提交 ff4208e6 编写于 作者: Q qiaolongfei

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into...

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into sgd-support-update-selected-rows
......@@ -36,11 +36,41 @@
- Trainer Count: 100
- Metrics: mini-batch / sec
| Batch Size | 32 | 64 | 128 | 256 |
| -- | -- | -- | -- | -- |
| PaddlePaddle Fluid | - | - | - | - |
| PaddlePaddle v2 | - | - | - | - |
| TensorFlow | - | - | - | - |
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
### Measure the Performance for Different PServer Count
......@@ -48,11 +78,41 @@
- Batch Size: 64
- Metrics: mini-batch / sec
| PServer Count | 10 | 20 | 40 | 60 |
| -- | -- | -- | -- | -- |
| PaddlePaddle Fluid | - | - | - | - |
| PaddlePaddle v2 | - | - | - | - |
| TensorFlow | - | - | - | - |
<table>
<thead>
<tr>
<th>PServer Count </th>
<th>10</th>
<th>20</th>
<th>40 </th>
<th>60</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
### Measure Parallel Efficiency By Increasing Trainer Count
......@@ -67,11 +127,69 @@ The parallel efficiency is:
$E = \div(S, N)$
| Trainer Counter | 1 | 10 | 20 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
| -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- |
| PaddlePaddle Fluid | - | - | - | - | - | - | - | - | - | - | - |
| PaddlePaddle v2 | - | - | - | - | - | - | - | - | - | - | - | - |
| TensorFlow | - | - | - | - | - | - | - | - | - | - | - | - | - |
<table>
<thead>
<tr>
<th>Trainer Counter </th>
<th>1</th>
<th>10</th>
<th>20 </th>
<th>30</th>
<th>40</th>
<th>50</th>
<th>60 </th>
<th>70</th>
<th>80</th>
<th>90</th>
<th>100 </th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
<tr>
<td>TensorFlow </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
<td>- </td>
<td>-</td>
<td>- </td>
<td>- </td>
</tr>
</tbody>
</table>
## Reproduce the benchmark
......
......@@ -16,11 +16,41 @@ Setting environment variable: `MKL_NUM_THREADS=1`.
- Metrics: samples / sec
| Batch Size | 32 | 64 | 128 | 256 |
| -- | -- | -- | -- | -- |
| PaddlePaddle Fluid | 15.44 | 16.32 | 16.74 | 16.79 |
| PaddlePaddle v2 | 15.97 | 17.04 | 17.60 | 17.83 |
| TensorFlow | 9.09 | 9.10 | 9.24 | 8.66 |
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 15.44 </td>
<td> 16.32 </td>
<td> 16.74 </td>
<td> 16.79 </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td> 15.97 </td>
<td> 17.04 </td>
<td> 17.60 </td>
<td> 17.83 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> 9.09 </td>
<td> 9.10 </td>
<td> 9.24 </td>
<td> 8.66 </td>
</tr>
</tbody>
</table>
### Different Batch Size
......@@ -28,12 +58,40 @@ Setting environment variable: `MKL_NUM_THREADS=1`.
- Trainer Count: 20
- Metrics: samples / sec
| Batch Size | 32 | 64 | 128 | 256 |
| -- | -- | -- | -- | -- |
| PaddlePaddle Fluid | 190.20 | 222.15 | 247.40 | 258.18 |
| PaddlePaddle v2 | 170.96 | 233.71 | 256.14 | 329.23 |
| TensorFlow | - | - | - | - |
<table>
<thead>
<tr>
<th>Batch Size </th>
<th> 32</th>
<th>64</th>
<th>128 </th>
<th>256</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 190.20 </td>
<td> 222.15 </td>
<td> 247.40 </td>
<td> 258.18 </td>
</tr>
<tr>
<td>PaddlePaddle v2 </td>
<td> 170.96 </td>
<td> 233.71 </td>
<td> 256.14 </td>
<td> 329.23 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
### Accelerate Rate
......@@ -41,11 +99,41 @@ Setting environment variable: `MKL_NUM_THREADS=1`.
- Batch Size: 128
- Metrics: samples / sec
| Trainer Count | 20 | 40 | 80 | 100 |
| -- | -- | -- | -- | -- |
| PaddlePaddle Fluid | 263.29 (78.64%) | 518.80 (77.47%) | 836.26 (62.44%) | 1019.29 (60.89%) |
| PaddlePaddle v2 (need more tests) | 326.85 (92.85%) | 534.58 (75.93%) | 853.30 (60.60%) | 1041.99 (59.20%) |
| TensorFlow | - | - | - | - |
<table>
<thead>
<tr>
<th>Trainer Count </th>
<th>20</th>
<th>40</th>
<th>80</th>
<th>100</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid</td>
<td> 263.29 (78.64%) </td>
<td> 518.80 (77.47%) </td>
<td> 836.26 (62.44%) </td>
<td> 1019.29 (60.89%) </td>
</tr>
<tr>
<td>PaddlePaddle v2 (need more tests) </td>
<td> 326.85 (92.85%) </td>
<td> 534.58 (75.93%) </td>
<td> 853.30 (60.60%) </td>
<td> 1041.99 (59.20%) </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
### Different Pserver Count
......@@ -53,11 +141,41 @@ Setting environment variable: `MKL_NUM_THREADS=1`.
- Batch Size: 128
- Metrics: samples/ sec
| PServer Count | 3 | 6 |10 | 20 |
| -- | -- | -- | -- | -- |
| PaddlePaddle Fluid(should fix in next PR) | 589.1 | 592.6 | 656.4 | 655.8 |
| PaddlePaddle v2 | 593.4 | 791.3 | 729.7 | 821.7 |
| TensorFlow | - | - | - | - |
<table>
<thead>
<tr>
<th>PServer Count </th>
<th>3</th>
<th>6</th>
<th>10</th>
<th>20</th>
</tr>
</thead>
<tbody>
<tr>
<td> PaddlePaddle Fluid(should fix in next PR) </td>
<td> 589.1 </td>
<td> 592.6 </td>
<td> 656.4 </td>
<td> 655.8 </td>
</tr>
<tr>
<td>PaddlePaddle v2 (need more tests) </td>
<td> 593.4 </td>
<td> 791.3 </td>
<td> 729.7 </td>
<td> 821.7 </td>
</tr>
<tr>
<td>TensorFlow </td>
<td> - </td>
<td> - </td>
<td> - </td>
<td> - </td>
</tr>
</tbody>
</table>
*The performance gap between Fuild and v2 comes from the network interference.*
......
......@@ -494,6 +494,12 @@ reshape
.. autofunction:: paddle.fluid.layers.reshape
:noindex:
pad
---
.. autofunction:: paddle.fluid.layers.pad
:noindex:
scale
-----
......
......@@ -5,9 +5,11 @@ In a large scale machine learning setup where the size of the training data is h
Polyak and Juditsky (1992) showed that the test performance of simple average of parameters obtained by Stochastic Gradient Descent (SGD) is as good as that of parameter values that are obtained by training the model over and over again, over the training dataset.
Hence, to accelerate the speed of Stochastic Gradient Descent, Averaged Stochastic Gradient Descent (ASGD) was proposed in Polyak and Juditsky (1992). For ASGD, the running average of parameters obtained by SGD, is used as the estimator for <img src="./images/theta_star.gif"/><br/> . The averaging is done as follows:
Hence, to accelerate the speed of Stochastic Gradient Descent, Averaged Stochastic Gradient Descent (ASGD) was proposed in Polyak and Juditsky (1992). For ASGD, the running average of parameters obtained by SGD, is used as the estimator for <img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/theta_star.gif"/><br/> . The averaging is done as follows:
<img src="./images/asgd.gif" align="center"/><br/>
<p align="center">
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/asgd.gif"><br />
</p>
We propose averaging for any optimizer similar to how ASGD performs it, as mentioned above.
......
......@@ -6,11 +6,33 @@ Here are some initial thoughts. Your comments are welcome!
I think we need only the following few CMake functions to make a project description mean and clean:
| C++ | CUDA C++ | Go |
|---|---|---|
| cc_library | nv_library | go_library |
| cc_binary | nv_binary | go_binary |
| cc_test | nv_test | go_test |
<table>
<thead>
<tr>
<th>C++</th>
<th>CUDA C++</th>
<th>Go</th>
</tr>
</thead>
<tbody>
<tr>
<td>cc_library </td>
<td>nv_library </td>
<td>go_library </td>
</tr>
<tr>
<td>cc_binary </td>
<td>nv_binary </td>
<td>go_binary </td>
</tr>
<tr>
<td> cc_test </td>
<td> nv_test </td>
<td> go_test </td>
</tr>
</tbody>
</table>
- The `_library` functions generate .a files from source code.
- The `_binary` functions generate executable binary files.
......
......@@ -14,11 +14,29 @@ In programming languages, a block is a pair of curly braces that includes local
Blocks work with control flow structures like `if`, `else`, and `for`, which have equivalents in deep learning:
| programming languages | PaddlePaddle |
|-----------------------|-----------------------|
| for, while loop | RNN, WhileOp |
| if, if-else, switch | IfElseOp, SwitchOp |
| sequential execution | a sequence of layers |
<table>
<thead>
<tr>
<th>programming languages</th>
<th>PaddlePaddle</th>
</tr>
</thead>
<tbody>
<tr>
<td>for, while loop </td>
<td>RNN, WhileOp </td>
</tr>
<tr>
<td>if, if-else, switch </td>
<td>IfElseOp, SwitchOp </td>
</tr>
<tr>
<td>sequential execution </td>
<td>a sequence of layers </td>
</tr>
</tbody>
</table>
A key difference is that a C++ program describes a one pass computation, whereas a deep learning program describes both the forward and backward passes.
......@@ -26,12 +44,33 @@ A key difference is that a C++ program describes a one pass computation, whereas
The existence of the backward pass makes the execution of a block of PaddlePaddle different from traditional programs:
| programming languages | PaddlePaddle |
|-----------------------|---------------------------------|
| stack | scope hierarchy |
| stack frame | scope |
| push at entering block| push at entering block |
| pop at leaving block | destroy when minibatch completes|
<table>
<thead>
<tr>
<th>programming languages</th>
<th>PaddlePaddle</th>
</tr>
</thead>
<tbody>
<tr>
<td>stack </td>
<td>scope hierarchy </td>
</tr>
<tr>
<td>stack frame </td>
<td>scope </td>
</tr>
<tr>
<td>push at entering block </td>
<td>push at entering block </td>
</tr>
<tr>
<td>pop at leaving block </td>
<td>destroy when minibatch completes </td>
</tr>
</tbody>
</table>
1. In traditional programs:
......
