提交 880b874b 编写于 作者: Y Yi Wang

Merge branch 'develop' of https://github.com/paddlepaddle/paddle into paddle_only_cpu

# Design Doc: Session
## Abstract
The *session* object encapsulates the environment in which the
computation graph is executed.
We will have the *local* session and *remote* session, they offer the
same [interface](#interface). The local session encapsulates the local
runtime environment and the remote session encapsulates the cluster
runtime environment.
The local runtime environment contains:
1. computation devices (i.e., CPU, GPU) handles, and
1. the [scope](../scope.md) which holds all variables.
The remote runtime environment contains:
1. computation devices (i.e., CPU and GPU on node 0, 1) in a cluster,
and
1. the distributed [scope](../scope.md) in a cluster which holds all
variables.
The user can create a remote session on Paddle Cloud and evaluate the
computation graph with it. In this way, the user can control the
remote computation resource in a cluster from his local computer.
## Background
The current design has an implicit global session in which
`paddle.eval()` is executed. The pain point is:
Since the user is not able to explicitly switch between runtime
environments, the user cannot run a topology in two independent
environments.
For example, in reinforcement learning, the user may want to have a
stale model for inference and a fresh model for training, and only
replace the stale model with the fresh model periodically.
Furthermore, we have no concept that encapsulates a remote environment
that executes a computation graph.
We need the session object to address above issues.
## Session
A session is an object that owns the runtime environment. All
computations are executed through `session.eval()`.
### Interface
```python
eval(
targets,
feed_dict=None,
)
```
Evaluates the target Operations or Variables in `targets`.
- *targets*: the evaluation targets. Can be a single Operation or
Variable, or a list with the Operations or Variables as
elements. The value returned by `eval()` has the same shape as the
`target` argument.
The PaddlePaddle program is represented by
the [ProgramDesc](../design/program.md), `eval()` will infer the
ProgramDesc from the given targets and run the PaddlePaddle
program. Please
see
[this graph](./distributed_architecture.md#local-training-architecture) for
the detailed illustration for the local session
and
[this graph](./distributed_architecture.md#distributed-training-architecture) for
the detailed illustration for the remote session.
- *feed_dict*: a dictionary that contains the tensors which override
the edges of the computation graph.
feed_dict not only can provide the input data, it can override any
OP's input as well:
```python
a = pd.constant(2.0, name="a")
b = pd.variable(name="b")
c = pd.mul(a,b)
sess.eval(targets=c, feed_dict={"b":3.0}) # returns 6.0
```
```python
close()
```
Closes the session and releases the scope that the session owns.
### Create a Local Session
```python
session(
devices=None
)
```
Creates a new session. One session owns one global scope, so creating
multiple sessions will create different scopes.
- *devices*: a single `string` or a list of `string` of device names,
the corresponding devices will be the computation devices for
`eval()`. If not specified, all available devices (e.g., all GPUs)
will be used. The user doesn't need to specify the CPU device since
it will be always used. Multiple sessions can use the same device.
#### Example
```Python
a = paddle.constant(1.0)
b = paddle.constant(2.0)
c = a + b
sess = paddle.session(devices=["gpu:0", "gpu:1", "fpga:0"])
sess.eval(c)
sess.close()
```
### Create a Remote Session
```python
create_cloud_job(
name,
num_trainer,
mem_per_trainer,
gpu_per_trainer,
cpu_per_trainer,
num_ps,
mem_per_ps,
cpu_per_ps,
)
```
Creates a Paddle Cloud job. Fails if the job name exists.
```python
get_cloud_job(
name
)
```
Gets a Paddle Cloud job.
```python
remote_session(
job
)
```
- *job*: the Paddle Cloud job.
