fix conflict

fix-typo
Dong Zhihong 7 years ago
commit 16a39d24f3

@ -12,27 +12,25 @@ 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`, the `name` of the tensor, 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,
|HeaderLength|ContentLength|**LoDTensorDesc**|**TensorValue**|
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.
```text
[offset] [type] [description]
0004 4 bytes integer HeaderLength, the length of LoDTensorDesc
0008 4 bytes integer ContentLength, the length of LodTensor Buffer
0009 1 bytes char TensorDesc
00010 1 bytes char TensorDesc
...
00100 1 bytes char TensorValue
00101 1 bytes char TensorValue
00102 1 bytes char TensorValue ..
...
```
|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() |
| ... | ... | ... |
## Summary
- We introduce a model format.
- The `ProgramDesc` describe the model **topology**.
- The model represented by its forward-pass computation procedure is saved in a **ProgramDesc** protobuf message.
- A bunch of specified format binary tensors describe the **parameters**.

@ -1,7 +1,7 @@
# Regularization in PaddlePaddle
## Introduction to Regularization
A central problem in machine learning is how to design an algorithm that will perform well not just on the training data, but also on new data. Many strategies are used by machine learning practitioners to reduce the test error, possibly at the expense of increased training error. These strategies are collectively known as **regularization**.
A central problem in machine learning is how to design an algorithm that will perform well not just on the training data, but also on new data. A frequently faced problem is the problem of **overfitting**, where the model does not make reliable predictions on new unseen data. **Regularization** is the process of introducing additional information in order to prevent overfitting. This is usually done by adding extra penalties to the loss function that restricts the parameter spaces that an optimization algorithm can explore.
### 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:
@ -18,52 +18,21 @@ The most commonly used norm penalties are the L2 norm penalty and the L1 norm pe
##### L1 Regularization
<img src="./images/l1_regularization.png" align="center"/><br/>
A much more detailed mathematical background of reguilarization can be found [here](http://www.deeplearningbook.org/contents/regularization.html).
A much more detailed mathematical background of regularization can be found [here](http://www.deeplearningbook.org/contents/regularization.html).
## Regularization Survey
## How to do Regularization in PaddlePaddle
On surveying existing frameworks like Tensorflow, PyTorch, Caffe, etc, it can be seen that there are 2 common approaches of doing regularization:
1. Making regularization a part of the optimizer using an attribute like `weight_decay` that is used to control the scale of the L2 Penalty. This approach is used in PyTorch as follows:
```python
opt = torch.optim.SGD(params, lr=0.2, weight_decay=0.2)
```
At every optimization step, this code will add the gradient of the L2 Norm of the params to the gradient of the params with respect to the loss function. This can seen in the following code snippet:
```python
if weight_decay != 0:
d_p.add_(weight_decay, p.data)
```
This is a very restyrictive way of doing regularization and does not give the users enough flexibility.
**Advantages**:
- It is easy to implement for us.
- Faster execution of backward. However, it can be done manually by advanced users too.
**Disadvantages**:
- Not flexible for other regularizations such as L1/L0 regularization.
- Does not allow for different regularization coefficient for different parameters. For example, in most models, ony the weight matrices are regularized and the bias vectors are unregularized.
- Tightly coupled optimizer and regularization implementation.
2. Adding regularization ops to the graph through Python API. This approach is used by Tensorflow and Caffe. Using this approach, we manually add regularization ops to the graph and then add the regularization loss to the final loss function before sending them to the optimizer.
**Advantages**:
- Allows for greater flexibility to the users of Paddle. Using this approach, the users can put different regularization to different parameters and also choose parameters that are not a part of regularization.
- Makes it easy for the users to customize and extend the framework.
**Disadvantages**:
- Implementation requires comprehensive design and time.
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
### Low-Level implementation
In the new design, we propose to create new operations for regularization. For now, we can add 2 ops thgat correspond to the most frequently used regularizations:
In the new design, we propose to create new operations for regularization. For now, we can add 2 ops that correspond to the most frequently used regularizations:
- 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.
@ -94,7 +63,7 @@ Since we want to create the regularization ops in a lazy manner, the regularizat
#### 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 lso 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).
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).

@ -67,7 +67,7 @@ func main() {
cp, err = pserver.LoadCheckpoint(e, idx)
if err != nil {
if err == pserver.ErrCheckpointNotFound {
log.Info("Could not find the pserver checkpoint.")
log.Info("load checkpoint error", "error", err)
} else {
panic(err)
}
@ -99,7 +99,7 @@ func main() {
candy.Must(err)
go func() {
log.Info("starting pserver", log.Ctx{"port": *port})
log.Info("serving pserver", log.Ctx{"port": *port})
err = http.Serve(l, nil)
candy.Must(err)
}()

@ -71,9 +71,15 @@ func newOptimizer(paramWithConfigs ParameterWithConfig, State []byte) *optimizer
cstate = unsafe.Pointer(&s[0])
}
var cptr (*C.uchar)
if len(c) > 0 {
cptr = (*C.uchar)(&c[0])
} else {
log.Error("empty config", "param name", paramWithConfigs.Param.Name)
}
o.config = c
o.opt = C.paddle_create_optimizer(
(*C.uchar)(&c[0]),
cptr,
C.int(len(c)),
C.paddle_element_type(p.ElementType),
cbuffer,

