt.epoch_size
time_monitor_steps = dataset_size
lr_schedule = BertLearningRate(
learning_rate=common_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=common_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(dataset_size * args_opt.epoch_size / 10),
decay_steps=int(dataset_size * args_opt.epoch_size),
power=common_cfg.AdamWeightDecay.power)
params = netwithloss.trainable_params()
decay_params = list(filter(common_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': common_cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}, {
'order_params': params
}]
optimizer = AdamWeightDecay(group_params,
learning_rate=lr_schedule,
eps=common_cfg.AdamWeightDecay.eps)
callback = [
TimeMonitor(time_monitor_steps),
LossCallBack(),
ModelSaveCkpt(netwithloss.bert, args_opt.save_ckpt_step,
args_opt.max_ckpt_num, save_ckpt_dir)
]
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=common_cfg.loss_scale_value,
scale_factor=common_cfg.scale_factor,
scale_window=common_cfg.scale_window)
netwithgrads = BertTrainWithLossScaleCell(netwithloss,
optimizer=optimizer,
scale_update_cell=update_cell)
model = Model(netwithgrads)
model.train(repeat_count,
dataset,
callbacks=callback,
dataset_sink_mode=(args_opt.enable_data_sink == "true"),
sink_size=args_opt.data_sink_steps)
def run_task_distill(ckpt_file):
"""
run task distill
"""
if ckpt_file == '':
raise ValueError("Student ckpt file should not be None")
cfg = phase2_cfg
load_teacher_checkpoint_path = args_opt.load_teacher_ckpt_path
load_student_checkpoint_path = ckpt_file
netwithloss = BertNetworkWithLoss_td(
teacher_config=td_teacher_net_cfg,
teacher_ckpt=load_teacher_checkpoint_path,
student_config=td_student_net_cfg,
student_ckpt=load_student_checkpoint_path,
is_training=True,
task_type=args_opt.task_type,
num_labels=task.num_labels,
is_predistill=False)
rank = 0
device_num = 1
train_dataset = create_tinybert_dataset('td',
cfg.batch_size,
device_num,
rank,
args_opt.do_shuffle,
args_opt.train_data_dir,
args_opt.schema_dir,
data_type=dataset_type)
dataset_size = train_dataset.get_dataset_size()
print('td2 train dataset size: ', dataset_size)
print('td2 train dataset repeatcount: ', train_dataset.get_repeat_count())
if args_opt.enable_data_sink == 'true':
repeat_count = args_opt.td_phase2_epoch_size * train_dataset.get_dataset_size(
) // args_opt.data_sink_steps
time_monitor_steps = args_opt.data_sink_steps
else:
repeat_count = args_opt.td_phase2_epoch_size
time_monitor_steps = dataset_size
optimizer_cfg = cfg.optimizer_cfg
lr_schedule = BertLearningRate(
learning_rate=optimizer_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=optimizer_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(dataset_size * args_opt.td_phase2_epoch_size / 10),
decay_steps=int(dataset_size * args_opt.td_phase2_epoch_size),
power=optimizer_cfg.AdamWeightDecay.power)
params = netwithloss.trainable_params()
decay_params = list(
filter(optimizer_cfg.AdamWeightDecay.decay_filter, params))
other_params = list(
filter(lambda x: not optimizer_cfg.AdamWeightDecay.decay_filter(x),
params))
group_params = [{
'params': decay_params,
'weight_decay': optimizer_cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}, {
'order_params': params
}]
optimizer = AdamWeightDecay(group_params,
learning_rate=lr_schedule,
eps=optimizer_cfg.AdamWeightDecay.eps)
eval_dataset = create_tinybert_dataset('td',
eval_cfg.batch_size,
device_num,
rank,
args_opt.do_shuffle,
args_opt.eval_data_dir,
args_opt.schema_dir,
data_type=dataset_type)
print('td2 eval dataset size: ', eval_dataset.get_dataset_size())
if args_opt.do_eval.lower() == "true":
callback = [
TimeMonitor(time_monitor_steps),
LossCallBack(),
EvalCallBack(netwithloss.bert, eval_dataset)
]
else:
callback = [
TimeMonitor(time_monitor_steps),
LossCallBack(),
ModelSaveCkpt(netwithloss.bert, args_opt.save_ckpt_step,
args_opt.max_ckpt_num, td_phase2_save_ckpt_dir)
]
if enable_loss_scale:
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=cfg.loss_scale_value,
scale_factor=cfg.scale_factor,
scale_window=cfg.scale_window)
netwithgrads = BertEvaluationWithLossScaleCell(
netwithloss, optimizer=optimizer, scale_update_cell=update_cell)
else:
netwithgrads = BertEvaluationCell(netwithloss, optimizer=optimizer)
model = Model(netwithgrads)
model.train(repeat_count,
train_dataset,
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_task_sink",
type=str,
default="true",
help="Enable task sink, default is true.")