#### Example
```Python
reader = paddle.reader.recordio("/pfs/home/peter/mnist-train-*") # data stored on Paddle Cloud
image = reader.column(0)
label = reader.column(1)
fc1 = paddle.op.fc(image, size=256, act="sigmoid")
fc2 = paddle.op.fc(fc1, size=10, act="softmax")
cost = paddle.op.cross_entropy(fc2, label)
opt = paddle.optimizer.sgd(cost)
job = paddle.create_cloud_job("test", 3, "1G", 1, 1, 2, "1G", 1)
sess = paddle.remote_ession(job)
for i in range(1000):
sess.eval(opt)
sess.close()
```
# Design Doc: LoD (Level-of-Detail) Tensor # Design Doc: LoD (Level-of-Detail) Tensor
PaddlePaddle's RNN doesn't require that all instances have the same length. To do so, we introduce an extension to Tensor, namely, LoD Tensor. 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.
## Challenge of Variable-length Inputs | | TensorFlow | PaddlePaddle |
|-----------------------|------------|--------------|
| RNN | Support | Support |
| recursive RNN | Support | Support |
| padding zeros | Must | No need |
| blob data type | Tensor | LoDTensor |
People usually represent a mini-batch by a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. So a transformation, T, of all images can be a matrix multiplication of the 10xOx32-dimensional tensor T and the 10x32x32 Tensor. 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.
Another example is that each mini-batch contains 32 sentences, where each word is a D-dimensional one-hot vector. If all sentences have the same length L, we can represent this mini-batch by a 32xLxD tensor. However, in most cases, sentences have variable lengths, and we will need an index data structure to record these variable lengths.
## LoD as a Solution ## The Challenge: Variable-length Sequences
### Mini-Batch of variable-length sentences Most deep learning systems represent a mini-batch as a Tensor. For example, a mini-batch of 10 images, each of size 32x32, is a 10x32x32 Tensor. Another example is that each mini-batch contains N sentences, where each word is a D-dimensional one-hot vector. Suppose that all sentences have the same length L, we can represent this mini-batch by a NxLxD tensor.
Let's imagine a mini-batch of 3 variable lengths sentences, containing 3, 1, and 2 words respectively. We can represent it by a (3+1+2)xD tensor plus some index information: Both examples show that the elements of sequences are usually of the same size. In the first example, all images are 32x32, and in the second one, all words are D-dimensional vectors. It doesn't make sense to allow variable-sized images, as that would require transformations like convolution to handle variable-sized Tensors.
The real challenge is that in most cases, sentences have variable lengths, and we will need an index data structure to segment the tensor into sequences. Also, sequences might consist of sub-sequences.
## A Solution: The LoD Index
To understand our solution, it is best to look at some examples.
### A Mini-Batch of Sentences
Let's imagine a mini-batch of 3 variable lengths sentences composed of 3, 1, and 2 words, respectively. We can represent the mini-batch by a (3+1+2)xD tensor plus some index information:
``` ```
3
3 1 2 3 1 2
||| | || ||| | ||
``` ```
Each `|` represents a D-dimensional word vectors. The number 3 on top indicate 3 sentences, and numbers 3, 1, and 2 on the second level represent the number of words in each sentence. where each `|` represents a D-dimensional word vector. The numbers, 3, 1, and 2, form a 1-level LoD.
### Recursive Sequences
Let check another example of a 2-level LoD Tensor. Consider a mini-batch of three articles with 3, 1, and 2 sentences, and each sentence consists of a variable number of words:
```
3 1 2
3 2 4 1 2 3
||| || |||| | || |||
```
### Mini-Batch of variable-length videos ### A Mini-Batch of Videos
This approach generalizes to the case where elements are not words, but higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. If a mini-batch contains 3 videos of 3, 1, and 2 frames respectively. The underlying tensor is of size (3+1+2)x640x480. The index information illustrates as: LoD tensors generalize to the case where elements are higher dimensional objects, like images. Suppose that a mini-batch contains videos of the same frame size 640x480. Here is a mini-batch of 3 videos with 3, 1, and 2 frames, respectively.
``` ```
3
3 1 2 3 1 2
口口口 口 口口 口口口 口 口口
``` ```
where each `口` represents an image. The underlying tensor is of size (3+1+2)x640x480, and each `口` represents a 640x480 image.
