提交 a2c017da 编写于 作者: M minqiyang

1. merge simple_dist_transpiler to distribute_transpiler

2. add align_var_to_block argument to func transpile
3. remove concat and spilt if align_var_to_block is False
4. unittests for simple_dist_transpiler
上级 580340ee
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
import unittest
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.layers as layers
from paddle.fluid.transpiler.distribute_transpiler import delete_ops
import numpy as np
class TestSimpleDistTranspiler(unittest.TestCase):
def setUp(self):
self.trainer_id = 0
self.trainers = 2
self.pservers = 2
self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175"
self.current_pserver_ep = "127.0.0.1:6175"
def net_conf(self):
x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
y_predict = fluid.layers.fc(input=x,
size=1000,
act=None,
param_attr=fluid.ParamAttr(name='fc_w'))
y = fluid.layers.data(name='y', shape=[1], dtype='float32')
cost = fluid.layers.square_error_cost(input=y_predict, label=y)
avg_cost = fluid.layers.mean(cost)
sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
return optimize_ops, params_grads
def test_simple_transpiler(self):
np.random.seed(1)
trainer = self.get_trainer()
pserver, startup = self.get_pserver(self.current_pserver_ep)
self.assertEqual([op.type for op in trainer.global_block().ops],
self.get_expect_trainer_ops())
self.assertEqual(len(pserver.blocks), 2)
# block0: listen_and_serv
self.assertEqual([op.type for op in pserver.blocks[0].ops],
["listen_and_serv"])
# block1: optimize pass
self.assertEqual([op.type for op in pserver.blocks[1].ops],
["sum", "scale", "sgd"])
print("xxx", [op.output_arg_names for op in startup.global_block().ops])
# confirm startup program
self.assertEqual([op.type for op in startup.global_block().ops],
["fill_constant", "uniform_random", "uniform_random"])
# the variable #fc_w will NOT be splited
fc_w_var = startup.global_block().var("fc_w@GRAD")
self.assertEqual(fc_w_var.shape, (1000, 1000))
fc_w_var = startup.global_block().var("fc_w@GRAD.trainer_0")
self.assertEqual(fc_w_var.shape, (1000, 1000))
def get_main_program(self):
main = fluid.Program()
with fluid.program_guard(main):
self.net_conf()
return main
def get_expect_trainer_ops(self):
trainer = fluid.Program()
with fluid.program_guard(trainer):
optimize_ops, params_grads = self.net_conf()
delete_ops(trainer.global_block(), optimize_ops)
ops = [op.type for op in trainer.global_block().ops] + [
"send_vars", "send_barrier", "recv", "recv", "fetch_barrier"
]
ops.insert(ops.index("elementwise_add_grad") + 1, "send_vars")
return ops
def get_trainer(self):
return self._transpiler_instance().get_trainer_program()
def get_pserver(self, ep):
t = self._transpiler_instance()
pserver = t.get_pserver_program(ep)
startup = t.get_startup_program(ep, pserver)
return pserver, startup
def _transpiler_instance(self):
main = self.get_main_program()
t = fluid.DistributeTranspiler()
t.transpile(
self.trainer_id,
program=main,
pservers=self.pserver_eps,
trainers=self.trainers,
align_var_to_block=False)
return t
if __name__ == "__main__":
unittest.main()
......@@ -15,6 +15,7 @@
from __future__ import print_function
import math
import numpy as np
from ps_dispatcher import RoundRobin, HashName, PSDispatcher
from .. import core, framework
......@@ -171,6 +172,7 @@ class DistributeTranspiler:
program=None,
pservers="127.0.0.1:6174",
trainers=1,
align_var_to_block=True,
split_method=RoundRobin,
sync_mode=True):
"""
......@@ -183,7 +185,8 @@ class DistributeTranspiler:
parameter servers.
Steps to transpile trainer:
1. split variable to multiple blocks, aligned by product(dim[1:]) (width).
1. split variable to multiple blocks, aligned by product(dim[1:]) (width)
if align_var_to_block is True
2. rename splited grad variables to add trainer_id suffix ".trainer_%d".
3. modify trainer program add split_op to each grad variable.
