Skip to content

Distributed Training API and Benchmark For PaddlePaddle

License

Notifications You must be signed in to change notification settings

zhaocaibei123/Fleet

 
 

Repository files navigation

Fleet

Fleet is High-Level API for distributed training in PaddlePaddle. The name of Fleet means that a large crowd of ships working together to finish a large scale job. The design of Fleet makes a trade-off between easy-to-use and algorithmic extensibility and is highly efficient. First, a user can shift from local machine paddlepaddle code to distributed code within ten lines of code. Second, different algorithms can be easily defined through distributed strategy through Fleet API. Finally, distributed training is extremely fast with Fleet and just enjoy it.

Note: all the examples here should be replicated from develop branch of Paddle

Fleet is Highly Efficient

Deep neural networks training with Fleet API is highly efficient in PaddlePaddle. We benchmark serveral standard models here.

Parameter Server Training

Parameter server training benchmark is performed on click through rate estimation task on Criteo Dataset and Semantic Representation Learning on One-billion word Dataset. Details of hardware and software information for this benchmark can be found in parameter server benchmark.

Collective Training

Collective Training is usually used in GPU training in PaddlePaddle. Benchmark of collective training with Fleet is as follows. Details of hardware and software information for this benchmark can be found in benchmark environment.

Fleet is Easy To Use

Fleet is easy to use for both collective training and parameter server training. Here is an example for collective training with Fleet.

Local Single GPU Cards Training

import paddle.fluid as fluid
from utils import gen_data
from nets import mlp

input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')

cost = mlp(input_x, input_y)
optimizer = fluid.optimizer.SGD(learning_rate=0.01)
optimizer.minimize(cost, fluid.default_startup_program())

train_prog = fluid.default_main_program()
place = fluid.CUDAPlace(0)

exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

step = 1001
for i in range(step):
    cost_val = exe.run(program=train_prog, feed=gen_data(), fetch_list=[cost.name])

Local Multiple GPU Cards Training

import paddle.fluid as fluid
from utils import gen_data
from nets import mlp
from paddle.fluid.incubate.fleet.collective import fleet, DistributedStrategy  # new line 1 
from paddle.fluid.incubate.fleet.base import role_maker # new line 2

input_x = fluid.layers.data(name="x", shape=[32], dtype='float32')
input_y = fluid.layers.data(name="y", shape=[1], dtype='int64')

cost = mlp(input_x, input_y)
optimizer = fluid.optimizer.SGD(learning_rate=0.01)

role = role_maker.PaddleCloudRoleMaker(is_collective=True) # new line 3
fleet.init(role) # new line 4

optimizer = fleet.distributed_optimizer(optimizer, strategy=DistributedStrategy()) # new line 5
optimizer.minimize(cost, fluid.default_startup_program())

train_prog = fleet.main_program # change line 1
place = fluid.CUDAPlace(0)

exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

step = 1001
for i in range(step):
    cost_val = exe.run(program=train_prog, feed=gen_data(), fetch_list=[cost.name])

Launch command:

python -m paddle.distributed.launch --selected_gpus="0,1,2,3" trainer.py

More Examples

About

Distributed Training API and Benchmark For PaddlePaddle

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages

  • C 62.0%
  • Python 31.7%
  • C++ 2.7%
  • Cuda 1.9%
  • Shell 1.3%
  • Makefile 0.3%
  • CMake 0.1%