# Copyright 2020-2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Checkpoint related classes and functions."""
from __future__ import absolute_import
import os
import stat
import time
import threading
import mindspore.context as context
from mindspore import log as logger
from mindspore import nn
from mindspore import _checkparam as Validator
from mindspore.train._utils import _make_directory
from mindspore.train.serialization import save_checkpoint, _save_graph
from mindspore.parallel._cell_wrapper import destroy_allgather_cell
from mindspore.parallel._recovery_context import _set_recovery_context, _get_recovery_context
from mindspore.parallel._auto_parallel_context import _get_auto_parallel_context
from mindspore.parallel._utils import _get_device_num
from mindspore.communication.management import get_rank
from mindspore.train._utils import get_parameter_redundancy, remove_param_redundancy
from mindspore.train.callback._callback import Callback, set_cur_net
from mindspore.common.tensor import Tensor
from mindspore.common.parameter import Parameter
from mindspore.common.generator import Generator
from mindspore.common.api import _cell_graph_executor
from mindspore._c_expression import collect_host_info, get_clock_syscnt
_cur_dir = os.getcwd()
SAVE_DIR = _cur_dir
_info_list = ["epoch_num", "step_num"]
def _get_dp_tp_from_redundancy(redundancy_tuple):
"""From redundancy get dp and tp"""
dp = []
tp = []
for dp_value in redundancy_tuple:
dp.append(list(dp_value))
for i in range(len(redundancy_tuple[0])):
tp.append([v[i] for v in redundancy_tuple])
return dp, tp
def _get_dp_tp_from_layout(parameter_redundancy_dict):
"""From layout dict get dp and tp"""
tp = []
dp = []
value_len = 0
for _, value in parameter_redundancy_dict.items():
if len(value) > value_len:
value_len = len(value)
dp, tp = _get_dp_tp_from_redundancy(value)
return dp, tp
def _chg_ckpt_file_name_if_same_exist(directory, prefix, exception=False):
"""Check if there is a file with the same name."""
if callable(prefix) or callable(directory):
return prefix
files = os.listdir(directory)
suffix_num = 0
pre_len = len(prefix)
for filename in files:
name_ext = os.path.splitext(filename)
if exception and filename[-16:] != "_breakpoint.ckpt":
continue
if not exception and (name_ext[-1] != ".ckpt" or filename[-16:] == "_breakpoint.ckpt"):
continue
# find same prefix file
if filename.find(prefix) == 0 and not filename[pre_len].isalpha():
# add the max suffix + 1
index = filename[pre_len:].find("-")
if index == 0:
suffix_num = max(suffix_num, 1)
elif index != -1:
num = filename[pre_len + 1:pre_len + index]
if num.isdigit():
suffix_num = max(suffix_num, int(num) + 1)
if suffix_num != 0:
prefix = f'{prefix}_{suffix_num}'
return prefix
def _check_format_and_other_params(format, enc_key, enc_mode, crc_check=False, async_save=False, exception_save=False,
map_param_inc=False, global_step_num=None):
param_not_default = (enc_key is not None or enc_mode != "AES-GCM" or crc_check or async_save
or exception_save or map_param_inc or global_step_num is not None)
if format == "safetensors" and param_not_default:
raise ValueError("For 'save_checkpoint', when format is 'safetensors', other param must be default.")
[文档]class CheckpointConfig:
"""
The configuration of model checkpoint.
Note:
- During the training process, if dataset is transmitted through the data channel,
it is suggested to set 'save_checkpoint_steps' to an integer multiple of loop_size.
Otherwise, the time to save the checkpoint may be biased.
It is recommended to set only one save strategy and one keep strategy at the same time.
If both `save_checkpoint_steps` and `save_checkpoint_seconds` are set,
`save_checkpoint_seconds` will be invalid.
If both `keep_checkpoint_max` and `keep_checkpoint_per_n_minutes` are set,
`keep_checkpoint_per_n_minutes` will be invalid.