@ -17,12 +17,11 @@ package pserver
import (
"bufio"
"bytes"
"crypto/md5"
"encoding/gob"
"encoding/hex"
"encoding/json"
"errors"
"fmt"
"hash/crc32"
"io/ioutil"
"os"
"path"
@ -40,7 +39,7 @@ type ElementType int
// ErrCheckpointNotFound indicates that the pserver checkpoint could
// not be found.
var ErrCheckpointNotFound = errors.New("checkpoint not found")
var ErrCheckpointNotFound = errors.New("checkpoint not found in etcd")
// RPC error message.
const (
@ -76,7 +75,7 @@ type ParameterWithConfig struct {
type checkpointMeta struct {
UUID string `json:"uuid"`
Path string `json:"path"`
MD5 string `json:"md5"`
CRC32 uint32 `json:"crc32"`
Timestamp int64 `json:"timestamp"`
}
@ -92,7 +91,7 @@ type Service struct {
idx int
checkpointInterval time.Duration
checkpointPath string
client *EtcdClient
client KVStore
mu sync.Mutex
optMap map[string]*optimizer
@ -104,7 +103,12 @@ type parameterCheckpoint struct {
State []byte
}
func loadMeta(e *EtcdClient, idx int) (meta checkpointMeta, err error) {
type KVStore interface {
GetKey(key string, timeout time.Duration) ([]byte, error)
PutKey(key string, value []byte, timeout time.Duration, withLease bool) error
}
func loadMeta(e KVStore, idx int) (meta checkpointMeta, err error) {
v, err := e.GetKey(PsCheckpoint+strconv.Itoa(idx), 3*time.Second)
if err != nil {
return
@ -123,7 +127,7 @@ func loadMeta(e *EtcdClient, idx int) (meta checkpointMeta, err error) {
}
// LoadCheckpoint loads checkpoint from file.
func LoadCheckpoint(e *EtcdClient, idx int) (Checkpoint, error) {
func LoadCheckpoint(e KVStore, idx int) (Checkpoint, error) {
log.Info("Loading checkpoint", "pserver index", idx)
defer traceTime(time.Now(), "load checkpoint")
@ -137,11 +141,8 @@ func LoadCheckpoint(e *EtcdClient, idx int) (Checkpoint, error) {
return nil, err
}
// TODO(helin): change MD5 to CRC since CRC is better for file
// checksum in our use case (emphasize speed over security).
h := md5.New()
md5 := hex.EncodeToString(h.Sum(content))
if md5 != cpMeta.MD5 {
crc32 := crc32.ChecksumIEEE(content)
if crc32 != cpMeta.CRC32 {
return nil, errors.New(WrongChecksum)
}
@ -150,12 +151,13 @@ func LoadCheckpoint(e *EtcdClient, idx int) (Checkpoint, error) {
if err = dec.Decode(&cp); err != nil {
return nil, err
}
return cp, nil
}
// NewService creates a new service, will bypass etcd registration if no
// endpoints specified. It will recovery from checkpoint file if a exists a specified checkpoint.
func NewService(idx int, interval time.Duration, path string, client *EtcdClient, cp Checkpoint) (*Service, error) {
func NewService(idx int, interval time.Duration, path string, client KVStore, cp Checkpoint) (*Service, error) {
s := &Service{
idx: idx,
checkpointInterval: interval,
@ -173,6 +175,7 @@ func NewService(idx int, interval time.Duration, path string, client *EtcdClient
}
s.optMap[p.Param.Name] = newOptimizer(p, item.State)
}
close(s.initialized)
}
return s, nil
}
@ -221,7 +224,7 @@ func (s *Service) FinishInitParams(_ int, _ *int) error {
for range t {
err := s.checkpoint()
if err != nil {
log.Error("finish init params error", log.Ctx{"error": err})
log.Error("checkpoint error", log.Ctx{"error": err})
}
}
}()
@ -274,6 +277,7 @@ func (s *Service) GetParam(name string, parameter *Parameter) error {
parameter.Name = name
parameter.ElementType = opt.elementType
parameter.Content = opt.GetWeights()
log.Info("sending parameter to the trainer", "name", parameter.Name, "size", len(parameter.Content), "type", parameter.ElementType)
return nil
}
@ -354,20 +358,29 @@ func (s *Service) checkpoint() (err error) {
oldMeta, err := loadMeta(s.client, s.idx)
if err == ErrCheckpointNotFound {
log.Info("Do not have existing checkpoint.")
log.Info("old meta not found, skip removing old meta")
err = nil
} else if err == nil {
log.Info("removing old meta")
if oldMeta.Path != "" {
rmErr := os.Remove(oldMeta.Path)
if rmErr != nil {
// log error, but still treat checkpoint as
// successful.
log.Error("remove old meta file error", log.Ctx{"error": rmErr})
}
}
}
if err != nil {
return
}
h := md5.New()
md5 := hex.EncodeToString(h.Sum(buf.Bytes()))
crc32 := crc32.ChecksumIEEE(buf.Bytes())
cpMeta := checkpointMeta{
UUID: id,
Timestamp: time.Now().UnixNano(),
MD5: md5,
CRC32: crc32,
Path: p,
}
@ -381,14 +394,5 @@ func (s *Service) checkpoint() (err error) {
return
}
if oldMeta.Path != "" {
rmErr := os.Remove(oldMeta.Path)
if rmErr != nil {
// log error, but still treat checkpoint as
// successful.
log.Error("remove old meta file error", log.Ctx{"error": rmErr})
}
}
return
}