4. append send_op to send splited variables to server and fetch
......@@ -293,9 +296,18 @@ class DistributeTranspiler:
for index in range(len(self.pserver_endpoints))
]
grad_blocks = split_dense_variable(grad_list, len(pserver_endpoints))
param_blocks = split_dense_variable(param_list, len(pserver_endpoints))
if align_var_to_block:
grad_blocks = split_dense_variable(grad_list,
len(pserver_endpoints))
param_blocks = split_dense_variable(param_list,
len(pserver_endpoints))
else:
# when we do NOT align var to block, we will always split params
# grads into one block.
grad_blocks = split_dense_variable(grad_list, 1)
param_blocks = split_dense_variable(param_list, 1)
assert (len(grad_blocks) == len(param_blocks))
# step2: Create new vars for the parameters and gradients blocks and
# add ops to do the split.
param_var_mapping = self._create_vars_from_blocklist(program,
......@@ -325,8 +337,22 @@ class DistributeTranspiler:
# step 3.1: insert send op to send gradient vars to parameter servers
ps_dispatcher.reset()
send_vars = []
for orig_varname, splited_vars in grad_var_mapping.items():
# in general cases, the number of pservers is times of 2, and this
# will lead to uneven distribution among weights and bias:
# fc_w@GRAD_trainer_0, fc_w@GRAD_trainer_1 --> pserver1
# fc_b@GRAD_trainer_0, fc_b@GRAD_trainer_1 --> pserver2
# shuffle the map will avoid the uneven distribution above
grad_var_mapping_items = grad_var_mapping.items()
if not align_var_to_block:
np.random.shuffle(grad_var_mapping_items)
for orig_varname, splited_vars in grad_var_mapping_items:
eplist = ps_dispatcher.dispatch(splited_vars)
if not align_var_to_block:
assert (len(splited_vars) == 1)
if len(splited_vars) == 1:
orig_varname = splited_vars[0].name
index = find_op_by_output_arg(program.global_block(),
......@@ -374,7 +400,7 @@ class DistributeTranspiler:
for i, ep in enumerate(eplist):
self.param_grad_ep_mapping[ep]["params"].append(recv_vars[i])
self.param_grad_ep_mapping[ep]["grads"].append(send_vars[i])
# step4: Concat the parameters splits together after recv.
for varname, splited_var in param_var_mapping.iteritems():
eps = []
for var in splited_var:
......@@ -399,6 +425,7 @@ class DistributeTranspiler:
RPC_OP_ROLE_ATTR_NAME: RPC_OP_ROLE_ATTR_VALUE
})
# step4: Concat the parameters splits together after recv.
for varname, splited_var in param_var_mapping.iteritems():
if len(splited_var) <= 1:
continue
......
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
from ..framework import Program, default_main_program, Parameter, Variable
from ..layer_helper import LayerHelper
def hash_name_to_server(params_grads, pserver_endpoints):
"""
:param param_grads:
:return: a map of pserver endpoint ->
params -> [param list]
grads -> [grad list]
"""
def _hash_param(param_name, total):
return hash(param_name) % total
param_grad_map = dict()
for param, grad in params_grads:
if param.trainable is True and grad is not None:
server_id = _hash_param(param.name, len(pserver_endpoints))
server_for_param = pserver_endpoints[server_id]
if not param_grad_map.has_key(server_for_param):
param_grad_map[server_for_param] = {"params": [], "grads": []}
param_grad_map[server_for_param]["params"].append(param)
param_grad_map[server_for_param]["grads"].append(grad)
return param_grad_map
def round_robin(params_grads, pserver_endpoints):
assert (len(params_grads) > len(pserver_endpoints))
param_grad_map = dict()
pserver_idx = 0
for param, grad in params_grads:
if param.trainable is True:
server_for_param = pserver_endpoints[pserver_idx]
if not param_grad_map.has_key(server_for_param):
param_grad_map[server_for_param] = {"params": [], "grads": []}
param_grad_map[server_for_param]["params"].append(param)
param_grad_map[server_for_param]["grads"].append(grad)
pserver_idx += 1
if pserver_idx >= len(pserver_endpoints):
pserver_idx = 0
return param_grad_map
class SimpleDistributeTranspiler:
def transpile(self,
optimize_ops,
params_grads,
program=None,
pservers="127.0.0.1:6174",
trainers=1,
split_method=round_robin):
"""
Transpile the program to a distributed data-parallelism programs.
The main_program will be transform to use a remote parameter server
to do parameter optimization. And the optimization graph will be put
in to a parameter server program.
Use different methods to split trainable varialbles to different
parameter servers.
Example to run:
exe = fluid.Executor(place)
t = fluid.DistributeTranspiler()
t.transpile(optimize_ops, params_grads, pservers="127.0.0.1:6174", trainers=1)
pserver_endpoint = os.getenv("PSERVER")
if pserver_endpoint:
pserver_prog = t.get_pserver_program(pserver_endpoint, optimize_ops)
exe.run(fluid.default_startup_program())
exe.run(pserver_prog)
else:
feeder = fluid.DataFeeder(feed_list=[images, label], place=place)
exe.run(fluid.default_startup_program())
for pass_id in range(PASS_NUM):
...
