def __init__(self, program_func, optimizer, param_path=None, place=None):
# 1. we need to generate a framework.Program by calling
# program_func. Reference: fluid.program_guard in
# test_word2vec.py
self.scope = core.Scope()
self.startup_program = framework.Program()
self.train_program = framework.Program()
with framework.program_guard(self.train_program, self.startup_program):
loss = program_func()
if not isinstance(optimizer, opt_module.Optimizer):
raise TypeError(
"The optimizer should be an instance of Optimizer")
optimize_ops, params_grads = optimizer.minimize(loss)
self.place = Trainer._check_and_get_place(place)
self.dist_transpile_if_necessary(optimize_ops, params_grads)
# 2. move the default_main_program to self.program and run the
# default_startup program on an empty core.Scope()
# Run startup program
with self._prog_and_scope_guard():
exe = executor.Executor(place)
exe.run(self.startup_program)
if param_path:
# load params from param_path into scope
io.load_persistables(exe, dirname=param_path)
def __init__(self, infer_func, param_path, place=None, parallel=False):
"""
:param infer_func: a function that will return predict Variable
:param param_path: the path where the inference model is saved by fluid.io.save_params
:param place: place to do the inference
:param parallel: use parallel_executor to run the inference, it will use multi CPU/GPU.
"""
self.param_path = param_path
self.scope = core.Scope()
self.parallel = parallel
self.place = check_and_get_place(place)
self.inference_program = framework.Program()
with framework.program_guard(self.inference_program):
with unique_name.guard():
self.predict_var = infer_func()
with self._prog_and_scope_guard():
# load params from param_path into scope
io.load_params(executor.Executor(self.place), param_path)
if parallel:
with self._prog_and_scope_guard():
self.exe = parallel_executor.ParallelExecutor(
use_cuda=isinstance(self.place, core.CUDAPlace),
loss_name=self.predict_var.name)
else:
self.exe = executor.Executor(self.place)
def __init__(self,
train_func,
optimizer,
param_path=None,
place=None,
parallel=False):
self.__stop = False
self.parallel = parallel
# 1. we need to generate a framework.Program by calling
# program_func. Reference: fluid.program_guard in
# test_word2vec.py
if not isinstance(optimizer, opt_module.Optimizer):
raise TypeError("The optimizer should be an instance of Optimizer")
self.scope = core.Scope()
self.startup_program = framework.Program()
self.train_program = framework.Program()
with framework.program_guard(self.train_program, self.startup_program):
program_func_outs = train_func()
self.train_func_outputs = program_func_outs if isinstance(
program_func_outs, list) else [program_func_outs]
self.test_program = self.train_program.clone()
if not isinstance(optimizer, opt_module.Optimizer):
raise TypeError(
"The optimizer should be an instance of Optimizer")
# The fisrt element of program_func_outs is loss.
loss = self.train_func_outputs[0]
optimize_ops, params_grads = optimizer.minimize(loss)
self.place = check_and_get_place(place)
self._dist_transpile_if_necessary(optimize_ops, params_grads)
# 2. move the default_main_program to self.program and run the
# default_startup program on an empty core.Scope()
# Run startup program
with self._prog_and_scope_guard():
exe = executor.Executor(place)
exe.run(self.startup_program)
if param_path:
# load params from param_path into scope
io.load_persistables(exe, dirname=param_path)
def _load_lookup_table_vars(executor, dirname, program, pserver_id,
table_name):
"""
The parameter server will load lookup table's local file in
selectedrows variable.
Args:
executor(Executor): The executor to run for loading persistable variables
dirname(str): The directory path
main_program(Program): Find the variable named table_name in main_program
pserver_id(int): the serial number in pserver_endpoints list
table_name(str): lookup table name
Returns:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
dirname = "./checkpoints/checkpoint_9/"
prog = fluid.default_main_program()
pserver_id = 1
table_name = "share_w"
_load_lookup_table_vars(executor=exe,
dirname=dirname, program=prog, pserver_id=pserver_id,
table_name=table_name)
"""
for var in program.list_vars():
if var.name == table_name:
lookup_table_var = var
break
assert lookup_table_var is not None
lookup_table_dir = os.path.join(dirname, LOOKUP_TABLE_DIR)
table_file = table_name + CHECKPOINT_SEPARATOR + str(pserver_id)
load_prog = framework.Program()
load_block = load_prog.global_block()
load_block.append_op(
type='load',
inputs={},
outputs={'Out': [lookup_table_var]},
attrs={'file_path': os.path.join(lookup_table_dir, table_file)})
executor.run(load_prog)
def _save_pserver_vars_by_notify(executor, dirname, lookup_table,
ps_endpoint_list):
"""
This function will send checkpoint notify message from Trainer 0
to all the pservers.
The checkpoint notify message contains lookup table name,
the absolute path on pserver to save lookup_table.