parser.add_argument("--enable_loop_sink",
type=str,
default="true",
help="Enable loop sink, default is true.")
parser.add_argument("--enable_mem_reuse",
type=str,
default="true",
help="Enable mem reuse, default is true.")
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(enable_task_sink=(args_opt.enable_task_sink == "true"),
enable_loop_sink=(args_opt.enable_loop_sink == "true"),
enable_mem_reuse=(args_opt.enable_mem_reuse == "true"))
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)
D.init()
rank = args_opt.device_id % device_num
else:
rank = 0
device_num = 1
ds = 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)
else:
raise ValueError(
"Don't support optimizer {}, only support [Lamb, Momentum, AdamWeightDecayDynamicLR]"
.format(cfg.optimizer))
callback = [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(ds.get_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(
"--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 do_train(dataset=None,
network=None,
load_checkpoint_path="",
save_checkpoint_path="",
epoch_num=1):
"""
Do train
Args:
dataset: the train dataset.
network: the network with loss
load_checkpoint_path: the file path which saved pretrain model checkpoint.
save_checkpoint_path: the file path which will save finetune model checkpoint.
epoch_num: the number of epoch
"""
if load_checkpoint_path == "":
raise ValueError(
"Pretrain model missed, finetune task must load pretrain model!")
steps_per_epoch = dataset.get_dataset_size(
) # samples / batch_size doing####
if cfg.optimizer == 'AdamWeightDecay':
lr_schedule = GPT2LearningRate(
learning_rate=cfg.AdamWeightDecay.learning_rate,
end_learning_rate=cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=cfg.AdamWeightDecay.power)
params = network.trainable_params(
) # return a list of all trainable parmeters of the network
# Use parameter groups and set different values
decay_params = list(filter(cfg.AdamWeightDecay.decay_filter,
params)) # without layernorm and bias
other_params = list(
filter(lambda x: not cfg.AdamWeightDecay.decay_filter(x),
params)) # with layernorm and bias
group_params = [{
'params': decay_params,
'weight_decay': cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}]
optimizer = AdamWeightDecay(group_params,
lr_schedule,
eps=cfg.AdamWeightDecay.eps)
elif cfg.optimizer == 'Lamb':
lr_schedule = GPT2LearningRate(
learning_rate=cfg.Lamb.learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=cfg.Lamb.power)
optimizer = Lamb(network.trainable_params(), lr_schedule)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(network.trainable_params(),
cfg.Momentum.learning_rate, cfg.Momentum.momentum)
else:
raise Exception(
"Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]"
)
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch,
keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(
prefix="gpt2_language_model_wiki2",
directory=None if save_checkpoint_path == "" else save_checkpoint_path,
config=ckpt_config)
param_dict = load_checkpoint(load_checkpoint_path)
final_param_dict = {}
for k, v in param_dict.items():
final_param_dict['gpt2_loss.gpt2.gpt2.' + k] = param_dict[k]
# set the weights of final linear weights to weights of gpt2 token embedding
final_param_dict['gpt2_loss.gpt2.dense1.weight'] = param_dict[
'gpt2_embedding_lookup.embedding_table']
load_param_into_net(network, final_param_dict)
print("Load new parameter successfully!\n")
