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

avx_docs
hedaoyuan 9 years ago
commit 283bee2883

@ -25,9 +25,9 @@ addons:
packages:
- gcc-4.8
- g++-4.8
- gfortran-4.8
- git
- build-essential
- libatlas-base-dev
- python
- python-pip
- python2.7-dev

@ -16,7 +16,7 @@
set(CBLAS_FOUND OFF)
## Find MKL First.
set(MKL_ROOT $ENV{MKL_ROOT} CACHE PATH "Folder contains MKL")
set(MKL_ROOT $ENV{MKLROOT} CACHE PATH "Folder contains MKL")
find_path(MKL_INCLUDE_DIR mkl.h PATHS
${MKL_ROOT}/include)

@ -15,7 +15,6 @@
INCLUDE(cblas)
IF(NOT ${CBLAS_FOUND})
MESSAGE(FATAL_ERROR "Please install OpenBlas, MKL or ATLAS.")
INCLUDE(ExternalProject)
SET(CBLAS_SOURCES_DIR ${THIRD_PARTY_PATH}/openblas)
@ -28,20 +27,40 @@ IF(NOT ${CBLAS_FOUND})
SET(CBLAS_LIBRARIES "${CBLAS_INSTALL_DIR}/lib/libopenblas.a" CACHE FILEPATH "openblas library" FORCE)
ENDIF(WIN32)
IF(CMAKE_COMPILER_IS_GNUCC)
ENABLE_LANGUAGE(Fortran)
LIST(APPEND CBLAS_LIBRARIES gfortran pthread)
ENDIF(CMAKE_COMPILER_IS_GNUCC)
IF(NOT CMAKE_Fortran_COMPILER)
MESSAGE(FATAL_ERROR "To build lapack in libopenblas, "
"you need to set gfortran compiler: cmake .. -DCMAKE_Fortran_COMPILER=...")
ENDIF(NOT CMAKE_Fortran_COMPILER)
ExternalProject_Add(
openblas
${EXTERNAL_PROJECT_LOG_ARGS}
URL "https://github.com/xianyi/OpenBLAS/archive/v0.2.19.tar.gz"
GIT_REPOSITORY https://github.com/xianyi/OpenBLAS.git
GIT_TAG v0.2.19
PREFIX ${CBLAS_SOURCES_DIR}
INSTALL_DIR ${CBLAS_INSTALL_DIR}
BUILD_IN_SOURCE 1
CONFIGURE_COMMAND ""
BUILD_COMMAND make CC=${CMAKE_C_COMPILER} FC=${CMAKE_Fortran_COMPILER}
INSTALL_COMMAND make install PREFIX=<INSTALL_DIR>
BUILD_COMMAND ${CMAKE_MAKE_PROGRAM} FC=${CMAKE_Fortran_COMPILER} CC=${CMAKE_C_COMPILER} HOSTCC=${CMAKE_C_COMPILER} NO_SHARED=1 libs netlib
INSTALL_COMMAND ${CMAKE_MAKE_PROGRAM} install NO_SHARED=1 PREFIX=<INSTALL_DIR>
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
)
ExternalProject_Add_Step(
openblas lapacke_install
COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke_mangling_with_flags.h" "${CBLAS_INSTALL_DIR}/include/lapacke_mangling.h"
COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke.h" "${CBLAS_INSTALL_DIR}/include/lapacke.h"
COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke_config.h" "${CBLAS_INSTALL_DIR}/include/lapacke_config.h"
COMMAND ${CMAKE_COMMAND} -E copy "${CBLAS_SOURCES_DIR}/src/openblas/lapack-netlib/LAPACKE/include/lapacke_utils.h" "${CBLAS_INSTALL_DIR}/include/lapacke_utils.h"
DEPENDEES install
)
LIST(APPEND external_project_dependencies openblas)
ENDIF()
ENDIF(NOT ${CBLAS_FOUND})
INCLUDE_DIRECTORIES(${CBLAS_INC_DIR})

@ -54,6 +54,7 @@ ExternalProject_Add(
CONFIGURE_COMMAND
${CMAKE_COMMAND} ${PROTOBUF_SOURCES_DIR}/src/protobuf/cmake
-Dprotobuf_BUILD_TESTS=OFF
-DZLIB_ROOT:FILEPATH=${ZLIB_ROOT}
-DCMAKE_POSITION_INDEPENDENT_CODE=ON
-DCMAKE_BUILD_TYPE=Release
-DCMAKE_INSTALL_PREFIX=${PROTOBUF_INSTALL_DIR}