@ -0,0 +1,86 @@
package pserver
import (
"bytes"
"encoding/binary"
"fmt"
"testing"
"time"
"github.com/stretchr/testify/assert"
)
const testDir = "./test_data"
type myKV struct {
m map[string][]byte
}
func (m *myKV) GetKey(key string, timeout time.Duration) ([]byte, error) {
if m.m == nil {
m.m = make(map[string][]byte)
}
return m.m[key], nil
}
func (m *myKV) PutKey(key string, value []byte, timeout time.Duration, withLease bool) error {
if m.m == nil {
m.m = make(map[string][]byte)
}
m.m[key] = value
return nil
}
func TestCheckpoint(t *testing.T) {
kv := &myKV{}
s, err := NewService(0, time.Hour, testDir, kv, nil)
assert.Nil(t, err)
err = s.checkpoint()
assert.Nil(t, err)
_, err = LoadCheckpoint(kv, 0)
assert.Nil(t, err)
}
func float32ToByte(f float32) []byte {
var buf bytes.Buffer
err := binary.Write(&buf, binary.LittleEndian, f)
if err != nil {
fmt.Println("binary.Write failed:", err)
}
return buf.Bytes()
}
func TestCheckpointWithData(t *testing.T) {
kv := &myKV{}
s, err := NewService(0, time.Hour, testDir, kv, nil)
assert.Nil(t, err)
var content []byte
for i := 0; i < 50000; i++ {
content = append(content, float32ToByte(float32(i))...)
}
p1 := Parameter{Name: "p1", ElementType: 1, Content: content}
err = s.InitParam(ParameterWithConfig{Param: p1}, nil)
assert.Nil(t, err)
err = s.FinishInitParams(0, nil)
assert.Nil(t, err)
var p2 Parameter
err = s.GetParam(p1.Name, &p2)
assert.Nil(t, err)
assert.Equal(t, p1, p2)
err = s.checkpoint()
assert.Nil(t, err)
cp, err := LoadCheckpoint(kv, 0)
assert.Nil(t, err)
s1, err := NewService(0, time.Hour, testDir, kv, cp)
assert.Nil(t, err)
var p3 Parameter
err = s1.GetParam(p1.Name, &p3)
assert.Nil(t, err)
assert.Equal(t, p1, p3)
}

@ -178,7 +178,3 @@ func TestBlockUntilInitialized(t *testing.T) {
wg.Wait()
}
func TestCheckpointSpeed(t *testing.T) {
//TODO(zhihong): test speed
}

@ -64,12 +64,18 @@ paddle_error paddle_gradient_machine_create_for_inference_with_parameters(
modelConfigProtobuf.resize(modelConfigSize);
is.read(&modelConfigProtobuf[0], modelConfigSize);
paddle::TrainerConfig config;
paddle::ModelConfig modelConfig;
if (!config.ParseFromString(modelConfigProtobuf) || !config.IsInitialized()) {
return kPD_PROTOBUF_ERROR;
if (!modelConfig.ParseFromString(modelConfigProtobuf) ||
!modelConfig.IsInitialized()) {
return kPD_PROTOBUF_ERROR;
}
} else {
modelConfig = config.model_config();
}
auto ptr = new paddle::capi::CGradientMachine();
ptr->machine.reset(paddle::GradientMachine::create(
config.model_config(), CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE}));
modelConfig, CREATE_MODE_TESTING, {paddle::PARAMETER_VALUE}));
std::vector<paddle::ParameterPtr>& parameters = ptr->machine->getParameters();
for (auto& para : parameters) {
para->load(is);

@ -1,6 +1,5 @@
# ddim lib
proto_library(framework_proto SRCS framework.proto)
proto_library(saver_proto SRCS framework.proto saver.proto)
cc_library(ddim SRCS ddim.cc DEPS eigen3)
cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
@ -10,7 +9,7 @@ cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor saver_proto framework_proto)
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor paddle_memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)

@ -15,6 +15,7 @@
#pragma once
#include <typeindex>
#include "paddle/framework/framework.pb.h"
#include "paddle/platform/enforce.h"
namespace paddle {
namespace framework {