......@@ -86,12 +86,40 @@ def layer.fc(X):
We'd like to have Python bindings to operators in package `paddle.operator`, and Python compositions of operators in package `paddle.layer`. So we have the following concepts in above illustrative example:
| C++ functions/functors | mul | add | | |
|------------------------|--------------|--------------|-------------|----------|
| C++ operator class | mulOp | addOp | FCOp | |
| Python binding | operator.mul | operator.add | operator.fc | |
| Python function | | | | layer.fc |
<table>
<thead>
<tr>
<th>C++ functions/functors</th>
<th>mul</th>
<th>add</th>
<th></th>
<th></th>
</tr>
</thead>
<tbody>
<tr>
<td>C++ operator class </td>
<td>mulOp</td>
<td>addOp </td>
<td>FCOp </td>
<td></td>
</tr>
<tr>
<td>Python binding </td>
<td>operator.mul</td>
<td> operator.add </td>
<td>operator.fc </td>
<td></td>
</tr>
<tr>
<td>Python function </td>
<td></td>
<td></td>
<td> </td>
<td>layer.fc</td>
</tr>
</tbody>
</table>
This is how we differentiate layer and operators in PaddlePaddle:
......
......@@ -2,12 +2,38 @@
Like other deep learning systems, PaddlePaddle supports training models from sequence data. Also, like other systems, PaddlePaddle represent a mini-batch of sequences as a Tensor. What is different is that PaddlePaddle doesn't require all sequences in a mini-batch to be of the same length. Thus no need for padding zeros.
| | TensorFlow | PaddlePaddle |
|-----------------------|------------|--------------|
| RNN | Support | Support |
| recursive RNN | Support | Support |
| padding zeros | Must | No need |
| blob data type | Tensor | LoDTensor |
<table>
<thead>
<tr>
<th></th>
<th>TensorFlow</th>
<th>PaddlePaddle</th>
</tr>
</thead>
<tbody>
<tr>
<td>RNN </td>
<td>Support </td>
<td>Support </td>
</tr>
<tr>
<td>recursive RNN </td>
<td>Support </td>
<td>Support </td>
</tr>
<tr>
<td>padding zeros </td>
<td> Must </td>
<td>No need </td>
</tr>
<tr>
<td> blob data type </td>
<td> Tensor</td>
<td> LoDTensor </td>
</tr>
</tbody>
</table>
PaddlePaddle achieves this flexibility by passing through a new data type, *LoD Tensor*, which is a Tensor attached with segmentation index known as *LoD*, between operators. The LoD index doesn't only segment a tensor, but also recursively segments sub-sequences. This document presents the design of LoD and LoDTensor.
......
......@@ -10,10 +10,27 @@ PaddlePaddle uses proto message to describe compile time program because :
The computation `Program` consists of nested `Blocks`. Each `Block` will consist of data(i.e. `Variable`) and `Operations`. The concept to represent them is in the table below.
| |compile time|runtime|
|---|---|---|
|Data|VarDesc(proto)|Variable(cpp)|
|Operation|OpDesc(proto)|Operator(cpp)|
<table>
<thead>
<tr>
<th></th>
<th>compile time</th>
<th>runtime</th>
</tr>
</thead>
<tbody>
<tr>
<td>Data </td>
<td>VarDesc(proto) </td>
<td>Variable(cpp) </td>
</tr>
<tr>
<td>Operation </td>
<td>OpDesc(proto) </td>
<td>Operator(cpp) </td>
</tr>
</tbody>
</table>
## Definition of VarType
......
......@@ -2,7 +2,7 @@
## Introduction
A Channel is a data structure that allows for synchronous interprocess
A Channel is a data structure that allows for synchronous interprocess
communication via message passing. It is a fundemental component of CSP
(communicating sequential processes), and allows for users to pass data
between threads without having to worry about synchronization.
......@@ -18,7 +18,7 @@ Creates a new channel that takes in variables of a specific dtype.
- **fluid.make_channel(dtype, capacity=0)**
- **dtype**: The data type of variables being sent/received through channel
- **capacity**: The capacity of the channel. A capacity of 0 represents
- **capacity**: The capacity of the channel. A capacity of 0 represents
an unbuffered channel. Capacity > 0 represents a buffered channel
```
......@@ -40,8 +40,8 @@ fluid.channel_close(ch)
### Send data to a channel
Sends a variable to a channel. Currently, variables of dtype `LoDTensor`,
`LoDRankTable`, `LoDTensorArray`, `SelectedRows`, `ReaderHolder`, and
Sends a variable to a channel. Currently, variables of dtype `LoDTensor`,
`LoDRankTable`, `LoDTensorArray`, `SelectedRows`, `ReaderHolder`, and
`ChannelHolder` are supported.
By default, the data of the Variable is moved from the sender to the receiver,
......@@ -52,7 +52,7 @@ however the user can optionally copy the data before performing the send.
- **variable**: The variable to send to the channel
- **is_copy**: If set to True, channel_send will perform a variable assign
to copy the source variable to a new variable to be sent.
```
ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
var = fill_constant(shape=[1],dtype=core.VarDesc.VarType.INT32, value=100)
......@@ -68,7 +68,7 @@ receiving variable.
- **channel**: The channel to receive the variable from
- **return_variable**: The destination variable used to store the data of the
variable received from the channel
```
ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
var = fill_constant(shape=[1],dtype=core.VarDesc.VarType.INT32, value=-1)
......@@ -84,9 +84,9 @@ internal queues, locks, and conditional variables.
### QueueMessage
QueueMessage encapsulates the state of the channel send/receive operation to be
put in the **sendq/recvq**. It contains a condition variable used to lock the
put in the **sendq/recvq**. It contains a condition variable used to lock the
thread (when there are no available sends/receives). In addition, it contains
a callback function to notify a thread when the QueueMessage is being
a callback function to notify a thread when the QueueMessage is being
processed by the channel.
### Queues
......@@ -108,21 +108,21 @@ channel_recv operation will put a new QueueMessage on the recvq and block the
current thread under two conditions:
1. The channel is buffered and there is no data on the buff_
2. The channel is unbuffered and does not have a sender
### State diagram
#### Channel Send
<p align="center">
<img src="./images/channel_send.png"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/channel_send.png"/><br/>
</p>
#### Channel Receive
<p align="center">
<img src="./images/channel_recv.png"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/channel_recv.png"/><br/>
</p>
## Limitations and Considerations
### Variable Copy
......@@ -135,5 +135,5 @@ be sent before it is sent.
Please note that this is acheived by adding an **assign** operator and creating
a temporary variable that is sent in place of the original variable. Please
note that **assign** operator has limited support for only certain variables
note that **assign** operator has limited support for only certain variables
datatypes.
......@@ -10,12 +10,42 @@ The answer relies on the fact that a `ProgramDesc` is similar to an abstract syn
The following table compares concepts in Fluid and Go
| Go | Fluid |
|----|-------|
|user-defined functions | [layers](https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid) |
| control-flow and built-in functions | [intrinsics/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators) |
| goroutines, channels | [class ThreadPool](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework/thread_pool.h) |
| runtime | [class Executor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h) |
<table>
<thead>
<tr>
<th></th>
<th>Go</th>
<th>Fluid</th>
</tr>
</thead>
<tbody>
<tr>
<td>user-defined functions </td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/tree/develop/python/paddle/fluid">layers</a></td>
<td></td>
</tr>
<tr>
<td>control-flow and built-in functions </td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators">intrinsics/operators</a></td>
<td></td>
</tr>
<tr>
<td>goroutines, channels </td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/framework/thread_pool.h">class ThreadPool</a></td>
<td></td>
</tr>
<tr>
<td>runtime </td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/executor.h">class Executor</a></td>
<td></td>
</tr>
</tbody>
</table>
## An Example Concurrent Program
......@@ -77,11 +107,11 @@ message ProgramDesc {
read(output = X)
kube_get_workers_addrs(output = L)
Y = tensor_array(len(L))
parallel_for(input = X, output = Y,
parallel_for(input = X, output = Y,
attrs = {L, block_id(1)}) # referring to block 1
]
}
block[1] = Block {
parent = 0,
vars = [x, y, index],
......@@ -102,7 +132,7 @@ func main() { //// block 0
X = fluid.read(...)
L = fluid.k8s.get_worker_addrs()
Y = fluid.tensor_array(len(L))
fluid.parallel_for(X, L,
fluid.parallel_for(X, L,
func(index int) { //// block 1
x = X[index]
fluid.send(L[index], x)
......@@ -116,7 +146,7 @@ An explanation of the above program:
- `fluid.k8s` is a package that provides access to Kubernetes API.
- `fluid.k8s.get_worker_addrs` returns the list of IP and ports of all pods of the current job except for the current one (the master pod).
- `fluid.tensor_array` creates a [tensor array](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor_array.h). `fluid.parallel_for` creates a `ParallelFor` intrinsic, which, when executed,
- `fluid.tensor_array` creates a [tensor array](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor_array.h). `fluid.parallel_for` creates a `ParallelFor` intrinsic, which, when executed,
1. creates `len(L)` scopes, each for the concurrent running of the sub-block (block 1 in this case), and initializes a variable named "index" in the scope to an integer value in the range `[0, len(L)-1]`, and
2. creates `len(L)` threads by calling into the `ThreadPool` singleton, each thread
......
......@@ -13,14 +13,41 @@ Most DL systems, including TensorFlow, Caffe2, and MxNet, can asynchronously exe
There were many concurrent programming models, implemented in various forms:
| concurrent programming model | implementation |
|-----|-----|
| mutex | types and functions in standard libraries |
| semaphore | types and functions in standard libraries |
| communicating sequential processes (CSP) | Go programming language |
| actor model | Erlang programming language |
| message passing | MPI |
| bulk synchronous parallel (BSP) | Pregel distributed programming framework |
<table>
<thead>
<tr>
<th>concurrent programming model</th>
<th>implementation</th>
</tr>
</thead>
<tbody>
<tr>
<td>mutex </td>
<td>types and functions in standard libraries </td>
</tr>
<tr>
<td>semaphore </td>
<td> types and functions in standard libraries </td>
</tr>
<tr>
<td> communicating sequential processes (CSP) </td>
<td> Go programming language </td>
</tr>
<tr>
<td> actor model </td>
<td> Erlang programming language </td>
</tr>
<tr>
<td> message passing </td>
<td> MPI </td>
</tr>
<tr>
<td> bulk synchronous parallel (BSP) </td>
<td> Pregel distributed programming framework </td>
</tr>
</tbody>
</table>
Since Fluid was designed to be a programming language, we would like to implement CSP in Fluid.