@ -31,6 +31,7 @@ IF(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
"please use pip to upgrade protobuf.")
ENDIF(${PY_GOOGLE.PROTOBUF_VERSION} VERSION_LESS "3.0.0")
ELSE(PYTHONLIBS_FOUND AND PYTHONINTERP_FOUND)
MESSAGE(FATAL_ERROR "Please install python 2.7 before building PaddlePaddle.")
##################################### PYTHON ########################################
SET(PYTHON_SOURCES_DIR ${THIRD_PARTY_PATH}/python)
SET(PYTHON_INSTALL_DIR ${THIRD_PARTY_PATH}/install/python)

@ -96,6 +96,7 @@ set(COMMON_FLAGS
-Wno-unused-parameter
-Wno-unused-function
-Wno-error=literal-suffix
-Wno-error=sign-compare
-Wno-error=unused-local-typedefs)
set(GPU_COMMON_FLAGS
@ -105,6 +106,7 @@ set(GPU_COMMON_FLAGS
-Wdelete-non-virtual-dtor
-Wno-unused-parameter
-Wno-unused-function
-Wno-error=sign-compare
-Wno-error=literal-suffix
-Wno-error=unused-local-typedefs
-Wno-error=unused-function # Warnings in Numpy Header.

@ -47,7 +47,7 @@ SET(EXTERNAL_PROJECT_LOG_ARGS
LOG_DOWNLOAD 0 # Wrap download in script to log output
LOG_UPDATE 1 # Wrap update in script to log output
LOG_CONFIGURE 1 # Wrap configure in script to log output
LOG_BUILD 1 # Wrap build in script to log output
LOG_BUILD 0 # Wrap build in script to log output
LOG_TEST 1 # Wrap test in script to log output
LOG_INSTALL 1 # Wrap install in script to log output
LOG_INSTALL 0 # Wrap install in script to log output
)