@ -13,7 +13,6 @@
limitations under the License. */
#include "paddle/framework/lod_tensor.h"
#include "paddle/framework/saver.pb.h"
#include "paddle/memory/memcpy.h"
#include "paddle/memory/memory.h"
@ -106,6 +105,15 @@ size_t LoDTensor::NumElements(size_t level, size_t idx) const {
return lod_[level][idx + 1] - lod_[level][idx];
}
size_t LoDTensor::NumInstancesInElement(size_t level, size_t idx) const {
PADDLE_ENFORCE_LT(level, NumLevels());
PADDLE_ENFORCE_LT(idx, NumElements(level));
auto abs_lod = ToAbsOffset(lod());
size_t begin = abs_lod[level][idx];
size_t end = abs_lod[level][idx + 1];
return end - begin;
}
void LoDTensor::ShrinkLevels(size_t level_begin, size_t level_end) {
auto new_lod = framework::SliceLevels(lod_, level_begin, level_end);
lod_ = new_lod;
@ -117,144 +125,15 @@ void LoDTensor::ShrinkInLevel(size_t level, size_t elem_begin,
PADDLE_ENFORCE_LT(elem_begin, NumElements(level));
PADDLE_ENFORCE_LT(elem_end, NumElements(level) + 1);
auto abs_lod = framework::ToAbsOffset(lod());
auto new_lod = framework::SliceInLevel(lod_, level, elem_begin, elem_end);
lod_ = new_lod;
}
std::string LoDTensor::SerializeToString() const {
LoDTensorProto desc;
// set data_type
if (this->type() == typeid(int8_t)) desc.set_data_type(DataType::BOOL);
if (this->type() == typeid(int16_t)) desc.set_data_type(DataType::INT16);
if (this->type() == typeid(int32_t)) desc.set_data_type(DataType::INT32);
if (this->type() == typeid(int64_t)) desc.set_data_type(DataType::INT64);
// FIXME(dzh): there is no fp16 in standard c++
if (this->type() == typeid(float)) // NOLINT
desc.set_data_type(DataType::FP32);
if (this->type() == typeid(double)) // NOLINT
desc.set_data_type(DataType::FP64);
for (int i = 0; i < dims().size(); ++i) {
desc.add_dims(dims()[i]);
}
// set lod information
desc.set_lod_level(this->NumLevels());
for (size_t i = 0; i < this->NumLevels(); ++i) {
LoDInfo* lod = desc.add_levels();
for (size_t j = 0; j < lod_[i].size(); ++j) {
lod->add_level(lod_[i][j]);
}
}
desc.set_version(0);
std::string desc_bytes = desc.SerializeAsString();
// FIXME(dzh) : implement fix chunk size buffer.
size_t DESC_SIZE = desc_bytes.size();
size_t DATA_SIZE = holder_->size() - offset_;
const size_t BUFFER_SIZE = DESC_SIZE + DATA_SIZE + 2 * sizeof(size_t);
char* buffer =
static_cast<char*>(memory::Alloc(platform::CPUPlace(), BUFFER_SIZE));
// format: desc_size data_size, desc_bytes, data_bytes.
platform::CPUPlace src_place;
platform::CPUPlace dst_place;
memory::Copy(dst_place, buffer, src_place, &BUFFER_SIZE, sizeof(size_t));
memory::Copy(dst_place, buffer + sizeof(size_t), src_place, &DESC_SIZE,
sizeof(size_t));
memory::Copy(dst_place, buffer + sizeof(size_t) * 2, src_place,
desc_bytes.c_str(), desc_bytes.size());
PADDLE_ENFORCE(this->numel() != 0, "Serialize a empty Tensor!");
platform::Place place = holder_->place();
int element_width = holder_->size() / this->numel();
if (platform::is_cpu_place(place)) {
memory::Copy(dst_place, buffer + sizeof(size_t) * 2 + desc_bytes.size(),
boost::get<platform::CPUPlace>(place),
static_cast<char*>(holder_->ptr()) + offset_ / element_width,
DATA_SIZE);
}
#ifdef PADDLE_WITH_GPU
if (platform::is_gpu_place(place)) {
memory::Copy(dst_place, buffer + sizeof(size_t) * 2 + desc_bytes.size(),
boost::get<platform::GPUPlace>(place),
static_cast<char*>(holder_->ptr()) + offset_ / element_width,
DATA_SIZE);
}
#endif
std::string ret(buffer, BUFFER_SIZE);
memory::Free(platform::CPUPlace(), buffer);
return ret;
// slice the underlying tensor
size_t begin = abs_lod[level][elem_begin];
size_t end = abs_lod[level][elem_end];
PADDLE_ENFORCE_LT(begin, end, "Cannot shrink, the result tensor is empty.");
ShareDataWith(Slice(begin, end));
}
void LoDTensor::DeserializeFromString(const std::string& s,
const platform::Place& dst_place) {
size_t DESC_SIZE, BUFFER_SIZE;
platform::CPUPlace src_place;
memory::Copy(src_place, &BUFFER_SIZE, src_place, s.c_str(), sizeof(size_t));
memory::Copy(src_place, &DESC_SIZE, src_place, s.c_str() + sizeof(size_t),
sizeof(size_t));
const size_t DATA_SIZE = BUFFER_SIZE - DESC_SIZE - sizeof(size_t) * 2;
// parse LoDTensorDesc
LoDTensorProto desc;
desc.ParseFromArray(s.c_str() + sizeof(size_t) * 2, DESC_SIZE);
std::vector<int64_t> dims;
std::copy(desc.dims().begin(), desc.dims().end(), std::back_inserter(dims));
this->Resize(make_ddim(dims));
// parse data type
void* ptr = nullptr;
if (desc.data_type() == DataType::BOOL)
ptr = this->mutable_data<bool>(dst_place);
if (desc.data_type() == DataType::INT16)
ptr = this->mutable_data<int16_t>(dst_place);
if (desc.data_type() == DataType::INT32)
ptr = this->mutable_data<int32_t>(dst_place);
if (desc.data_type() == DataType::INT64)
ptr = this->mutable_data<int64_t>(dst_place);
// FIXME(dzh): there is no fp16 in standard c++
if (desc.data_type() == DataType::FP32)
ptr = this->mutable_data<float>(dst_place);
if (desc.data_type() == DataType::FP64)
ptr = this->mutable_data<double>(dst_place);
LoD lod;
std::vector<size_t> levels;
for (int i = 0; i < desc.levels().size(); ++i) {
auto current_level = desc.levels()[i].level();
std::copy(current_level.begin(), current_level.end(),
std::back_inserter(levels));
lod.emplace_back(levels);
levels.clear();
}
this->set_lod(lod);
if (platform::is_cpu_place(dst_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), ptr, src_place,
s.c_str() + sizeof(size_t) * 2 + DESC_SIZE, DATA_SIZE);
}
#ifdef PADDLE_WITH_GPU
if (platform::is_gpu_place(dst_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), ptr, src_place,
s.c_str() + sizeof(size_t) * 2 + DESC_SIZE, DATA_SIZE);
}
#endif
}
} // namespace framework
} // namespace paddle

@ -85,7 +85,9 @@ class LoDTensor : public Tensor {
void set_lod(const LoD& lod) { lod_ = lod; }
LoD lod() const { return lod_; }
const LoD& lod() const { return lod_; }
LoD* mutable_lod() { return &lod_; }
/*
* Get the start offset and end offset of an element from LoD.
@ -122,6 +124,12 @@ class LoDTensor : public Tensor {
*/
size_t NumElements(size_t level, size_t idx) const;
/*
* Get the number of instances in the underlying tensor in the `idx`-th
* element.
*/
size_t NumInstancesInElement(size_t level, size_t idx) const;
/*
* Shrink levels[level_begin:level_end]
*/
@ -133,29 +141,45 @@ class LoDTensor : public Tensor {
*/
void ShrinkInLevel(size_t level, size_t elem_begin, size_t elem_end);
/**
* @brief Serialize tensor to char bytes.
* Please check model_format.md for the format detail.
* NOTE: GPUTensor will copy data to cpu implicitly.
* @return return string
*/
// FIXME(dzh) : Currently, this interface should only be used in
// save/restore model and checkpoint. ParameterServer do not use shape
// information to do the optimization, as a result, when we serialize
// parameter/gradient to string, we should serialize the tensor
// to string in the ps trainer instead of LoDTensor.
std::string SerializeToString() const;
/**
* @brief Deserialize char bytes to tensor.
* @return return string
*/
void DeserializeFromString(const std::string& s,
const platform::Place& dst_place);
private:
LoD lod_;
};
/*
* Expand the `source` to fit the LoD of `lod`. For example, a `source`
* LoDTensor is
* - LoD: [0, 2]
* - tensor: [a0, a1]
* a `lod` is
* - LoD: [0 3 5]
* returns a new LoDTensor
* - [a0 a0 a0 a1 a1]
*/
template <typename T>
LoDTensor LodExpand(const LoDTensor& source, const LoD& lod, size_t level,
const platform::Place& place) {
LoD abs_lod = ToAbsOffset(lod);
const auto& lod_level = lod[level];
size_t num_instances = source.dims()[0];
// new tensor
LoDTensor tensor;
tensor.set_lod(lod);
auto dims = source.dims();
dims[0] = lod_level.back();
tensor.Resize(dims);
tensor.mutable_data<T>(place);
PADDLE_ENFORCE_EQ(num_instances, lod_level.size() - 1);
for (size_t ins = 0; ins < num_instances; ins++) {
for (size_t elem = lod_level[ins]; elem < lod_level[ins + 1]; elem++) {
tensor.Slice(elem, elem + 1)
.CopyFrom(source.Slice(ins, ins + 1), platform::CPUPlace(),
platform::CPUDeviceContext());
}
}
return tensor;
}
} // namespace framework
} // namespace paddle

