@@ -96,7 +96,7 @@ There are two modes in term of memory management in `PaddleBuf` :
In the two modes, the first is more convenient while the second strictly controls memory management to facilitate integration with `tcmalloc` and other libraries.
### Upgrade performance based on contrib::AnalysisConfig (Prerelease)
### Upgrade performance based on contrib::AnalysisConfig
AnalyisConfig is at the stage of pre-release and protected by `namespace contrib` , which may be adjusted in the future.
...
...
@@ -106,9 +106,11 @@ The usage of `AnalysisConfig` is similiar with that of `NativeConfig` but the fo
```c++
AnalysisConfigconfig;
config.model_dir=xxx;
config.use_gpu=false;// GPU optimization is not supported at present
config.specify_input_name=true;// it needs to set name of input
config.SetModel(dirname);// set the directory of the model
config.EnableUseGpu(100,0/*gpu id*/);// use GPU,or
config.DisableGpu();// use CPU
config.SwitchSpecifyInputNames(true);// need to appoint the name of your input
config.SwitchIrOptim();// turn on the optimization switch,and a sequence of optimizations will be executed in operation
```
Note that input PaddleTensor needs to be allocated. Previous examples need to be revised as follows:
NVIDIA TensorRT is a is a platform for high-performance deep learning inference. It delivers low latency and high throughput for deep learning inference application.
Subgraph is used in Paddle 1.0 to preliminarily integrate TensorRT, which enables TensorRT module to enhance inference performance of paddle models. The module is still under development. Currently supported models are AlexNet, MobileNet, ResNet50, VGG19, ResNext, Se-ReNext, GoogleNet, DPN, ICNET, MobileNet-SSD and so on. We will introduce the obtaining, usage and theory of Paddle-TensorRT library in this documentation.
## Build inference libraries with `TensorRT`
Subgraph is used in PaddlePaddle to preliminarily integrate TensorRT, which enables TensorRT module to enhance inference performance of paddle models. The module is still under development. Currently supported models are AlexNet, MobileNet, ResNet50, VGG19, ResNext, Se-ReNext, GoogleNet, DPN, ICNET, Deeplabv3 Mobile, Net-SSD and so on. We will introduce the obtaining, usage and theory of Paddle-TensorRT library in this documentation.
[`paddle_inference_api.h`]('https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/api/paddle_inference_api.h') defines all APIs of TensorRT.
...
...
@@ -58,17 +64,17 @@ A complete process is shown below:
The parameters of the neural network are redundant to some extent. In many tasks, we can turn the Float32 model into Int8 model on the premise of precision. At present, Paddle-TRT supports to turn the trained Float32 model into Int8 model off line. The specific processes are as follows: 1）**Create the calibration table**. We prepare about 500 real input data, and input the data to the model. Paddle-TRT will count the range information of each op input and output value in the model, and record in the calibration table. The information can reduce the information loss during model transformation. 2）After creating the calibration table, run the model again, **Paddle-TRT will load the calibration table automatically**, and conduct the inference in the INT8 mode.
2. compile and test the INT8 example
```shell
cd SAMPLE_BASE_DIR/sample
# sh run_impl.sh {the address of inference libraries} {the name of test script} {model directories}
# We generate 500 input data to simulate the process, and it's suggested that you use real example for experiment.
sh run_impl.sh BASE_DIR/fluid_inference_install_dir/ fluid_generate_calib_test SAMPLE_BASE_DIR/sample/mobilenetv1
```
After the running period, there will be a new file named trt_calib_* under the `SAMPLE_BASE_DIR/sample/build/mobilenetv1` model directory, which is the calibration table.
``` shell
# conduct INT8 inference
# copy the model file with calibration tables to a specific address
Subgraph is used to integrate TensorRT in PaddlePaddle. After model is loaded, neural network can be represented as a computing graph composed of variables and computing nodes. Functions Paddle TensorRT implements are to scan the whole picture, discover subgraphs that can be optimized with TensorRT and replace them with TensorRT nodes. During the inference of model, Paddle will call TensorRT library to optimize TensorRT nodes and call native library of Paddle to optimize other nodes. During the inference, TensorRT can integrate Op horizonally and vertically to filter redundant Ops and is able to choose appropriate kernel for specific Op in specific platform to speed up the inference of model.
