提交 01626be9 编写于 作者: S Superjom

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

......@@ -22,9 +22,11 @@
hooks:
- id: clang-formater
- repo: https://github.com/PaddlePaddle/pre-commit-golang
sha: 16398aeccf263adaf53b2495eed0406347d76281
sha: 8337620115c25ff8333f1b1a493bd031049bd7c0
hooks:
- id: go-fmt
types: [go]
- id: gometalinter
types: [go]
- id: go-fmt
types:
- go
- id: gometalinter
types:
- go
......@@ -18,7 +18,6 @@ package main
#include <stdlib.h>
#include <string.h>
#include <stdio.h>
#define PADDLE_MASTER_OK 0
#define PADDLE_MASTER_ERROR -1
......@@ -101,6 +100,12 @@ func paddle_release_master_client(client C.paddle_master_client) {
remove(client)
}
//export paddle_start_get_records
func paddle_start_get_records(client C.paddle_master_client, pass C.int) {
c := get(client)
c.StartGetRecords(int(pass))
}
//export paddle_set_dataset
func paddle_set_dataset(client C.paddle_master_client, path **C.char, size C.int) C.int {
c := get(client)
......@@ -121,15 +126,19 @@ func paddle_set_dataset(client C.paddle_master_client, path **C.char, size C.int
// paddle_next_record gets the nexts training record.
//
// returns number of bytes of the records if success, -1 if failed.
// returns number of bytes of the records if success, -1 if failed, -2 if pass end.
//
//export paddle_next_record
func paddle_next_record(client C.paddle_master_client, record **C.uchar) C.int {
c := get(client)
r, err := c.NextRecord()
if err != nil {
// Error
// TODO: return the type of error?
// NOTE: use errors to indicate pass ends
if err.Error() == master.ErrAllTaskFailed.Error() ||
err.Error() == master.ErrNoMoreAvailable.Error() ||
err.Error() == master.ErrPassBefore.Error() {
return -2
}
*record = (*C.uchar)(nil)
return -1
}
......
......@@ -16,7 +16,6 @@ package master
import (
"os"
"sync"
"time"
"github.com/PaddlePaddle/Paddle/go/connection"
......@@ -27,9 +26,9 @@ import (
// Client is the client of the master server.
type Client struct {
conn *connection.Conn
ch chan record
initChOnce sync.Once
conn *connection.Conn
ch chan record
bufSize int
}
type record struct {
......@@ -46,11 +45,7 @@ func WithBuffer(bufSize int) func(*Client) error {
if bufSize <= 0 {
return nil
}
c.initChOnce.Do(func() {
c.ch = make(chan record, bufSize)
go c.getRecords()
})
c.bufSize = bufSize
return nil
}
}
......@@ -104,25 +99,41 @@ func NewClient(opts ...func(*Client) error) (*Client, error) {
if err != nil {
return nil, err
}
}
c.ch = make(chan record, c.bufSize)
// FIXME: connection is created asyncrosly in monitorMaster go routine,
// ensure the connection is ready for use before calling c.addClient.
time.Sleep(time.Second)
return c, nil
}
func (c *Client) getRecords() {
// StartGetRecords must be called at beginning of each pass
func (c *Client) StartGetRecords(passID int) {
go c.getRecords(passID)
}
func (c *Client) getRecords(passID int) {
for {
t, err := c.getTask()
t, err := c.getTask(passID)
if err != nil {
log.Errorf("Get task failed, sleep 3 seconds and continue, %s", err)
time.Sleep(3 * time.Second)
continue
if err.Error() == ErrPassBefore.Error() ||
err.Error() == ErrNoMoreAvailable.Error() ||
err.Error() == ErrAllTaskFailed.Error() {
c.ch <- record{nil, err}
break
}
if err.Error() == ErrPassAfter.Error() {
// wait util last pass finishes
time.Sleep(time.Second * 3)
continue
}
log.Errorf("getTask error: %s", err)
}
for _, chunk := range t.Chunks {
f, err := os.Open(chunk.Path)
if err != nil {
log.Errorln(err)
f, e := os.Open(chunk.Path)
if e != nil {
log.Errorln(e)
continue
}
......@@ -178,18 +189,21 @@ func (c *Client) monitorMaster(addrCh <-chan string) {
}
}
// SetDataset set dataset for the master server to dispatch.
// SetDataset sets dataset to dispatch for the master server.
//
// SetDataset can be call multiple times at one pass. But only the first call
// will be honored.
//
// SetDataset can be call multiple times from different nodes. But
// only the first call will be honored.
// After all tasks are done, another call of SetDataset will start another pass.
func (c *Client) SetDataset(globPaths []string) error {
return c.conn.Call("Service.SetDataset", globPaths, nil)
err := c.conn.Call("Service.SetDataset", globPaths, nil)
return err
}
// getTask gets a new task from the master server.
func (c *Client) getTask() (Task, error) {
func (c *Client) getTask(passID int) (Task, error) {
var t Task
err := c.conn.Call("Service.GetTask", 0, &t)
err := c.conn.Call("Service.GetTask", passID, &t)
return t, err
}
......@@ -208,12 +222,6 @@ func (c *Client) taskFailed(meta TaskMeta) error {
// NextRecord will block until the next record is available. It is
// thread-safe.
func (c *Client) NextRecord() ([]byte, error) {
c.initChOnce.Do(func() {
// initialize with in case WithBuffer is not used.
c.ch = make(chan record, 0)
go c.getRecords()
})
r := <-c.ch
return r.r, r.err
}
......
......@@ -54,22 +54,22 @@ func TestGetFinishTask(t *testing.T) {
panic(err)
}
go func(l net.Listener) {
s, err := NewService(&InMemStore{}, chunkPerTask, time.Second, 1)
if err != nil {
panic(err)
s, sErr := NewService(&InMemStore{}, chunkPerTask, time.Second, 1)
if sErr != nil {
panic(sErr)
}
server := rpc.NewServer()
err = server.Register(s)
if err != nil {
panic(err)
sErr = server.Register(s)
if sErr != nil {
panic(sErr)
}
mux := http.NewServeMux()
mux.Handle(rpc.DefaultRPCPath, server)
err = http.Serve(l, mux)
if err != nil {
panic(err)
sErr = http.Serve(l, mux)
if sErr != nil {
panic(sErr)
}
}(l)
......@@ -103,6 +103,7 @@ func TestGetFinishTask(t *testing.T) {
ch := make(chan string, 1)
ch <- addr
go c.monitorMaster(ch)
err = c.SetDataset([]string{path})
if err != nil {
panic(err)
......@@ -111,44 +112,47 @@ func TestGetFinishTask(t *testing.T) {
checkOnePass := func(i int) {
var tasks []Task
for idx := 0; idx < totalTask; idx++ {
task, err := c.getTask()
if err != nil {
t.Fatalf("Error: %v, pass: %d\n", err, i)
task, cErr := c.getTask(i)
if cErr != nil && cErr.Error() != ErrNoMoreAvailable.Error() && cErr.Error() != ErrPassAfter.Error() {
t.Fatalf("error: %v, pass: %d\n", cErr, i)
}
tasks = append(tasks, task)
}
_, err = c.getTask()
if err == nil {
// getting task before task finishes should return error
_, cErr := c.getTask(i)
if cErr == nil {
t.Fatalf("Should get error, pass: %d\n", i)
}
err = c.taskFinished(tasks[0].Meta.ID)
if err != nil {
t.Fatalf("Error: %v, pass: %d\n", err, i)
cErr = c.taskFinished(tasks[0].Meta.ID)
if cErr != nil {
t.Fatalf("Error: %v, pass: %d\n", cErr, i)
}
err = c.taskFailed(tasks[0].Meta)
if err != nil {
t.Fatalf("Error: %v, pass: %d\n", err, i)
// call taskFailed once won't put the task to failed queue, just ensure
// the call
cErr = c.taskFailed(tasks[0].Meta)
if cErr != nil {
t.Fatalf("Error: %v, pass: %d\n", cErr, i)
}
tasks = tasks[1:]
task, err := c.getTask()
if err != nil {
t.Fatal(err)
_, cErr = c.getTask(i)