- The `enc_mode` and `crc_check` parameters are mutually exclusive and cannot be configured simultaneously.
Args:
save_checkpoint_steps (int): Steps to save checkpoint. Default: ``1`` .
save_checkpoint_seconds (int): Seconds to save checkpoint.
Can't be used with save_checkpoint_steps at the same time. Default: ``0`` .
keep_checkpoint_max (int): Maximum number of checkpoint files can be saved. Default: ``5`` .
keep_checkpoint_per_n_minutes (int): Save the checkpoint file every `keep_checkpoint_per_n_minutes` minutes.
Can't be used with keep_checkpoint_max at the same time. Default: ``0`` .
integrated_save (bool): Whether to merge and save the split Tensor in the automatic parallel scenario.
Integrated save function is only supported in automatic parallel scene, not supported
in manual parallel. Default: ``True`` .
async_save (bool): Whether asynchronous execution saves the checkpoint to a file. Default: ``False`` .
saved_network (Cell): Network to be saved in checkpoint file. If the saved_network has no relation
with the network in training, the initial value of saved_network will be saved. Default: ``None`` .
append_info (list): The information save to checkpoint file. Support "epoch_num", "step_num" and
dict. The key of dict must be str, the value of dict must be one of int, float, bool, Parameter or Tensor.
Default: ``None`` .
enc_key (Union[None, bytes]): Byte type key used for encryption. If the value is None, the encryption
is not required. Default: ``None`` .
enc_mode (str): This parameter is valid only when enc_key is not set to None. Specifies the encryption
mode, currently supports 'AES-GCM', 'AES-CBC' and 'SM4-CBC'. Default: ``'AES-GCM'`` .
exception_save (bool): Whether to save the current checkpoint when an exception occurs. Default: ``False`` .
crc_check (bool): Whether to perform crc32 calculation when saving checkpoint and save the calculation
result to the end of ckpt. Default: ``False`` .
remove_redundancy (bool): Whether to enable saving the checkpoint with redundancy removal.
Redundancy removal refers to eliminating redundant data in data parallelism mode. Default: ``False`` , means
redundant-free saving is not enabled.
format (str): Format of the output file, can be "ckpt" or "safetensors". Default: "ckpt".
kwargs (dict): Configuration options dictionary.
Raises:
ValueError: If input parameter is not the correct type.
Examples:
>>> from mindspore import nn
>>> from mindspore.train import Model, CheckpointConfig, ModelCheckpoint
>>>
>>> # Define the network structure of LeNet5. Refer to
>>> # https://gitee.com/mindspore/docs/blob/master/docs/mindspore/code/lenet.py
>>> net = LeNet5()
>>> loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
>>> optim = nn.Momentum(net.trainable_params(), 0.01, 0.9)
>>> model = Model(net, loss_fn=loss, optimizer=optim)
>>> # Create the dataset taking MNIST as an example. Refer to
>>> # https://gitee.com/mindspore/docs/blob/master/docs/mindspore/code/mnist.py
>>> dataset = create_dataset()
>>> config = CheckpointConfig(save_checkpoint_seconds=100, keep_checkpoint_per_n_minutes=5, saved_network=net)
>>> config.save_checkpoint_steps
1
>>> config.save_checkpoint_seconds
>>> config.keep_checkpoint_max
5
>>> config.keep_checkpoint_per_n_minutes
>>> config.integrated_save
True
>>> config.async_save
False
>>> config.saved_network
>>> config.enc_key
>>> config.enc_mode
'AES-GCM'
>>> config.append_dict
>>> config.get_checkpoint_policy
>>> ckpoint_cb = ModelCheckpoint(prefix='LeNet5', directory='./checkpoint', config=config)
>>> model.train(10, dataset, callbacks=ckpoint_cb)
"""
def __init__(self,
save_checkpoint_steps=1,
save_checkpoint_seconds=0,
keep_checkpoint_max=5,
keep_checkpoint_per_n_minutes=0,
integrated_save=True,
async_save=False,
saved_network=None,
append_info=None,
enc_key=None,
enc_mode='AES-GCM',
exception_save=False,
crc_check=False,
remove_redundancy=False,
format="ckpt",
**kwargs):
if save_checkpoint_steps is not None:
save_checkpoint_steps = Validator.check_non_negative_int(save_checkpoint_steps)
if save_checkpoint_seconds is not None:
save_checkpoint_seconds = Validator.check_non_negative_int(save_checkpoint_seconds)
if keep_checkpoint_max is not None:
keep_checkpoint_max = Validator.check_non_negative_int(keep_checkpoint_max)
if keep_checkpoint_per_n_minutes is not None:
keep_checkpoint_per_n_minutes = Validator.check_non_negative_int(keep_checkpoint_per_n_minutes)
if saved_network is not None and not isinstance(saved_network, nn.Cell):
raise TypeError(f"For 'CheckpointConfig', the type of 'saved_network' must be None or Cell, "
f"but got {str(type(saved_network))}.")