@ -64,7 +64,8 @@ As a simple example, consider the following:
1. **BLAS Dependencies(optional)**
Paddle will find BLAS from system's default path. But you can specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
CMake will search BLAS libraries from system. If not found, OpenBLAS will be downloaded, built and installed automatically.
To utilize preinstalled BLAS you can simply specify MKL, OpenBLAS or ATLAS via `MKL_ROOT`, `OPENBLAS_ROOT` or `ATLAS_ROOT`.
```bash
# specify MKL
@ -99,7 +100,7 @@ As a simple example, consider the following:
```bash
# necessary
sudo apt-get update
sudo apt-get install -y g++ make cmake build-essential libatlas-base-dev python python-pip libpython-dev git
sudo apt-get install -y g++ make cmake build-essential python python-pip libpython-dev git
sudo pip install wheel numpy
sudo pip install 'protobuf>=3.0.0'
```

@ -2,15 +2,9 @@
## Create AWS Account and IAM Account
To use AWS, we need to sign up an AWS account on Amazon's Web site.
An AWS account allows us to login to the AWS Console Web interface to
create IAM users and user groups. Usually, we create a user group with
privileges required to run PaddlePaddle, and we create users for
those who are going to run PaddlePaddle and add these users into the
group. IAM users can identify themselves using password and tokens,
where passwords allows users to log in to the AWS Console, and tokens
make it easy for users to submit and inspect jobs from the command
line.
AWS account allow us to manage AWS from Web Console. Amazon AMI enable us to manage AWS from command line interface.
We need to create an AMI user with sufficient privilege to create kubernetes cluster on AWS.
To sign up an AWS account, please
follow
@ -19,8 +13,7 @@ To create users and user groups under an AWS account, please
follow
[this guide](http://docs.aws.amazon.com/IAM/latest/UserGuide/id_users_create.html).
Please be aware that this tutorial needs the following privileges in
the user group:
Please be aware that this tutorial needs the following privileges for the user in AMI:
- AmazonEC2FullAccess
- AmazonS3FullAccess
@ -31,6 +24,7 @@ the user group:
- IAMUserSSHKeys
- IAMFullAccess
- NetworkAdministrator
- AWSKeyManagementServicePowerUser
By the time we write this tutorial, we noticed that Chinese AWS users
@ -46,9 +40,11 @@ it.
Here we will show you step by step on how to run PaddlePaddle training on AWS cluster.
###Download kube-aws and kubectl
### Download kube-aws and kubectl
#### kube-aws
####kube-aws
[kube-aws](https://github.com/coreos/kube-aws) is a CLI tool to automate cluster deployment to AWS.
Import the CoreOS Application Signing Public Key:
@ -88,24 +84,22 @@ mv ${PLATFORM}/kube-aws /usr/local/bin
```
####kubectl
#### kubectl
[kubectl](https://kubernetes.io/docs/user-guide/kubectl-overview/) is a command line interface for running commands against Kubernetes clusters.
Go to the [releases](https://github.com/kubernetes/kubernetes/releases) and download the latest release tarball.
Extract the tarball and then concate the kubernetes binaries directory into PATH:
```
export PATH=<path/to/kubernetes-directory>/platforms/linux/amd64:$PATH
export PATH=<path/to/kubernetes-directory>/platforms/linux/amd64:$PATH # The exact path depend on your platform
```
User credentials and security tokens will be generated later in user directory, not in `~/.kube/config`, they will be necessary to use the CLI or the HTTP Basic Auth.
###Configure AWS Credentials
First check out [this](http://docs.aws.amazon.com/cli/latest/userguide/installing.html) for installing the AWS command line interface, if you use ec2 instance with default amazon AMI, the cli tool has already been installed on your machine.
### Configure AWS Credentials
First check out [this](http://docs.aws.amazon.com/cli/latest/userguide/installing.html) for installing the AWS command line interface.
And then configure your AWS account information:
@ -126,33 +120,35 @@ Default output format: json
```
Test that your credentials work by describing any instances you may already have running on your account:
Verify that your credentials work by describing any instances you may already have running on your account:
```
aws ec2 describe-instances
```
###Define Cluster Parameters
### Define Cluster Parameters
####EC2 key pair
#### EC2 key pair
The keypair that will authenticate SSH access to your EC2 instances. The public half of this key pair will be configured on each CoreOS node.
After creating a key pair, you will use the name you gave the keys to configure the cluster. Key pairs are only available to EC2 instances in the same region. More info in the [EC2 Keypair docs](http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-key-pairs.html).
Follow [EC2 Keypair docs](http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/ec2-key-pairs.html) to create a EC2 key pair
####KMS key
After creating a key pair, you will use the name you gave the keys to configure the cluster. Key pairs are only available to EC2 instances in the same region.
#### KMS key
Amazon KMS keys are used to encrypt and decrypt cluster TLS assets. If you already have a KMS Key that you would like to use, you can skip creating a new key and provide the Arn string for your existing key.
You can create a KMS key in the AWS console, or with the aws command line tool:
```
$ aws kms --region=us-west-2 create-key --description="kube-aws assets"
$ aws kms --region=us-west-1 create-key --description="kube-aws assets"
{
"KeyMetadata": {
"CreationDate": 1458235139.724,
"KeyState": "Enabled",
"Arn": "arn:aws:kms:us-west-2:xxxxxxxxx:key/xxxxxxxxxxxxxxxxxxx",
"Arn": "arn:aws:kms:us-west-1:xxxxxxxxx:key/xxxxxxxxxxxxxxxxxxx",
"AWSAccountId": "xxxxxxxxxxxxx",
"Enabled": true,
"KeyUsage": "ENCRYPT_DECRYPT",
@ -166,7 +162,9 @@ You will use the `KeyMetadata.Arn` string to identify your KMS key in the init s
And then you need to add several inline policies in your user permission.
kms inline policy:
Go to AMI user page, click on `Add inline policy` button, and then select `Custom Policy`
paste into following inline policies:
```
{
@ -182,16 +180,8 @@ kms inline policy:
"Resource": [
"arn:aws:kms:*:xxxxxxxxx:key/*"
]
}
]
}
```
cloudformation inline policy:
```
"Version": "2012-10-17",
"Statement": [
{
},
{
"Sid": "Stmt1482205746000",
"Effect": "Allow",
"Action": [
@ -200,10 +190,11 @@ cloudformation inline policy:
"cloudformation:DeleteStack",
"cloudformation:DescribeStacks",
"cloudformation:DescribeStackResource",
"cloudformation:GetTemplate"
"cloudformation:GetTemplate",
"cloudformation:DescribeStackEvents"
],
"Resource": [
"arn:aws:cloudformation:us-west-2:xxxxxxxxx:stack/YOUR_CLUSTER_NAME/*"
"arn:aws:cloudformation:us-west-1:xxxxxxxxx:stack/YOUR_CLUSTER_NAME/*"
]
}
]
@ -211,15 +202,23 @@ cloudformation inline policy:
```
####External DNS name
#### External DNS name
When the cluster is created, the controller will expose the TLS-secured API on a public IP address. You will need to create an A record for the external DNS hostname you want to point to this IP address. You can find the API external IP address after the cluster is created by invoking kube-aws status.
####S3 bucket
#### S3 bucket
You need to create an S3 bucket before startup the Kubernetes cluster.
####Initialize an asset directory
command (need to have a global unique name):
```
paddle aws s3api --region=us-west-1 create-bucket --bucket bucket-name
```
If you get an error message, try a different bucket name. The bucket name needs to be globally unique.
#### Initialize an asset directory
Create a directory on your local machine to hold the generated assets:
@ -237,12 +236,16 @@ $ kube-aws init \
--region=us-west-1 \
--availability-zone=us-west-1c \
--key-name=key-pair-name \
--kms-key-arn="arn:aws:kms:us-west-2:xxxxxxxxxx:key/xxxxxxxxxxxxxxxxxxx"
--kms-key-arn="arn:aws:kms:us-west-1:xxxxxxxxxx:key/xxxxxxxxxxxxxxxxxxx"
```
Here `us-west-1c` is used for parameter `--availability-zone`, but supported availability zone varies among AWS accounts.
Please check if `us-west-1c` is supported by `aws ec2 --region us-west-1 describe-availability-zones`, if not switch to other supported availability zone. (e.g., `us-west-1a`, or `us-west-1b`)
There will now be a cluster.yaml file in the asset directory. This is the main configuration file for your cluster.
####Render contents of the asset directory
#### Render contents of the asset directory
In the simplest case, you can have kube-aws generate both your TLS identities and certificate authority for you.
@ -285,21 +288,21 @@ $ tree
These assets (templates and credentials) are used to create, update and interact with your Kubernetes cluster.
###Kubernetes Cluster Start Up
### Kubernetes Cluster Start Up
####Create the instances defined in the CloudFormation template
#### Create the instances defined in the CloudFormation template
Now for the exciting part, creating your cluster:
Now for the exciting part, creating your cluster (choose any `<prefix>`):
```
$ kube-aws up --s3-uri s3://<your-bucket-name>/<prefix>
```
####Configure DNS
#### Configure DNS
You can invoke `kube-aws status` to get the cluster API endpoint after cluster creation, if necessary. This command can take a while. And then dig the load balancer hostname to get the ip address, use this ip to setup an A record for your external dns name.
You can invoke `kube-aws status` to get the cluster API endpoint after cluster creation, if necessary. This command can take a while. And use command `dig` to check the load balancer hostname to get the ip address, use this ip to setup an A record for your external dns name.
####Access the cluster
#### Access the cluster
Once the API server is running, you should see:
@ -312,7 +315,7 @@ ip-10-0-0-xx.us-west-1.compute.internal Ready,SchedulingDisabled 5m
```
###Setup PaddlePaddle Environment on AWS
### Setup PaddlePaddle Environment on AWS
Now, we've created a cluster with following network capability:

@ -20,23 +20,27 @@ limitations under the License. */
namespace paddle {
const SequenceArg& BufferArg::sequence() const {
// CHECK_EQ(bufferType_, TENSOR_SEQUENCE_DATA);
CHECK_EQ(bufferType_, TENSOR_SEQUENCE_DATA);
return dynamic_cast<const SequenceArg&>(*this);
}
const SparseMatrixArg& BufferArg::sparse() const {
// CHECK_EQ(bufferType_, TENSOR_SPARSE);
CHECK_EQ(bufferType_, TENSOR_SPARSE);
return dynamic_cast<const SparseMatrixArg&>(*this);
}
SparseMatrixArg::SparseMatrixArg(const CpuSparseMatrix& sparse, ArgType argType)
: BufferArg(sparse, argType),
row_(reinterpret_cast<void*>(sparse.getRows()), VALUE_TYPE_INT32),
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {}
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {
bufferType_ = TENSOR_SPARSE;
}
SparseMatrixArg::SparseMatrixArg(const GpuSparseMatrix& sparse, ArgType argType)
: BufferArg(sparse, argType),
row_(reinterpret_cast<void*>(sparse.getRows()), VALUE_TYPE_INT32),
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {}
col_(reinterpret_cast<void*>(sparse.getCols()), VALUE_TYPE_INT32) {
bufferType_ = TENSOR_SPARSE;
}
} // namespace paddle