if cErr != nil && cErr.Error() != ErrNoMoreAvailable.Error() && cErr.Error() != ErrPassAfter.Error() {
t.Fatalf("Should be ErrNoMoreAvailable or ErrPassAfter: %s", cErr)
}
tasks = append(tasks, task)
for _, task := range tasks {
err = c.taskFinished(task.Meta.ID)
if err != nil {
t.Fatalf("Error: %v, pass: %d\n", err, i)
cErr = c.taskFinished(task.Meta.ID)
if cErr != nil {
t.Fatal(cErr)
}
}
}
for i := 0; i < 10; i++ {
// init pass data
c.StartGetRecords(i)
checkOnePass(i)
}
}
......@@ -20,8 +20,10 @@ import (
"net/http"
"net/rpc"
"os"
"runtime"
"strconv"
"strings"
"sync"
"testing"
"time"
......@@ -29,6 +31,18 @@ import (
"github.com/PaddlePaddle/recordio"
)
// tool function for testing output goroutine ids
func goid() int {
var buf [64]byte
n := runtime.Stack(buf[:], false)
idField := strings.Fields(strings.TrimPrefix(string(buf[:n]), "goroutine "))[0]
id, err := strconv.Atoi(idField)
if err != nil {
panic(fmt.Sprintf("cannot get goroutine id: %v", err))
}
return id
}
func TestNextRecord(t *testing.T) {
const (
path = "/tmp/master_client_TestFull"
......@@ -45,7 +59,7 @@ func TestNextRecord(t *testing.T) {
panic(err)
}
go func(l net.Listener) {
s, err := master.NewService(&master.InMemStore{}, 10, time.Second, 1)
s, err := master.NewService(&master.InMemStore{}, 1, time.Second*60, 1)
if err != nil {
panic(err)
}
......@@ -69,7 +83,7 @@ func TestNextRecord(t *testing.T) {
panic(err)
}
w := recordio.NewWriter(f, -1, -1)
w := recordio.NewWriter(f, 1, -1)
for i := 0; i < total; i++ {
_, err = w.Write([]byte{byte(i)})
if err != nil {
......@@ -87,32 +101,49 @@ func TestNextRecord(t *testing.T) {
panic(err)
}
c, err := master.NewClient(master.WithAddr(fmt.Sprintf(":%d", p)), master.WithBuffer(10))
if err != nil {
panic(err)
}
err = c.SetDataset([]string{path})
if err != nil {
panic(err)
}
for pass := 0; pass < 50; pass++ {
received := make(map[byte]bool)
for i := 0; i < total; i++ {
r, err := c.NextRecord()
if err != nil {
t.Fatal(pass, i, "Read error:", err)
// start several client to test task fetching
var wg sync.WaitGroup
for i := 0; i < 4; i++ {
wg.Add(1)
// test for multiple concurrent clients
go func() {
defer wg.Done()
// each go-routine needs a single client connection instance
c, e := master.NewClient(master.WithAddr(fmt.Sprintf(":%d", p)), master.WithBuffer(1))
if e != nil {
t.Fatal(e)
}
if len(r) != 1 {
t.Fatal(pass, i, "Length should be 1.", r)
e = c.SetDataset([]string{path})
if e != nil {
panic(e)
}
if received[r[0]] {
t.Fatal(pass, i, "Received duplicate.", received, r)
// test for n passes
for pass := 0; pass < 10; pass++ {
c.StartGetRecords(pass)
received := make(map[byte]bool)
taskid := 0
for {
r, e := c.NextRecord()
if e != nil {
// ErrorPassAfter will wait, else break for next pass
if e.Error() == master.ErrPassBefore.Error() ||
e.Error() == master.ErrNoMoreAvailable.Error() {
break
}
t.Fatal(pass, taskid, "Read error:", e)
}
if len(r) != 1 {
t.Fatal(pass, taskid, "Length should be 1.", r)
}
if received[r[0]] {
t.Fatal(pass, taskid, "Received duplicate.", received, r)
}
taskid++
received[r[0]] = true
}
}
received[r[0]] = true
}
}()
}
wg.Wait()
}
......@@ -19,6 +19,7 @@ import (
"compress/gzip"
"encoding/gob"
"errors"
"math/rand"
"os"
"path/filepath"
"sync"
......@@ -33,6 +34,18 @@ const (
dialTimeout = 5 * time.Second
)
// ErrAllTaskFailed occur when tasks are in done or failed state.
var ErrAllTaskFailed = errors.New("all task finished")
// ErrNoMoreAvailable occur when no task in todo and yet not all done or fail.
var ErrNoMoreAvailable = errors.New("no more available task")
// ErrPassBefore client side pass number does not match with master counter.
var ErrPassBefore = errors.New("pass number smaller than master")
// ErrPassAfter client side pass number does not match with master counter.
var ErrPassAfter = errors.New("pass number larger than master")
// Store is the interface for save and load the master state.
type Store interface {
Save([]byte) error
......@@ -75,17 +88,26 @@ type Service struct {
chunksPerTask int
timeoutDur time.Duration
failureMax int
ready chan struct{}
store Store
mu sync.Mutex
initDone bool
taskQueues taskQueues
ready chan struct{}
initDone bool
mu sync.Mutex
taskQueues taskQueues
currPass int
jobTasks []taskEntry
savingTrainer string
}
func partition(chunks []Chunk, chunksPerTask int) []taskEntry {
id := 0
// generate uniq id across job using nanosecond + randint + counter
// FIXME(typhoonzero): this is a workaround, use uuid
randStart := rand.Int()
counter := 0
timestamp := time.Now().Nanosecond()
id := timestamp + randStart + counter
if chunksPerTask <= 0 {
chunksPerTask = 1
}
......@@ -95,7 +117,8 @@ func partition(chunks []Chunk, chunksPerTask int) []taskEntry {
for i, c := range chunks {
if i%chunksPerTask == 0 && len(cur.Task.Chunks) > 0 {
cur.Task.Meta.ID = id
id++
counter++
id = timestamp + randStart + counter
result = append(result, cur)
cur.Task.Chunks = nil
}
......@@ -266,19 +289,21 @@ func (s *Service) SetDataset(globPaths []string, _ *int) error {
return err
}
s.taskQueues.Todo = partition(chunks, s.chunksPerTask)
s.jobTasks = partition(chunks, s.chunksPerTask)
s.taskQueues.Todo = s.jobTasks
err = s.snapshot()
if err != nil {
log.Errorln(err)
return err
}
close(s.ready)
s.initDone = true
return nil
}
// processFailedTask retry s.failureMax times for failed task.
// return true if all task are done or failed.
func (s *Service) processFailedTask(t taskEntry, epoch int) {
if t.Task.Meta.Epoch != epoch {
// new epoch, task launched after the
......@@ -302,8 +327,9 @@ func (s *Service) processFailedTask(t taskEntry, epoch int) {
return
}
log.Warningf("Task %v failed %d times, discard.", t.Task, t.NumFailure)
log.Warningf("Task %v failed %d times, re-dispatch.", t.Task, t.NumFailure)
s.taskQueues.Todo = append(s.taskQueues.Todo, t)
return
}
func (s *Service) checkTimeoutFunc(taskID int, epoch int) func() {
......@@ -331,37 +357,30 @@ func (s *Service) logFields() log.Fields {
}
// GetTask gets a new task from the service.
func (s *Service) GetTask(_ int, task *Task) error {
// passID is the client side pass count
func (s *Service) GetTask(passID int, task *Task) error {
select {
case <-s.ready:
}
s.mu.Lock()
defer s.mu.Unlock()
if passID < s.currPass {
return ErrPassBefore
}
if passID > s.currPass {
// Client may get run to pass after master when one client faster than the
// other
return ErrPassAfter
}
if len(s.taskQueues.Todo) == 0 {
if len(s.taskQueues.Done) == 0 {
if len(s.taskQueues.Pending) == 0 {
err := errors.New("all task failed")
log.WithFields(s.logFields()).Warningln("All tasks failed.")
return err
}
// TODO(helin): client need to retry in this
// error case. Gotcha: RPC client can't
// compare returned error with predefined
// errors like io.EOF, because the error
// instance deserialized from RPC is a
// different instance than the error defined
// in package. So we need to figure out a way
// for client to check this error correctly.
