fluid_cluster_train_en.md 6.0 KB
Newer Older
1 2 3 4
# Fluid Distributed Training

## Introduction

5
In this article, we'll explain how to configure and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
6 7 8

## Preparations

9
### Getting the cluster ready
10

11
Prepare the compute nodes in the cluster. Nodes in this cluster can be of any specification that runs PaddlePaddle, and with a unique IP address assigned to it. Make sure they can communicate to each other.
12 13 14 15 16

### Have PaddlePaddle installed

PaddlePaddle must be installed on all nodes. If you have GPU cards on your nodes, be sure to properly install drivers and CUDA libraries.

17
PaddlePaddle build and installation guide can be found  [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
18

19 20 21 22 23 24
In addition to above, the `cmake` command should be run with the option `WITH_DISTRIBUTE` set to on. An example bare minimum `cmake` command would look as follows:

``` bash
cmake .. -DWITH_DOC=OFF -DWITH_GPU=OFF -DWITH_DISTRIBUTE=ON -DWITH_SWIG_PY=ON -DWITH_PYTHON=ON
```

25
### Update the training script
26 27 28 29 30

#### Non-cluster training script

Let's take [Deep Learning 101](http://www.paddlepaddle.org/docs/develop/book/01.fit_a_line/index.html)'s first chapter: "fit a line" as an example.

31
The non-cluster version of this demo with fluid API is as follows:
32 33 34

``` python
import paddle.v2 as paddle
35
import paddle.fluid as fluid
36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73

x = fluid.layers.data(name='x', shape=[13], dtype='float32')
y_predict = fluid.layers.fc(input=x, size=1, act=None)
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(x=cost)

sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
sgd_optimizer.minimize(avg_cost)

BATCH_SIZE = 20

train_reader = paddle.batch(
    paddle.reader.shuffle(
        paddle.dataset.uci_housing.train(), buf_size=500),
    batch_size=BATCH_SIZE)

place = fluid.CPUPlace()
feeder = fluid.DataFeeder(place=place, feed_list=[x, y])
exe = fluid.Executor(place)

exe.run(fluid.default_startup_program())

PASS_NUM = 100
for pass_id in range(PASS_NUM):
    fluid.io.save_persistables(exe, "./fit_a_line.model/")
    fluid.io.load_persistables(exe, "./fit_a_line.model/")
    for data in train_reader():
        avg_loss_value, = exe.run(fluid.default_main_program(),
                                  feed=feeder.feed(data),
                                  fetch_list=[avg_cost])

        if avg_loss_value[0] < 10.0:
            exit(0)  # if avg cost less than 10.0, we think our code is good.
exit(1)
```

74
We created a simple fully-connected neural network training program and handed it to the fluid executor to run for 100 passes.
75

76
Now let's try to convert it to a distributed version to run on a cluster.
77 78 79

#### Introducing parameter server

80
As we can see from the non-cluster version of training script, there is only one role in the script: the trainer, that performs the computing as well as holds the parameters. In cluster training, since multi-trainers are working on the same task, they need one centralized place to hold and distribute parameters. This centralized place is called the Parameter Server in PaddlePaddle.
81

82
![parameter server architecture](src/trainer.png)
83

84
Parameter Server in fluid not only holds the parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more technical details, please refer to  [this document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md).
85

86
Now we need to create programs for both: trainers and parameter servers, the question is how?
87 88 89

#### Slice the program

90
Fluid provides a tool called "Distributed Transpiler" that automatically converts the non-cluster program into cluster program.
91

92
The idea behind this tool is to find the optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP.
93 94 95 96 97 98 99 100 101 102

Optimize OPs and gradient parameters can be found from the return values of optimizer's minimize function.

To put them together:

``` python
... #define the program, cost, and create sgd optimizer

optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost) #get optimize OPs and gradient parameters

103
t = fluid.DistributeTranspiler() # create the transpiler instance
104
# slice the program into 2 pieces with optimizer_ops and gradient parameters list, as well as pserver_endpoints, which is a comma separated list of [IP:PORT] and number of trainers
105
t.transpile(optimize_ops, params_grads, pservers=pserver_endpoints, trainers=2)
106 107 108 109 110

... #create executor

# in pserver, run this
#current_endpoint here means current pserver IP:PORT you wish to run on
111 112 113 114
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint, pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
115 116 117 118 119 120 121 122 123 124 125 126 127

# in trainer, run this
... # define data reader
exe.run(fluid.default_startup_program())
for pass_id in range(100):
    for data in train_reader():
        exe.run(t.get_trainer_program())


```

### E2E demo

128
Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/fluid/tests/book_distribute/notest_dist_fit_a_line.py).
129 130 131
First `cd` into the folder that contains the `python` files. In this case:

```bash
132
cd /paddle/python/paddle/fluid/tests/book_distribute
133 134 135
```

In parameter server node run the following in the command line:
136 137 138 139 140 141 142 143 144

``` bash
PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=PSERVER python notest_dist_fit_a_line.py
```

*please note we assume that your parameter server runs at 192.168.1.2:6174*

Wait until the prompt `Server listening on 192.168.1.2:6174`

145
Then in 2 of your trainer nodes run this:
146 147 148 149 150

``` bash
PSERVERS=192.168.1.2:6174 SERVER_ENDPOINT=192.168.1.2:6174 TRAINING_ROLE=TRAINER python notest_dist_fit_a_line.py
```

151
*the reason you need to run this command twice in 2 nodes is because: in the script we set the trainer count to be 2. You can change this setting on line 50*
152 153

Now you have 2 trainers and 1 parameter server up and running.