if not save_checkpoint_steps and not save_checkpoint_seconds and \
not keep_checkpoint_max and not keep_checkpoint_per_n_minutes:
raise ValueError("For 'CheckpointConfig', the input arguments 'save_checkpoint_steps', "
"'save_checkpoint_seconds', "
"'keep_checkpoint_max' and 'keep_checkpoint_per_n_minutes' can't be all None or 0.")
Validator.check_bool(exception_save)
self.exception_save = exception_save
self._save_checkpoint_steps = save_checkpoint_steps
self._save_checkpoint_seconds = save_checkpoint_seconds
if self._save_checkpoint_steps and self._save_checkpoint_steps > 0:
self._save_checkpoint_seconds = None
self._keep_checkpoint_max = keep_checkpoint_max
self._keep_checkpoint_per_n_minutes = keep_checkpoint_per_n_minutes
if self._keep_checkpoint_max and self._keep_checkpoint_max > 0:
self._keep_checkpoint_per_n_minutes = None
else:
if not self._keep_checkpoint_per_n_minutes or self._keep_checkpoint_per_n_minutes == 0:
self._keep_checkpoint_max = 1
self._integrated_save = Validator.check_bool(integrated_save)
self._async_save = Validator.check_bool(async_save)
self._saved_network = saved_network
self._append_dict = self._handle_append_info(append_info)
self._enc_key = Validator.check_isinstance('enc_key', enc_key, (type(None), bytes))
self._enc_mode = Validator.check_isinstance('enc_mode', enc_mode, str)
self._crc_check = Validator.check_isinstance('crc_check', crc_check, bool)
self._format = Validator.check_isinstance('format', format, str)
self._map_param_inc = kwargs.get('incremental', False)
self.enable_redundance = kwargs.get('enable_redundance', False)
self.remove_redundancy = Validator.check_isinstance('remove_redundancy', remove_redundancy, bool)
_check_format_and_other_params(format, enc_key, enc_mode, crc_check, async_save, exception_save,
self._map_param_inc)
@property
def save_checkpoint_steps(self):
"""
Get the value of steps to save checkpoint.
Returns:
int, steps to save checkpoint.
"""
return self._save_checkpoint_steps
@property
def save_checkpoint_seconds(self):
"""Get the value of _save_checkpoint_seconds.
Returns:
int, seconds to save the checkpoint file.
"""
return self._save_checkpoint_seconds
@property
def keep_checkpoint_max(self):
"""
Get the value of maximum number of checkpoint files can be saved.
Returns:
int, Maximum number of checkpoint files can be saved.
"""
return self._keep_checkpoint_max
@property
def keep_checkpoint_per_n_minutes(self):
"""
Get the value of save the checkpoint file every n minutes.