err := errors.New("no more available task")
log.WithFields(s.logFields()).Warningln("No more available task.")
return err
if len(s.taskQueues.Done) == 0 && len(s.taskQueues.Pending) == 0 {
log.WithFields(s.logFields()).Warningln("All tasks failed, may start next pass")
return ErrAllTaskFailed
}
s.taskQueues.Todo = s.taskQueues.Done
s.taskQueues.Done = nil
log.WithFields(s.logFields()).Infoln("No more todo task, but trainer is requesting task to do. Move all done task to todo.")
log.WithFields(s.logFields()).Warningln("No more available task.")
return ErrNoMoreAvailable
}
t := s.taskQueues.Todo[0]
......@@ -381,7 +400,7 @@ func (s *Service) GetTask(_ int, task *Task) error {
}
// TaskFinished tell the service that a task is finished.
func (s *Service) TaskFinished(taskID int, _ *int) error {
func (s *Service) TaskFinished(taskID int, dummy *int) error {
select {
case <-s.ready:
}
......@@ -401,11 +420,14 @@ func (s *Service) TaskFinished(taskID int, _ *int) error {
delete(s.taskQueues.Pending, taskID)
log.WithFields(s.logFields()).Infof("Task #%d finished.", taskID)
if len(s.taskQueues.Pending) == 0 && len(s.taskQueues.Todo) == 0 {
log.WithFields(s.logFields()).Infoln("No more todo and pending task, start a new pass.")
s.taskQueues.Todo = append(s.taskQueues.Todo, s.taskQueues.Done...)
s.taskQueues.Done = nil
if len(s.taskQueues.Todo) == 0 && len(s.taskQueues.Pending) == 0 {
// increase master side pass count if all tasks finished
s.currPass++
s.taskQueues.Todo = s.jobTasks
s.taskQueues.Done = []taskEntry{}
// TODO(typhoonzero): deal with failed tasks
s.taskQueues.Failed = []taskEntry{}
log.WithFields(s.logFields()).Warningf("all task finished, add new pass data, newpass: %d.", s.currPass)
}
err := s.snapshot()
......@@ -416,7 +438,7 @@ func (s *Service) TaskFinished(taskID int, _ *int) error {
}
// TaskFailed tells the service that a task is failed.
func (s *Service) TaskFailed(meta TaskMeta, _ *int) error {
func (s *Service) TaskFailed(meta TaskMeta, dummy *int) error {
select {
case <-s.ready:
}
......
......@@ -44,7 +44,8 @@ func TestPartionIndex(t *testing.T) {
cs := make([]Chunk, 100)
ts := partition(cs, 20)
for i := range ts {
if ts[i].Task.Meta.ID != i {
// test auto increament ids
if i > 0 && ts[i].Task.Meta.ID != ts[i-1].Task.Meta.ID+1 {
t.Error(ts[i], i)
}
}
......
......@@ -6,16 +6,19 @@ import cPickle as pickle
etcd_ip = os.getenv("MASTER_IP", "127.0.0.1")
etcd_endpoint = "http://" + etcd_ip + ":2379"
print "connecting to master, etcd endpoints: ", etcd_endpoint
master_client = master.client(etcd_endpoint, 5, 64)
def cloud_reader():
print "connecting to master, etcd endpoints: ", etcd_endpoint
master_client = master.client(etcd_endpoint, 5, 64)
global master_client
master_client.set_dataset(
["/pfs/dlnel/public/dataset/uci_housing/uci_housing-*-of-*"])
["/pfs/dlnel/public/dataset/uci_housing/uci_housing-*"], passes=30)
while 1:
r, e = master_client.next_record()
if not r:
if e != -2: # other errors
print "get record error:", e
break
yield pickle.loads(r)
......@@ -27,10 +30,12 @@ def main():
# network config
x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(13))
y_predict = paddle.layer.fc(input=x,
param_attr=paddle.attr.Param(name='w'),
param_attr=paddle.attr.Param(
name='w', learning_rate=1e-3),
size=1,
act=paddle.activation.Linear(),
bias_attr=paddle.attr.Param(name='b'))
bias_attr=paddle.attr.Param(
name='b', learning_rate=1e-3))
y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
cost = paddle.layer.mse_cost(input=y_predict, label=y)
......@@ -38,9 +43,8 @@ def main():
parameters = paddle.parameters.create(cost)
# create optimizer of new remote updater to pserver
optimizer = paddle.optimizer.Momentum(momentum=0)
optimizer = paddle.optimizer.Momentum(momentum=0, learning_rate=1e-3)
print "etcd endoint: ", etcd_endpoint
trainer = paddle.trainer.SGD(cost=cost,
parameters=parameters,
update_equation=optimizer,
......@@ -51,6 +55,8 @@ def main():
# event_handler to print training and testing info
def event_handler(event):
if isinstance(event, paddle.event.EndIteration):
# FIXME: for cloud data reader, pass number is managed by master
# should print the server side pass number
if event.batch_id % 100 == 0:
print "Pass %d, Batch %d, Cost %f" % (
event.pass_id, event.batch_id, event.cost)
......
......@@ -37,7 +37,7 @@ std::vector<std::string> Evaluator::getNames() const {
double Evaluator::getValue(const std::string name) const {
paddle::Error err;
double v = m->rawPtr->getValue(name, &err);
if (err) {
if (!err.isOK()) {
throw std::runtime_error(err.msg());
}
return v;
......
......@@ -3,7 +3,7 @@ cc_library(ddim SRCS ddim.cc DEPS eigen3)
cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
nv_test(dim_test SRCS dim_test.cu DEPS ddim)
cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory)
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)
......@@ -29,7 +29,5 @@ py_proto_compile(framework_py_proto SRCS attr_type.proto op_proto.proto op_desc.