Subgraph is used to integrate TensorRT in PaddlePaddle. After model is loaded, neural network can be represented as a computing graph composed of variables and computing nodes. Functions Paddle TensorRT implements are to scan the whole picture, discover subgraphs that can be optimized with TensorRT and replace them with TensorRT nodes. During the inference of model, Paddle will call TensorRT library to optimize TensorRT nodes and call native library of Paddle to optimize other nodes. During the inference, TensorRT can integrate Op horizonally and vertically to filter redundant Ops and is able to choose appropriate kernel for specific Op in specific platform to speed up the inference of model.
A simple model expresses the process :
...
...
@@ -121,6 +150,6 @@ A simple model expresses the process :
We can see in the Original Network that the green nodes represent nodes supported by TensorRT, the red nodes represent variables in network and yellow nodes represent nodes which can only be operated by native functions in Paddle. Green nodes in original network are extracted to compose subgraph which is replaced by a single TensorRT node to be transformed into `block-25` node in network. When such nodes are encountered during the runtime, TensorRT library will be called to execute them.
We can see in the Original Network that the green nodes represent nodes supported by TensorRT, the red nodes represent variables in network and yellow nodes represent nodes which can only be operated by native functions in Paddle. Green nodes in original network are extracted to compose subgraph which is replaced by a single TensorRT node to be transformed into `block-25` node in network. When such nodes are encountered during the runtime, TensorRT library will be called to execute them.
为了使`OpKernel`的计算过程书写更加简单，并且CPU、CUDA的代码可以复用，我们通常借助 Eigen unsupported Tensor模块来实现`Compute`接口。关于在PaddlePaddle中如何使用Eigen库，请参考[使用文档](https://github.com/PaddlePaddle/FluidDoc/blob/release/1.2/doc/fluid/dev/use_eigen_cn.md)。
@@ -29,4 +29,8 @@ The following content describes the APIs related to the learning rate scheduler:
* :code:`piecewise_decay`: Piecewise decay. That is, the stair-like decay for a given number of steps, the learning rate stays the same within each step. For related API Reference please refer to :ref:`api_fluid_layers_piecewise_decay`
* :code:`append_LARS`: The learning rate is obtained by the Layer-wise Adaptive Rate Scaling algorithm. For related algorithms, please refer to `Train Feed forward Neural Network with Layerwise Adaptive Rate via Approximating Back-matching Propagation <https://arxiv. Org/abs/1802.09750>`_ . For related API Reference please refer to :ref:`api_fluid_layers_append_LARS`
\ No newline at end of file
* :code:`append_LARS`: The learning rate is obtained by the Layer-wise Adaptive Rate Scaling algorithm. For related algorithms, please refer to `Train Feed forward Neural Network with Layerwise Adaptive Rate via Approximating Back-matching Propagation <https://arxiv. Org/abs/1802.09750>`_ . For related API Reference please refer to :ref:`api_fluid_layers_append_LARS`
* :code:`cosine_decay`: Cosine attenuation. It means the learning rate changes with the number of steps in the form of a cosine function. For related API Reference please refer to :ref:`api_fluid_layers_cosine_decay`
* :code:`linear_lr_warmup`: The learning rate increases linearly to an appointed rate with the number of steps. For related API Reference please refer to :ref:`api_fluid_layers_linear_lr_warmup`
@@ -9,9 +9,8 @@ This instruction will show you how to install PaddlePaddle on Windows. The foll
**Note** :
* The current version does not support NCCL, distributed training, AVX, warpctc and MKL related functions.
* The current version does not support NCCL, distributed training related functions.
* Currently, only PaddlePaddle for CPU is supported on Windows.
...
...
@@ -30,14 +29,20 @@ Version of pip or pip3 should be equal to or above 9.0.1 .
* Install PaddlePaddle
****CPU version of PaddlePaddle***:
Execute `pip install paddlepaddle` or `pip3 install paddlepaddle` to download and install PaddlePaddle.
****GPU version of PaddlePaddle***:
Execute `pip install paddlepaddle-gpu`(python2.7) or `pip3 install paddlepaddle-gpu`(python3.x) to download and install PaddlePaddle.
## ***Verify installation***
After completing the installation, you can use `python` or `python3` to enter the python interpreter and then use `import paddle.fluid` to verify that the installation was successful.