Returns:
Int, save the checkpoint file every n minutes.
"""
return self._keep_checkpoint_per_n_minutes
@property
def integrated_save(self):
"""
Get the value of whether to merge and save the split Tensor in the automatic parallel scenario.
Returns:
bool, whether to merge and save the split Tensor in the automatic parallel scenario.
"""
return self._integrated_save
@property
def async_save(self):
"""
Get the value of whether asynchronous execution saves the checkpoint to a file.
Returns:
bool, whether asynchronous execution saves the checkpoint to a file.
"""
return self._async_save
@property
def saved_network(self):
"""
Get the value of network to be saved in checkpoint file.
Returns:
Cell, network to be saved in checkpoint file.
"""
return self._saved_network
@property
def enc_key(self):
"""
Get the value of byte type key used for encryption.
Returns:
(None, bytes), byte type key used for encryption.
"""
return self._enc_key
@property
def enc_mode(self):
"""
Get the value of the encryption mode.
Returns:
str, encryption mode.
"""
return self._enc_mode
@property
def crc_check(self):
"""
Get the value of the whether to enable crc check.
Returns:
bool, whether to enable crc check.
"""
return self._crc_check
@property
def format(self):
return self._format
@property
def append_dict(self):
"""
Get the value of information dict saved to checkpoint file.
Returns:
dict, the information saved to checkpoint file.
"""
return self._append_dict
@property
def map_param_inc(self):
"""
Get the value of whether to save map Parameter incrementally.
Returns:
bool, whether to save map Parameter incrementally.
"""
return self._map_param_inc
[文档] def get_checkpoint_policy(self):
"""
Get the policy of checkpoint.
Returns:
dict, the information of checkpoint policy.
"""
checkpoint_policy = {'save_checkpoint_steps': self.save_checkpoint_steps,
'save_checkpoint_seconds': self.save_checkpoint_seconds,
'keep_checkpoint_max': self.keep_checkpoint_max,
'keep_checkpoint_per_n_minutes': self.keep_checkpoint_per_n_minutes,
'saved_network': self.saved_network}
return checkpoint_policy
@staticmethod
def _handle_append_info(append_info):
"""Handle ckpt append info."""
if append_info is None or append_info == []:
return None
if not isinstance(append_info, list):
raise TypeError(f"For 'CheckpointConfig', the type of 'append_info' must be list,"
f"but got {str(type(append_info))}.")
handle_append_info = {}
if "epoch_num" in append_info:
handle_append_info["epoch_num"] = 0
if "step_num" in append_info:
handle_append_info["step_num"] = 0
if "random_op" in append_info:
handle_append_info["random_op"] = 0
dict_num = 0
for element in append_info:
if not isinstance(element, str) and not isinstance(element, dict):
raise TypeError(f"For 'CheckpointConfig', the type of 'append_info' element must be str or dict,"
f"but got {str(type(element))}.")
if isinstance(element, str) and element not in _info_list:
raise ValueError(f"For 'CheckpointConfig', the value of element in the argument 'append_info' "
f"must be in {_info_list}, "
f"but got {element}.")
if isinstance(element, dict):
dict_num += 1
if dict_num > 1:
raise TypeError(f"For 'CheckpointConfig', the element of 'append_info' must has only one dict, "
"but got {dict_num}")
for key, value in element.items():
if isinstance(key, str) and isinstance(value,
(int, float, bool, str, Parameter, Tensor, Generator)):
handle_append_info[key] = value
else:
raise TypeError(f"For 'CheckpointConfig', the key type of the dict 'append_info' "
f"must be string, the value type must be int or float or bool, "
f"but got key type {type(key)}, value type {type(value)}")
return handle_append_info
[文档]class ModelCheckpoint(Callback):
"""
The checkpoint callback class.
It is called to combine with train process and save the model and network parameters after training.
Note:
In the distributed training scenario, please specify different directories for each training process
to save the checkpoint file. Otherwise, the training may fail.