add_custom_target(framework_py_proto_init ALL COMMAND ${CMAKE_COMMAND} -E touch __init__.py)
add_dependencies(framework_py_proto framework_py_proto_init)
proto_library(net_proto SRCS net_proto.proto DEPS op_proto)
# cc_library(net SRCS net.cc DEPS operator net_proto op_registry fc_op)
cc_library(net SRCS net.cc DEPS operator net_proto op_registry)
cc_library(net SRCS net.cc DEPS op_registry)
cc_test(net_op_test SRCS net_op_test.cc DEPS net add_op mul_op sigmoid_op softmax_op fc_op)
/* 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. */
#pragma once
#include "paddle/memory/memcpy.h"
namespace paddle {
namespace framework {
template <typename T>
inline void Tensor::check_memory_size() const {
PADDLE_ENFORCE(holder_ != nullptr,
"Tenosr holds no memory. Call Tensor::mutable_data first.");
PADDLE_ENFORCE(holder_->size() >= product(dims_) * sizeof(T) + offset_,
"Tensor's dims_ is out of bound. Call Tensor::mutable_data "
"first to re-allocate memory.");
}
template <typename T>
inline const T* Tensor::data() const {
check_memory_size<T>();
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}
template <typename T>
inline T* Tensor::data() {
check_memory_size<T>();
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
template <typename T>
inline T* Tensor::mutable_data(DDim dims, platform::Place place) {
static_assert(std::is_pod<T>::value, "T must be POD");
Resize(dims);
return mutable_data<T>(place);
}
template <typename T>
inline T* Tensor::mutable_data(platform::Place place) {
static_assert(std::is_pod<T>::value, "T must be POD");
PADDLE_ENFORCE(product(dims_) > 0,
"Tensor's numel must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
/* some versions of boost::variant don't have operator!= */
size_t size = product(dims_) * sizeof(T);
if (holder_ == nullptr || !(holder_->place() == place) ||
holder_->size() < size + offset_) {
if (platform::is_cpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), size));
}
#ifndef PADDLE_ONLY_CPU
else if (platform::is_gpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
boost::get<platform::GPUPlace>(place), size));
}
#endif
offset_ = 0;
}
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
template <typename T>
inline void Tensor::ShareDataWith(const Tensor& src) {
src.check_memory_size<T>();
*this = src;
}
template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
const platform::CPUDeviceContext& ctx) {
src.check_memory_size<T>();
Resize(src.dims());
auto src_place = src.holder_->place();
auto src_ptr = static_cast<const void*>(src.data<T>());
auto dst_place = ctx.GetPlace();
auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
auto size = product(src.dims_) * sizeof(T);
if (platform::is_cpu_place(src_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size);
}
#ifndef PADDLE_ONLY_CPU
else if (platform::is_gpu_place(src_place)) {
memory::Copy(boost::get<platform::CPUPlace>(dst_place), dst_ptr,
boost::get<platform::GPUPlace>(src_place), src_ptr, size, 0);
}
#endif
}
#ifndef PADDLE_ONLY_CPU
template <typename T>
inline void Tensor::CopyFrom(const Tensor& src,
const platform::CUDADeviceContext& ctx) {
src.check_memory_size<T>();
Resize(src.dims());
auto src_place = src.holder_->place();
auto src_ptr = static_cast<const void*>(src.data<T>());
auto dst_place = ctx.GetPlace();
auto dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
auto size = product(src.dims_) * sizeof(T);
if (platform::is_cpu_place(src_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
boost::get<platform::CPUPlace>(src_place), src_ptr, size,
ctx.stream());
} else if (platform::is_gpu_place(src_place)) {
memory::Copy(boost::get<platform::GPUPlace>(dst_place), dst_ptr,
boost::get<platform::GPUPlace>(src_place), src_ptr, size,
ctx.stream());
}
}
#endif
template <typename T>
inline Tensor Tensor::Slice(const int& begin_idx, const int& end_idx) const {
check_memory_size<T>();
PADDLE_ENFORCE(begin_idx >= 0, "Slice begin index is less than zero.");
PADDLE_ENFORCE(end_idx <= dims_[0], "Slice end index is out of bound.");
PADDLE_ENFORCE(begin_idx < end_idx,
"Begin index must be less than end index.");
PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1.");
int base = product(dims_) / dims_[0];
Tensor dst;
dst.holder_ = holder_;
DDim dst_dims = dims_;
dst_dims[0] = end_idx - begin_idx;
dst.Resize(dst_dims);
dst.offset_ = offset_ + begin_idx * base * sizeof(T);
return dst;
}
inline void Tensor::Resize(const DDim& dims) { dims_ = dims; }
inline const DDim& Tensor::dims() const { return dims_; }
} // namespace framework
} // namespace paddle
......@@ -20,17 +20,7 @@
namespace paddle {
namespace framework {
std::shared_ptr<PlainNet> AddBackwardOp(std::shared_ptr<PlainNet> ForwardOps) {
auto grad_ops = std::make_shared<PlainNet>();
for (auto& op : ForwardOps->ops_) {
auto op_grad = OpRegistry::CreateGradOp(op);
grad_ops->AddOp(op_grad);
}
grad_ops->CompleteAddOp();
return grad_ops;
}
void PlainNet::CompleteAddOp(bool calc) {
void NetOp::CompleteAddOp(bool calc) {
add_op_done_ = true;
if (!calc) return;
std::unordered_set<std::string> input_set;
......@@ -70,7 +60,7 @@ void PlainNet::CompleteAddOp(bool calc) {
attrs_["temporary_index"] = tmp_index;
}
std::string PlainNet::DebugString() const {
std::string NetOp::DebugString() const {
std::ostringstream os;
os << OperatorBase::DebugString() << std::endl;
for (auto& op : ops_) {
......@@ -82,5 +72,7 @@ std::string PlainNet::DebugString() const {
return os.str();
}
bool NetOp::IsNetOp() const { return true; }
} // namespace framework
} // namespace paddle
......@@ -37,21 +37,7 @@ namespace framework {
* This is the base class of network, all the networks should implement the APIs
* it defines.
*/
class Net : public OperatorBase {
public:
virtual void AddOp(const std::shared_ptr<OperatorBase>& op) = 0;
virtual void CompleteAddOp(bool calc) = 0;
};
using NetPtr = std::shared_ptr<Net>;
/**
* @brief a basic implementation of Net.
*
* PlainNet is a very simple Net, it create a list of operators, and run them
* sequentially following the order they added.
*/
class PlainNet : public Net {
class NetOp : public OperatorBase {
public:
/**
* Infer all the operators' input and output variables' shapes, will be called
......@@ -80,15 +66,17 @@ class PlainNet : public Net {
/**
* @brief Add an operator by ptr
*/
void AddOp(const std::shared_ptr<OperatorBase>& op) override {
void AddOp(const std::shared_ptr<OperatorBase>& op) {
PADDLE_ENFORCE(!add_op_done_, "Cannot AddOp when this network is sealed");
ops_.push_back(op);
}
void CompleteAddOp(bool calculate = true) override;
void CompleteAddOp(bool calculate = true);
std::string DebugString() const override;
bool IsNetOp() const override;
std::vector<std::shared_ptr<OperatorBase>> ops_;
private:
......@@ -100,7 +88,5 @@ class PlainNet : public Net {
}
};
std::shared_ptr<PlainNet> AddBackwardOp(std::shared_ptr<PlainNet> ForwardOps);
} // namespace framework
} // namespace paddle
......@@ -40,7 +40,7 @@ void AssertSameVectorWithoutOrder(const std::vector<T>& expected,
}
TEST(OpKernel, all) {
auto net = std::make_shared<PlainNet>();
auto net = std::make_shared<NetOp>();
ASSERT_NE(net, nullptr);
auto op1 = std::make_shared<TestOp>();
......@@ -69,30 +69,23 @@ TEST(OpKernel, all) {
net->Run(scope, dev_ctx);
ASSERT_EQ(2, infer_shape_cnt);
ASSERT_EQ(2, run_cnt);
ASSERT_THROW(net->AddOp(op2), std::runtime_error);
}
TEST(AddBackwardOp, TestGradOp) {
auto net = std::make_shared<PlainNet>();
ASSERT_NE(net, nullptr);
net->AddOp(framework::OpRegistry::CreateOp("mul", {"X", "Y"}, {"Out"}, {}));
net->AddOp(
framework::OpRegistry::CreateOp("add_two", {"X", "Y"}, {"Out"}, {}));
net->AddOp(framework::OpRegistry::CreateOp("add_two", {"X", "Y"}, {""}, {}));
auto grad_ops = AddBackwardOp(net);
for (auto& op : grad_ops->ops_) {
op->DebugString();
}
ASSERT_THROW(net->AddOp(op2), paddle::platform::EnforceNotMet);
}
// TODO(zhihong): add fc grad without registering.
// TEST(AddBackwardOp, TestNoGradOp) {
// auto net = std::make_shared<PlainNet>();
// ASSERT_NE(net, nullptr);
// net->AddOp(framework::OpRegistry::CreateOp("fc", {"X", "W", "b"}, {"Y"},
// {})); auto grad_ops = AddBackwardOp(net); for (auto& op : grad_ops->ops_) {
// op->DebugString();
// }
// }
//! TODO(yuyang18): Refine Backward Op.
// TEST(AddBackwardOp, TestGradOp) {
// auto net = std::make_shared<NetOp>();
// ASSERT_NE(net, nullptr);
// net->AddOp(framework::OpRegistry::CreateOp("mul", {"X", "Y"}, {"Out"}, {}));
// net->AddOp(
// framework::OpRegistry::CreateOp("add_two", {"X", "Y"}, {"Out"}, {}));
// net->AddOp(framework::OpRegistry::CreateOp("add_two", {"X", "Y"}, {""},
// {}));
// auto grad_ops = AddBackwardOp(net);
// for (auto& op : grad_ops->ops_) {
// op->DebugString();
// }
//}
} // namespace framework
} // namespace paddle
syntax="proto2";
package paddle.framework;
import "op_proto.proto";
message NetDesc {
// network identification
optional string name = 1;
// operator contains in network
repeated OpProto operators = 2;
// network type to run with. e.g "plainNet", "DAG"
optional string net_type = 3;
// num worker always
optional int32 num_workers = 4;
}
......@@ -403,15 +403,16 @@ class GradOpRegisterHelper {
STATIC_ASSERT_GLOBAL_NAMESPACE( \
__reg_op_kernel_##type##_##DEVICE_TYPE##__, \
"REGISTER_OP_KERNEL must be in global namespace"); \
struct __op_kernel_register__##type##__ { \
__op_kernel_register__##type##__() { \
struct __op_kernel_register__##type##__##DEVICE_TYPE##__ { \
__op_kernel_register__##type##__##DEVICE_TYPE##__() { \
::paddle::framework::OperatorWithKernel::OpKernelKey key; \
key.place_ = PlaceType(); \
::paddle::framework::OperatorWithKernel::AllOpKernels()[#type][key] \
.reset(new __VA_ARGS__()); \
} \
}; \
static __op_kernel_register__##type##__ __reg_kernel_##type##__; \
static __op_kernel_register__##type##__##DEVICE_TYPE##__ \
__reg_kernel_##type##__##DEVICE_TYPE##__; \
int __op_kernel_register_##type##_handle_##DEVICE_TYPE##__() { return 0; }
// (type, KernelType)
......