@ -23,10 +23,11 @@ limitations under the License. */
namespace paddle {
enum BufferType {
TENSOR_NORMAL = 0,
TENSOR_SEQUENCE_ID = 1,
TENSOR_SEQUENCE_DATA = 2,
TENSOR_SPARSE = 3
TENSOR_UNKNOWN = 0,
TENSOR_NORMAL = 1,
TENSOR_SEQUENCE_ID = 2,
TENSOR_SEQUENCE_DATA = 3,
TENSOR_SPARSE = 4
};
enum SparseDataType {
@ -98,6 +99,7 @@ public:
valueType_(DataType<real>::value),
shape_(2),
argType_(argType) {
bufferType_ = TENSOR_NORMAL;
shape_.setDim(0, matrix.getHeight());
shape_.setDim(1, matrix.getWidth());
}
@ -110,6 +112,7 @@ public:
valueType_(DataType<real>::value),
shape_(shape),
argType_(argType) {
bufferType_ = TENSOR_NORMAL;
CHECK_EQ(matrix.getElementCnt(), shape.getElements());
}
@ -119,6 +122,7 @@ public:
valueType_(DataType<real>::value),
shape_(1),
argType_(argType) {
bufferType_ = TENSOR_NORMAL;
shape_.setDim(0, vector.getSize());
}
@ -128,6 +132,7 @@ public:
valueType_(VALUE_TYPE_INT32),
shape_(1),
argType_(argType) {
bufferType_ = TENSOR_NORMAL;
shape_.setDim(0, vector.getSize());
}
@ -162,6 +167,8 @@ public:
ValueType valueType() const { return valueType_; }
BufferType bufferType() const { return bufferType_; }
const TensorShape& shape() const { return shape_; }
bool isSparse() const { return (TENSOR_SPARSE == bufferType_); }
bool isSequenceArg() const { return TENSOR_SEQUENCE_DATA == bufferType_; }
const SequenceArg& sequence() const;
const SparseMatrixArg& sparse() const;
@ -170,8 +177,8 @@ protected:
void* buf_;
ValueType valueType_;
TensorShape shape_;
BufferType bufferType_;
ArgType argType_ = UNSPECIFIED;
BufferType bufferType_{TENSOR_UNKNOWN};
ArgType argType_{UNSPECIFIED};
// leading dimensions. The size is dims_.size()
// Dims lds_;
};
@ -192,11 +199,13 @@ public:
const TensorShape& shape,
ArgType argType = UNSPECIFIED)
: BufferArg(buf, VALUE_TYPE_INT32, shape, argType) {
bufferType_ = TENSOR_SEQUENCE_ID;
CHECK_EQ(shape_.ndims(), (size_t)1);
numSeqs_ = shape_[0] - 1;
}
SequenceIdArg(const IVector& vector) : BufferArg(vector) {
bufferType_ = TENSOR_SEQUENCE_ID;
numSeqs_ = shape_[0] - 1;
}
@ -226,12 +235,16 @@ public:
const SequenceIdArg& startPositions,
ArgType argType = UNSPECIFIED)
: BufferArg(buf, valueType, shape, argType),
startPositions_(startPositions) {}
startPositions_(startPositions) {
bufferType_ = TENSOR_SEQUENCE_DATA;
}
SequenceArg(const Matrix& matrix,
const IVector& vector,
ArgType argType = UNSPECIFIED)
: BufferArg(matrix, argType), startPositions_(vector) {}
: BufferArg(matrix, argType), startPositions_(vector) {
bufferType_ = TENSOR_SEQUENCE_DATA;
}
~SequenceArg() {}
@ -264,6 +277,7 @@ public:
nnz_(nnz),
format_(format),
type_(type) {
bufferType_ = TENSOR_SPARSE;
CHECK((valueType == VALUE_TYPE_FLOAT) || (valueType == VALUE_TYPE_DOUBLE));
CHECK_EQ(shape_.ndims(), (size_t)2);
CHECK_EQ(row_.shape().ndims(), (size_t)1);

@ -24,7 +24,7 @@ if(WITH_TESTING)
add_simple_unittest(TensorTypeTest)
add_simple_unittest(BufferArgTest)
add_simple_unittest(FunctionTest)
# add_simple_unittest(ContextProjectionOpTest)
add_simple_unittest(ContextProjectionOpTest)
endif()
endif()

File diff suppressed because it is too large Load Diff

@ -21,14 +21,14 @@ namespace paddle {
/**
* \brief Context Projection Forward.
*
* \param[out] outputs output data.
* \param[in] input input data.
* \param[in] weight input weight.
* \param[in] sequence input data.
* \param[in] context_length consecutive rows for concatenation.
* \param[in] context_start context start position.
* \param[in] begin_pad begining pad position.
* \param[in] is_padding whether padding 0 or not.
* \param[in/out] outputs output data.
* \param[in] input input data.
* \param[in] weight input weight.
* \param[in] sequence input data.
* \param[in] context_length consecutive rows for concatenation.
* \param[in] context_start context start position.
* \param[in] begin_pad begining pad position.
* \param[in] is_padding whether padding 0 or not.
*
*/
template <DeviceType DType>
@ -56,7 +56,7 @@ void ContextProjectionForward(
*/
template <DeviceType DType>
void ContextProjectionBackward(
typename Tensor<real, DType>::Matrix& out_grad,
const typename Tensor<real, DType>::Matrix& out_grad,
typename Tensor<real, DType>::Matrix& in_grad,
typename Tensor<real, DType>::Matrix& w_grad,
const typename Tensor<int, DType>::Vector& seq_vec,
@ -68,7 +68,7 @@ void ContextProjectionBackward(
template <DeviceType DType>
void ContextProjectionBackwardData(
typename Tensor<real, DType>::Matrix& out_grad,
const typename Tensor<real, DType>::Matrix& out_grad,
typename Tensor<real, DType>::Matrix& in_grad,
const typename Tensor<int, DType>::Vector& sequence,
size_t context_length,
@ -76,7 +76,7 @@ void ContextProjectionBackwardData(
template <DeviceType DType>
void ContextProjectionBackwardWeight(
typename Tensor<real, DType>::Matrix& out_grad,
const typename Tensor<real, DType>::Matrix& out_grad,
typename Tensor<real, DType>::Matrix& w_grad,
const typename Tensor<int, DType>::Vector& seq_vec,
size_t context_length,