If this callback is used in the `model` function, the checkpoint file will saved
parameters of the optimizer by default.
Args:
prefix (Union[str, callable object]): The prefix name or callable object to generate name of checkpoint files.
Default: ``'CKP'`` .
directory (Union[str, callable object]): The folder path where the checkpoint is stored, or the callable object
used to generate the path. By default, the file is saved in the current directory.
Default: ``None`` .
config (CheckpointConfig): Checkpoint strategy configuration. Default: ``None`` .
Raises:
ValueError: If `prefix` is not str or contains the '/' character and is not a callable object.
ValueError: If `directory` is not str and is not a callable object.
TypeError: If the config is not CheckpointConfig type.
Examples:
>>> import numpy as np
>>> import mindspore.dataset as ds
>>> from mindspore import nn
>>> from mindspore.train import Model, ModelCheckpoint
>>>
>>> data = {"x": np.float32(np.random.rand(64, 10)), "y": np.random.randint(0, 5, (64,))}
>>> train_dataset = ds.NumpySlicesDataset(data=data).batch(32)
>>> net = nn.Dense(10, 5)
>>> crit = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
>>> opt = nn.Momentum(net.trainable_params(), 0.01, 0.9)
>>> ckpt_callback = ModelCheckpoint(prefix="myckpt")
>>> model = Model(network=net, optimizer=opt, loss_fn=crit)
>>> model.train(2, train_dataset, callbacks=[ckpt_callback])
"""
def __init__(self, prefix='CKP', directory=None, config=None):
super(ModelCheckpoint, self).__init__()
self._latest_ckpt_file_name = ""
self._init_time = time.time()
self._last_time = time.time()
self._last_time_for_keep = time.time()
self._last_triggered_step = 0
"""a callable for users to set self-defined prefix."""
self._prefix_func = None
"""a callable for users to set self-defined directory."""
self._directory_func = None
if not callable(prefix) and (not isinstance(prefix, str) or prefix.find('/') >= 0):
raise ValueError("For 'ModelCheckpoint', the argument 'prefix' "
"for checkpoint file name is invalid, it must be "
"callable or string that does not contain '/', but got {}.".format(prefix))
self._prefix = prefix
self._exception_prefix = prefix
if directory is not None:
if callable(directory):
self._directory_func = directory
else:
self._directory = _make_directory(directory)
else:
self._directory = _cur_dir
if callable(prefix):
self._prefix_func = prefix
if _get_recovery_context("enable_recovery"):
_set_recovery_context(ckpt_path=self._directory)
if config is None:
self._config = CheckpointConfig()
else:
if not isinstance(config, CheckpointConfig):
raise TypeError("For 'ModelCheckpoint', the type of argument 'config' should be "
"'CheckpointConfig', "
"but got {}.".format(type(config)))
self._config = config
self._aiturbo_init_flag = os.getenv("AITURBO") == "1"
# get existing checkpoint files
if self._aiturbo_init_flag:
from aiturbo.checkpoint.aiturbo_mindspore_ckpt import CheckpointShmManager
self._manager = CheckpointShmManager()
else:
self._manager = CheckpointManager(self._config.format)
if not callable(directory) and not callable(prefix):
self._prefix = _chg_ckpt_file_name_if_same_exist(self._directory, self._prefix)
self._append_dict = self._config.append_dict or {}
self._append_epoch_num = self._append_dict.get("epoch_num") if "epoch_num" in self._append_dict else 0
self._append_step_num = self._append_dict.get("step_num") if "step_num" in self._append_dict else 0
self._graph_saved = False
self._need_flush_from_cache = True
self._map_param_inc = self._config.map_param_inc
[文档] def step_end(self, run_context):
"""
Save the checkpoint at the end of step.
Args:
run_context (RunContext): Context of the train running.