......@@ -90,7 +90,7 @@ TEST(OpRegistry, IllegalAttr) {
bool caught = false;
try {
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (std::runtime_error& err) {
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg = "larger_than check fail";
const char* err_msg = err.what();
......@@ -136,7 +136,7 @@ TEST(OpRegistry, CustomChecker) {
bool caught = false;
try {
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (std::runtime_error& err) {
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg = "Attribute 'test_attr' is required!";
const char* err_msg = err.what();
......@@ -154,7 +154,7 @@ TEST(OpRegistry, CustomChecker) {
caught = false;
try {
paddle::framework::OpRegistry::CreateOp(op_desc);
} catch (std::runtime_error& err) {
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg = "'test_attr' must be even!";
const char* err_msg = err.what();
......@@ -192,7 +192,7 @@ TEST(ProtoMaker, DuplicatedAttr) {
pd::OpProto op_proto;
pd::OpAttrChecker op_checker;
auto proto_maker = TestAttrProtoMaker(&op_proto, &op_checker);
ASSERT_THROW(proto_maker.Validate(), std::runtime_error);
ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
}
class TestInOutProtoMaker : public pd::OpProtoAndCheckerMaker {
......@@ -208,5 +208,5 @@ TEST(ProtoMaker, DuplicatedInOut) {
pd::OpProto op_proto;
pd::OpAttrChecker op_checker;
auto proto_maker = TestInOutProtoMaker(&op_proto, &op_checker);
ASSERT_THROW(proto_maker.Validate(), std::runtime_error);
ASSERT_THROW(proto_maker.Validate(), paddle::platform::EnforceNotMet);
}
......@@ -90,15 +90,17 @@ class OperatorBase {
virtual void Run(const std::shared_ptr<Scope>& scope,
const platform::DeviceContext& dev_ctx) const = 0;
// Get a input with argument's name described in `op_proto`
virtual bool IsNetOp() const { return false; }
//! Get a input with argument's name described in `op_proto`
const std::string& Input(const std::string& name) const;
// Get a input which has multiple variables.
// TODO add a vector_view to prevent memory copy.
//! Get a input which has multiple variables.
//! TODO add a vector_view to prevent memory copy.
std::vector<std::string> Inputs(const std::string& name) const;
// Get a output with argument's name described in `op_proto`
//! Get a output with argument's name described in `op_proto`
const std::string& Output(const std::string& name) const;
// Get an output which has multiple variables.
// TODO add a vector_view to prevent memory copy.
//! Get an output which has multiple variables.
//! TODO add a vector_view to prevent memory copy.
std::vector<std::string> Outputs(const std::string& name) const;
public:
......@@ -199,7 +201,9 @@ class OperatorWithKernel : public OperatorBase {
place_ = dev_ctx.GetPlace();
}
bool operator==(const OpKernelKey& o) const { return place_ == o.place_; }
bool operator==(const OpKernelKey& o) const {
return platform::places_are_same_class(place_, o.place_);
}
};
struct OpKernelHash {
......
......@@ -12,7 +12,7 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include <paddle/framework/tensor.h>
#include "paddle/framework/tensor.h"
namespace paddle {
namespace framework {}
......
......@@ -20,6 +20,7 @@ limitations under the License. */
#include <typeindex>
#include "paddle/framework/ddim.h"
#include "paddle/memory/memory.h"
#include "paddle/platform/device_context.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "unsupported/Eigen/CXX11/Tensor"
......@@ -31,9 +32,11 @@ template <bool less, size_t i, typename... args>
struct CastToPyBufferImpl;
} // namespace details
} // namespace pybind
namespace framework {
class Tensor {
public:
template <bool less, size_t i, typename... args>
friend struct paddle::pybind::details::CastToPyBufferImpl;
......@@ -46,143 +49,129 @@ class Tensor {
public:
Tensor() : offset_(0) {}
/*! Return a pointer to mutable memory block. */
template <typename T>
const T* data() const {
EnforceSufficientMemory<T>();
return reinterpret_cast<const T*>(
reinterpret_cast<uintptr_t>(holder_->ptr()) + offset_);
}
inline T* data();
/*! Return a pointer to constant memory block. */
template <typename T>
T* data() {
EnforceSufficientMemory<T>();
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
template <typename T, // must be POD types
typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
T* mutable_data(DDim dims, platform::Place place) {
Resize(dims);
return mutable_data<T>(place);
}
template <typename T, // must be POD types
typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
T* mutable_data(platform::Place place) {
PADDLE_ENFORCE(product(dims_) > 0,
"Tensor's numel must be larger than zero to call "
"Tensor::mutable_data. Call Tensor::set_dim first.");
if (holder_ == nullptr ||
!(holder_->place() ==
place) /* some versions of boost::variant don't have operator!= */
|| holder_->size() < product(dims_) * sizeof(T) + offset_) {
if (platform::is_cpu_place(place)) {
holder_.reset(new PlaceholderImpl<T, platform::CPUPlace>(
boost::get<platform::CPUPlace>(place), product(dims_) * sizeof(T)));
} else if (platform::is_gpu_place(place)) {
#ifdef PADDLE_ONLY_CPU
PADDLE_THROW("'GPUPlace' is not supported in CPU only device.");
#else
holder_.reset(new PlaceholderImpl<T, platform::GPUPlace>(
boost::get<platform::GPUPlace>(place), product(dims_) * sizeof(T)));
#endif
} else {
PADDLE_THROW("Unknown 'place'.");
}
offset_ = 0;
}
return reinterpret_cast<T*>(reinterpret_cast<uintptr_t>(holder_->ptr()) +
offset_);
}
inline const T* data() const;
/**
* @brief Return a pointer to mutable memory block.
* @note If not exist, then allocation.
*/
template <typename T>
void ShareDataWith(const Tensor& src) {
src.EnforceSufficientMemory<T>();
*this = src;
}
inline T* mutable_data(platform::Place place);
/**
* @brief Return a pointer to mutable memory block.
*
* @param[in] dims The dimensions of the memory block.
* @param[in] place The place of the memory block.
*
* @note If not exist, then allocation.
*/
template <typename T>
inline T* mutable_data(DDim dims, platform::Place place);
/*! Return the dimensions of the memory block. */
inline const DDim& dims() const;
/*! Resize the dimensions of the memory block. */
inline void Resize(const DDim& dims);
/*! The internal of two tensors share the same memory block. */
template <typename T>
void CopyFrom(const Tensor& src, platform::Place dst_place) {
PADDLE_ENFORCE(platform::is_cpu_place(src.holder_->place()) &&
platform::is_cpu_place(dst_place),
"Tensor::CopyFrom only support CPU now.");
src.EnforceSufficientMemory<T>();
size_t size = product(src.dims_) * sizeof(T);
Resize(src.dims());
const void* src_ptr = static_cast<const void*>(src.data<T>());
void* dst_ptr = static_cast<void*>(mutable_data<T>(dst_place));
memcpy(dst_ptr, src_ptr, size);
}
inline void ShareDataWith(const Tensor& src);
/**
* @brief Copy the content of external tensor to a new place.
*
* @param[in] src The external tensor.
* @param[in] ctx The device context contains place where to store.
*
* @note CopyFrom supports CPU <-> GPU, GPU <-> GPU.