:param optimize_ops: op list of optimization, should be the
return value of Optimizer.minimize
:type optimize_ops: list
:param program: program to optimize, default default_main_program
:param pservers: parameter server endpoints like "m1:6174,m2:6174"
:type pservers: string
:return: return a list of programs
"""
if program is None:
program = default_main_program()
self.program = program
self.trainers = trainers
self.optimize_ops = optimize_ops
self._optimize_distributed(
optimize_ops,
program,
params_grads,
pservers=pservers,
trainers=trainers,
split_method=split_method)
def _clone_param(self, block, v):
assert isinstance(v, Parameter)
new_p = Parameter(
block=block,
shape=v.shape,
dtype=v.dtype,
type=v.type,
lod_level=v.lod_level,
stop_gradient=v.stop_gradient,
trainable=v.trainable,
optimize_attr=v.optimize_attr,
regularizer=v.regularizer,
name=v.name)
block.vars[new_p.name] = new_p
def _clone_var(self, block, var):
assert isinstance(var, Variable)
return block.create_var(
name=var.name,
shape=var.shape,
dtype=var.dtype,
type=var.type,
lod_level=var.lod_level,
persistable=var.persistable)
def _optimize_distributed(self, optimize_ops, program, params_and_grads,
**kwargs):
if kwargs.has_key("split_method"):
split_method = kwargs["split_method"]
else:
split_method = round_robin
assert (callable(split_method))
pserver_endpoints = kwargs["pservers"].split(",")
self.param_grad_map = split_method(params_and_grads, pserver_endpoints)
send_op_ordered_inputs = []
send_op_ordered_outputs = []
epmap = []
for ep, v in self.param_grad_map.iteritems():
send_op_ordered_inputs.extend(v["grads"])
send_op_ordered_outputs.extend(v["params"])
for i in v["grads"]:
epmap.append(ep)
send_op = program.global_block().append_op(
type="send",
inputs={"X": send_op_ordered_inputs
}, # inputs is a list of tensors to be send
outputs={"Out": send_op_ordered_outputs},
attrs={"endpoints": pserver_endpoints,
"epmap": epmap})
def get_trainer_program(self):
# remove optimize ops and add a send op to main_program
self.program.global_block().delete_ops(self.optimize_ops)
return self.program
def _create_var_for_trainers(self, block, var, trainers):
var_list = []
for i in xrange(trainers):
var_each = block.create_var(
name="%s.trainer_%d" % (var.name, i),
psersistable=var.persistable,
dtype=var.dtype,
shape=var.shape)
var_list.append(var_each)
return var_list
def get_pserver_program(self, endpoint, optimize_ops):
pserver_program = Program()
for v in self.param_grad_map[endpoint]["params"]:
self._clone_param(pserver_program.global_block(), v)
optimize_sub_program = Program()
grad_var_names = [
var.name for var in self.param_grad_map[endpoint]["grads"]
]
for opt_op in optimize_ops:
for _, var in opt_op.inputs.iteritems():
# NOTE: append operators to merge gradients from multiple
# trainers. If trainers == 1, this is not needed.
if self.trainers > 1 and var.name in grad_var_names:
vars2merge = self._create_var_for_trainers(
optimize_sub_program.global_block(), var, self.trainers)
merged_var = optimize_sub_program.global_block().create_var(
name=var.name,
persistable=var.persistable,
dtype=var.dtype,
shape=var.shape)
optimize_sub_program.global_block().append_op(
type="sum",
inputs={"X": vars2merge},
outputs={"Out": merged_var})
optimize_sub_program.global_block().append_op(
type="scale",
inputs={"X": merged_var},
outputs={"Out": merged_var},
attrs={"scale": 1.0 / float(self.trainers)})
else:
optimize_sub_program.global_block().create_var(
name=var.name,
persistable=var.persistable,
dtype=var.dtype,
shape=var.shape)
if opt_op.inputs.has_key("Grad"):
if opt_op.inputs["Grad"].name in grad_var_names:
optimize_sub_program.global_block().append_op(
type=opt_op.type,
inputs=opt_op.inputs,
outputs=opt_op.outputs,
attrs=opt_op.attrs)
else:
optimize_sub_program.global_block().append_op(
type=opt_op.type,
inputs=opt_op.inputs,
outputs=opt_op.outputs,
attrs=opt_op.attrs)
pserver_program.global_block().append_op(
type="recv",
inputs={"RX":
self.param_grad_map[endpoint]["grads"]}, # grads to recv
outputs={},
attrs={
"OptimizeBlock": optimize_sub_program.global_block(),
"endpoint": endpoint,
"ParamList":
[p.name for p in self.param_grad_map[endpoint]["params"]],
"GradList":
[p.name for p in self.param_grad_map[endpoint]["grads"]],
"Trainers": self.trainers
})
pserver_program.sync_with_cpp()
return pserver_program
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册