"""
cb_params = run_context.original_args()
if self._aiturbo_init_flag:
from aiturbo.checkpoint import aiturbo_mindspore as aiturbo
ckpt_storage_path = self._directory
rank_id = get_rank()
stage_num = _get_auto_parallel_context("pipeline_stages")
stage_rank_num = _get_device_num() // stage_num
param_layout = cb_params.train_network.parameter_layout_dict
if not param_layout:
layout = {"stage_num": stage_num, "stage_rank_num": stage_rank_num, "stage_layout": None}
aiturbo.init(ckpt_storage_path, rank_id, layout, None, False, None)
else:
device_num = _get_device_num()
chunk_size = device_num // stage_num
initial_rank = (rank_id // chunk_size) * chunk_size
param_redundancy_dict = get_parameter_redundancy(param_layout, initial_rank)
dp, _ = _get_dp_tp_from_layout(param_redundancy_dict)
layout = {"stage_num": stage_num, "stage_rank_num": stage_rank_num,
"stage_layout": param_redundancy_dict}
single_params = remove_param_redundancy(param_redundancy_dict)
single_params = {device_id: list(params) for device_id, params in single_params.items()}
aiturbo.init(ckpt_storage_path, rank_id, layout, single_params, not self._config.enable_redundance, dp)
self._aiturbo_init_flag = False
if self._prefix_func:
self._prefix = self._prefix_func(cb_params)
if not isinstance(self._prefix, str) or self._prefix.find('/') >= 0:
raise ValueError("For 'ModelCheckpoint', the argument 'prefix' "
"for checkpoint file name is callable, it must return a "
"string that does not contain '/', but got {}.".format(self._prefix))
if self._directory_func:
self._directory = self._directory_func(cb_params)
collect_host_info("Callback", "ModelCheckpoint", "step_end", start_time=get_clock_syscnt(), level=1)
# In disaster recovery scenario, the training process may be rolled back to the last step where
# the ckpt was successfully saved, so the _last_triggered_step should be updated.
if _get_recovery_context("enable_recovery") and cb_params.last_save_ckpt_step is not None:
self._last_triggered_step = cb_params.last_save_ckpt_step
cb_params.last_save_ckpt_step = None
_make_directory(self._directory)
# save graph (only once)
if not self._graph_saved:
graph_file_name = os.path.join(self._directory, self._prefix + '-graph.meta')
if os.path.isfile(graph_file_name) and context.get_context("mode") == context.GRAPH_MODE:
os.remove(graph_file_name)
_save_graph(cb_params.train_network, graph_file_name)
self._graph_saved = True
thread_list = threading.enumerate()
for thread in thread_list:
if thread.getName() == "asyn_save_ckpt":
thread.join()
self._save_ckpt(cb_params)
[文档] def end(self, run_context):
"""
Save the last checkpoint after training finished.
Args:
run_context (RunContext): Context of the train running.
"""
cb_params = run_context.original_args()
collect_host_info("Callback", "ModelCheckpoint", "end", start_time=get_clock_syscnt(), level=1)
_to_save_last_ckpt = True
self._save_ckpt(cb_params, _to_save_last_ckpt)
thread_list = threading.enumerate()
for thread in thread_list:
if thread.getName() == "asyn_save_ckpt":
thread.join()
destroy_allgather_cell()
def _check_save_ckpt(self, cb_params, force_to_save):
"""Check whether save checkpoint files or not."""
if self._config.save_checkpoint_steps and self._config.save_checkpoint_steps > 0:
if cb_params.cur_step_num >= self._last_triggered_step + self._config.save_checkpoint_steps \
or force_to_save is True:
return True
elif self._config.save_checkpoint_seconds and self._config.save_checkpoint_seconds > 0:
self._cur_time = time.time()
if (self._cur_time - self._last_time) > self._config.save_checkpoint_seconds or force_to_save:
self._last_time = self._cur_time
return True
return False
def _append_dict_content(self, epoch_num, step_num):
"""Append append_dict content."""