## ***How to uninstall***
****CPU version of PaddlePaddle***:
Use the following command to uninstall PaddlePaddle : `pip uninstallpaddlepaddle `or `pip3 uninstall paddlepaddle`
****GPU version of PaddlePaddle***:
Use the following command to uninstall PaddlePaddle : `pip uninstall paddlepaddle-gpu` or `pip3 uninstall paddlepaddle-gpu`
Paddle Fluid supports PyReader, which implements feeding data from Python to C++. Different from :ref:`user_guide_use_numpy_array_as_train_data_en` , the process of loading data to Python is asynchronous with the process of :code:`Executor::Run()` reading data when PyReader is in use.
Moreover, PyReader is able to work with :code:`double_buffer_reader` to upgrade the performance of reading data.
Besides Python Reader, we provide PyReader. The performance of PyReader is better than :ref:`user_guide_use_numpy_array_as_train_data` , because the process of loading data is asynchronous with the process of training model when PyReader is in use. And PyReader can coordinate with :code:`double_buffer_reader` to improve the performance of reading data. What's more, :code:`double_buffer_reader` can achieve the transformation from CPU Tensor to GPU Tensor, which improve the efficiency of reading data to some extent.
Create PyReader Object
################################
...
...
@@ -17,7 +16,7 @@ You can create PyReader object as follows:
import paddle.fluid as fluid
py_reader = fluid.layers.py_reader(capacity=64,
shapes=[(-1,3,224,224), (-1,1)],
shapes=[(-1,784), (-1,1)],
dtypes=['float32', 'int64'],
name='py_reader',
use_double_buffer=True)
...
...
@@ -28,14 +27,14 @@ In the code, ``capacity`` is buffer size of PyReader;
``name`` is name of PyReader instance;
``use_double_buffer`` is True by default, which means :code:`double_buffer_reader` is used.
To create some different PyReader objects (Usually, you have to create two different PyReader objects for training and testing phase), the names of objects must be different. For example, In the same task, PyReader objects in training and testing period are created as follows:
Attention: If you want to create multiple PyReader objects（such as two different PyReader in training and inference period respectively）， you have to appoint different names for different PyReader objects,since PaddlePaddle uses different names to distinguish different variables, and `Program.clone()` (reference to :ref:`api_fluid_Program_clone` ）can't copy PyReader objects.
@@ -46,11 +45,10 @@ To create some different PyReader objects (Usually, you have to create two diffe
name='test',
use_double_buffer=True)
Note: You could not copy PyReader object with :code:`Program.clone()` so you have to create PyReader objects in training and testing phase with the method mentioned above
While using PyReader, if you need to share the model parameters of training and test periods, you can use :code:`fluid.unique_name.guard()` .
Notes: Paddle use different names to distinguish different variables, and the names are generated by the counter in :code:`unique_name` module. By the way, the counts rise by one every time a variable name is generated. :code:`fluid.unique_name.guard()` aims to reset the counter in :code:`unique_name` module, in order to ensure that the variable names are the same when calling :code:`fluid.unique_name.guard()` repeatedly, so that parameters can be shared.
Because you could not copy PyReader with :code:`Program.clone()` so you have to share the parameters of training phase with testing phase through :code:`fluid.unique_name.guard()` .
Details are as follows:
An example of configuring networks during the training and test periods by PyReader is as follows:
.. code-block:: python
...
...
@@ -61,41 +59,97 @@ Details are as follows:
import numpy
def network(is_train):
# Create py_reader object and give different names
# when is_train = True and is_train = False
reader = fluid.layers.py_reader(
capacity=10,
shapes=((-1, 784), (-1, 1)),
dtypes=('float32', 'int64'),
name="train_reader" if is_train else "test_reader",
use_double_buffer=True)
# Use read_file() method to read out the data from py_reader
img, label = fluid.layers.read_file(reader)
...
# Here, we omitted the definition of loss of the model
return loss , reader
# Create main program and startup program for training
train_prog = fluid.Program()
train_startup = fluid.Program()
with fluid.program_guard(train_prog, train_startup):
# Use fluid.unique_name.guard() to share parameters with test network
with fluid.unique_name.guard():
train_loss, train_reader = network(True)
adam = fluid.optimizer.Adam(learning_rate=0.01)
adam.minimize(train_loss)
# Create main program and startup program for testing
test_prog = fluid.Program()
test_startup = fluid.Program()
with fluid.program_guard(test_prog, test_startup):
# Use fluid.unique_name.guard() to share parameters with train network
with fluid.unique_name.guard():
test_loss, test_reader = network(False)
Configure data source of PyReader objects
##########################################
PyReader provides :code:`decorate_tensor_provider` and :code:`decorate_paddle_reader` , both of which receieve Python :code:`generator` as data source.The difference is:
PyReader object sets the data source by :code:`decorate_paddle_reader()` or :code:`decorate_tensor_provider()` :code:`decorate_paddle_reader()` and :code:`decorate_tensor_provider()` both receive the Python generator :code:`generator` as parameters. :code:`generator` generates a batch of data every time by yield ways inside.