*/
template <typename T>
inline void CopyFrom(const Tensor& src,
const platform::CPUDeviceContext& ctx);
#ifndef PADDLE_ONLY_CPU
template <typename T>
inline void CopyFrom(const Tensor& src,
const platform::CUDADeviceContext& ctx);
#endif
/**
* @brief Return the slice of the tensor.
*
* @param[in] begin_idx The begin index of the slice.
* @param[in] end_idx The end index of the slice.
*/
template <typename T>
Tensor Slice(const int& begin_idx, const int& end_idx) const {
EnforceSufficientMemory<T>();
PADDLE_ENFORCE(begin_idx >= 0, "Slice begin index is less than zero.");
PADDLE_ENFORCE(end_idx <= dims_[0], "Slice end index is out of bound.");
PADDLE_ENFORCE(begin_idx < end_idx,
"Begin index must be less than end index.");
PADDLE_ENFORCE(dims_[0] != 1, "Can not slice a tensor with dims_[0] = 1.");
int base = product(dims_) / dims_[0];
Tensor dst;
dst.holder_ = holder_;
DDim dst_dims = dims_;
dst_dims[0] = end_idx - begin_idx;
dst.Resize(dst_dims);
dst.offset_ = offset_ + begin_idx * base * sizeof(T);
return dst;
}
void Resize(const DDim& dims) { dims_ = dims; }
const DDim& dims() const { return dims_; }
inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
private:
// Placeholder hides type T, so it doesn't appear as a template
// parameter of Variable.
template <typename T>
inline void check_memory_size() const;
private:
/**
* @note Placeholder hides type T, so it doesn't appear as a template
* parameter of Variable.
*/
struct Placeholder {
virtual ~Placeholder() {}
virtual void* ptr() const = 0;
virtual platform::Place place() const = 0;
virtual size_t size() const = 0;
virtual std::type_index type() const = 0;
virtual platform::Place place() const = 0;
};
template <typename T, typename PlaceType>
template <typename T, typename Place>
struct PlaceholderImpl : public Placeholder {
PlaceholderImpl(PlaceType place, size_t size)
PlaceholderImpl(Place place, size_t size)
: ptr_(static_cast<T*>(memory::Alloc(place, size)),
memory::PODDeleter<T, PlaceType>(place)),
memory::PODDeleter<T, Place>(place)),
place_(place),
size_(size) {}
size_(size) {
PADDLE_ENFORCE(ptr_ != nullptr, "Insufficient %s memory to allocation.",
is_cpu_place(place_) ? "CPU" : "GPU");
}
virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
virtual size_t size() const { return size_; }
virtual paddle::platform::Place place() const { return place_; }
virtual platform::Place place() const { return place_; }
virtual void* ptr() const { return static_cast<void*>(ptr_.get()); }
virtual std::type_index type() const { return std::type_index(typeid(T)); }
std::unique_ptr<T, memory::PODDeleter<T, PlaceType>> ptr_;
platform::Place place_; // record the place of ptr_.
size_t size_; // size of the memory block.
/*! the pointer of memory block. */
std::unique_ptr<T, memory::PODDeleter<T, Place>> ptr_;
/*! the place of memory block. */
platform::Place place_;
/*! the size of memory block. */
size_t size_;
};
template <typename T>
inline void EnforceSufficientMemory() const {
PADDLE_ENFORCE(holder_ != nullptr,
"Tenosr holds no memory. Call Tensor::mutable_data first.");
PADDLE_ENFORCE(holder_->size() >= product(dims_) * sizeof(T) + offset_,
"Tensor's dims_ is out of bound. Call Tensor::mutable_data "
"first to re-allocate memory.");
}
std::shared_ptr<Placeholder> holder_; // holds the memory block if allocated.
/*! holds the memory block if allocated. */
std::shared_ptr<Placeholder> holder_;
/*! points to dimensions of memory block. */
DDim dims_;
// A PlaceHolder may be shared by more than one tensor. Some of them may be
// slices of the others. So the offset_ is introduced here to indicate the
// byte offset between PlaceHolder::ptr_ and where tensor's data really
// begins.
/**
* @brief A PlaceHolder may be shared by more than one tensor.
*
* @note Some of them may be slices of the others. So the offset_
* is introduced here to indicate the byte offset between
* PlaceHolder::ptr_ and where the tensor data really begins.
*/
size_t offset_;
};
} // namespace framework
} // namespace paddle
#include "paddle/framework/detail/tensor-inl.h"
......@@ -33,7 +33,7 @@ TEST(Tensor, DataAssert) {
bool caught = false;
try {
src_tensor.data<double>();
} catch (std::runtime_error& err) {
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg =
"Tenosr holds no memory. Call Tensor::mutable_data first.";
......@@ -72,7 +72,8 @@ TEST(Tensor, MutableData) {
p2 = src_tensor.mutable_data<float>(make_ddim({2, 2}), CPUPlace());
EXPECT_EQ(p1, p2);
}
#ifdef __CUDACC__
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
float* p1 = nullptr;
......@@ -107,7 +108,7 @@ TEST(Tensor, ShareDataWith) {
bool caught = false;
try {
dst_tensor.ShareDataWith<float>(src_tensor);
} catch (std::runtime_error& err) {
} catch (paddle::platform::EnforceNotMet err) {
caught = true;
std::string msg =
"Tenosr holds no memory. Call Tensor::mutable_data first.";
......@@ -123,7 +124,7 @@ TEST(Tensor, ShareDataWith) {
ASSERT_EQ(src_tensor.data<int>(), dst_tensor.data<int>());
}
#ifdef __CUDACC__
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
Tensor dst_tensor;
......@@ -160,7 +161,7 @@ TEST(Tensor, Slice) {
EXPECT_EQ(src_data_address + 3 * 4 * 1 * sizeof(int), slice_data_address);
}
#ifdef __CUDACC__
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
src_tensor.mutable_data<double>(make_ddim({6, 9}), GPUPlace());
......@@ -188,25 +189,74 @@ TEST(Tensor, Slice) {
TEST(Tensor, CopyFrom) {
using namespace paddle::framework;
using namespace paddle::platform;
{
Tensor src_tensor;
Tensor dst_tensor;
int* src_ptr = src_tensor.mutable_data<int>(make_ddim({3, 3}), CPUPlace());
int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
memcpy(src_ptr, arr, 9 * sizeof(int));
Tensor src_tensor;
int* src_ptr = src_tensor.mutable_data<int>(make_ddim({3, 3}), CPUPlace());
int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
memcpy(src_ptr, arr, 9 * sizeof(int));
Tensor dst_tensor;
dst_tensor.CopyFrom<int>(src_tensor, CPUPlace());
const int* dst_ptr = dst_tensor.data<int>();
ASSERT_NE(src_ptr, dst_ptr);
for (size_t i = 0; i < 9; ++i) {
EXPECT_EQ(src_ptr[i], dst_ptr[i]);
auto* cpu_ctx = new paddle::platform::CPUDeviceContext();
dst_tensor.CopyFrom<int>(src_tensor, *cpu_ctx);
const int* dst_ptr = dst_tensor.data<int>();
ASSERT_NE(src_ptr, dst_ptr);
for (size_t i = 0; i < 9; ++i) {
EXPECT_EQ(src_ptr[i], dst_ptr[i]);
}
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
dst_tensor.CopyFrom<int>(slice_tensor, *cpu_ctx);
const int* slice_ptr = slice_tensor.data<int>();
dst_ptr = dst_tensor.data<int>();
ASSERT_NE(dst_ptr, slice_ptr);
for (size_t i = 0; i < 3; ++i) {
EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
}
}
#ifndef PADDLE_ONLY_CPU
{
Tensor src_tensor;
Tensor gpu_tensor;
Tensor dst_tensor;
int* src_ptr = src_tensor.mutable_data<int>(make_ddim({3, 3}), CPUPlace());
int arr[9] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
memcpy(src_ptr, arr, 9 * sizeof(int));
// CPU Tensor to GPU Tensor
auto gpu_ctx = new paddle::platform::CUDADeviceContext(0);
gpu_tensor.CopyFrom<int>(src_tensor, *gpu_ctx);
// GPU Tensor to CPU Tensor
auto cpu_ctx = new paddle::platform::CPUDeviceContext();
dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_ctx);
// Compare Tensors
const int* dst_ptr = dst_tensor.data<int>();
ASSERT_NE(src_ptr, dst_ptr);
for (size_t i = 0; i < 9; ++i) {
EXPECT_EQ(src_ptr[i], dst_ptr[i]);
}
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
// CPU Slice Tensor to GPU Tensor
gpu_tensor.CopyFrom<int>(slice_tensor, *gpu_ctx);
Tensor slice_tensor = src_tensor.Slice<int>(1, 2);
dst_tensor.CopyFrom<int>(slice_tensor, CPUPlace());
const int* slice_ptr = slice_tensor.data<int>();
dst_ptr = dst_tensor.data<int>();
ASSERT_NE(dst_ptr, slice_ptr);
for (size_t i = 0; i < 3; ++i) {
EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
// GPU Tensor to CPU Tensor
dst_tensor.CopyFrom<int>(gpu_tensor, *cpu_ctx);
// Compare Slice Tensors
const int* slice_ptr = slice_tensor.data<int>();
dst_ptr = dst_tensor.data<int>();
ASSERT_NE(dst_ptr, slice_ptr);
for (size_t i = 0; i < 3; ++i) {
EXPECT_EQ(dst_ptr[i], slice_ptr[i]);
}
}
#endif
}
......@@ -207,8 +207,8 @@ Error __must_check backward(Argument& act) {
argument_.value->setData(act.value->getData() + offset, 1UL, size);
argument_.grad->setData(act.grad->getData() + offset, 1UL, size);
Error status = softmax_.backward(argument_);
if (!status) return status;
Error err = softmax_.backward(argument_);
if (!err.isOK()) return err;
}
return Error();
}
......