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32,
scale_factor=2,
scale_window=1000)
netwithgrads = GPT2FinetuneCell(network,
optimizer=optimizer,
scale_update_cell=update_cell)
netwithgrads.set_train(True)
loss_cb = LossMonitor()
model = Model(netwithgrads)
# callbacks = [TimeMonitor(dataset.get_dataset_size()), LossCallBack(dataset.get_dataset_size()), ckpoint_cb]
callbacks = [TimeMonitor(dataset.get_dataset_size()), loss_cb, ckpoint_cb]
print("============== Starting Training ==============")
model.train(epoch_num, dataset, callbacks=callbacks)
print("============== Training Success ==============")
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_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 do_train(dataset=None,
network=None,
load_checkpoint_path="",
save_checkpoint_path="",
epoch_num=1):
"""
Do train
Args:
dataset: the train dataset.
network: the network with loss
load_checkpoint_path: the file path which saved pretrained model checkpoint.
save_checkpoint_path: the file path which will save finetuned model checkpoint.
epoch_num: the number of epoch.
"""
if load_checkpoint_path == "":
raise ValueError(
"Pretrain model missed, finetune task must load pretrain model!")
steps_per_epoch = dataset.get_dataset_size()
# optimizer
if cfg.optimizer == 'AdamWeightDecay':
lr_schedule = GPT2LearningRate(
learning_rate=cfg.AdamWeightDecay.learning_rate,
end_learning_rate=cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=cfg.AdamWeightDecay.power)
params = network.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
}]
optimizer = AdamWeightDecay(group_params,
lr_schedule,
eps=cfg.AdamWeightDecay.eps)
elif cfg.optimizer == 'Lamb':
lr_schedule = GPT2LearningRate(
learning_rate=cfg.Lamb.learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
warmup_steps=int(steps_per_epoch * epoch_num * 0.1),
decay_steps=steps_per_epoch * epoch_num,
power=cfg.Lamb.power)
optimizer = Lamb(network.trainable_params(), lr_schedule)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(network.trainable_params(),
cfg.Momentum.learning_rate, cfg.Momentum.momentum)
else:
raise Exception(
"Optimizer not supported. support: [AdamWeightDecay, Lamb, Momentum]"
)
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch,
keep_checkpoint_max=1)
prefix_name = "gpt2_translation_" + str(cfg.gpt2_network) + "_" + str(cfg.optimizer) + "_" \
+ str(epoch_num) + "_bs" + str(gpt2_net_cfg.batch_size)
ckpoint_cb = ModelCheckpoint(
prefix=prefix_name,
directory=None if save_checkpoint_path == "" else save_checkpoint_path,
config=ckpt_config)
param_dict = load_checkpoint(load_checkpoint_path)
final_param_dict = {}
for name, _ in param_dict.items():
final_param_dict['gpt2.gpt2.' + name] = param_dict[name]
final_param_dict['gpt2.dense1.weight'] = param_dict[
'gpt2_embedding_lookup.embedding_table']
load_param_into_net(network, final_param_dict)
print("Load the pretrained parameter successfully! \n")
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32,
scale_factor=2,
scale_window=1000)
netwithgrads = GPT2FinetuneCell(network,
optimizer=optimizer,
scale_update_cell=update_cell)
netwithgrads.set_train(True)
loss_cb = LossMonitor(per_print_times=1)
model = Model(netwithgrads)
callbacks = [TimeMonitor(dataset.get_dataset_size()), loss_cb, ckpoint_cb]
print(
"=================== Starting Training For Translation Task ===================="
)
model.train(epoch_num,
dataset,
callbacks=callbacks,