Args:
executor(Executor): The executor to run for send checkpoint notify.
dirname(str): The folder where to save checkpoints.
lookup_table(string): the lookup table name, when use distribute
lookup table, we can get lookup table name by DistributeTranspiler.
table_name
ps_endpoint_list(list): the parameter server ip:port list.
when use distribute lookup table, we can get ps_endpoint_list by
distribute arguments.
Return:
None
Examples:
.. code-block:: python
exe = fluid.Executor(fluid.CPUPlace())
param_path = "./my_paddle_model"
prog = fluid.default_main_program()
table_name = "share_w"
ps_endpoints = ["127.0.0.1:6000","127.0.0.1:6001"]
_save_pserver_vars_by_notify(executor=exe,
dirname=param_path, lookup_table=table_name,
ps_endpoint_list=ps_endpoints)
"""
cur_dir = _get_lookuptable_dir(dirname)
checkpoint_notify_program = framework.Program()
checkpoint_notify_block = checkpoint_notify_program.global_block()
attrs = {}
attrs['epmap'] = ps_endpoint_list
attrs['dir'] = cur_dir
attrs['lookup_table'] = lookup_table
checkpoint_notify_block.append_op(type='checkpoint_notify',
inputs={},
outputs={},
attrs=attrs)
executor.run(checkpoint_notify_program)
def __init__(self, network_func, param_path=None, place=None):
# 1. we need to generate a framework.Program by calling
# network_func. Reference: fluid.program_guard in test_word2vec.py
# 2. move the default_main_program to self.program.
# 3. run the default_startup program.
# 4. load params from param_path into scope
self.scope = core.Scope()
self.place = place
self.startup_program = framework.Program()
# TODO: generate the startup_program with network_func
exe = executor.Executor(place)
exe.run(self.startup_program, scope=self.scope)
if param_path:
# load params from param_path into scope
io.load_persistables(exe, dirname=param_path)
def __init__(self, infer_func, param_path, place=None, parallel=False):
self.param_path = param_path
self.scope = core.Scope()
self.parallel = parallel
self.place = check_and_get_place(place)
self.inference_program = framework.Program()
with framework.program_guard(self.inference_program):
with unique_name.guard():
self.predict_var = infer_func()
with self._prog_and_scope_guard():
# load params from param_path into scope
io.load_params(executor.Executor(self.place), param_path)
if parallel:
with self._prog_and_scope_guard():
self.exe = parallel_executor.ParallelExecutor(
use_cuda=isinstance(self.place, core.CUDAPlace),
loss_name=self.predict_var.name)
else:
self.exe = executor.Executor(self.place)
self.inference_program = self.inference_program.clone(for_test=True)
def __init__(self,
train_func,
optimizer_func,
param_path=None,
place=None,
parallel=False,
checkpoint_config=None):
self.__stop = False
self.parallel = parallel
# config for checkpoint
# only chief worker will save variables
self.trainer_id = 0
self.checkpoint_cfg = checkpoint_config
if self.checkpoint_cfg:
assert isinstance(self.checkpoint_cfg, CheckpointConfig)
serial = io.get_latest_checkpoint_serial(
self.checkpoint_cfg.checkpoint_dir)
self.checkpoint_cfg.load_serial = serial if serial >= 0 else None
self.scope = core.Scope()
# 1. we need to generate a framework.Program by calling
# program_func. Reference: fluid.program_guard in
# test_word2vec.py
self.startup_program = framework.Program()
self.train_program = framework.Program()
with framework.program_guard(self.train_program, self.startup_program):
program_func_outs = train_func()
self.train_func_outputs = program_func_outs if isinstance(
program_func_outs, list) else [program_func_outs]
self.test_program = self.train_program.clone(for_test=True)
# The first element of program_func_outs is loss.
loss = self.train_func_outputs[0]
optimizer = optimizer_func()
if not isinstance(optimizer, opt_module.Optimizer):
raise TypeError(
"The optimizer should be an instance of Optimizer")
optimize_ops, params_grads = optimizer.minimize(loss)
self.place = check_and_get_place(place)
self._dist_transpile_if_necessary(optimize_ops, params_grads)
# 2. move the default_main_program to self.program and run the
# default_startup program on an empty core.Scope()
# Run startup program
with self._prog_and_scope_guard():
exe = executor.Executor(place)
exe.run(self.startup_program)
if self.checkpoint_cfg and self.checkpoint_cfg.load_serial:
with self._prog_and_scope_guard():
exe = executor.Executor(place)
io.load_checkpoint(exe, self.checkpoint_cfg.checkpoint_dir,
self.checkpoint_cfg.load_serial,
self.startup_program)
if not self.checkpoint_cfg.is_pserver:
epoch_id, step_id = io.load_trainer_args(
self.checkpoint_cfg.checkpoint_dir,
self.checkpoint_cfg.load_serial, self.trainer_id,
self._get_checkpoint_load_args())
self.checkpoint_cfg.epoch_id = int(epoch_id)
self.checkpoint_cfg.step_id = int(step_id)
if param_path and os.path.isdir(param_path):
# load params from param_path into scope
io.load_persist_vars_without_grad(
exe, dirname=param_path, program=self.startup_program)