......@@ -118,9 +145,9 @@ There are four types of actions with a channel:
```go
close(ch)
```
Please be aware that a closed channel is not a nil channel, which is `var ch chan int`.
There are some [axioms with channels](https://dave.cheney.net/2014/03/19/channel-axioms):
1. A send to a nil channel blocks forever
......
......@@ -2,13 +2,13 @@
## Introduction
In golang, the [**select**](https://golang.org/ref/spec#Select_statements)
statement lets a goroutine wait on multiple communication operations at the
same time. The **select** blocks until one of its cases can run, then
executes the case. If multiple cases are ready to run, then one case is
In golang, the [**select**](https://golang.org/ref/spec#Select_statements)
statement lets a goroutine wait on multiple communication operations at the
same time. The **select** blocks until one of its cases can run, then
executes the case. If multiple cases are ready to run, then one case is
choosen at random to be executed.
With the introduction of CSP for Paddle, we mimic this behavior by
With the introduction of CSP for Paddle, we mimic this behavior by
creating a ***select_op***.
## How to use it
......@@ -17,11 +17,11 @@ The **select_op** is available as a c++ operator. However most users
will prefer to use the much simplier Python API.
- **fluid.Select()**: Creates a select operator and adds it to the current
block within the main program. Also creates a sub block and adds it to the
main program. This sub block is used to hold all variables and operators
block within the main program. Also creates a sub block and adds it to the
main program. This sub block is used to hold all variables and operators
used by the case statements.
Within the select block, users can add cases by
Within the select block, users can add cases by
calling **select.case** or **select.default** method.
- **fluid.Select.case(channel_action, channel, result_variable)**: Represents
......@@ -37,13 +37,13 @@ execute.
```
ch1 = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
quit_ch = fluid.make_channel(dtype=core.VarDesc.VarType.LOD_TENSOR)
x = fill_constant(shape=[1], dtype=core.VarDesc.VarType.INT32, value=0)
y = fill_constant(shape=[1], dtype=core.VarDesc.VarType.INT32, value=1)
while_cond = fill_constant(shape=[1], dtype=core.VarDesc.VarType.BOOL, value=True)
while_op = While(cond=while_cond)
with while_op.block():
with fluid.Select() as select:
with select.case(fluid.channel_send, channel, x):
......@@ -99,17 +99,17 @@ blocks {
}
}
// Create "select" operator.
// inputs:
// inputs:
// X: All input variables used by operators within the select block
// case_to_execute: Variable filled in by select_op when it determines
// which case to execute.
//
// outputs:
// Out: All output variables referenced by operators within select block.
//
// Out: All output variables referenced by operators within select block.
//
// attrs:
// sub_block: The block id containing the select "cases"
// cases: Serialized list of all cases in the select op.
// cases: Serialized list of all cases in the select op.
// Each case is serialized as: '<index>,<type>,<channel>,<value>'
// where type is 0 for default, 1 for send, and 2 for receive.
// No channel and values are needed for default cases.
......@@ -150,7 +150,7 @@ into **X**. It will also create a temp variable called **case_to_execute**. Th
filled in by the select_op after it has completed processing the case statements.
If there are no available cases to execute (ie: all cases are blocked on channel operations, and
there is no default statement), then the select_op will block the current thread. The thread will
there is no default statement), then the select_op will block the current thread. The thread will
unblock once there is a channel operation affecting one of the case statements, at which point, the
**select_op** will set the **case_to_execute** variable to the index of the case to execute.
......@@ -247,17 +247,17 @@ blocks {
```
Cases are represented by a **conditional_block operator**, whose's condition is set as the output of
equal(**case_to_execute**, **case_index**). Since each case index is unique in this sub-block,
Cases are represented by a **conditional_block operator**, whose's condition is set as the output of
equal(**case_to_execute**, **case_index**). Since each case index is unique in this sub-block,
only one case will be executed.
### select_op flow
<p align="center">
<img src="./images/select_op_workflow.png"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/select_op_workflow.png"/><br/>
</p>
The select algorithm is inspired by golang's select routine. Please refer to
The select algorithm is inspired by golang's select routine. Please refer to
http://www.tapirgames.com/blog/golang-concurrent-select-implementation for more information.
## Backward Pass
......
......@@ -40,11 +40,11 @@ computation is only specified in Python code which sits outside of PaddlePaddle,
Similar to how a compiler uses an intermediate representation (IR) so that the programmer does not need to manually optimize their code for most of the cases, we can have an intermediate representation in PaddlePaddle as well. The compiler optimizes the IR as follows:
<img src="src/compiler.png"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/compiler.png"/>
PaddlePaddle can support model parallelism by converting the IR so that the user no longer needs to manually perform the computation and operations in the Python component:
<img src="src/paddle-compile.png"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/paddle-compile.png"/>
The IR for PaddlePaddle after refactoring is called a `Block`, it specifies the computation dependency graph and the variables used in the computation.
......@@ -60,7 +60,7 @@ For a detailed explanation, refer to this document -
The revamped distributed training architecture can address the above discussed limitations. Below is the illustration of how it does so:
<img src="src/distributed_architecture.png"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/distributed_architecture.png"/>
The major components are: *Python API*, *Distribute Transpiler* and *Remote Executor*.
......@@ -152,7 +152,7 @@ for data in train_reader():
`JobDesc` object describe the distributed job resource specification to run on
Cluster environment.
<img src="src/remote_executor.png" width="500" align="center" />
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/remote_executor.png" width="500" align="center" />
`RemoteExecutor.run` sends the `ProgramDesc` and
[TrainingJob](https://github.com/PaddlePaddle/cloud/blob/unreleased-tpr/doc/autoscale/README.md#training-job-resource)
......@@ -171,7 +171,7 @@ In the future, a more general placement algorithm should be implemented, which m
The local training architecture will be the same as the distributed training architecture, the difference is that everything runs locally, and there is just one PaddlePaddle runtime:
<img src="src/local_architecture.png"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/local_architecture.png"/>
### Training Data
......
......@@ -8,11 +8,11 @@ Op graph to a multi-CPU Op graph, and run `ParallelDo` Op to run the graph.
## Transpiler
<img src="src/multi-threads/single-thread@3x.png" width="300">
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/single-thread@3x.png" width="300">
After converted:
<img src="src/multi-threads/multi-threads@3x.png" width="1000">
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/multi-threads@3x.png" width="1000">
## Implement
......
......@@ -41,11 +41,11 @@ We will need these OPs: *Send*, *Recv*, *Enqueue*, *Dequeue*.
Below is an example of converting the user defined graph to the
subgraphs for the trainer and the parameter server:
<img src="src/local-graph.png" width="300"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/local-graph.png" width="300"/>
After converting:
<img src="src/dist-graph.png" width="700"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/dist-graph.png" width="700"/>
1. The parameter variable W and its optimizer program are placed on the parameter server.
1. Operators are added to the program.
......@@ -69,8 +69,7 @@ In Fluid, we introduce [SelectedRows](../selected_rows.md) to represent a list o
non-zero gradient data. So when we do parameter optimization both locally and remotely,
we only need to send those non-zero rows to the optimizer operators:
<img src="src/sparse_update.png" width="700" />
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/sparse_update.png" width="700" />
### Benefits
- Model parallelism becomes easier to implement: it is an extension to
......
......@@ -5,7 +5,7 @@ This document describes the RNN (Recurrent Neural Network) operator and how it i
## RNN Algorithm Implementation
<p align="center">
<img src="./rnn.jpg"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/rnn.jpg"/>
</p>
The above diagram shows an RNN unrolled into a full network.
......@@ -22,7 +22,7 @@ There are several important concepts here:
There could be local variables defined in each step-net. PaddlePaddle runtime realizes these variables in *step-scopes* which are created for each step.
<p align="center">
<img src="./rnn.png"/><br/>
<img src="https://github.com/PaddlePaddle/Paddle/tree/develop/doc/fluid/images/rnn.png"/><br/>
Figure 2 illustrates the RNN's data flow
</p>
......@@ -93,7 +93,7 @@ For example, we could have a 2-level RNN, where the top level corresponds to par
The following figure illustrates feeding in text into the lower level, one sentence at a step, and the feeding in step outputs to the top level. The final top level output is about the whole text.
<p align="center">
<img src="./2_level_rnn.png"/>
<img src="https://github.com/PaddlePaddle/Paddle/tree/develop/doc/fluid/images/2_level_rnn.png"/>
</p>
```python
......@@ -149,5 +149,5 @@ If the `output_all_steps` is set to False, it will only output the final time st
<p align="center">
<img src="./rnn_2level_data.png"/>
<img src="https://github.com/PaddlePaddle/Paddle/tree/develop/doc/fluid/images/rnn_2level_data.png"/>
</p>
......@@ -2,7 +2,7 @@
## What is batch normalization
Batch normalization is a frequently-used method in deep network training. It adjusts the mean and variance of a layer's output, and make the data distribution easier for next layer's training.
Batch normalization is a frequently-used method in deep network training. It adjusts the mean and variance of a layer's output, and make the data distribution easier for next layer's training.
The principle of batch normalization can be summarized into a simple function:
......@@ -66,7 +66,7 @@ As most C++ operators do, `batch_norm_op` is defined by inputs, outputs, attribu
The following graph showes the training computational process of `batch_norm_op`:
<img src="../images/batch_norm_op_kernel.png" width="800"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/batch_norm_op_kernel.png" width="800"/>
cudnn provides APIs to finish the whole series of computation, we can use them in our GPU kernel.
......@@ -74,13 +74,13 @@ cudnn provides APIs to finish the whole series of computation, we can use them i
`batch_norm_op` is warpped as a layer in Python:
```python
def batch_norm_layer(net,
```python
def batch_norm_layer(net,
input,
output,
scale,
bias,
use_global_est = False,
output,
scale,
bias,
use_global_est = False,
epsilon = 1e-6,
momentum = 0.99):
mean_cache = scope.new_var(name = 'estimated_mean', trainable = False)
......@@ -119,15 +119,15 @@ for pass_id in range(PASS_NUM):
if pass_id % 100 == 0:
net.infer(test_image) # run inferencing model
# ...
```
```
`is_infer` is an attribute. Once an operator is created, its attributes can not be changed. It suggests us that we shall maintain two `batch_norm_op` in the model, one's `is_infer` is `True`(we call it `infer_batch_norm_op`) and the other one's is `False`(we call it `train_batch_norm_op`). They share all parameters and variables, but be placed in two different branches. That is to say, if a network contains a `batch_norm_op`, it will fork into two branches, one go through `train_batch_norm_op` and the other one go through `infer_batch_norm_op`:
<div align=center>
<img src="../images/batch_norm_fork.png" width="500"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/batch_norm_fork.png" width="500"/>
</div>
Just like what is shown in the above graph, the net forks before `batch_norm_op` and will never merge again. All the operators after `batch_norm_op` will duplicate.