dataset_sink_mode=False)
print(
"=================== Translation Training Success ===================="
)
from tests.dataset_mock import MindData
GRADIENT_CLIP_TYPE = 1
GRADIENT_CLIP_VALUE = 1.0
clip_grad = C.MultitypeFuncGraph("clip_grad")
grad_scale = C.MultitypeFuncGraph("grad_scale")
reciprocal = P.Reciprocal()
@grad_scale.register("Tensor", "Tensor")
def tensor_grad_scale(scale, grad):
return grad * reciprocal(scale)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=65536,
scale_factor=2,
scale_window=1000)
@clip_grad.register("Number", "Number", "Tensor")
def _clip_grad(clip_type, clip_value, grad):
dt = F.dtype(grad)
if clip_type == 0:
new_grad = C.clip_by_value(
grad, F.cast(F.tuple_to_array((-clip_value, )), dt),
F.cast(F.tuple_to_array((clip_value, )), dt))
else:
new_grad = nn.ClipByNorm()(grad,
F.cast(F.tuple_to_array((clip_value, )),
dt))
return new_grad
def run_general_distill():
"""
run general distill
"""
args_opt = get_argument()
context.set_context(mode=context.GRAPH_MODE,
device_target=args_opt.device_target,
reserve_class_name_in_scope=False)
if args_opt.device_target == "Ascend":
context.set_context(device_id=args_opt.device_id)
save_ckpt_dir = os.path.join(
args_opt.save_ckpt_path,
datetime.datetime.now().strftime('%Y-%m-%d_time_%H_%M_%S'))
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()
save_ckpt_dir = save_ckpt_dir + '_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)
else:
rank = 0
device_num = 1
if not os.path.exists(save_ckpt_dir):
os.makedirs(save_ckpt_dir)
enable_loss_scale = True
if args_opt.device_target == "GPU":
if bert_student_net_cfg.compute_type != mstype.float32:
logger.warning(
'Compute about the student only support float32 temporarily, run with float32.'
)
bert_student_net_cfg.compute_type = mstype.float32
# Backward of the network are calculated using fp32,
# and the loss scale is not necessary
enable_loss_scale = False
if args_opt.device_target == "CPU":
logger.warning(
'CPU only support float32 temporarily, run with float32.')
bert_teacher_net_cfg.dtype = mstype.float32
bert_teacher_net_cfg.compute_type = mstype.float32
bert_student_net_cfg.dtype = mstype.float32
bert_student_net_cfg.compute_type = mstype.float32
enable_loss_scale = False
netwithloss = BertNetworkWithLoss_gd(
teacher_config=bert_teacher_net_cfg,
teacher_ckpt=args_opt.load_teacher_ckpt_path,
student_config=bert_student_net_cfg,
is_training=True,
use_one_hot_embeddings=False)
if args_opt.dataset_type == "tfrecord":
dataset_type = DataType.TFRECORD
elif args_opt.dataset_type == "mindrecord":
dataset_type = DataType.MINDRECORD
else:
raise Exception("dataset format is not supported yet")
dataset = create_tinybert_dataset('gd',
common_cfg.batch_size,
device_num,
rank,
args_opt.do_shuffle,
args_opt.data_dir,
args_opt.schema_dir,
data_type=dataset_type)
dataset_size = dataset.get_dataset_size()
print('dataset size: ', dataset_size)
print("dataset repeatcount: ", dataset.get_repeat_count())
if args_opt.enable_data_sink == "true":
repeat_count = args_opt.epoch_size * dataset_size // args_opt.data_sink_steps
time_monitor_steps = args_opt.data_sink_steps
else:
repeat_count = args_opt.epoch_size
time_monitor_steps = dataset_size
lr_schedule = BertLearningRate(
learning_rate=common_cfg.AdamWeightDecay.learning_rate,
end_learning_rate=common_cfg.AdamWeightDecay.end_learning_rate,
warmup_steps=int(dataset_size * args_opt.epoch_size / 10),