add_subdirectory(detail)
cc_library(memory SRCS memory.cc)
cc_library(memcpy SRCS memcpy.cc DEPS device_context)
cc_library(memcpy SRCS memcpy.cc)
cc_library(paddle_memory
DEPS
......
......@@ -27,12 +27,11 @@ BuddyAllocator::BuddyAllocator(SystemAllocator* system_allocator,
system_allocator_(std::move(system_allocator)) {}
BuddyAllocator::~BuddyAllocator() {
DLOG(INFO) << "BuddyAllocator Disconstructor makes sure that all of these "
"have actually been freed";
VLOG(3) << "BuddyAllocator Disconstructor makes sure that all of these "
"have actually been freed";
while (!pool_.empty()) {
auto block = static_cast<MemoryBlock*>(std::get<2>(*pool_.begin()));
DLOG(INFO) << "Free from block (" << block << ", " << max_chunk_size_
<< ")";
VLOG(3) << "Free from block (" << block << ", " << max_chunk_size_ << ")";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
......@@ -52,12 +51,11 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
// acquire the allocator lock
std::lock_guard<std::mutex> lock(mutex_);
DLOG(INFO) << "Allocate " << unaligned_size << " bytes from chunk size "
<< size;
VLOG(3) << "Allocate " << unaligned_size << " bytes from chunk size " << size;
// if the allocation is huge, send directly to the system allocator
if (size > max_chunk_size_) {
DLOG(INFO) << "Allocate from system allocator.";
VLOG(3) << "Allocate from system allocator.";
return SystemAlloc(size);
}
......@@ -72,9 +70,9 @@ void* BuddyAllocator::Alloc(size_t unaligned_size) {
return nullptr;
}
} else {
DLOG(INFO) << "Allocation from existing memory block " << std::get<2>(*it)
<< " at address "
<< reinterpret_cast<MemoryBlock*>(std::get<2>(*it))->data();
VLOG(3) << "Allocation from existing memory block " << std::get<2>(*it)
<< " at address "
<< reinterpret_cast<MemoryBlock*>(std::get<2>(*it))->data();
}
total_used_ += size;
......@@ -91,10 +89,10 @@ void BuddyAllocator::Free(void* p) {
// Acquire the allocator lock
std::lock_guard<std::mutex> lock(mutex_);
DLOG(INFO) << "Free from address " << block;
VLOG(3) << "Free from address " << block;
if (block->type(cache_) == MemoryBlock::HUGE_CHUNK) {
DLOG(INFO) << "Free directly from system allocator";
VLOG(3) << "Free directly from system allocator";
system_allocator_->Free(block, block->total_size(cache_),
block->index(cache_));
......@@ -111,8 +109,8 @@ void BuddyAllocator::Free(void* p) {
// Trying to merge the right buddy
if (block->has_right_buddy(cache_)) {
DLOG(INFO) << "Merging this block " << block << " with its right buddy "
<< block->right_buddy(cache_);
VLOG(3) << "Merging this block " << block << " with its right buddy "
<< block->right_buddy(cache_);
auto right_buddy = block->right_buddy(cache_);
......@@ -129,8 +127,8 @@ void BuddyAllocator::Free(void* p) {
// Trying to merge the left buddy
if (block->has_left_buddy(cache_)) {
DLOG(INFO) << "Merging this block " << block << " with its left buddy "
<< block->left_buddy(cache_);
VLOG(3) << "Merging this block " << block << " with its left buddy "
<< block->left_buddy(cache_);
auto left_buddy = block->left_buddy(cache_);
......@@ -146,8 +144,8 @@ void BuddyAllocator::Free(void* p) {
}
// Dumping this block into pool
DLOG(INFO) << "Inserting free block (" << block << ", "
<< block->total_size(cache_) << ")";
VLOG(3) << "Inserting free block (" << block << ", "
<< block->total_size(cache_) << ")";
pool_.insert(
IndexSizeAddress(block->index(cache_), block->total_size(cache_), block));
......@@ -166,7 +164,7 @@ void* BuddyAllocator::SystemAlloc(size_t size) {
size_t index = 0;
void* p = system_allocator_->Alloc(index, size);
DLOG(INFO) << "Allocated " << p << " from system allocator.";
VLOG(3) << "Allocated " << p << " from system allocator.";
if (p == nullptr) return nullptr;
......@@ -192,8 +190,8 @@ BuddyAllocator::PoolSet::iterator BuddyAllocator::RefillPool() {
if (p == nullptr) return pool_.end();
DLOG(INFO) << "Creating and inserting new block " << p
<< " from system allocator";
VLOG(3) << "Creating and inserting new block " << p
<< " from system allocator";
static_cast<MemoryBlock*>(p)->init(cache_, MemoryBlock::FREE_CHUNK, index,
max_chunk_size_, nullptr, nullptr);
......@@ -237,19 +235,19 @@ void* BuddyAllocator::SplitToAlloc(BuddyAllocator::PoolSet::iterator it,
auto block = static_cast<MemoryBlock*>(std::get<2>(*it));
pool_.erase(it);
DLOG(INFO) << "Split block (" << block << ", " << block->total_size(cache_)
<< ") into";
VLOG(3) << "Split block (" << block << ", " << block->total_size(cache_)
<< ") into";
block->split(cache_, size);
DLOG(INFO) << "Left block (" << block << ", " << block->total_size(cache_)
<< ")";
VLOG(3) << "Left block (" << block << ", " << block->total_size(cache_)
<< ")";
block->set_type(cache_, MemoryBlock::ARENA_CHUNK);
// the rest of memory if exist
if (block->has_right_buddy(cache_)) {
if (block->right_buddy(cache_)->type(cache_) == MemoryBlock::FREE_CHUNK) {
DLOG(INFO) << "Insert right block (" << block->right_buddy(cache_) << ", "
<< block->right_buddy(cache_)->total_size(cache_) << ")";
VLOG(3) << "Insert right block (" << block->right_buddy(cache_) << ", "
<< block->right_buddy(cache_)->total_size(cache_) << ")";
pool_.insert(
IndexSizeAddress(block->right_buddy(cache_)->index(cache_),
......@@ -276,7 +274,7 @@ void BuddyAllocator::CleanIdleFallBackAlloc() {
return;
}
DLOG(INFO) << "Return block " << block << " to fallback allocator.";
VLOG(3) << "Return block " << block << " to fallback allocator.";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
......@@ -312,7 +310,7 @@ void BuddyAllocator::CleanIdleNormalAlloc() {
MemoryBlock* block = static_cast<MemoryBlock*>(std::get<2>(*pool));
DLOG(INFO) << "Return block " << block << " to base allocator.";
VLOG(3) << "Return block " << block << " to base allocator.";
system_allocator_->Free(block, max_chunk_size_, block->index(cache_));
cache_.invalidate(block);
......