Just like what is shown in the above graph, the net forks before `batch_norm_op` and will never merge again. All the operators after `batch_norm_op` will duplicate.
When the net runs in training mode, the end of the left branch will be set as the running target, so the dependency tracking process will ignore right branch automatically. When the net runs in inferencing mode, the process is reversed.
......
......@@ -2,12 +2,33 @@
Due to the refactorization of the PaddlePaddle core, we need Python classes to construct corresponding protobuf messages that describe a DL program.
| Python classes | Protobuf messages |
| --- | --- |
| Program | ProgramDesc |
| Block | BlockDesc |
| Operator | OpDesc |
| Variable | VarDesc |
<table>
<thead>
<tr>
<th>Python classes</th>
<th>Protobuf messages</th>
</tr>
</thead>
<tbody>
<tr>
<td>Program </td>
<td>ProgramDesc </td>
</tr>
<tr>
<td>Block </td>
<td>BlockDesc </td>
</tr>
<tr>
<td>Operator </td>
<td>OpDesc </td>
</tr>
<tr>
<td>Variable </td>
<td>VarDesc </td>
</tr>
</tbody>
</table>
Please be aware that these Python classes need to maintain some construction-time information, which are not part of the protobuf messages.
......
......@@ -6,23 +6,23 @@ A central problem in machine learning is how to design an algorithm that will pe
### Parameter Norm Penalties
Most common regularization approaches in deep learning are based on limiting the capacity of the models by adding a parameter norm penalty to the objective function `J`. This is given as follows:
<img src="./images/loss_equation.png" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/loss_equation.png" align="center"/><br/>
The parameter `alpha` is a hyperparameter that weights the relative contribution of the norm penalty term, `omega`, relative to the standard objective function `J`.
The most commonly used norm penalties are the L2 norm penalty and the L1 norm penalty. These are given as follows:
##### L2 Regularization:
<img src="./images/l2_regularization.png" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/l2_regularization.png" align="center"/><br/>
##### L1 Regularization
<img src="./images/l1_regularization.png" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/l1_regularization.png" align="center"/><br/>
A much more detailed mathematical background of regularization can be found [here](http://www.deeplearningbook.org/contents/regularization.html).
## Regularization Survey
A detailed survey of regularization in various deep learning frameworks can be found [here](https://github.com/PaddlePaddle/Paddle/wiki/Regularization-Survey).
A detailed survey of regularization in various deep learning frameworks can be found [here](https://github.com/PaddlePaddle/Paddle/wiki/Regularization-Survey).
## Proposal for Regularization in PaddlePaddle
......@@ -32,41 +32,35 @@ In the new design, we propose to create new operations for regularization. For n
- L2_regularization_op
- L1_regularization_op
These ops can be like any other ops with their own CPU/GPU implementations either using Eigen or separate CPU and GPU kernels. As the initial implementation, we can implement their kernels using Eigen following the abstraction pattern implemented for [Activation Ops](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/accuracy_op.h). This abstraction pattern can make it very easy to implement new regularization schemes other than L1 and L2 norm penalties.
These ops can be like any other ops with their own CPU/GPU implementations either using Eigen or separate CPU and GPU kernels. As the initial implementation, we can implement their kernels using Eigen following the abstraction pattern implemented for [Activation Ops](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/accuracy_op.h). This abstraction pattern can make it very easy to implement new regularization schemes other than L1 and L2 norm penalties.
The idea of building ops for regularization is in sync with the refactored Paddle philosophy of using operators to represent any computation unit. The way these ops will be added to the computation graph, will be decided by the [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) in Python API.
The idea of building ops for regularization is in sync with the refactored Paddle philosophy of using operators to represent any computation unit. The way these ops will be added to the computation graph, will be decided by the [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) in Python API.
### Computation Graph
Below is an example of a really simple feed forward neural network.
<img src="./images/feed_forward.png" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/feed_forward.png" align="center"/><br/>
The Python API will modify this computation graph to add regularization operators. The modified computation graph will look as follows:
<img src="./images/feed_forward_regularized.png" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/feed_forward_regularized.png" align="center"/><br/>
   
### Python API implementation for Regularization
Using the low level ops, `L2_regularization_op` and `L1_regularization_op`, any user can add regularization to their computation graphs. However, this will require a lot of lines of code and we should design Python APIs that support regularization. An example of such an API can be seen in [Keras](https://keras.io/regularizers/). As per the PaddlePaddle [Python API design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md), the layer functions are responsible for creating operators, operator parameters and variables. Since regularization is a property of parameters, it makes sense to create these in the layer functions.
Using the low level ops, `L2_regularization_op` and `L1_regularization_op`, any user can add regularization to their computation graphs. However, this will require a lot of lines of code and we should design Python APIs that support regularization. An example of such an API can be seen in [Keras](https://keras.io/regularizers/). As per the PaddlePaddle [Python API design](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md), the layer functions are responsible for creating operators, operator parameters and variables. Since regularization is a property of parameters, it makes sense to create these in the layer functions.
#### Creation of Regularization ops
There are two possibilities for creating the regularization ops:
1. We create these ops immediately while building the computation graph.
2. We add these ops in a lazy manner, just before the backward, similar to the way the optimization ops are added.
1. We create these ops immediately while building the computation graph.
2. We add these ops in a lazy manner, just before the backward, similar to the way the optimization ops are added.
The proposal is to add these ops in a lazy manner just before the backward pass.
The proposal is to add these ops in a lazy manner just before the backward pass.
#### Storage of Regularization attributes
Since we want to create the regularization ops in a lazy manner, the regularization attributes (type of regularization and weight of regularization penalty) can be stored as attributes of the [`Parameter`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/framework.py#L421) class. This is because regularization is a property of the parameters and storing regularization properties with Parameters also allows for shared parameters.
Since we want to create the regularization ops in a lazy manner, the regularization attributes (type of regularization and weight of regularization penalty) can be stored as attributes of the [`Parameter`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/framework.py#L421) class. This is because regularization is a property of the parameters and storing regularization properties with Parameters also allows for shared parameters.
#### High-level API
In PaddlePaddle Python API, users will primarily rely on [layer functions](https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/python_api.md#layer-function) to create neural network layers. Hence, we also need to provide regularization functionality in layer functions. The design of these APIs can be postponed for later right now. A good reference for these APIs can be found in [Keras](https://keras.io/regularizers/) and also by looking at Tensorflow in [`tf.contrib.layers`](https://www.tensorflow.org/api_guides/python/contrib.layers).
......@@ -10,11 +10,37 @@ Fluid is the answer. Fluid is similar to PyTorch and TensorFlow Eager Execution
Deep learning infrastructure is one of the fastest evolving technologies. Within four years, there have already been three generations of technologies invented.
| Existed since | model as sequence of layers | model as graph of operators | No model |
|--|--|--|--|
| 2013 | Caffe, Theano, Torch, PaddlePaddle | | |
| 2015 | | TensorFlow, MxNet, Caffe2, ONNX, n-graph | |
| 2016 | | | PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid |
<table>
<thead>
<tr>
<th>Existed since</th>
<th>model as sequence of layers</th>
<th>model as graph of operators</th>
<th>No model</th>
</tr>
</thead>
<tbody>
<tr>
<td>2013 </td>
<td>Caffe, Theano, Torch, PaddlePaddle </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>2015 </td>
<td> </td>
<td>TensorFlow, MxNet, Caffe2, ONNX, n-graph </td>
<td> </td>
</tr>
<tr>
<td>2016 </td>
<td> </td>
<td> </td>
<td> PyTorch, TensorFlow Eager Execution, PaddlePaddle Fluid</td>
</tr>
</tbody>
</table>
From the above table, we see that the deep learning technology is evolving towards getting rid of the concept of a model. To understand the reasons behind this direction, a comparison of the *programming paradigms* or the ways to program deep learning applications using these systems, would be helpful. The following section goes over these.
......
......@@ -36,11 +36,37 @@ At compile time, the Python program generates a protobuf message representation
At runtime, the C++ program realizes the graph and runs it.
| | Representation (protobuf messages) | Realization (C++ class objects) |
|---|---|---|
|Data|[VarDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107)|[Variable](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24)|
|Operation|[OpDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35)|[Operator](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64)|
|Block|BlockDesc|Block|
<table>
<thead>
<tr>
<th></th>
<th>Representation (protobuf messages)</th>
<th>Realization (C++ class objects) </th>
</tr>
</thead>
<tbody>
<tr>
<td>Data</td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L107">VarDesc</a></td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/variable.h#L24">Variable</a></td>
</tr>
<tr>
<td>Operation </td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L35">OpDesc</a></td>
<td>
<a href="https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/operator.h#L64">Operator</a></td>
</tr>
<tr>
<td>Block </td>
<td>BlockDesc </td>
<td>Block </td>
</tbody>
</table>
The word *graph* is interchangeable with *block* in this document. A graph consists of computation steps and local variables similar to a C++/Java program block, or a pair of parentheses(`{` and `}`).
......
# DeepSpeech2 on PaddlePaddle: Design Doc
# DeepSpeech2 on PaddlePaddle: Design Doc
We are planning to build Deep Speech 2 (DS2) \[[1](#references)\], a powerful Automatic Speech Recognition (ASR) engine, on PaddlePaddle. For the first-stage plan, we have the following short-term goals:
......@@ -68,11 +68,33 @@ We roughly break down the project into 14 tasks:
Tasks parallelizable within phases:
Roadmap | Description | Parallelizable Tasks
----------- | :------------------------------------ | :--------------------
Phase I | Simplified model & components | *Task 1* ~ *Task 8*
Phase II | Standard model & benchmarking & profiling | *Task 9* ~ *Task 12*
Phase III | Documentations | *Task13* ~ *Task14*
<table>
<thead>
<tr>
<th>Roadmap</th>
<th>Description</th>
<th> Parallelizable Tasks</th>
</tr>
</thead>
<tbody>
<tr>
<td>Phase I </td>
<td>Simplified model & components </td>
<td>Task 1 ~ Task 8</td>
</tr>
<tr>
<td>Phase II </td>
<td> Standard model & benchmarking & profiling</td>
<td>Task 9 ~ Task 12 </td>
</tr>
<tr>
<td>Phase III </td>
<td> Documentations</td>
<td> Task13 ~ Task14 </td>
</tr>
</tbody>
</table>
Issue for each task will be created later. Contributions, discussions and comments are all highly appreciated and welcomed!