decay_steps=int(dataset_size * args_opt.epoch_size),
power=common_cfg.AdamWeightDecay.power)
params = netwithloss.trainable_params()
decay_params = list(filter(common_cfg.AdamWeightDecay.decay_filter,
params))
other_params = list(
filter(lambda x: not common_cfg.AdamWeightDecay.decay_filter(x),
params))
group_params = [{
'params': decay_params,
'weight_decay': common_cfg.AdamWeightDecay.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}, {
'order_params': params
}]
optimizer = AdamWeightDecay(group_params,
learning_rate=lr_schedule,
eps=common_cfg.AdamWeightDecay.eps)
callback = [
TimeMonitor(time_monitor_steps),
LossCallBack(),
ModelSaveCkpt(netwithloss.bert, args_opt.save_ckpt_step,
args_opt.max_ckpt_num, save_ckpt_dir)
]
if enable_loss_scale:
update_cell = DynamicLossScaleUpdateCell(
loss_scale_value=common_cfg.loss_scale_value,
scale_factor=common_cfg.scale_factor,
scale_window=common_cfg.scale_window)
netwithgrads = BertTrainWithLossScaleCell(
netwithloss, optimizer=optimizer, scale_update_cell=update_cell)
else:
netwithgrads = BertTrainCell(netwithloss, optimizer=optimizer)
model = Model(netwithgrads)
model.train(repeat_count,
dataset,
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",
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 test_train():
'''
finetune function
'''
target = args_opt.device_target
if target == "Ascend":
devid = int(os.getenv('DEVICE_ID'))
context.set_context(mode=context.GRAPH_MODE,
device_target="Ascend",
device_id=devid)
elif target == "GPU":
context.set_context(mode=context.GRAPH_MODE, device_target="GPU")
else:
raise Exception("Target error, GPU or Ascend is supported.")
#BertCLSTrain for classification
#BertNERTrain for sequence labeling
if cfg.task == 'NER':
if cfg.use_crf:
netwithloss = BertNER(bert_net_cfg,
True,
num_labels=len(tag_to_index),
use_crf=True,
tag_to_index=tag_to_index,
dropout_prob=0.1)
else:
netwithloss = BertNER(bert_net_cfg,
True,
num_labels=cfg.num_labels,
dropout_prob=0.1)
elif cfg.task == 'SQUAD':
netwithloss = BertSquad(bert_net_cfg, True, 2, dropout_prob=0.1)
else:
netwithloss = BertCLS(bert_net_cfg,
True,
num_labels=cfg.num_labels,
dropout_prob=0.1)
if cfg.task == 'SQUAD':
dataset = get_squad_dataset(bert_net_cfg.batch_size, cfg.epoch_num)
else:
dataset = get_dataset(bert_net_cfg.batch_size, cfg.epoch_num)
# optimizer
steps_per_epoch = dataset.get_dataset_size()
if cfg.optimizer == 'AdamWeightDecayDynamicLR':
optimizer = AdamWeightDecayDynamicLR(
netwithloss.trainable_params(),
decay_steps=steps_per_epoch * cfg.epoch_num,
learning_rate=cfg.AdamWeightDecayDynamicLR.learning_rate,
end_learning_rate=cfg.AdamWeightDecayDynamicLR.end_learning_rate,
power=cfg.AdamWeightDecayDynamicLR.power,
warmup_steps=int(steps_per_epoch * cfg.epoch_num * 0.1),
weight_decay=cfg.AdamWeightDecayDynamicLR.weight_decay,
eps=cfg.AdamWeightDecayDynamicLR.eps)
elif cfg.optimizer == 'Lamb':
optimizer = Lamb(netwithloss.trainable_params(),
decay_steps=steps_per_epoch * cfg.epoch_num,
start_learning_rate=cfg.Lamb.start_learning_rate,
end_learning_rate=cfg.Lamb.end_learning_rate,
power=cfg.Lamb.power,
weight_decay=cfg.Lamb.weight_decay,
warmup_steps=int(steps_per_epoch * cfg.epoch_num *
0.1),
decay_filter=cfg.Lamb.decay_filter)
elif cfg.optimizer == 'Momentum':
optimizer = Momentum(netwithloss.trainable_params(),
learning_rate=cfg.Momentum.learning_rate,
momentum=cfg.Momentum.momentum)
else:
raise Exception("Optimizer not supported.")