### Mini-Batch of fixed-size images ### A Mini-Batch of Images
Let's get back to a typical example, image classification, where each mini-batch has M fixed-sized images. The LoD Tensor representation is In traditional cases like a mini-batch with N fixed-sized images, the LoD Tensor representation is as
``` ```
M
1 1 1 1 1 1 1 1 1 1
口口口口 ... 口 口口口口 ... 口
``` ```
The many 1's on the second level seem duplicated. For this particular case of 2 levels and the second level always have length 1, we can ignore the LoD index. In this case, we don't lose any information by ignoring the many 1's in the index and simply considering this LoD Tensor as a usual Tensor:
### Design and summarization
In summary, as long as that the essential elements (words or images) have the same size, we can represent mini-batches by a LoD Tensor: ```
口口口口 ... 口
```
- The underlying tensor has size LxD1xD2x..., where D1xD2... is the size of the essential elements, and ### Model Parameters
- The first dimension size L has an additonal property -- a LoD index as a nested vector:
```c++ A model parameter is just a usual Tensor, which, just like the above example, is a **0-level LoD Tensor**.
typedef std::vector<std::<vector>> LoD;
```
- The LoD index is not necessary when there are only two levels and all elements of the second level have length 1.
## Slicing of LoD Tensor ## The LoD Tensor
Consider that we have a network with three levels of RNN: the top level one handles articles, the second level one handles sentences, and the basic level one handles words. This network requires that mini-batches represented by 3 level LoD Tensor, for example, Let us revisit above example of the 2-level LoD Tensor
``` ```
3
3 1 2 3 1 2
3 2 4 1 2 3 3 2 4 1 2 3
||| || |||| | || ||| ||| || |||| | || |||
``` ```
To allow each level of RNN to handle its input, we define **the slicing of a LoD Tensor is defined as getting the j-th sequence on level i, or the <i,j>-slice** It is indeed a tree, where leaves are elementary sequences identified by **branches**.
For example, the third sentence in above example is identified by branch <0,2>, where 0 indicates the first article with length 3, and 2 indicates the third sentence in this article with length 4.
### The LoD Index
For example, the <2,1>-slice of above slice is We can save the LoD index in the above example
``` ```
2 3 1 2
|| 3 2 4 1 2 3
``` ```
and the <1,2>-slice of above example is in a not-full 2D matrix:
```c++
typedef std::vector<std::vector<int> > LoD;
``` ```
2
2 3
|| |||
```
Let's go on slicing this slice. Its <1,1>-slice is where
- `LoD.size()` is the number of levels, or the maximum length of branches,
- `LoD[i][j]` is the length of the j-th segment at the i-th level.
## The Offset Representation
To quickly access elementary sequences, we adopt an offset representation -- instead of saving the lengths, we save the beginning and ending elements of sequences.
In the above example, we accumulate the length of elementary sequences:
``` ```
1 3 2 4 1 2 3
1
|
``` ```
### The Slicing Algorithm into offsets
The algorithm, with over-simplified data structure, is defined as ```
0 3 5 9 10 12 15
= = = = = =
3 2+3 4+5 1+9 2+10 3+12
```
```c++ so we know that the first sentence is from word 0 to word 3, and the second sentence from work 3 to word 5.
typedef std::vector<std::vector<int>> LoD;
struct LoDTensor { Similarly, the lengths in the top level LoD
LoD lod_;
float* tensor_;
};
LoDTensor Slice(const LoDTensor& lodt, int level, int sequence); ```
3 1 2
``` ```
Let us revisit the example above are transformed into offsets of elements/words as follows:
``` ```
3 0 9 10 15
3 1 2 = = =
3 2 4 1 2 3 3+2+4 1+9 2+3+10
||| || |||| | || |||
``` ```
Suppose that we want to retrieve the <1,2>-slice so we can tell that the first article is from word 0 to word 9, and the second article is from word 9 to word 10.
The complete offset representation is as follows:
``` ```
2 0 9 10 15
2 3 0 3 5 9 10 12 15
|| ||| ||| || |||| | || |||
``` ```
we will need to find out the starting position of this slice by summing over all leaf nodes in `LoD` to the left of the slice, i.e., 3 + 2 + 4 + 1 = 10. ## Slicing of LoD Tensors
When we use the above 2-level LoD Tensor as the input to a nested-RNN, we need to retrieve certain sequences. Here we define the sequence identified by branch <i,j,...> as the **<i,j,...>-slice**.