......@@ -94,7 +116,7 @@ The classical DS2 network contains 15 layers (from bottom to top):
- **One** CTC-loss layer
<div align="center">
<img src="images/ds2_network.png" width=350><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/ds2_network.png" width=350><br/>
Figure 1. Archetecture of Deep Speech 2 Network.
</div>
......@@ -102,37 +124,82 @@ We don't have to persist on this 2-3-7-1-1-1 depth \[[2](#references)\]. Similar
Key ingredients about the layers:
- **Data Layers**:
- **Data Layers**:
- Frame sequences data of audio **spectrogram** (with FFT).
- Token sequences data of **transcription** text (labels).
- Token sequences data of **transcription** text (labels).
- These two type of sequences do not have the same lengthes, thus a CTC-loss layer is required.
- **2D Convolution Layers**:
- **2D Convolution Layers**:
- Not only temporal convolution, but also **frequency convolution**. Like a 2D image convolution, but with a variable dimension (i.e. temporal dimension).
- With striding for only the first convlution layer.
- No pooling for all convolution layers.
- **Uni-directional RNNs**
- **Uni-directional RNNs**
- Uni-directional + row convolution: for low-latency inference.
- Bi-direcitional + without row convolution: if we don't care about the inference latency.
- **Row convolution**:
- For looking only a few steps ahead into the feature, instead of looking into a whole sequence in bi-directional RNNs.
- Not nessesary if with bi-direcitional RNNs.
- Not nessesary if with bi-direcitional RNNs.
- "**Row**" means convolutions are done within each frequency dimension (row), and no convolution kernels shared across.
- **Batch Normalization Layers**:
- Added to all above layers (except for data and loss layer).
- Sequence-wise normalization for RNNs: BatchNorm only performed on input-state projection and not state-state projection, for efficiency consideration.
Required Components | PaddlePaddle Support | Need to Develop
:------------------------------------- | :-------------------------------------- | :-----------------------
Data Layer I (Spectrogram) | Not supported yet. | TBD (Task 3)
Data Layer II (Transcription) | `paddle.data_type.integer_value_sequence` | -
2D Convolution Layer | `paddle.layer.image_conv_layer` | -
DataType Converter (vec2seq) | `paddle.layer.block_expand` | -
Bi-/Uni-directional RNNs | `paddle.layer.recurrent_group` | -
Row Convolution Layer | Not supported yet. | TBD (Task 4)
CTC-loss Layer | `paddle.layer.warp_ctc` | -
Batch Normalization Layer | `paddle.layer.batch_norm` | -
CTC-Beam search | Not supported yet. | TBD (Task 6)
<table>
<thead>
<tr>
<th>Required Components</th>
<th> PaddlePaddle Support</th>
<th> Need to Develop</th>
</tr>
</thead>
<tbody>
<tr>
<td>Data Layer I (Spectrogram) </td>
<td>Not supported yet.</td>
<td>TBD (Task 3)</td>
</tr>
<tr>
<td>Data Layer II (Transcription) </td>
<td> paddle.data_type.integer_value_sequence</td>
<td> - </td>
</tr>
<tr>
<td>2D Convolution Layer </td>
<td> paddle.layer.image_conv_layer</td>
<td> - </td>
</tr>
<tr>
<td>DataType Converter (vec2seq)</td>
<td> paddle.layer.block_expand</td>
<td> - </td>
</tr>
<tr>
<td>Bi-/Uni-directional RNNs </td>
<td>paddle.layer.recurrent_group</td>
<td> - </td>
</tr>
<tr>
<td>Row Convolution Layer </td>
<td>Not supported yet.</td>
<td>TBD (Task 4)</td>
</tr>
<tr>
<td>CTC-loss Layer </td>
<td>paddle.layer.warp_ctc</td>
<td> - </td>
</tr>
<tr>
<td>Batch Normalization Layer </td>
<td>paddle.layer.batch_norm</td>
<td> - </td>
</tr>
<tr>
<td>CTC-Beam search </td>
<td>Not supported yet.</td>
<td> TBD (Task 6) </td>
</tr>
</tbody>
</table>
### Row Convolution
......@@ -141,18 +208,18 @@ TODO by Assignees
### Beam Search with CTC and LM
<div align="center">
<img src="images/beam_search.png" width=600><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/beam_search.png" width=600><br/>
Figure 2. Algorithm for CTC Beam Search Decoder.
</div>
- The **Beam Search Decoder** for DS2 CTC-trained network follows the similar approach in \[[3](#references)\] as shown in Figure 2, with two important modifications for the ambiguous parts:
- 1) in the iterative computation of probabilities, the assignment operation is changed to accumulation for one prefix may comes from different paths;
- The **Beam Search Decoder** for DS2 CTC-trained network follows the similar approach in \[[3](#references)\] as shown in Figure 2, with two important modifications for the ambiguous parts:
- 1) in the iterative computation of probabilities, the assignment operation is changed to accumulation for one prefix may comes from different paths;
- 2) the if condition ```if l^+ not in A_prev then``` after probabilities' computation is deprecated for it is hard to understand and seems unnecessary.
- An **external scorer** would be passed into the decoder to evaluate a candidate prefix during decoding whenever a white space appended in English decoding and any character appended in Mandarin decoding.
- Such external scorer consists of language model, word count or any other custom scorers.
- The **language model** is built from Task 5, with parameters should be carefully tuned to achieve minimum WER/CER (c.f. Task 7)
- This decoder needs to perform with **high efficiency** for the convenience of parameters tuning and speech recognition in reality.
- This decoder needs to perform with **high efficiency** for the convenience of parameters tuning and speech recognition in reality.
## Future Work
......
......@@ -199,7 +199,7 @@ Packing the `selected_generation_scores` will get a `LoDTensor`, and each tail i
## LoD and shape changes during decoding
<p align="center">
<img src="./images/LOD-and-shape-changes-during-decoding.jpg"/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/LOD-and-shape-changes-during-decoding.jpg"/>
</p>
According to the image above, the only phase that changes the LoD is beam search.
......
# Design for GAN
GAN (General Adversarial Net [https://arxiv.org/abs/1406.2661]) is an important model for unsupervised learning and widely used in many areas.
GAN (General Adversarial Net [https://arxiv.org/abs/1406.2661]) is an important model for unsupervised learning and widely used in many areas.
It applies several important concepts in machine learning system design, including building and running subgraphs, dependency tracing, different optimizers in one executor and so forth.
In our GAN design, we wrap it as a user-friendly easily customized python API to design different models. We take the conditional DC-GAN (Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks [https://arxiv.org/abs/1511.06434]) as an example due to its good performance on image generation.
<p align="center">
<img src="./test.dot.png" width = "35%" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/test.dot.png" width = "35%" align="center"/><br/>
Figure 1. The overall running logic of GAN. The black solid arrows indicate the forward pass; the green dashed arrows indicate the backward pass of generator training; the red dashed arrows indicate the backward pass of the discriminator training. The BP pass of the green (red) arrow should only update the parameters in the green (red) boxes. The diamonds indicate the data providers. d\_loss and g\_loss marked in red and green are the two targets we would like to run.
</p>
The operators, layers and functions required/optional to build a GAN demo is summarized in https://github.com/PaddlePaddle/Paddle/issues/4563.
<p align="center">
<img src="./dcgan.png" width = "90%" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/dcgan.png" width = "90%" align="center"/><br/>
Figure 2. Photo borrowed from the original DC-GAN paper.
</p>
## The Conditional-GAN might be a class.
## The Conditional-GAN might be a class.
This design we adopt the popular open source design in https://github.com/carpedm20/DCGAN-tensorflow and https://github.com/rajathkmp/DCGAN. It contains following data structure:
- DCGAN(object): which contains everything required to build a GAN model. It provides following member functions methods as API:
......@@ -29,7 +29,7 @@ This design we adopt the popular open source design in https://github.com/carped
Returns a generated image.
- discriminator(image):
Given an image, decide if it is from a real source or a fake one.
Given an image, decide if it is from a real source or a fake one.
Returns a 0/1 binary label.
- build_model(self):
......@@ -47,7 +47,7 @@ To be more detailed, we introduce our design of DCGAN as following:
```python
class DCGAN(object):
def __init__(self, y_dim=None):
# hyper parameters
self.y_dim = y_dim # conditional gan or not
self.batch_size = 100
......@@ -82,18 +82,18 @@ class DCGAN(object):
# input z: the random noise
# input y: input data label (optional)
# output G_im: generated fake images
if not self.y_dim:
z = pd.layer.concat(1, [z, y])
G_h0 = pd.layer.fc(z, self.G_w0, self.G_b0)
G_h0_bn = pd.layer.batch_norm(G_h0)
G_h0_relu = pd.layer.relu(G_h0_bn)
G_h1 = pd.layer.deconv(G_h0_relu, self.G_w1, self.G_b1)
G_h1_bn = pd.layer.batch_norm(G_h1)
G_h1_relu = pd.layer.relu(G_h1_bn)
G_h2 = pd.layer.deconv(G_h1_relu, self.G_W2, self.G_b2))
G_im = pd.layer.tanh(G_im)
return G_im
......@@ -111,11 +111,11 @@ class DCGAN(object):
D_h0 = pd.layer.conv2d(image, w=self.D_w0, b=self.D_b0)
D_h0_bn = pd.layer.batchnorm(h0)
D_h0_relu = pd.layer.lrelu(h0_bn)
D_h1 = pd.layer.conv2d(D_h0_relu, w=self.D_w1, b=self.D_b1)
D_h1_bn = pd.layer.batchnorm(D_h1)
D_h1_relu = pd.layer.lrelu(D_h1_bn)
D_h2 = pd.layer.fc(D_h1_relu, w=self.D_w2, b=self.D_b2)
return D_h2
```
......@@ -123,7 +123,7 @@ class DCGAN(object):
### Class member function: Build the model
- Define data readers as placeholders to hold the data;
- Build generator and discriminators;
- Define two training losses for discriminator and generator, respectively.