The differences of :code:`decorate_paddle_reader()` and :code:`decorate_tensor_provider()` ways are:
- :code:`generator` of :code:`decorate_paddle_reader()` should return data of Numpy Array type, but :code:`generator` of :code:`decorate_tensor_provider()` should return LoDTensor type.
- :code:`decorate_tensor_provider()` requires that the returned data type and size of LoDTensor of :code:`generator` have to match the appointed dtypes and shapes parameters while configuring py_reader, but :code:`decorate_paddle_reader()` doesn't have the requirements, since the data type and size can transform inside.
Specific ways are as follows:
.. code-block:: python
import paddle.fluid as fluid
import numpy as np
1. :code:`decorate_tensor_provider` : :code:`generator` generates a :code:`list` or :code:`tuple` each time, with each element of :code:`list` or :code:`tuple` being :code:`LoDTensor` or Numpy array, and :code:`LoDTensor` or :code:`shape` of Numpy array must be the same as :code:`shapes` stated while PyReader is created.
BATCH_SIZE = 32
# Case 1: Use decorate_paddle_reader() method to set the data source of py_reader
2. :code:`decorate_paddle_reader` : :code:`generator` generates a :code:`list` or :code:`tuple` each time, with each element of :code:`list` or :code:`tuple` being Numpy array,but the :code:`shape` of Numpy array doesn't have to be the same as :code:`shape` stated while PyReader is created. :code:`decorate_paddle_reader` will :code:`reshape` Numpy array internally.
@@ -35,7 +35,9 @@ In the training of data parallelism mode, Fluid uses two communication modes to
The pserver process can be on a compute node that is completely different from the trainer, or it can share the same node with a trainer. The number of pserver processes required for a distributed task usually needs to be adjusted according to the actual situation to achieve the best performance. However, usually pserver processes are no more than trainer processes.
When using GPU training, the pserver can choose to use the GPU or only use the CPU. If the pserver also uses the GPU, it will result in the extra overhead of copying the gradient data received from the CPU to the GPU. In some cases, the overall training performance will be degraded.
**Note:** When using GPU training, the pserver can choose to use the GPU or only use the CPU. If the pserver also uses the GPU, it will result in the extra overhead of copying the gradient data received from the CPU to the GPU. In some cases, the overall training performance will be degraded.
**Note:** When using GPU training, if there are multiple GPU cards in each trainer node, the gradient polymerization will execute in NCCL2 way among the cards in one node, and then in multiple nodes through pserver.
- Structure of NCCL2 communication method:
...
...
@@ -178,8 +180,10 @@ Distributed training in NCCL2 mode, because there is no parameter server role, t
* Configure :code:`mode="nccl2"` in :code:`fluid.DistributeTranspilerConfig` .
* When calling :code:`transpile`, :code:`trainers` is fed with the endpoints of all trainer nodes, and passed with the argument :code:`current_endpoint` .
In this step, :code:`gen_nccl_id_op` will add in :code:`startup program` to synchronize NCCLID information during the multi-computer program initialization.
* Initialize :code:`ParallelExecutor` with :code:`num_trainers` and :code:`trainer_id` .
In this step, :code:`ParallelExecutor` will initialize NCCL2 by the multi-computer way and do the operations :code:`allreduce` across the nodes for the gradient of every parameter to execute muti-computer training
For example:
.. code-block:: python
...
...
@@ -198,24 +202,45 @@ For example:
.. csv-table:: Description of the necessary parameters for NCCL2 mode
:header: "parameter", "description"
"trainer_id", "The unique ID of each trainer node in the task, starting at 0, there cannot be any duplication"
"trainers", "endpoints of all trainer nodes in the task, used to broadcast NCCL IDs when NCCL2 is initialized"
"current_endpoint", "endpoint of current node"
"trainer_id", "(int)The unique ID of each trainer node in the task, starting at 0, there cannot be any duplication"
"trainers", "(int)endpoints of all trainer nodes in the task, used to broadcast NCCL IDs when NCCL2 is initialized"
"current_endpoint", "(string)endpoint of current node"
Currently, distributed training using NCCL2 only supports synchronous training. The distributed training using NCCL2 mode is more suitable for the model which is relatively large and needs \
synchronous training and GPU training. If the hardware device supports RDMA and GPU Direct, this can achieve high distributed training performance.