@ -138,10 +138,10 @@ void ContextProjectionForward<DEVICE_TYPE_GPU>(GpuMatrix& output,
begin_pad);
}
__global__ void KeContextProjectionBackwardData(real* out_grad,
__global__ void KeContextProjectionBackwardData(const real* out_grad,
const int* sequence,
real* in_grad,
int input_dim,
size_t input_dim,
int context_length,
int context_start) {
int idx = threadIdx.x;
@ -152,7 +152,8 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
real value = 0;
int instances = seq_end - seq_start + context_length - 1;
out_grad += seq_start * input_dim * context_length;
auto out = const_cast<real*>(out_grad);
out += seq_start * input_dim * context_length;
in_grad += seq_start * input_dim;
for (int k = 0; k <= input_dim / block_size; k++) {
if (idx < input_dim) {
@ -169,7 +170,7 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
int outx = (i - context_length) < 0 ? i : (context_length - 1);
int outy = (i - context_length) < 0 ? 0 : (i - (context_length - 1));
real* output_r =
out_grad + outy * input_dim * context_length + outx * input_dim;
out + outy * input_dim * context_length + outx * input_dim;
for (int j = outy; j < seq_end - seq_start; j++) {
value += output_r[idx];
if (j - outy == outx) break;
@ -194,7 +195,7 @@ __global__ void KeContextProjectionBackwardData(real* out_grad,
* @param[in] context_start context start.
*
*/
void hl_context_projection_backward_data(real* out_grad,
void hl_context_projection_backward_data(const real* out_grad,
const int* sequence,
real* input_grad,
size_t num_sequences,
@ -216,7 +217,7 @@ void hl_context_projection_backward_data(real* out_grad,
}
template <>
void ContextProjectionBackwardData<DEVICE_TYPE_GPU>(GpuMatrix& out_grad,
void ContextProjectionBackwardData<DEVICE_TYPE_GPU>(const GpuMatrix& out_grad,
GpuMatrix& in_grad,
const GpuIVector& sequence,
size_t context_length,
@ -231,7 +232,7 @@ void ContextProjectionBackwardData<DEVICE_TYPE_GPU>(GpuMatrix& out_grad,
}
template<int THREADS_X, int THREADS_Y>
__global__ void KeContextProjectionBackwardWeight(real* out_grad,
__global__ void KeContextProjectionBackwardWeight(const real* out_grad,
const int* sequence,
real* w_grad,
int num_sequences,
@ -254,7 +255,8 @@ __global__ void KeContextProjectionBackwardWeight(real* out_grad,
for (int seqId = idy; seqId < num_sequences; seqId += THREADS_Y) {
int seq_start = sequence[seqId];
int seq_end = sequence[seqId+1];
output_r = out_grad + seq_start * w_dim * context_length;
output_r = const_cast<real*>(out_grad)
+ seq_start * w_dim * context_length;
if (context_start < 0) {
if (padId + context_start < 0) {
@ -318,7 +320,7 @@ __global__ void KeContextProjectionBackwardWeight(real* out_grad,
* beginning.
*
*/
void hl_context_projection_backward_weight(real* out_grad,
void hl_context_projection_backward_weight(const real* out_grad,
const int* sequence,
real* w_grad,
size_t num_sequences,
@ -346,7 +348,7 @@ void hl_context_projection_backward_weight(real* out_grad,
template <>
void ContextProjectionBackwardWeight<DEVICE_TYPE_GPU>(
GpuMatrix& out_grad,
const GpuMatrix& out_grad,
GpuMatrix& w_grad,
const GpuIVector& seq_vec,
size_t context_length,
@ -365,7 +367,7 @@ void ContextProjectionBackwardWeight<DEVICE_TYPE_GPU>(
}
template <>
void ContextProjectionBackward<DEVICE_TYPE_GPU>(GpuMatrix& out_grad,
void ContextProjectionBackward<DEVICE_TYPE_GPU>(const GpuMatrix& out_grad,
GpuMatrix& in_grad,
GpuMatrix& w_grad,
const GpuIVector& sequence,