- Define two training losses for discriminator and generator, respectively.
If we have execution dependency engine to back-trace all tensors, the module building our GAN model will be like this:
```python
class DCGAN(object):
......@@ -133,7 +133,7 @@ class DCGAN(object):
self.images = pd.data(pd.float32, [self.batch_size, self.im_size, self.im_size])
self.faked_images = pd.data(pd.float32, [self.batch_size, self.im_size, self.im_size])
self.z = pd.data(tf.float32, [None, self.z_size])
# step 1: generate images by generator, classify real/fake images with discriminator
if self.y_dim: # if conditional GAN, includes label
self.G = self.generator(self.z, self.y)
......@@ -147,12 +147,12 @@ class DCGAN(object):
# generate fake images
self.sampled = self.sampler(self.z)
self.D_f = self.discriminator(self.images)
# step 2: define the two losses
self.d_loss_real = pd.reduce_mean(pd.cross_entropy(self.D_t, np.ones(self.batch_size))
self.d_loss_fake = pd.reduce_mean(pd.cross_entropy(self.D_f, np.zeros(self.batch_size))
self.d_loss = self.d_loss_real + self.d_loss_fake
self.g_loss = pd.reduce_mean(pd.cross_entropy(self.D_f, np.ones(self.batch_szie))
```
......@@ -176,7 +176,7 @@ class DCGAN(object):
self.G = self.generator(self.z)
self.D_g = self.discriminator(self.G, self.y)
self.g_loss = pd.reduce_mean(pd.cross_entropy(self.D_g, np.ones(self.batch_szie))
with pd.default_block().d_block():
if self.y_dim: # if conditional GAN, includes label
self.D_t = self.discriminator(self.images, self.y)
......@@ -217,7 +217,7 @@ if __name__ == "__main__":
# load mnist data
data_X, data_y = self.load_mnist()
# Two subgraphs required!!!
with pd.block().d_block():
d_optim = pd.train.Adam(lr = .001, beta= .1)
......@@ -228,7 +228,7 @@ if __name__ == "__main__":
# executor
sess = pd.executor()
# training
for epoch in xrange(10000):
for batch_id in range(N / batch_size):
......@@ -239,7 +239,7 @@ if __name__ == "__main__":
batch_z = np.random.uniform(-1., 1., [batch_size, z_dim])
if batch_id % 2 == 0:
sess.run(d_step,
sess.run(d_step,
feed_dict = {dcgan.images: batch_im,
dcgan.y: batch_label,
dcgan.z: batch_z})
......
......@@ -4,9 +4,9 @@
.. toctree::
:maxdepth: 1
new_op_en.md
new_op_kernel_en.md
use_eigen_en.md
new_op_cn.md
new_op_kernel.md
use_eigen_cn.md
name_convention.md
support_new_device.md
releasing_process.md
......
......@@ -5,7 +5,7 @@ Development
:maxdepth: 1
new_op_en.md
new_op_kernel_en.md
new_op_kernel.md
use_eigen_en.md
name_convention.md
support_new_device.md
......
......@@ -26,13 +26,32 @@
依据是否包含kernel,可以将Op分为两种:包含Kernel的Op和不包含kernel的Op,前者Op的定义继承自`OperatorWithKernel`,后者继承自`OperatorBase`。本教程主要介绍带Kernel的Op如何写,简单总结Op需要包含的内容如下:
内容 | 定义位置
-------------- | :----------------------
OpProtoMake定义 | `.cc`文件,Backward Op不需要定义OpProtoMake
Op定义 | `.cc`文件
Kernel实现 | CPU、CUDA共享Kernel实现在`.h`文件中,否则,CPU 实现在`.cc`文件中,CUDA 实现在`.cu`文件中。
注册Op | Op注册实现在`.cc`文件;Kernel注册CPU实现在`.cc`文件中,CUDA实现在`.cu`文件中
<table>
<thead>
<tr>
<th>内容</th>
<th>定义位置</th>
</tr>
</thead>
<tbody>
<tr>
<td>OpProtoMake定义 </td>
<td>`.cc`文件,Backward Op不需要定义OpProtoMake </td>
</tr>
<tr>
<td>Op定义 </td>
<td> `.cc`文件</td>
</tr>
<tr>
<td>Kernel实现 </td>
<td> CPU、CUDA共享Kernel实现在`.h`文件中,否则,CPU 实现在`.cc`文件中,CUDA 实现在`.cu`文件中。</td>
</tr>
<tr>
<td>注册Op </td>
<td> Op注册实现在`.cc`文件;Kernel注册CPU实现在`.cc`文件中,CUDA实现在`.cu`文件中</td>
</tr>
</tbody>
</table>
实现新的op都添加至目录[paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators)下,文件命名以`*_op.h`(如有) 、 `*_op.cc``*_op.cu`(如有)结尾。**系统会根据文件名自动构建op和其对应的Python扩展。**
......
......@@ -33,6 +33,33 @@ Op definition | `.cc` files
Kernel implementation | The kernel methods shared between CPU and CUDA are defined in `.h` files. CPU-specific kernels live in `.cc` files, while CUDA-specific kernels are implemented in `.cu`files.
Registering the Op | Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation.
<table>
<thead>
<tr>
<th>Information</th>
<th> Where is it defined</th>
</tr>
</thead>
<tbody>
<tr>
<td>OpProtoMake definition </td>
<td> `.cc`files, Backward Op does not need an OpProtoMake interface. </td>
</tr>
<tr>
<td>Op definition </td>
<td> `.cc` files</td>
</tr>
<tr>
<td>Kernel implementation </td>
<td> The kernel methods shared between CPU and CUDA are defined in `.h` files. CPU-specific kernels live in `.cc` files, while CUDA-specific kernels are implemented in `.cu`files.</td>
</tr>
<tr>
<td>Registering the Op </td>
<td> Ops are registered in `.cc` files; For Kernel registration, `.cc` files contain the CPU implementation, while `.cu` files contain the CUDA implementation.</td>
</tr>
</tbody>
</table>
New Operator implementations are added to the list [paddle/operators](https://github.com/PaddlePaddle/Paddle/tree/develop/paddle/operators), with file names in the format `*_op.h` (if applicable), `*_op.cc`, `*_op.cu` (if applicable).** The system will use the naming scheme to automatically build operators and their corresponding Python extensions.**
......@@ -279,7 +306,7 @@ A forward operator unit test inherits `unittest.TestCase` and defines metaclass
def test_check_output(self):
self.check_output()
def test_check_grad_normal(self):
self.check_grad(['X', 'Y'], 'Out', max_relative_error=0.5)
......
......@@ -37,7 +37,7 @@ PaddlePaddle每次发新的版本,遵循以下流程:
可以在此页面的"Artifacts"下拉框中找到生成的3个二进制文件,分别对应CAPI,`cp27m``cp27mu`的版本。然后按照上述的方法
使用`twine`工具上传即可。
<img src="ci_build_whl.png">
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/ci_build_whl.png">
* 注:CI环境使用 https://github.com/PaddlePaddle/buildtools 这里的DockerImage作为编译环境以支持更多的Linux
发型版,如果需要手动编译,也可以使用这些镜像。这些镜像也可以从 https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/ 下载得到。
......@@ -66,7 +66,7 @@ PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-
* 建议,开发者fork的版本库使用`develop`分支同步主版本库的`develop`分支
* 建议,开发者fork的版本库中,再基于`develop`版本fork出自己的功能分支。
* 当功能分支开发完毕后,向PaddlePaddle的主版本库提交`Pull Reuqest`,进而进行代码评审。
* 在评审过程中,开发者修改自己的代码,可以继续在自己的功能分支提交代码。
* 在评审过程中,开发者修改自己的代码,可以继续在自己的功能分支提交代码。
* BugFix分支也是在开发者自己的fork版本库维护,与功能分支不同的是,BugFix分支需要分别给主版本库的`master``develop`与可能有的`release/版本号`分支,同时提起`Pull Request`
......@@ -78,13 +78,116 @@ PaddlePaddle开发过程使用[git-flow](http://nvie.com/posts/a-successful-git-
PaddlePaddle每次发版本首先要保证PaddlePaddle Book中所有章节功能的正确性。功能的正确性包括验证PaddlePaddle目前的`paddle_trainer`训练和纯使用`Python`训练模型正确性。
| | 新手入门章节 | 识别数字 | 图像分类 | 词向量 | 情感分析 | 语意角色标注 | 机器翻译 | 个性化推荐 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| API.V2 + Docker + GPU | | | | | | | | |
| API.V2 + Docker + CPU | | | | | | | | |
| `paddle_trainer` + Docker + GPU | | | | | | | | |
| `paddle_trainer` + Docker + CPU | | | | | | | | |
| API.V2 + Ubuntu + GPU | | | | | | | | |
| API.V2 + Ubuntu + CPU | | | | | | | | |
| `paddle_trainer` + Ubuntu + GPU | | | | | | | | |
| `paddle_trainer` + Ubuntu + CPU | | | | | | | | |
<table>
<thead>
<tr>
<th></th>
<th>新手入门章节 </th>
<th> 识别数字</th>
<th> 图像分类</th>
<th>词向量</th>
<th> 情感分析</th>
<th>语意角色标注</th>
<th> 机器翻译</th>
<th>个性化推荐</th>
</tr>
</thead>
<tbody>
<tr>
<td>API.V2 + Docker + GPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> API.V2 + Docker + CPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>`paddle_trainer` + Docker + GPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>`paddle_trainer` + Docker + CPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> API.V2 + Ubuntu + GPU</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td>API.V2 + Ubuntu + CPU </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> `paddle_trainer` + Ubuntu + GPU</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
<tr>
<td> `paddle_trainer` + Ubuntu + CPU</td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
<td> </td>
</tr>
</tbody>
</table>
......@@ -4,30 +4,70 @@
A model is an output of the training process. One complete model consists of two parts, the **topology** and the **parameters**. In order to support industrial deployment, the model format must be self-complete and must not expose any training source code.
As a result, In PaddlePaddle, the **topology** is represented as a [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/1c0a4c901c9fc881d120249c703b15d1c50dae7d/doc/design/program.md), which describes the model structure. The **parameters** contain all the trainable weights in the model. We must support large size parameters and efficient serialization/deserialization of parameters.
As a result, In PaddlePaddle, the **topology** is represented as a [ProgramDesc](https://github.com/PaddlePaddle/Paddle/blob/1c0a4c901c9fc881d120249c703b15d1c50dae7d/doc/design/program.md), which describes the model structure. The **parameters** contain all the trainable weights in the model. We must support large size parameters and efficient serialization/deserialization of parameters.