# load checkpoint into network
ckpt_config = CheckpointConfig(save_checkpoint_steps=steps_per_epoch,
keep_checkpoint_max=1)
ckpoint_cb = ModelCheckpoint(prefix=cfg.ckpt_prefix,
directory=cfg.ckpt_dir,
config=ckpt_config)
param_dict = load_checkpoint(cfg.pre_training_ckpt)
load_param_into_net(netwithloss, param_dict)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=2**32,
scale_factor=2,
scale_window=1000)
if cfg.task == 'SQUAD':
netwithgrads = BertSquadCell(netwithloss,
optimizer=optimizer,
scale_update_cell=update_cell)
else:
netwithgrads = BertFinetuneCell(netwithloss,
optimizer=optimizer,
scale_update_cell=update_cell)
model = Model(netwithgrads)
model.train(cfg.epoch_num, dataset, callbacks=[LossCallBack(), ckpoint_cb])
def run_train_pipeline(args_opt):
device_id = int(os.getenv("DEVICE_ID"))
rank_id = int(os.getenv("RANK_ID"))
local_rank = rank_id
print('local_rank:{}, device id:{} start to run...'.format(
local_rank, device_id),
flush=True)
context.set_context(save_graphs=False,
mode=context.GRAPH_MODE,
device_target="Ascend",
device_id=device_id)
context.set_context(variable_memory_max_size="31GB")
strategy_ckpt_save_file = "/cache/" + "strategy" + str(
local_rank) + ".ckpt"
if args_opt.distribute == "true":
D.init()
device_num = D.get_group_size()
rank = D.get_rank()
print("device_id is {}, rank_id is {}, device_num is {}".format(
device_id, rank, device_num))
context.reset_auto_parallel_context()
context.set_auto_parallel_context(
parallel_mode=ParallelMode.SEMI_AUTO_PARALLEL,
gradients_mean=False,
device_num=device_num,
full_batch=True,
loss_repeated_mean=True,
enable_parallel_optimizer=bool(args_opt.optimizer_shard),
pipeline_stages=args_opt.stage_num,
strategy_ckpt_save_file=strategy_ckpt_save_file)
set_algo_parameters(elementwise_op_strategy_follow=True)
_set_multi_subgraphs()
else:
rank = 0
device_num = 1
model_parallel_num = args_opt.tensor_model_parallel_num
stage_device_num = int(device_num / args_opt.stage_num)
data_parallel_num = int(stage_device_num / model_parallel_num)
per_batch_size = args_opt.per_batch_size
batch_size = per_batch_size * data_parallel_num * args_opt.micro_size
config = PANGUALPHAConfig(data_parallel_num=data_parallel_num,
model_parallel_num=model_parallel_num,
batch_size=batch_size,
seq_length=args_opt.seq_length,
vocab_size=args_opt.vocab_size,
embedding_size=args_opt.embedding_size,
num_layers=args_opt.num_layers,
num_heads=args_opt.num_heads,
expand_ratio=4,
post_layernorm_residual=False,
dropout_rate=0.1,
compute_dtype=mstype.float16,
use_past=False,
self_layernorm=True,
forward_reduce_scatter=True,
stage_num=args_opt.stage_num,
micro_size=args_opt.micro_size,
word_emb_dp=False)
print("===config is: ", config, flush=True)
pangu_alpha = PANGUALPHAPipeline(config)
loss = CrossEntropyLoss(config)
pangu_alpha_with_loss = PANGUALPHAWithLossPipeline(config, pangu_alpha,
loss)
pangu_alpha_with_loss = VirtualDatasetOneInputCell(pangu_alpha_with_loss)
print("=====args_opt is: ", args_opt, flush=True)
lr = LearningRate(learning_rate=args_opt.start_lr,
end_learning_rate=args_opt.end_lr,
warmup_steps=args_opt.warmup_step,
decay_steps=args_opt.decay_steps)
per_stage_layers = config.num_layers // config.stage_num
per_stage_devices = device_num // config.stage_num
self_stage = rank_id // per_stage_devices
range_min = self_stage * per_stage_layers
range_max = range_min + per_stage_layers
if self_stage == 0:
params = [pangu_alpha.embedding_table]
params.extend(pangu_alpha.backbone.pangu_alpha_embedding.