@ -56,22 +56,25 @@ void testMatrixProjectionForward(int context_start,
cpu_out.randomizeUniform();
gpu_out.copyFrom(cpu_out);
compare.getCpuFunction()->calc(
{Tensor(cpu_in.getData(), Dims{batch_size, input_dim}),
Tensor(cpu_weight ? cpu_weight->getData() : nullptr,
Dims{pad, input_dim}),
Tensor(reinterpret_cast<real*>(cpu_seq->getData()),
Dims{cpu_seq->getSize()})},
{Tensor(cpu_out.getData(), Dims{batch_size, input_dim * context_length})},
{});
compare.getGpuFunction()->calc(
{Tensor(gpu_in.getData(), Dims{batch_size, input_dim}),
Tensor(gpu_weight ? gpu_weight->getData() : nullptr,
Dims{pad, input_dim}),
Tensor(reinterpret_cast<real*>(gpu_seq->getData()),
Dims{gpu_seq->getSize()})},
{Tensor(gpu_out.getData(), Dims{batch_size, input_dim * context_length})},
{});
BufferArgs cpu_inputs;
BufferArgs cpu_outputs;
cpu_inputs.addArg(cpu_in, *cpu_seq);
if (cpu_weight) {
cpu_inputs.addArg(*cpu_weight, *cpu_seq);
}
cpu_outputs.addArg(cpu_out, *cpu_seq, ADD_TO);
compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
BufferArgs gpu_inputs;
BufferArgs gpu_outputs;
gpu_inputs.addArg(gpu_in, *gpu_seq);
if (gpu_weight) {
gpu_inputs.addArg(*gpu_weight, *gpu_seq);
}
gpu_outputs.addArg(gpu_out, *gpu_seq, ADD_TO);
compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
autotest::TensorCheckEqual(cpu_out, gpu_out);
}
@ -117,25 +120,23 @@ void testMatrixProjectionBackward(int context_start,
gpu_w_grad->copyFrom(*cpu_w_grad);
}
compare.getCpuFunction()->calc(
{Tensor(cpu_in_grad.getData(), Dims{batch_size, input_dim}),
Tensor(cpu_w_grad ? cpu_w_grad->getData() : nullptr,
Dims{pad, input_dim}),
Tensor(reinterpret_cast<real*>(cpu_seq->getData()),
Dims{cpu_seq->getSize()})},
{Tensor(cpu_out_grad.getData(),
Dims{batch_size, input_dim * context_length})},
{});
compare.getGpuFunction()->calc(
{Tensor(gpu_in_grad.getData(), Dims{batch_size, input_dim}),
Tensor(gpu_w_grad ? gpu_w_grad->getData() : nullptr,
Dims{pad, input_dim}),
Tensor(reinterpret_cast<real*>(gpu_seq->getData()),
Dims{gpu_seq->getSize()})},
{Tensor(gpu_out_grad.getData(),
Dims{batch_size, input_dim * context_length})},
{});
BufferArgs cpu_inputs;
BufferArgs cpu_outputs;
cpu_inputs.addArg(cpu_out_grad, *cpu_seq);
cpu_outputs.addArg(cpu_in_grad, *cpu_seq, ADD_TO);
cpu_outputs.addArg(
cpu_w_grad ? *cpu_w_grad : CpuMatrix(nullptr, 0, input_dim), ADD_TO);
compare.getCpuFunction()->calc(cpu_inputs, cpu_outputs);
BufferArgs gpu_inputs;
BufferArgs gpu_outputs;
gpu_inputs.addArg(gpu_out_grad, *gpu_seq);
gpu_outputs.addArg(gpu_in_grad, *gpu_seq, ADD_TO);
gpu_outputs.addArg(
gpu_w_grad ? *gpu_w_grad : GpuMatrix(nullptr, 0, input_dim), ADD_TO);
compare.getGpuFunction()->calc(gpu_inputs, gpu_outputs);
autotest::TensorCheckErr(cpu_in_grad, gpu_in_grad);
if (is_padding) {

@ -93,6 +93,12 @@ void BufferArgs::addArg(const GpuSparseMatrix& arg, ArgType argType) {
addArg(*_args_.back());
}
void BufferArgs::addArg(const Matrix& matrix,
const IVector& vector,
ArgType argType) {
args_.push_back(std::make_shared<SequenceArg>(matrix, vector, argType));
}
ClassRegistrar<FunctionBase> FunctionBase::funcRegistrar_;
} // namespace paddle

@ -102,6 +102,10 @@ public:
void addArg(const CpuSparseMatrix& arg, ArgType argType = UNSPECIFIED);
void addArg(const GpuSparseMatrix& arg, ArgType argType = UNSPECIFIED);
void addArg(const Matrix& matrix,
const IVector& vector,
ArgType argType = UNSPECIFIED);
// get argument
const BufferArg& operator[](size_t num) const {
CHECK_LT(num, args_.size());