@ -92,11 +92,14 @@ TEST_F(LoDTensorTester, ShrinkInLevel) {
size_t level = 0;
LoDTensor new_lod_tensor = lod_tensor_;
new_lod_tensor.ShrinkInLevel(level, 0, 1);
EXPECT_EQ(new_lod_tensor.NumLevels(), 3UL);
EXPECT_EQ(new_lod_tensor.NumElements(0), 1UL);
EXPECT_EQ(new_lod_tensor.NumElements(1), 2UL);
EXPECT_EQ(new_lod_tensor.NumElements(2), 5UL);
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(2), 5UL);
ASSERT_EQ(new_lod_tensor.dims()[0], 12);
for (int i = 0; i < 12 * 128; i++) {
ASSERT_EQ(new_lod_tensor.data<float>()[i], i);
}
level = 1;
new_lod_tensor = lod_tensor_;
@ -104,23 +107,41 @@ TEST_F(LoDTensorTester, ShrinkInLevel) {
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 3UL);
ASSERT_EQ(new_lod_tensor.data<float>(), lod_tensor_.data<float>());
ASSERT_EQ(new_lod_tensor.dims()[0], 7);
for (int i = 5 * 128; i < 12 * 128; i++) {
ASSERT_EQ(new_lod_tensor.data<float>()[i - 5 * 128], i);
}
LoDTensor t1;
t1.set_lod(lod_tensor_.lod());
t1.ShareDataWith(lod_tensor_);
LoDTensor t2;
t2.set_lod(lod_tensor_.lod());
t2.ShareDataWith(lod_tensor_);
t1.ShrinkInLevel(0, 1, 2);
t2.ShrinkInLevel(0, 0, 1);
EXPECT_NE(t1.data<float>(), t2.data<float>());
EXPECT_NE(t1.data<float>(), lod_tensor_.data<float>());
}
TEST_F(LoDTensorTester, SerializeDeserialize) {
LoDTensor new_lod_tensor = lod_tensor_;
float* src_ptr = lod_tensor_.data<float>();
std::string s = lod_tensor_.SerializeToString();
LoDTensor dst;
dst.DeserializeFromString(s, platform::CPUPlace());
float* dst_ptr = dst.data<float>();
for (int i = 0; i < kLodTensorSize; ++i) {
EXPECT_EQ(dst_ptr[i], src_ptr[i]);
TEST(LodExpand, test) {
LoD lod{{0, 2}};
LoDTensor tensor;
tensor.set_lod(lod);
tensor.Resize({2, 1});
tensor.mutable_data<float>(platform::CPUPlace());
tensor.data<float>()[0] = 0;
tensor.data<float>()[1] = 1;
LoD target;
target.emplace_back(std::vector<size_t>{0, 3, 5});
auto new_tensor = LodExpand<float>(tensor, target, 0UL, platform::CPUPlace());
std::vector<int> result{{0, 0, 0, 1, 1}};
for (size_t i = 0; i < 5; i++) {
ASSERT_EQ(new_tensor.data<float>()[i], result[i]);
}
ASSERT_EQ(dst.NumElements(0), 2UL);
ASSERT_EQ(dst.NumElements(1), 3UL);
ASSERT_EQ(dst.NumElements(2), 8UL);
}
} // namespace framework

@ -47,31 +47,4 @@ TEST(LoDTensor, LoDInGPU) {
for (size_t i = 0; i < src_lod[0].size(); ++i) {
CHECK_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);
}
}
TEST(LoDTensor, SerializeDeserialize) {
paddle::framework::LoDTensor lod_tensor;
paddle::platform::GPUPlace place(0);
paddle::framework::LoD src_lod;
src_lod.push_back(std::vector<size_t>{0, 2, 4, 6, 8, 10, 12, 14});
lod_tensor.Resize({14, 16});
lod_tensor.mutable_data<float>(place);
lod_tensor.set_lod(src_lod);
CHECK_EQ(lod_tensor.lod_element(0, 2).first, 4UL);
CHECK_EQ(lod_tensor.lod_element(0, 4).first, 8UL);
test<<<1, 8>>>(src_lod[0].data(), src_lod[0].size());
cudaDeviceSynchronize();
std::string s = lod_tensor.SerializeToString();
paddle::framework::LoDTensor dst;
dst.DeserializeFromString(s, place);
paddle::framework::LoD dst_lod = dst.lod();
for (size_t i = 0; i < dst_lod[0].size(); ++i) {
CHECK_EQ(src_lod[0].data()[i], dst_lod[0].data()[i] * 2);
}
}
}