position_embedding.trainable_params())
elif self_stage == config.stage_num - 1:
params = [pangu_alpha.embedding_table]
params.extend(pangu_alpha.backbone.layernorm.trainable_params())
params.extend(
pangu_alpha.backbone.top_query_embedding.trainable_params())
else:
params = []
for i in range(range_min, range_max):
params.extend(pangu_alpha.backbone.blocks[i].trainable_params())
decay_filter = lambda x: 'layernorm' not in x.name.lower(
) and "bias" not in x.name.lower()
decay_params = list(filter(decay_filter, params))
other_params = list(filter(lambda x: not decay_filter(x), params))
group_params = [{
'params': decay_params,
'weight_decay': args_opt.weight_decay
}, {
'params': other_params,
'weight_decay': 0.0
}, {
'order_params': params
}]
if args_opt.optimizer == "lamb":
optimizer = nn.Lamb(group_params, learning_rate=lr)
else:
optimizer = nn.AdamWeightDecay(group_params,
learning_rate=lr,
beta1=0.9,
beta2=0.95,
eps=1e-8)
save_steps = args_opt.save_steps
ckpt_dir = os.path.join(args_opt.ckpt_save_sir, f"rank_{str(local_rank)}")
if not os.path.exists(ckpt_dir):
Path(ckpt_dir).mkdir(parents=True, exist_ok=True)
ds = create_dataset(config.batch_size,
data_path=args_opt.data_url,
data_start_index=0)
epoch_num = args_opt.epoch_size
step_per_epoch = ds.get_dataset_size()
callback_size = args_opt.sink_size
actual_epoch_num = int(epoch_num * step_per_epoch / callback_size)
callback = [
TimeMonitor(callback_size),
LossCallBack(callback_size, local_rank, config.stage_num)
]
config_ck = CheckpointConfig(save_checkpoint_steps=save_steps,
keep_checkpoint_max=1,
integrated_save=False,
filter_prefix="accu_grads")
ckpoint_cb = ModelCheckpoint(prefix="PanguAlpha",
directory=ckpt_dir,
config=config_ck)
callback.append(ckpoint_cb)
loss_scale_value = math.pow(2, 32)
update_cell = DynamicLossScaleUpdateCell(loss_scale_value=loss_scale_value,
scale_factor=2,
scale_window=1000)
pangu_alpha_with_grads = PANGUALPHATrainPipelineWithLossScaleCell(
pangu_alpha_with_loss,
optimizer=optimizer,
config=config,
scale_update_cell=update_cell)
model = Model(pangu_alpha_with_grads)
de.config.set_sending_batches(2 * args_opt.sink_size)
model.train(actual_epoch_num,
ds,
callbacks=callback,
sink_size=callback_size,
dataset_sink_mode=True)