## Implementation
The topology is saved as a plain text in a detailed self-contain protobuf file.
The topology is saved as a plain text in a detailed self-contain protobuf file.
The parameters are saved as a binary file. As we all know, the protobuf message has a limit of [64M size](https://developers.google.com/protocol-buffers/docs/reference/cpp/google.protobuf.io.coded_stream#CodedInputStream.SetTotalBytesLimit.details). We have done a [benchmark experiment](https://github.com/PaddlePaddle/Paddle/pull/4610), which shows that protobuf is not fit for the task.
As a result, we design a particular format for tensor serialization. By default, an arbitrary tensor in Paddle is a [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md), and has a description information proto of [LoDTensorDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L99). We save the DescProto as the byte string header. It contains all the necessary information, such as the `dims`, and the `LoD` information in [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/1c0a4c901c9fc881d120249c703b15d1c50dae7d/paddle/framework/lod_tensor.md). A tensor stores values in a continuous memory buffer. For speed we dump the raw memory to disk and save it as the byte string content. So, the binary format of one tensor is,
As a result, we design a particular format for tensor serialization. By default, an arbitrary tensor in Paddle is a [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/lod_tensor.md), and has a description information proto of [LoDTensorDesc](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/framework.proto#L99). We save the DescProto as the byte string header. It contains all the necessary information, such as the `dims`, and the `LoD` information in [LoDTensor](https://github.com/PaddlePaddle/Paddle/blob/1c0a4c901c9fc881d120249c703b15d1c50dae7d/paddle/framework/lod_tensor.md). A tensor stores values in a continuous memory buffer. For speed we dump the raw memory to disk and save it as the byte string content. So, the binary format of one tensor is,
The table below shows a tensor's byte view in detail. Note that all the signed values are written in the little-endian format.
|field name | type | description |
| --- | --- | --- |
| version | uint32_t | Version of saved file. Always 0 now. |
| tensor desc length | uint32_t | TensorDesc(Protobuf message) length in bytes. |
| tensor desc | void* | TensorDesc protobuf binary message |
| tensor data | void* | Tensor's data in binary format. The length of `tensor_data` is decided by `TensorDesc.dims()` and `TensorDesc.data_type()` |
| lod_level | uint64_t | Level of LoD |
| length of lod[0] | uint64_t | [Optional] length of lod[0] in bytes. |
| data of lod[0] | uint64_t* | [Optional] lod[0].data() |
| ... | ... | ... |
<table>
<thead>
<tr>
<th>field name</th>
<th>type </th>
<th>description </th>
</tr>
</thead>
<tbody>
<tr>
<td> version</td>
<td> uint32_t </td>
<td> Version of saved file. Always 0 now.</td>
</tr>
<tr>
<td> tensor desc length </td>
<td> uint32_t </td>
<td> TensorDesc(Protobuf message) length in bytes. </td>
</tr>
<tr>
<td>tensor desc </td>
<td> void*</td>
<td> TensorDesc protobuf binary message </td>
</tr>
<tr>
<td> tensor data </td>
<td> void* </td>
<td> Tensor's data in binary format. The length of `tensor_data` is decided by `TensorDesc.dims()` and `TensorDesc.data_type()` </td>
</tr>
<tr>
<td> lod_level</td>
<td> uint64_t </td>
<td> Level of LoD </td>
</tr>
<tr>
<td> length of lod[0] </td>
<td> uint64_t </td>
<td> [Optional] length of lod[0] in bytes. </td>
</tr>
<tr>
<td> data of lod[0] </td>
<td> uint64_t* </td>
<td> [Optional] lod[0].data() </td>
</tr>
<tr>
<td>... </td>
<td> ... </td>
<td> ... </td>
</tr>
</tbody>
</table>
## Summary
......
......@@ -65,10 +65,10 @@ exit(1)
**因此,在分布式的Fluid环境中,我们有两个角色需要创建,分别是Parameter Server和Trainer。**
### 分布式训练
### 分布式训练
Fliud专门提供了工具[Distributed Transpiler](https://github.com/PaddlePaddle/Paddle/blob/ba65d54d9d3b41cd3c5171b00f476d4e60133ddb/doc/fluid/design/dist_train/distributed_architecture.md#distributed-transpiler)用于将单机版的训练程序转换为分布式版本的训练程序。工具背后的理念是找出程序的优化算子和梯度参数,将他们分隔为两部分,通过send/recv 操作算子进行连接,优化算子和梯度参数可以在优化器的minimize函数的返回值中获取到。
```python
optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
```
将Distributed Transpiler、优化算子和梯度函数放在一个代码中如下:
```python
......@@ -99,15 +99,51 @@ for pass_id in range(100):
### 分布式训练脚本运行说明
分布式任务的运行需要将表格中说明的多个参数进行赋值:
| 参数名 | 值类型 | 说明 | 示例 |
|:-------------|:------|:---------------------------------------|:-------------|
| trainer_id | int | 当前训练节点的ID,训练节点ID编号为0 - n-1, n为trainers的值 | 0/1/2/3 |
| pservers | str | parameter server 列表 | 127.0.0.1:6710,127.0.0.1:6711 |
| trainers | int | 训练节点的总个数,>0的数字 | 4 |
| server_endpoint | str | 当前所起的服务节点的IP:PORT | 127.0.0.1:8789 |
| training_role | str | 节点角色, TRAINER/PSERVER | PSERVER |
**注意:** ```training_role```是用来区分当前所起服务的角色的,用于训练程序中,用户可根据需要自行定义,其他参数为fluid.DistributeTranspiler的transpile函数所需要,需要在调用函数前进行定义,样例如下:
<table>
<thead>
<tr>
<th>参数名</th>
<th> 值类型</th>
<th>说明</th>
<th> 示例</th>
</tr>
</thead>
<tbody>
<tr>
<td>trainer_id </td>
<td> int</td>
<td> 当前训练节点的ID,训练节点ID编号为0 - n-1, n为trainers的值 </td>
<td> 0/1/2/3 </td>
</tr>
<tr>
<td>pservers </td>
<td> str</td>
<td> parameter server 列表 </td>
<td> 127.0.0.1:6710,127.0.0.1:6711 </td>
</tr>
<tr>
<td>trainers </td>
<td>int </td>
<td> 训练节点的总个数,>0的数字 </td>
<td> 4 </td>
</tr>
<tr>
<td> server_endpoint</td>
<td> str </td>
<td> 当前所起的服务节点的IP:PORT </td>
<td> 127.0.0.1:8789 </td>
</tr>
<tr>
<td> training_role</td>
<td>str </td>
<td> 节点角色, TRAINER/PSERVER </td>
<td> PSERVER </td>
</tr>
</tbody>
</table>
**注意:** ```training_role```是用来区分当前所起服务的角色的,用于训练程序中,用户可根据需要自行定义,其他参数为fluid.DistributeTranspiler的transpile函数所需要,需要在调用函数前进行定义,样例如下:
```python
t = fluid.DistributeTranspiler()
......
......@@ -42,14 +42,40 @@ cprofilev -a 0.0.0.0 -p 3214 -f profile.out main.py
每一列的含义是:
| 列名 | 含义 |
| --- | --- |
| ncalls | 函数的调用次数 |
| tottime | 函数实际使用的总时间。该时间去除掉本函数调用其他函数的时间 |
| percall | tottime的每次调用平均时间 |
| cumtime | 函数总时间。包含这个函数调用其他函数的时间 |
| percall | cumtime的每次调用平均时间 |
| filename:lineno(function) | 文件名, 行号,函数名 |
<table>
<thead>
<tr>
<th>列名</th>
<th>含义 </th>
</tr>
</thead>
<tbody>
<tr>
<td> ncalls</td>
<td> 函数的调用次数</td>
</tr>
<tr>
<td>tottime</td>
<td> 函数实际使用的总时间。该时间去除掉本函数调用其他函数的时间</td>
</tr>
<tr>
<td> percall </td>
<td> tottime的每次调用平均时间</td>
</tr>
<tr>
<td> cumtime</td>
<td> 函数总时间。包含这个函数调用其他函数的时间</td>
</tr>
<tr>
<td> percall</td>
<td> cumtime的每次调用平均时间</td>
</tr>
<tr>
<td> filename:lineno(function) </td>
<td> 文件名, 行号,函数名 </td>
</tr>
</tbody>
</table>
### 寻找性能瓶颈
......
......@@ -57,14 +57,40 @@ port, we will see the output like the following:
where each line corresponds to Python function, and the meaning of
each column is as follows:
| column | meaning |
| --- | --- |
| ncalls | the number of calls into a function |
| tottime | the total execution time of the function, not including the execution time of other functions called by the function |
| percall | tottime divided by ncalls |
| cumtime | the total execution time of the function, including the execution time of other functions being called |
| percall | cumtime divided by ncalls |
| filename:lineno(function) | where the function is defined |
<table>
<thead>
<tr>
<th>column</th>
<th>meaning </th>
</tr>
</thead>
<tbody>
<tr>
<td> ncalls</td>
<td> the number of calls into a function</td>
</tr>
<tr>
<td>tottime</td>
<td> the total execution time of the function, not including the execution time of other functions called by the function</td>
</tr>
<tr>
<td> percall </td>
<td> tottime divided by ncalls</td>
</tr>
<tr>
<td> cumtime</td>
<td> the total execution time of the function, including the execution time of other functions being called</td>
</tr>
<tr>
<td> percall</td>
<td> cumtime divided by ncalls</td>
</tr>
<tr>
<td> filename:lineno(function) </td>
<td> where the function is define </td>
</tr>
</tbody>
</table>
### Identify Performance Bottlenecks
......
......@@ -23,7 +23,7 @@ But how to record the time for the mixed C++ and CUDA program? There many C++ A
The overall flow is shown as the following figure.
<img src="./images/profiler.png" align="center"/><br/>
<img src="https://raw.githubusercontent.com/PaddlePaddle/Paddle/develop/doc/fluid/images/profiler.png" align="center"/><br/>
### Event
......@@ -36,10 +36,10 @@ enum EventKind {
kPopRange};
```
- kMark: only a marker without time range.
- kPushRange: mark the starting event for time range.
- kPushRange: mark the starting event for time range.
- kPopRange: mark the ending event for time range.
For the CPU code, the events only need to record the current time. For the CUDA code, the [event management functions of CUDA](http://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__EVENT.html#group__CUDART__EVENT) are used. For many pieces of code, an event lists are used to record each piece.