To avoid the traversal of the LoD tree at slicing time, we can do it at the construction time -- instead of saving the lengths of the next level in the LoD tree, we can save the starting offset of the next level. For example, above LoD Tensor can be transformed into For example, the <2>-slice of above example is
``` ```
0 10 15
0 9 10 10 12 15
0 3 5 9 10 12 || |||
||| || |||| | || |||
``` ```
We don't really need the 0 on top, so the LoD Tensor could be and the <2,0>-slice of above slice is
``` ```
0 9 10 10 12
0 3 5 9 10 12 ||
||| || |||| | || |||
``` ```
...@@ -77,7 +77,7 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) { ...@@ -77,7 +77,7 @@ BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
// GPU buddy allocator initialization // GPU buddy allocator initialization
std::call_once(gpu_allocator_flag, [&]() { std::call_once(gpu_allocator_flag, [&]() {
int gpu_num = platform::GetDeviceCount(); int gpu_num = platform::GetCUDADeviceCount();
allocators.reserve(gpu_num); allocators.reserve(gpu_num);
for (int gpu = 0; gpu < gpu_num; gpu++) { for (int gpu = 0; gpu < gpu_num; gpu++) {
platform::SetDeviceId(gpu); platform::SetDeviceId(gpu);
......
...@@ -20,7 +20,7 @@ TEST(Device, Init) { ...@@ -20,7 +20,7 @@ TEST(Device, Init) {
using paddle::platform::CUDADeviceContext; using paddle::platform::CUDADeviceContext;
using paddle::platform::GPUPlace; using paddle::platform::GPUPlace;
int count = paddle::platform::GetDeviceCount(); int count = paddle::platform::GetCUDADeviceCount();
for (int i = 0; i < count; i++) { for (int i = 0; i < count; i++) {
DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); DeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
Eigen::GpuDevice* gpu_device = Eigen::GpuDevice* gpu_device =
...@@ -34,7 +34,7 @@ TEST(Device, CUDADeviceContext) { ...@@ -34,7 +34,7 @@ TEST(Device, CUDADeviceContext) {
using paddle::platform::CUDADeviceContext; using paddle::platform::CUDADeviceContext;
using paddle::platform::GPUPlace; using paddle::platform::GPUPlace;
int count = paddle::platform::GetDeviceCount(); int count = paddle::platform::GetCUDADeviceCount();
for (int i = 0; i < count; i++) { for (int i = 0; i < count; i++) {
CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i)); CUDADeviceContext* device_context = new CUDADeviceContext(GPUPlace(i));
Eigen::GpuDevice* gpu_device = device_context->eigen_device(); Eigen::GpuDevice* gpu_device = device_context->eigen_device();
......
...@@ -26,11 +26,11 @@ DEFINE_double(fraction_of_gpu_memory_to_use, 0.95, ...@@ -26,11 +26,11 @@ DEFINE_double(fraction_of_gpu_memory_to_use, 0.95,
namespace paddle { namespace paddle {
namespace platform { namespace platform {
int GetDeviceCount() { int GetCUDADeviceCount() {
int count; int count;
PADDLE_ENFORCE( PADDLE_ENFORCE(
cudaGetDeviceCount(&count), cudaGetDeviceCount(&count),
"cudaGetDeviceCount failed in paddle::platform::GetDeviceCount"); "cudaGetDeviceCount failed in paddle::platform::GetCUDADeviceCount");
return count; return count;
} }
......
...@@ -28,7 +28,7 @@ const std::string kEnvFractionGpuMemoryToUse = ...@@ -28,7 +28,7 @@ const std::string kEnvFractionGpuMemoryToUse =
"PADDLE_FRACTION_GPU_MEMORY_TO_USE"; "PADDLE_FRACTION_GPU_MEMORY_TO_USE";
//! Get the total number of GPU devices in system. //! Get the total number of GPU devices in system.
int GetDeviceCount(); int GetCUDADeviceCount();
//! Get the current GPU device id in system. //! Get the current GPU device id in system.
int GetCurrentDeviceId(); int GetCurrentDeviceId();
......
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