In this article, we'll explain how to config and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
In this article, we'll explain how to configure and run distributed training jobs with PaddlePaddle Fluid in a bare metal cluster.
## Preparations
### Get your cluster ready
### Getting the cluster ready
Prepare your computer 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 with each other.
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.
### 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.
PaddlePaddle build and installation guide can be found from[here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
PaddlePaddle build and installation guide can be found [here](http://www.paddlepaddle.org/docs/develop/documentation/en/getstarted/build_and_install/index_en.html).
### Update training script
### Update the training script
#### 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.
This demo's non-cluster version with fluid API is as follows:
The non-cluster version of this demo with fluid API is as follows:
``` python
importpaddle.v2aspaddle
...
...
@@ -65,25 +65,25 @@ for pass_id in range(PASS_NUM):
exit(1)
```
We created a simple fully connected neural networks training program and handed it to the fluid executor to run for 100 passes.
We created a simple fully-connected neural network training program and handed it to the fluid executor to run for 100 passes.
Now let's try to convert it to a distributed version to run in a cluster.
Now let's try to convert it to a distributed version to run on a cluster.
#### Introducing parameter server
As you see from the non-cluster version of training script, there is only one role in it: the trainer, who does the computing as well as holding 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.
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.
Parameter Server in fluid does not only hold parameters but is also assigned with a part of the program. Trainers communicate with parameter servers via send/receive OPs. For more tech detail, please refer to this [document](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/dist_refactor/distributed_architecture.md).
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).
Now we need to create program for both trainers and parameter servers, the question is how?
Now we need to create programs for both: trainers and parameter servers, the question is how?
#### Slice the program
Fluid provides a tool called "Distribute Transpiler" to automatically convert the non-cluster program into cluster program.
Fluid provides a tool called "Distributed Transpiler" that automatically converts the non-cluster program into cluster program.
The idea behind this tool is to find optimize OPs and gradient parameters, slice the program into 2 pieces and connect them with send/receive OP.
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.
Optimize OPs and gradient parameters can be found from the return values of optimizer's minimize function.
...
...
@@ -94,9 +94,9 @@ To put them together:
optimize_ops,params_grads=sgd_optimizer.minimize(avg_cost)#get optimize OPs and gradient parameters
t=fluid.DistributeTranspiler()# create the transpiler instance
# 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
Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run this in the command line:
Please find the complete demo from [here](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/fluid/tests/book_distribute/notest_dist_fit_a_line.py). In parameter server node run the following in the command line:
*the reason you need to run this command twice in 2 nodes is: in the script we set the trainer count to be 2. You can change this setting on line 50*
*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*
Now you have 2 trainers and 1 parameter server up and running.
@@ -1866,6 +1867,146 @@ def matmul(x, y, transpose_x=False, transpose_y=False, name=None):
returnout
defedit_distance(input,
label,
normalized=False,
ignored_tokens=None,
name=None):
"""
EditDistance operator computes the edit distances between a batch of hypothesis strings and their references. Edit distance, also called Levenshtein distance, measures how dissimilar two strings are by counting the minimum number of operations to transform one string into anthor. Here the operations include insertion, deletion, and substitution. For example, given hypothesis string A = "kitten" and reference B = "sitting", the edit distance is 3 for A will be transformed into B at least after two substitutions and one insertion:
"kitten" -> "sitten" -> "sittin" -> "sitting"
Input(Hyps) is a LoDTensor consisting of all the hypothesis strings with the total number denoted by `batch_size`, and the separation is specified by the LoD information. And the `batch_size` reference strings are arranged in order in the same way in the LoDTensor Input(Refs).
Output(Out) contains the `batch_size` results and each stands for the edit stance for a pair of strings respectively. If Attr(normalized) is true, the edit distance will be divided by the length of reference string.
Args:
input(Variable): The indices for hypothesis strings.
label(Variable): The indices for reference strings.
normalized(bool): Indicated whether to normalize the edit distance by the length of reference string.
ignored_tokens(list of int): Tokens that should be removed before calculating edit distance.
Returns:
Variable: sequence-to-sequence edit distance in shape [batch_size, 1].
Examples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[8], dtype='float32')
y = fluid.layers.data(name='y', shape=[7], dtype='float32')
This op is used to decode sequences by greedy policy by below steps:
1. Get the indexes of max value for each row in input. a.k.a. numpy.argmax(input, axis=0).
2. For each sequence in result of step1, merge repeated tokens between two blanks and delete all blanks.
A simple example as below:
.. code-block:: text
Given:
input.data = [[0.6, 0.1, 0.3, 0.1],
[0.3, 0.2, 0.4, 0.1],
[0.1, 0.5, 0.1, 0.3],
[0.5, 0.1, 0.3, 0.1],
[0.5, 0.1, 0.3, 0.1],
[0.2, 0.2, 0.2, 0.4],
[0.2, 0.2, 0.1, 0.5],
[0.5, 0.1, 0.3, 0.1]]
input.lod = [[0, 4, 8]]
Then:
output.data = [[2],
[1],
[3]]
output.lod = [[0, 2, 3]]
Args:
input(Variable): (LoDTensor<float>), the probabilities of variable-length sequences, which is a 2-D Tensor with LoD information. It's shape is [Lp, num_classes + 1], where Lp is the sum of all input sequences' length and num_classes is the true number of classes. (not including the blank label).
blank(int): the blank label index of Connectionist Temporal Classification (CTC) loss, which is in thehalf-opened interval [0, num_classes + 1).
Returns:
Variable: CTC greedy decode result.
Examples:
.. code-block:: python
x = fluid.layers.data(name='x', shape=[8], dtype='float32')