@ -1,39 +0,0 @@
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
syntax = "proto2";
option optimize_for = LITE_RUNTIME;
package paddle.framework;
import "framework.proto";
/**
* This file contains necessary information for model, checkpoint.
* etc.
*/
message LoDInfo { repeated int64 level = 1; }
/**
* Save the LoDTensorDesc information through LoDTensorProto, its data memory
* is copyed to c buffer immediately. See model_format.md for details.
*/
message LoDTensorProto {
optional DataType data_type = 1;
repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
repeated LoDInfo levels = 3;
optional int32 lod_level = 4 [ default = 0 ];
optional int32 version = 5;
}

@ -132,6 +132,8 @@ class Tensor {
std::type_index type() const { return holder_->type(); }
size_t memory_size() const;
private:
inline void check_memory_size() const;

@ -20,6 +20,8 @@
#include <algorithm>
#include <limits>
#include "paddle/framework/eigen.h"
namespace paddle {
namespace framework {
@ -104,10 +106,10 @@ void TensorArray::Write(size_t index, const LoDTensor& value) {
values_.resize(index + 1);
}
values_[index].set_lod(value.lod());
values_[index].Resize(value.dims());
values_[index].mutable_data<value_type>(platform::CPUPlace());
values_[index].CopyFrom(value, platform::CPUPlace(),
platform::CPUDeviceContext());
values_[index].mutable_data<value_type>(value.place());
values_[index].CopyFrom(value, value.place(), platform::CPUDeviceContext());
}
void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
@ -116,6 +118,7 @@ void TensorArray::WriteShared(size_t index, const LoDTensor& value) {
values_.resize(index + 1);
}
values_[index].set_lod(value.lod());
values_[index].ShareDataWith(value);
}
@ -144,6 +147,155 @@ DySeqMetaBatch TensorArray::Unpack(const LoDTensor& source, int level,
return unpacker.meta;
}
LoDTensor TensorArray::LodPack(size_t level) const {
PADDLE_ENFORCE_GT(size(), 0UL, "no time step exists");
// the levels should be no less than 2
LoDTensor merged;
const LoDTensor *pre, *cur;
pre = &Read(0);
for (size_t step = 1; step < size(); step++) {
cur = &Read(step);
PADDLE_ENFORCE_GT(cur->NumLevels(), 0);
PADDLE_ENFORCE_GT(pre->NumLevels(), 0);
PADDLE_ENFORCE_EQ(pre->NumLevels(), cur->NumLevels());
PADDLE_ENFORCE_EQ(pre->NumElements(level), cur->NumElements(level));
merged = LodPackTwo(*pre, *cur, level);
pre = &merged;
}
return merged;
}
/*
* NOTE currently, only the lowest level supports packing.
* The lowest LoD will be changed, while the relative offsets in levels above
* stay unchanged.
*
* previous step : [0] [1] [3]
* current step: [0 1 2] [2 3] []
* packed to
* [0 0] [0 1] [0 2] [1 2] [1 3] [3]
*/
LoDTensor TensorArray::LodPackTwo(const LoDTensor& pre, const LoDTensor& cur,
size_t level) const {
PADDLE_ENFORCE_EQ(pre.NumLevels(), cur.NumLevels());
PADDLE_ENFORCE_EQ(pre.NumLevels(), level + 1,
"Only the lowest LoD level supports pack temporarily.");
// calculate the result tensor's shape first
size_t num_instances = 0;
for (size_t elem = 0; elem < pre.NumElements(level); elem++) {
size_t prefix_size = pre.NumElements(level, elem);
size_t num_candidates = cur.NumElements(level, elem);
if (num_candidates > 0) {
num_instances += num_candidates * (prefix_size + 1);
} else {
num_instances += prefix_size;
}
}
auto res_dims = pre.dims();
res_dims[0] = num_instances;
LoDTensor result;
result.Resize(res_dims);
result.mutable_data<value_type>(cur.place());
Vector<size_t> last_lod_level;
// copy data
size_t index = 0;
last_lod_level.push_back(index);
for (size_t elem = 0; elem < pre.NumElements(level); elem++) {
size_t prefix_size = pre.NumElements(level, elem);
size_t num_candidates = cur.NumElements(level, elem);
// slice the prefix Tensor
LoDTensor prefix = pre;
prefix.ShrinkInLevel(level, elem, elem + 1);
LoDTensor candidate = cur;
if (num_candidates > 0) {
candidate.ShrinkInLevel(level, elem, elem + 1);
} else { // just push prefix
result.Slice(index, index + prefix_size)
.CopyFrom(prefix, result.place(), platform::CPUDeviceContext());
index += prefix_size;
last_lod_level.push_back(index);
}
for (size_t candi = 0; candi < num_candidates; candi++) {
// TODO(superjom) support GPU
result.Slice(index, index + prefix_size)
.CopyFrom(prefix, result.place(), platform::CPUDeviceContext());
index += prefix_size;
// copy candidate record
result.Slice(index, index + 1)
.CopyFrom(candidate.Slice(candi, candi + 1), result.place(),
platform::CPUDeviceContext());
index++;
last_lod_level.push_back(index);
}
}
// update lod
auto lod = cur.lod();
lod.back() = last_lod_level;
result.set_lod(lod);
return result;
}
/*
* source [0 1 2] [3 4] [5 6 7] will be transformd to a list of LoDTensors such
* as
* [0 3 5] [1 4 6] [2 7] with 1-level LoDs:
* - [0 1 2 3]
* - [0 1 2 3]
* - [0 1 1 2], the [1,1) here means the second sequence is empty
*
* NOTE Unpack a LoDTensor in this approach may result in a big LoD.
*/
void TensorArray::LodUnpack(const LoDTensor& source, size_t level) {
PADDLE_ENFORCE_EQ(level, source.NumLevels() - 1,
"only the lowest LoD level supports unpack.");
const size_t non_empty_instances = source.dims()[0];
size_t index = 0;
Vector<size_t> lowest_lod_level;
lowest_lod_level.push_back(index);
for (size_t step = 0; step < non_empty_instances; step++) {
size_t num_instances = 0;
for (size_t id = 0; id < source.NumElements(level); id++) {
auto instance = source;
instance.ShrinkInLevel(level, id, id + 1);
if (static_cast<size_t>(instance.dims()[0]) > step) {
num_instances++;
index++;
}
lowest_lod_level.push_back(index);
}
// create tensor for this time step
LoDTensor tensor;
auto dims = source.dims();
dims[0] = num_instances;
// set lod
auto lod = source.lod();
lod.back() = lowest_lod_level;
tensor.set_lod(lod);
index = 0;
for (size_t id = 0; id < source.NumElements(level); id++) {
auto instance = source;
instance.ShrinkInLevel(level, id, id + 1);
if (static_cast<size_t>(instance.dims()[0]) > step) {
// copy this instance
tensor.Slice(index, index + 1)
.CopyFrom(instance.Slice(step, step + 1), tensor.place(),
platform::CPUDeviceContext());
index++;
}
}
Write(step, tensor);
}
}
LoDTensor TensorArray::Stack() const {
LoDTensor result;
if (size() == 0) return result;