@ -118,16 +118,15 @@ void ContextProjection::forward() {
/// first use state_, otherwise use weight_(padding false === w nullptr)
auto w_ptr =
state_ ? state_.get() : is_padding ? weight_->getW().get() : nullptr;
auto start_pos = in_->sequenceStartPositions;
const auto start_pos = in_->sequenceStartPositions->getVector(useGpu_);
BufferArgs inputs;
BufferArgs outputs;
inputs.addArg(*in_->value);
inputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr,
w_ptr ? w_ptr->getHeight() : 0,
input_dim));
inputs.addArg(*in_->sequenceStartPositions->getVector(useGpu_));
outputs.addArg(*out_->value, ADD_TO);
inputs.addArg(*in_->value, *start_pos);
if (w_ptr) {
inputs.addArg(CpuMatrix(w_ptr->getData(), w_ptr->getHeight(), input_dim),
*start_pos);
}
outputs.addArg(*out_->value, *start_pos, ADD_TO);
forward_[0]->calc(inputs, outputs);
if (state_ && config_.context_start() < 0) {
@ -166,13 +165,16 @@ void ContextProjection::backward(const UpdateCallback& callback) {
BufferArgs inputs;
BufferArgs outputs;
inputs.addArg(CpuMatrix(
in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim));
inputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr,
w_ptr ? w_ptr->getHeight() : 0,
input_dim));
inputs.addArg(*in_->sequenceStartPositions->getVector(useGpu_));
outputs.addArg(*out_->grad, ADD_TO);
inputs.addArg(*out_->grad, *in_->sequenceStartPositions->getVector(useGpu_));
outputs.addArg(
CpuMatrix(
in_->grad ? in_->grad->getData() : nullptr, batch_size, input_dim),
*in_->sequenceStartPositions->getVector(useGpu_),
ADD_TO);
outputs.addArg(CpuMatrix(w_ptr ? w_ptr->getData() : nullptr,
w_ptr ? w_ptr->getHeight() : 0,
input_dim),
ADD_TO);
backward_[0]->calc(inputs, outputs);
if (config_.trainable_padding()) {

@ -34,6 +34,10 @@ class IScanner(object):
class DenseScanner(IScanner):
"""
:type __mat__: numpy.ndarray
"""
def __init__(self, input_type, pos):
IScanner.__init__(self, input_type, pos)
self.__mat__ = None
@ -47,6 +51,8 @@ class DenseScanner(IScanner):
def finish_scan(self, argument):
assert isinstance(argument, swig_paddle.Arguments)
assert isinstance(self.input_type, dp2.InputType)
if self.__mat__.dtype != numpy.float32:
self.__mat__ = self.__mat__.astype(numpy.float32)
m = swig_paddle.Matrix.createDenseFromNumpy(self.__mat__, True, False)
argument.setSlotValue(self.pos, m)

@ -1,6 +1,4 @@
#!/bin/bash
brew update
brew tap homebrew/science
brew install python
sudo pip install --upgrade protobuf
brew install swig openblas md5sha1sum protobuf
brew install openblas swig md5sha1sum

@ -6,7 +6,7 @@ if [[ "$TRAVIS_OS_NAME" == "linux" ]]; then
export PYTHONPATH=/opt/python/2.7.12/lib/python2.7/site-packages
export PYTHONHOME=/opt/python/2.7.12
export PATH=/opt/python/2.7.12/bin:${PATH}
cmake .. -DON_TRAVIS=ON -DON_COVERALLS=ON -DCOVERALLS_UPLOAD=ON ${EXTRA_CMAKE_OPTS}
cmake .. -DCMAKE_Fortran_COMPILER=/usr/bin/gfortran-4.8 -DON_TRAVIS=ON -DON_COVERALLS=ON -DCOVERALLS_UPLOAD=ON ${EXTRA_CMAKE_OPTS}
NRPOC=`nproc`
make -j $NPROC
make coveralls

@ -4,7 +4,7 @@
source ./common.sh
# Compile Documentation only.
cmake .. -DCMAKE_BUILD_TYPE=Debug -DWITH_GPU=OFF -DWITH_DOC=ON ${EXTRA_CMAKE_OPTS}
cmake .. -DCMAKE_BUILD_TYPE=Debug -DCMAKE_Fortran_COMPILER=/usr/bin/gfortran-4.8 -DWITH_GPU=OFF -DWITH_DOC=ON ${EXTRA_CMAKE_OPTS}
make paddle_docs paddle_docs_cn
# check websites for broken links

Loading…
Cancel
Save