if "epoch_num" in self._append_dict:
self._append_dict["epoch_num"] = self._append_epoch_num + epoch_num
if "step_num" in self._append_dict:
self._append_dict["step_num"] = self._append_step_num + step_num
def _save_ckpt(self, cb_params, force_to_save=False):
"""Save checkpoint files."""
if cb_params.cur_step_num == self._last_triggered_step:
return
# if param is cache enable, flush data from cache to host before save_ckpt
if self._need_flush_from_cache:
self._flush_from_cache(cb_params)
save_ckpt = self._check_save_ckpt(cb_params, force_to_save)
step_num_in_epoch = int((cb_params.cur_step_num - 1) % cb_params.batch_num + 1)
if save_ckpt:
if self._prefix_func:
cur_ckpoint_file = self._prefix + f".{self._config.format}"
else:
cur_ckpoint_file = self._prefix + "-" + str(cb_params.cur_epoch_num) + "_" \
+ str(step_num_in_epoch) + f".{self._config.format}"
# update checkpoint file list.
self._manager.update_ckpoint_filelist(self._directory, self._prefix)
# keep checkpoint files number equal max number.
if self._config.keep_checkpoint_max and 0 < self._config.keep_checkpoint_max <= self._manager.ckpoint_num:
self._manager.remove_oldest_ckpoint_file()
elif self._config.keep_checkpoint_per_n_minutes and self._config.keep_checkpoint_per_n_minutes > 0:
self._cur_time_for_keep = time.time()
if (self._cur_time_for_keep - self._last_time_for_keep) \
< self._config.keep_checkpoint_per_n_minutes * 60:
self._manager.keep_one_ckpoint_per_minutes(self._config.keep_checkpoint_per_n_minutes,
self._cur_time_for_keep)
# generate the new checkpoint file and rename it.
global SAVE_DIR
SAVE_DIR = self._directory
cur_file = os.path.join(self._directory, cur_ckpoint_file)
self._last_time_for_keep = time.time()
self._last_triggered_step = cb_params.cur_step_num
# TODO(MS_DISABLE_REF_MODE): Delete when remove MS_DISABLE_REF_MODE env.
if context.get_context("enable_ge") and os.getenv('MS_DISABLE_REF_MODE') \
and context.get_context("mode") == context.GRAPH_MODE:
set_cur_net(cb_params.train_network)
cb_params.train_network.add_flags(ge_sync_data=True)
_cell_graph_executor(cb_params.train_network, phase='save')
self._append_dict_content(cb_params.cur_epoch_num, cb_params.cur_step_num)
network = self._config.saved_network if self._config.saved_network is not None else cb_params.train_network
if os.getenv("AITURBO") == "1":
save_checkpoint(network, cur_file, self._config.integrated_save, self._config.async_save,
self._append_dict, self._config.enc_key, self._config.enc_mode,
crc_check=self._config.crc_check, incremental=self._map_param_inc,
global_step_num=cb_params.cur_step_num)
elif self._config.remove_redundancy:
parallel_mode = context.get_auto_parallel_context("parallel_mode")
if parallel_mode == "stand_alone":
raise TypeError(f"The deduplication feature for saving checkpoint can only be used "
f"in parallel scenarios, but got {parallel_mode}.")
param_layout = network.parameter_layout_dict
rank_id = get_rank()
if param_layout:
device_num = _get_device_num()
stage_num = _get_auto_parallel_context("pipeline_stages")
chunk_size = device_num // stage_num
initial_rank = (rank_id // chunk_size) * chunk_size
param_redundancy_dict = get_parameter_redundancy(param_layout, initial_rank)
single_params = remove_param_redundancy(param_redundancy_dict)
save_param_names = single_params.get(rank_id)
param_layout_set = set(param_layout.keys())
if save_param_names == param_layout.keys():
logger.warning(
f"For remove_redundancy save checkpoint, the saved parameters are non-redundant.")