@ -86,6 +86,16 @@ class TensorArray {
*/
DySeqMetaBatch Unpack(const LoDTensor &source, int level, bool length_desend);
/*
* Pack an array of LoDTensors to a LoDTensor.
*/
LoDTensor LodPack(size_t level) const;
/*
* Unpack a LoDTensor to an array of LoDTensors.
*/
void LodUnpack(const LoDTensor &source, size_t level);
/*
* Pack the values into a tensor with rank one higher than each tensor in
* values.
@ -111,6 +121,9 @@ class TensorArray {
protected:
void Unstack(const LoDTensor &source, bool data_shared) const;
LoDTensor LodPackTwo(const LoDTensor &pre, const LoDTensor &cur,
size_t level) const;
private:
mutable std::vector<LoDTensor> values_;
}; // class TensorArray

@ -126,5 +126,57 @@ TEST_F(TensorArrayTester, size) {
ASSERT_EQ(ta.size(), static_cast<size_t>(batch_size));
}
TEST(TensorArray, LodPack) {
// three time steps, each step stores a LoDTensors
// - [0] [1]
// - [2 3], [4 5]
// - [6 7] [] [8], [9, 10]
// try to get a LoDTensor with content:
// - [0 2 6]
// - [0 2 7]
// - [0 3]
// - [1 4 8]
// - [1 5 9]
// - [1 5 10]
std::array<LoDTensor, 3> tensors;
tensors[0].Resize(make_ddim({2, 1}));
tensors[1].Resize(make_ddim({4, 1}));
tensors[2].Resize(make_ddim({5, 1}));
int index = 0;
for (auto& t : tensors) {
t.mutable_data<int>(platform::CPUPlace());
for (int i = 0; i < t.dims()[0]; i++) {
t.data<int>()[i] = index;
index++;
}
}
std::array<LoD, 3> lods;
std::vector<std::vector<size_t>> levels{
{0, 1, 2}, {0, 2, 4}, {0, 2, 2, 3, 5}};
for (int i = 0; i < 3; i++) {
lods[i].emplace_back(levels[i].begin(), levels[i].end());
}
TensorArray ta;
for (int i = 0; i < 3; i++) {
tensors[i].set_lod(lods[i]);
ta.Write(i, tensors[i]);
}
auto merged = ta.LodPack(0);
std::vector<int> target_tensor_data{{0, 2, 6, // 0
0, 2, 7, // 1
0, 3, // 2
1, 4, 8, // 3
1, 5, 9, // 5
1, 5, 10}};
EXPECT_EQ(merged.dims()[0], (int)target_tensor_data.size());
for (size_t i = 0; i < target_tensor_data.size(); i++) {
EXPECT_EQ(target_tensor_data[i], merged.data<int>()[i]);
}
}
} // namespace framework
} // namespace paddle

@ -62,12 +62,16 @@ inline void Tensor::check_memory_size() const {
PADDLE_ENFORCE_NOT_NULL(
holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");
PADDLE_ENFORCE_GE(
holder_->size(), numel() * SizeOfType(type()) + offset_,
holder_->size(), memory_size() + offset_,
"Tensor's dims_ is out of bound. Call Tensor::mutable_data "
"first to re-allocate memory.\n"
"or maybe the required data-type mismatches the data already stored.");
}
inline size_t Tensor::memory_size() const {
return holder_ == nullptr ? 0UL : numel() * SizeOfType(type());
}
template <typename T>
inline const T* Tensor::data() const {
check_memory_size();

@ -46,6 +46,8 @@ class Variable {
std::type_index(typeid(T)) == std::type_index(holder_->Type());
}
void Clear() { holder_.reset(); }
private:
struct Placeholder {
virtual ~Placeholder() {}