For the CPU code, the events only need to record the current time. For the CUDA code, the [event management functions of CUDA](http://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__EVENT.html#group__CUDART__EVENT) are used. For many pieces of code, an event lists are used to record each piece.
```c++
class Event {
......@@ -66,11 +66,11 @@ struct EventList {
};
```
As mentioned above, there is no need to record the timeline when disabling the profiler. So there is a global state to enable or disable the profiler.
As mentioned above, there is no need to record the timeline when disabling the profiler. So there is a global state to enable or disable the profiler.
```c++
enum ProfilerState {
kDisabled,
kDisabled,
kCPU,
kCUDA
};
......
digraph G {
rnn [label="1st level RNN" shape=box]
subgraph cluster0 {
label = "time step 0"
sent0 [label="sentence"]
sent1 [label="sentence"]
rnn1 [label="2nd level RNN" shape=box]
sent0 -> rnn1
sent1 -> rnn1
}
subgraph cluster1 {
label = "time step 1"
sent2 [label="sentence"]
sent3 [label="sentence"]
rnn2 [label="2nd level RNN" shape=box]
sent2 -> rnn2
sent3 -> rnn2
}
subgraph cluster2 {
label = "time step 2"
sent4 [label="sentence"]
sent5 [label="sentence"]
rnn3 [label="2nd level RNN" shape=box]
sent4 -> rnn3
sent5 -> rnn3
}
para0 [label="paragraph info 0"]
para1 [label="paragraph info 1"]
para2 [label="paragraph info 2"]
rnn1 -> para0
rnn2 -> para1
rnn3 -> para2
para0 -> rnn
para1 -> rnn
para2 -> rnn
chapter [label="chapter info"]
rnn -> chapter
}
digraph ImageBatchNormForkGragh {
subgraph cluster_before {
Prev [label="...", shape=plaintext];
Rnn [label="rnn_op", shape=box];
BatchNorm [label="batch_norm_op", shape=box];
Fc [label="fc_op", shape=box];
After [label="...", shape=plaintext];
Prev -> Rnn -> BatchNorm -> Fc -> After;
label="original";
}
subgraph cluster_after {
Prev2 [label="...", shape=plaintext];
Rnn2 [label="rnn_op", shape=box];
BatchNorm2_1 [label="train_batch_norm_op", shape=box];
BatchNorm2_2 [label="infer_batch_norm_op", shape=box];
Fc2_1 [label="fc_op", shape=box];
Fc2_2 [label="fc_op", shape=box];
After2_1 [label="...", shape=plaintext];
After2_2 [label="...", shape=plaintext];
Prev2 -> Rnn2 -> BatchNorm2_1 -> Fc2_1 -> After2_1;
Rnn2 -> BatchNorm2_2 ->Fc2_2 ->After2_2
label="forked";
}
}
cat ./graph_construction_example.dot | \
sed 's/color=red/color=red, style=invis/g' | \
sed 's/color=green/color=green, style=invis/g' | \
dot -Tpng > graph_construction_example_forward_only.png
cat ./graph_construction_example.dot | \
sed 's/color=green/color=green, style=invis/g' | \
dot -Tpng > graph_construction_example_forward_backward.png
cat ./graph_construction_example.dot | \
dot -Tpng > graph_construction_example_all.png
digraph ImageClassificationGraph {
///////// The forward part /////////
FeedX [label="Feed", color=blue, shape=box];
FeedY [label="Feed", color=blue, shape=box];
InitW [label="Init", color=blue, shape=diamond];
Initb [label="Init", color=blue, shape=diamond];
FC [label="FC", color=blue, shape=box];
MSE [label="MSE", color=blue, shape=box];
x [label="x", color=blue, shape=oval];
l [label="l", color=blue, shape=oval];
y [label="y", color=blue, shape=oval];
W [label="W", color=blue, shape=doublecircle];
b [label="b", color=blue, shape=doublecircle];
cost [label="cost", color=blue, shape=oval];
FeedX -> x -> FC -> y -> MSE -> cost [color=blue];
FeedY -> l [color=blue];
InitW -> W [color=blue];
Initb -> b [color=blue];
W -> FC [color=blue];
b -> FC [color=blue];
l -> MSE [color=blue];
////////// The backward part /////////
MSE_Grad [label="MSE_grad", color=red, shape=box];
FC_Grad [label="FC_grad", color=red, shape=box];
d_cost [label="d cost", color=red, shape=oval];
d_y [label="d y", color=red, shape=oval];
d_b [label="d b", color=red, shape=oval];
d_W [label="d W", color=red, shape=oval];
cost -> MSE_Grad [color=red];
d_cost -> MSE_Grad [color=red];
l -> MSE_Grad [color=red];
y -> MSE_Grad -> d_y [color=red];
x -> FC_Grad [color=red];
y -> FC_Grad [color=red];
d_y -> FC_Grad [color=red];
W -> FC_Grad -> d_W [color=red];
b -> FC_Grad -> d_b [color=red];
////////// The optimizaiton part //////////
OPT_W [label="SGD", color=green, shape=box];
OPT_b [label="SGD", color=green, shape=box];
W -> OPT_W [color=green];
b -> OPT_b [color=green];
d_W -> OPT_W -> W [color=green];
d_b -> OPT_b -> b [color=green];
////////// Groupings //////////
subgraph clusterMSE {
style=invis;
MSE;
MSE_Grad;
}
subgraph clusterFC {
style=invis;
FC;
FC_Grad;
}
}
digraph G {
label = "simple RNN implementation"
ranksep=2;
//graph [nodesep=1, ranksep=1];
node[nodesep=1]
subgraph cluster0 {
label = "global scope"
rankdir = TB
W
boot_memory
input
output
}
subgraph cluster1 {
label = "step-scope 0"
rankdir = TB
memory0[label="memory"]
prememory0[label="pre-memory"]
step_input0[label="step input"]
step_output0[label="step output"]
}
subgraph cluster2 {
label = "step-scope 1"
rankdir = TB
memory1[label="memory"]
prememory1[label="pre-memory"]
step_input1[label="step input"]
step_output1[label="step output"]
}
subgraph cluster3 {
label = "step-scope 2"
rankdir = TB
memory2[label="memory"]
prememory2[label="pre-memory"]
step_input2[label="step input"]
step_output2[label="step output"]
}
stepnet [shape=box]
stepnet0 [shape=box, style=dashed]
stepnet1 [shape=box, style=dashed]
stepnet2 [shape=box, style=dashed]
edge[color=blue]
boot_memory -> prememory0 [label="init" color="blue"]
memory0 -> prememory1 [label="copy/reference" color="blue"]
memory1 -> prememory2 [label="copy/reference" color="blue"]
edge[color=black]
W -> stepnet0[constraint=false, style=dashed]
W -> stepnet1[constraint=false, style=dashed]
W -> stepnet2[constraint=false, style=dashed]
memory0 -> stepnet0[style=dashed]
prememory0 -> stepnet0 -> step_output0[style=dashed]
memory1 -> stepnet1[style=dashed]
prememory1 -> stepnet1 -> step_output1[style=dashed]
memory2 -> stepnet2[style=dashed]
prememory2 -> stepnet2 -> step_output2[style=dashed]
input -> step_input0
input -> step_input1
input -> step_input2
step_input0 -> stepnet0 [style=dashed]
step_input1 -> stepnet1[style=dashed]
step_input2 -> stepnet2[style=dashed]
step_output0 -> output
step_output1 -> output
step_output2 -> output
stepnet0 -> stepnet[style=dashed]
stepnet1 -> stepnet[style=dashed]
stepnet2 -> stepnet[style=dashed]
}
digraph G {
chapter [label="chapter"]
subgraph cluster0 {
label = "paragraph 0"
top_rnn0[label="top rnn step 0" shape=box]
p0 [label="paragraph 0"]
p1 [label="paragraph 1"]
}
subgraph cluster1{
label = "paragraph 1"
top_rnn1[label="top rnn step 1" shape=box]
p2 [label="paragraph 0"]
p3 [label="paragraph 1"]
}
subgraph cluster_p0 {
label = "sentence 0"
low_rnn0 [label="low rnn step 0" shape=box]
s00 [label="sentence 0"]
s01 [label="sentence 1"]
low_rnn0 -> s00
low_rnn0 -> s01
}
subgraph cluster_p1 {
label = "sentence 1"
low_rnn1 [label="low rnn step 1" shape=box]
s10 [label="sentence 0"]
s11 [label="sentence 1"]
low_rnn1 -> s10
low_rnn1 -> s11
}
subgraph cluster_p2 {
label = "sentence 1"
low_rnn2 [label="low rnn step 0" shape=box]
s20 [label="sentence 0"]
s21 [label="sentence 1"]
low_rnn2 -> s20
low_rnn2 -> s21
}
subgraph cluster_p3 {
label = "sentence 1"
low_rnn3 [label="low rnn step 1" shape=box]
s30 [label="sentence 0"]
s31 [label="sentence 1"]
low_rnn3 -> s30
low_rnn3 -> s31
}
chapter -> top_rnn0
chapter -> top_rnn1
top_rnn0 -> p0
top_rnn0 -> p1
top_rnn1 -> p2
top_rnn1 -> p3
p0 -> low_rnn0
p1 -> low_rnn1
p2 -> low_rnn2
p3 -> low_rnn3
}
digraph Test {
z -> generator -> G_img;
G_img -> discriminator -> D_f -> d_loss_f;
label0 -> d_loss_f -> d_loss;
img -> discriminator -> D_t -> d_loss_t;
label1 -> d_loss_t -> d_loss;
d_loss -> d_loss_t[color=red, style=dashed];
d_loss -> d_loss_f[color=red, style=dashed];
d_loss_t -> D_t[color=red, style=dashed];
d_loss_f -> D_f[color=red, style=dashed];
D_t -> discriminator[color=red, style=dashed];
D_f -> discriminator[color=red, style=dashed];
D_f -> g_loss;
label2 -> g_loss;
g_loss -> D_f[color=green, style=dashed];
D_f -> discriminator[color=green, style=dashed];
discriminator -> G_img[color=green, style=dashed];
G_img -> generator[color=green, style=dashed];
discriminator [color=red, shape=box];
generator [color=green, shape=box];
z [shape=diamond];
img [shape=diamond];
label0 [shape=diamond];
label1 [shape=diamond];
label2 [shape=diamond];
d_loss [color=red];
g_loss [color=green];
}
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
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