Start Up NCCL2 Distributed Training in Muti-Process Mode
++++++++++++++++++++++++++++++++++++++++++++++
Usually you can get better multi-training performance by using multi-process mode to start up NCCL2 distributed training assignment. Paddle provides :code:`paddle.distributed.launch` module to start up multi-process assignment, after which each training process will use an independent GPU device.
Attention during usage:
* set the number of nodes: set the number of nodes of an assignment by the environment variable :code:`PADDLE_NUM_TRAINERS` , and this variable will also be set in every training process.
* set the number of devices of each node: by activating the parameter :code:`--gpus` , you can set the number of GPU devices of each node, and the sequence number of each process will be set in the environment variable :code:`PADDLE_TRAINER_ID` automatically.
* data segment: mult-process mode means one process in each device. Generally, each process manages a part of training data, in order to make sure that all processes can manage the whole data set.
* entrance file: entrance file is the training script for actual startup.
* journal: for each training process, the joural is saved in the default :code:`./mylog` directory, and you can assign by the parameter :code:`--log_dir` .
**Note** : Please ensure each node has the same amount of data to train in NCCL2 mode distributed training, which prevents
**Note:** When using distributed training in NCCL2 mode, if you only want to use a part of cards in one node, you can appoint by configuring the environment variable :code:`export CUDA_VISIBLE_DEVICES=0,1,2,3` .
**Note:** Please ensure each node has the same amount of data to train in NCCL2 mode distributed training, which prevents
exit at the final iteration. There are two common ways:
- Randomly sample some data to complement nodes where less data are distributed. (We recommend this method for sake of a complete dataset to be trained)
- Each node only trains fixed number of batches per pass, which is controlled by python codes. If a node has more data than this fixed amount, then these
marginal data will not be trained.
**Note** : If there are multiple network devices in the system, you need to manually specify the devices used by NCCL2.
**Note** : If there are multiple network devices in the system, you need to manually specify the devices used by NCCL2.
Assuming you need to use :code:`eth2` as the communication device, you need to set the following environment variables:
@@ -97,7 +97,7 @@ Machine Translation transforms a natural language (source language) into another
The Transformer implemented in this example is a machine translation model based on the self-attention mechanism, in which there is no more RNN or CNN structure, but fully utilizes Attention to learn the context dependency. Compared with RNN/CNN, in a single layer, this structure has lower computational complexity, easier parallelization, and easier modeling for long-range dependencies, and finally achieves the best translation effect among multiple languages.
@@ -131,7 +131,7 @@ In many scenarios of natural language processing, it is necessary to measure the
The DAM (Deep Attention Matching Network) introduced in this example is the work of Baidu Natural Language Processing Department published in ACL-2018, which is used for the selection of responses in multi-round dialogue of retrieval chat robots. Inspired by Transformer, DAM is based entirely on the attention mechanism. It uses the stack-type self-attention structure to learn the semantic representations of responses and contexts at different granularities, and then uses cross-attention to obtain relativity between responses and contexts. The performance on the two large-scale multi-round dialogue datasets is better than other models.
@@ -151,7 +151,7 @@ Machine Reading Comprehension (MRC) is one of the core tasks in Natural Language
Baidu reading comprehension dataset is an open-source real-world dataset publicized by Baidu Natural Language Processing Department. All the questions and original texts are derived from actual data (Baidu search engine data and Baidu know Q&A community), and the answer is given by humans. Each question corresponds to multiple answers. The dataset contains 200k questions, 1000k original text and 420k answers. It is currently the largest Chinese MRC dataset. Baidu also publicized the corresponding open-source reading comprehension model, called DuReader. DuReader adopts the current common network hierarchical structure, and captures the interaction between the problems and the original texts through the double attention mechanism to generate the original representation of the query-aware. Finally, based on the original text of query-aware, the answer scope is predicted by point network.
- `DuReader in PaddlePaddle Fluid <https://github.com/PaddlePaddle/models/blob/develop/PaddleNLP/machine_reading_comprehension/README.md>`__
- `DuReader in PaddlePaddle Fluid <https://github.com/PaddlePaddle/models/blob/develop/PaddleNLP/reading_comprehension>`__