File diff suppressed because it is too large Load Diff

@ -0,0 +1,138 @@
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "MKLDNNLayer.h"
#include "mkldnn.hpp"
namespace paddle {
typedef mkldnn::batch_normalization_forward bn_fwd;
typedef mkldnn::batch_normalization_backward bn_bwd;
/**
* @brief A subclass of MKLDNNLayer BatchNorm layer.
*
* The config file api is mkldnn_batch_norm
*/
class MKLDNNBatchNormLayer : public MKLDNNLayer {
protected:
// save forward primitive_desc, which can be used backward
std::shared_ptr<bn_fwd::primitive_desc> fwdPD_;
// Epsilon value used in the batch normalization formula.
static const real EPS;
// weight and bias in paddle
std::unique_ptr<Weight> weight_;
std::unique_ptr<Weight> biases_;
// mkldnn use a large buffer store both scale and shift
// which are weight and bias in paddle corresponding.
MatrixPtr valueScaleShift_;
MatrixPtr gradScaleShift_;
// Moving average of mean.
std::unique_ptr<Weight> movingMean_;
// Moving average of variance.
std::unique_ptr<Weight> movingVar_;
// if useGlobalStats_ is true, will use the loaded mean and variance.
// otherwise, calculate mean and variance in every mini-batch.
bool useGlobalStats_;
// used in MKLDNN primitive desc
unsigned flags_;
// use to compute moving mean and variance.
real movingAvgFraction_;
// whether the weight has been init
bool hasInitedWgt_;
// local mean and variance
// when useGlobalStats_ they are loaded from moving mean and variance
// when do not useGlobalStats_ they are calculated from this mini-batch
MKLDNNMatrixPtr mean_;
MKLDNNMatrixPtr var_;
public:
explicit MKLDNNBatchNormLayer(const LayerConfig& config)
: MKLDNNLayer(config), useGlobalStats_(true), hasInitedWgt_(false) {}
~MKLDNNBatchNormLayer() {}
bool init(const LayerMap& layerMap,
const ParameterMap& parameterMap) override;
void forward(PassType passType) override;
void reshape(
int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;
void resetFwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void resetBwd(std::vector<mkldnn::primitive>& pipeline,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& bias,
MKLDNNMatrixPtr& out) override;
void updateWeights(const UpdateCallback& callback) override;
void convertWeightsFromPaddle() override;
protected:
void initWeight();
/**
* cal moving mean and variance.
* moving = moving * AvgFraction + local * (1 - AvgFraction)
*/
void calMovingMeanAndVar();
/**
* Forward functions: reset buffers(input, weight, output),
* reset primitive descriptor,
* reset pipeline.
*/
void resetFwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& out);
void resetFwdPD(std::shared_ptr<bn_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr in,
MKLDNNMatrixPtr wgt,
MKLDNNMatrixPtr out);
void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<bn_fwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& out);
/**
* Backward functions: reset buffers(input, weight, output),
* reset primitive descriptor,
* reset pipeline.
*/
void resetBwdBuffers(MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& out);
void resetBwdPD(std::shared_ptr<bn_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& out);
void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
std::shared_ptr<bn_bwd::primitive_desc>& pd,
MKLDNNMatrixPtr& in,
MKLDNNMatrixPtr& wgt,
MKLDNNMatrixPtr& out);
};
} // namespace paddle

@ -91,10 +91,16 @@ void MKLDNNTester::setInputImgSize() {
// init randome parameters of ref, and copy to mkldnn
void MKLDNNTester::randomWgtDatas() {
EXPECT_EQ(parameters_[DNN].size(), parameters_[REF].size());
const bool isBN = refLayer_->getType() == "batch_norm";
for (size_t i = 0; i < parameters_[REF].size(); ++i) {
const VectorPtr& dnnValue = parameters_[DNN][i]->getBuf(PARAMETER_VALUE);
const VectorPtr& refValue = parameters_[REF][i]->getBuf(PARAMETER_VALUE);
parameters_[REF][i]->randomize();
if (isBN && i == 2) {
// this param is moving average in batch norm, which must larger than 0
real offset = fabs(refValue->getMin()) + 1.0;
refValue->add(offset);
}
dnnValue->copyFrom(*refValue);
VLOG(MKLDNN_TESTS) << "Random weight " << parameters_[DNN][i]->getName();
@ -132,8 +138,7 @@ void MKLDNNTester::checkForward() {
void MKLDNNTester::checkBackwardData() {
VLOG(MKLDNN_TESTS) << "Check Backward Data";
// TODO(TJ): uncomment me when batch norm ready
// const bool isBN = dnnLayer_->getType() == "mkldnn_batch_norm";
const bool isBN = refLayer_->getType() == "batch_norm";
for (size_t i = 0; i < dataLayers_[DNN].size(); ++i) {
const MatrixPtr& dnnDiff = dataLayers_[DNN][i]->getOutputGrad();
const MatrixPtr& refDiff = dataLayers_[REF][i]->getOutputGrad();
@ -144,11 +149,11 @@ void MKLDNNTester::checkBackwardData() {
double delta = compareMatrix(dnnDiff, refDiff);
EXPECT_LE(fabs(delta), eps_);
// TODO(TJ): uncomment me when batch norm ready
// if (isBN) {
// // the other two inputs in batch norm are for moving mean and var
// break;
// }
if (isBN) {
// the other two inputs in batch norm are for moving mean and var
// do not have grad to compare
break;
}
}
}
@ -308,10 +313,14 @@ double MKLDNNTester::compareVector(const VectorPtr& v1, const VectorPtr& v2) {
void MKLDNNTester::runOnce() {
// test forward
randomBotDatas();
dnnLayer_->forward(PASS_TRAIN);
refLayer_->forward(PASS_TRAIN);
dnnLayer_->forward(passType_);
refLayer_->forward(passType_);
checkForward();
if (passType_ == PASS_TEST) {
return;
}
// test backward
// simple updater
UpdateCallback updateCallback = [](Parameter* para) {
@ -343,6 +352,7 @@ void MKLDNNTester::run(const TestConfig& dnn,
size_t batchSize,
size_t inputImgH,
size_t inputImgW,
PassType passType,
bool printDetails,
size_t iter,
float epsilon) {
@ -361,6 +371,7 @@ void MKLDNNTester::run(const TestConfig& dnn,
ih_ = inputImgH;
iw_ = inputImgW;
passType_ = passType;
log_ = printDetails;
iter_ = iter;
eps_ = epsilon;

@ -62,12 +62,15 @@ protected:
float eps_;
/// input image size, default 1
size_t ih_, iw_;
/// passType, PASS_TRAIN, PASS_TEST or PASS_GC (Gradient Check pass)
PassType passType_;
public:
explicit MKLDNNTester(size_t iter = 3, float epsilon = 1e-4) {
iter_ = iter;
eps_ = epsilon;
log_ = false;
passType_ = PASS_TRAIN;
}
~MKLDNNTester() {}
@ -78,6 +81,7 @@ public:
size_t batchSize,
size_t inputImgH = 1,
size_t inputImgW = 1,
PassType passType = PASS_TRAIN,
bool printDetails = false,
size_t iter = 3,
float epsilon = 1e-4);

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save