def choice_func(x):
return x not in param_layout_set or x in save_param_names
else:
param_redundancy_dict = get_parameter_redundancy(network)
single_params = remove_param_redundancy(param_redundancy_dict)
save_param_names = single_params.get(rank_id)
def choice_func(x):
return x in save_param_names
save_checkpoint(network, cur_file, False, self._config.async_save,
self._append_dict, self._config.enc_key, self._config.enc_mode,
crc_check=self._config.crc_check, format=self._config.format,
incremental=self._map_param_inc, choice_func=choice_func)
else:
save_checkpoint(network, cur_file, self._config.integrated_save, self._config.async_save,
self._append_dict, self._config.enc_key, self._config.enc_mode,
crc_check=self._config.crc_check, format=self._config.format,
incremental=self._map_param_inc)
self._latest_ckpt_file_name = cur_file
def _flush_from_cache(self, cb_params):
"""Flush cache data to host if tensor is cache enable."""
has_cache_params = False
params = cb_params.train_network.get_parameters()
for param in params:
if param.cache_enable:
has_cache_params = True
Tensor(param).flush_from_cache()
if not has_cache_params:
self._need_flush_from_cache = False
@property
def latest_ckpt_file_name(self):
"""Return the latest checkpoint path and file name."""
return self._latest_ckpt_file_name
@property
def _get_save_checkpoint_steps(self):
"""Return save ckpt steps"""
return self._config.save_checkpoint_steps
@property
def _get_last_trigger_step(self):
"""Return last triggered steps"""
return self._last_triggered_step
class CheckpointManager:
"""Manage checkpoint files according to train_config of checkpoint."""
def __init__(self, format='ckpt'):
self._ckpoint_filelist = []
self._format = format
@property
def ckpoint_filelist(self):
"""Get all the related checkpoint files managed here."""
return self._ckpoint_filelist
@property
def ckpoint_num(self):
"""Get the number of the related checkpoint files managed here."""
return len(self._ckpoint_filelist)
def update_ckpoint_filelist(self, directory, prefix):
"""Update the checkpoint file list."""
self._ckpoint_filelist = []
format = self._format
format_length = len(format) + 1
files = os.listdir(directory)
for filename in files:
if os.path.splitext(filename)[-1] == f".{format}" and filename.startswith(prefix + "-"):
mid_name = filename[len(prefix):-format_length]
flag = not (True in [char.isalpha() for char in mid_name])
if flag:
self._ckpoint_filelist.append(os.path.join(directory, filename))
def remove_ckpoint_file(self, file_name):
"""Remove the specified checkpoint file from this checkpoint manager and also from the directory."""
try:
os.chmod(file_name, stat.S_IWRITE)
os.remove(file_name)
self._ckpoint_filelist.remove(file_name)
except OSError:
logger.warning("OSError, failed to remove the older ckpt file %s.", file_name)
except ValueError:
logger.warning("ValueError, failed to remove the older ckpt file %s.", file_name)
def remove_oldest_ckpoint_file(self):
"""Remove the oldest checkpoint file from this checkpoint manager and also from the directory."""
ckpoint_files = sorted(self._ckpoint_filelist, key=os.path.getmtime)
self.remove_ckpoint_file(ckpoint_files[0])
def keep_one_ckpoint_per_minutes(self, minutes, cur_time):
"""Only keep the latest one ckpt file per minutes, remove other files generated in [last_time, cur_time]."""
del_list = []
oldest_file = ''
oldest_time = cur_time
for ck_file in self._ckpoint_filelist:
modify_time = os.path.getmtime(ck_file)
if cur_time - modify_time < 60 * minutes:
del_list.append(ck_file)
if modify_time < oldest_time:
oldest_time = modify_time
oldest_file = ck_file
for mv_file in del_list:
if mv_file == oldest_file:
continue
self.remove_ckpoint_file(mv_file)