......@@ -35,7 +35,7 @@ void Copy<platform::CPUPlace, platform::GPUPlace>(platform::CPUPlace dst_place,
platform::GPUPlace src_place,
const void* src, size_t num,
cudaStream_t stream) {
platform::GPUPlaceGuard g(src_place.device);
platform::SetDeviceId(src_place.device);
platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToHost, stream);
}
......@@ -45,7 +45,7 @@ void Copy<platform::GPUPlace, platform::CPUPlace>(platform::GPUPlace dst_place,
platform::CPUPlace src_place,
const void* src, size_t num,
cudaStream_t stream) {
platform::GPUPlaceGuard g(dst_place.device);
platform::SetDeviceId(dst_place.device);
platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyHostToDevice, stream);
}
......@@ -56,7 +56,7 @@ void Copy<platform::GPUPlace, platform::GPUPlace>(platform::GPUPlace dst_place,
const void* src, size_t num,
cudaStream_t stream) {
if (dst_place == src_place) {
platform::GPUPlaceGuard g(src_place.device);
platform::SetDeviceId(src_place.device);
platform::GpuMemcpyAsync(dst, src, num, cudaMemcpyDeviceToDevice, stream);
} else {
platform::GpuMemcpyPeer(dst, dst_place.device, src, src_place.device, num,
......
......@@ -20,13 +20,39 @@ limitations under the License. */
namespace paddle {
namespace memory {
/**
* \brief Copy memory from one place to another place.
*
* \param[in] DstPlace Destination allocation place (CPU).
* \param[in] dst Destination memory address.
* \param[in] SrcPlace Source allocation place (CPU).
* \param[in] src Source memory address.
* \param[in] num memory size in bytes to copy.
*
*/
template <typename DstPlace, typename SrcPlace>
void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num);
#ifndef PADDLE_ONLY_CPU
/**
* \brief Copy memory from one place to another place.
*
* \param[in] DstPlace Destination allocation place (CPU or GPU).
* \param[in] dst Destination memory address.
* \param[in] SrcPlace Source allocation place (CPU or GPU).
* \param[in] src Source memory address.
* \param[in] num memory size in bytes to copy.
* \param[in] stream CUDA stream.
*
* \note For GPU memory copy, CUDA stream need to be specified
* for asynchronously memory copy.
*
*/
template <typename DstPlace, typename SrcPlace>
void Copy(DstPlace, void* dst, SrcPlace, const void* src, size_t num,
cudaStream_t stream);
#endif // PADDLE_ONLY_CPU
} // namespace memory
......
......@@ -60,6 +60,7 @@ detail::BuddyAllocator* GetGPUBuddyAllocator(int gpu_id) {
platform::GpuMaxChunkSize());
}
}
platform::SetDeviceId(gpu_id);
return as[gpu_id];
}
......
......@@ -20,19 +20,53 @@ limitations under the License. */
namespace paddle {
namespace memory {
/**
* \brief Allocate memory block in one place.
*
* \param[in] place Allocation place (CPU or GPU).
* \param[in] size Allocation size.
*
* \return Allocated memory block address.
*
* \note If return nullptr, it indicates memory allocation failed
* because insufficient memory in current system. When Alloc
* function is invoked, you must check the returned memory
* address is valid or not.
*/
template <typename Place>
void* Alloc(Place, size_t);
void* Alloc(Place place, size_t size);
/**
* \brief Free memory block in one place.
*
* \param[in] place Allocation place (CPU or GPU).
* \param[in] ptr Memory block address to free.
*
*/
template <typename Place>
void Free(Place, void*);
void Free(Place place, void* ptr);
/**
* \brief Total size of used memory in one place.
*
* \param[in] place Allocation place (CPU or GPU).
*
*/
template <typename Place>
size_t Used(Place);
size_t Used(Place place);
template <typename T, /* must be POD types */
typename Place /* platform::GPUPlace or platform::CPUPlace */,
typename std::enable_if<std::is_pod<T>::value>::type* = nullptr>
/**
* \brief Free memory block in one place.
*
* \note In some cases, custom deleter is used to
* deallocate the memory automatically for
* std::unique_ptr<T> in tensor.h.
*
*/
template <typename T, typename Place>
class PODDeleter {
static_assert(std::is_pod<T>::value, "T must be POD");
public:
PODDeleter(Place place) : place_(place) {}
void operator()(T* ptr) { Free(place_, static_cast<void*>(ptr)); }
......
......@@ -13,17 +13,14 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/add_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor.h"
namespace paddle {
namespace operators {
class AddOp : public framework::OperatorWithKernel {
class AddOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 2, "Input size of AddOp must be two");
PADDLE_ENFORCE(outputs.size() == 1, "Output size of AddOp must be one");
PADDLE_ENFORCE(
......@@ -35,10 +32,10 @@ protected:
}
};
class AddOpMaker : public framework::OpProtoAndCheckerMaker {
class AddOpMaker : public OpProtoAndCheckerMaker {
public:
AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of add op");
AddInput("Y", "The second input of add op");
AddOutput("Out", "The output of add op");
......@@ -50,11 +47,10 @@ The equation is: Out = X + Y
}
};
class AddOpGrad : public framework::OperatorWithKernel {
class AddOpGrad : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {}
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {}
std::string DebugString() const override {
LOG(INFO) << "AddOpGrad";
return "";
......@@ -64,7 +60,6 @@ protected:
} // namespace operators
} // namespace paddle
REGISTER_OP(add_two, paddle::operators::AddOp, paddle::operators::AddOpMaker);
REGISTER_GRADIENT_OP(add_two, add_two_grad, paddle::operators::AddOpGrad);
REGISTER_OP_CPU_KERNEL(
add_two, paddle::operators::AddKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP(add_two, ops::AddOp, ops::AddOpMaker);
REGISTER_GRADIENT_OP(add_two, add_two_grad, ops::AddOpGrad);
REGISTER_OP_CPU_KERNEL(add_two, ops::AddKernel<ops::CPUPlace, float>);
#include "paddle/operators/add_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/add_op.h"
REGISTER_OP_GPU_KERNEL(add_two,
paddle::operators::AddKernel<paddle::platform::GPUPlace, float>);
\ No newline at end of file
REGISTER_OP_GPU_KERNEL(add_two, ops::AddKernel<ops::GPUPlace, float>);
......@@ -13,27 +13,24 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "glog/logging.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class AddKernel : public framework::OpKernel {
class AddKernel : public OpKernel {
public:
void Compute(const framework::KernelContext& context) const override {
auto input0 = context.Input(0)->Get<framework::Tensor>();
auto input1 = context.Input(1)->Get<framework::Tensor>();
auto* output = context.Output(0)->GetMutable<framework::Tensor>();
void Compute(const KernelContext& context) const override {
auto input0 = context.Input(0)->Get<Tensor>();
auto input1 = context.Input(1)->Get<Tensor>();
auto output = context.Output(0)->GetMutable<Tensor>();
output->mutable_data<T>(context.GetPlace());
framework::EigenVector<T>::Flatten(*output).device(
EigenVector<T>::Flatten(*output).device(
*(context.GetEigenDevice<Place>())) =
framework::EigenVector<T>::Flatten(input0) +
framework::EigenVector<T>::Flatten(input1);
EigenVector<T>::Flatten(input0) + EigenVector<T>::Flatten(input1);
}
};
......
......@@ -13,17 +13,14 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/cross_entropy_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor.h"
namespace paddle {
namespace operators {
class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
class OnehotCrossEntropyOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 2,
"Input size of OnehotCrossEntropyOp must be two");
PADDLE_ENFORCE(outputs.size() == 1,
......@@ -35,15 +32,14 @@ protected:
PADDLE_ENFORCE(inputs[0]->dims().size() == 2, "X's dimension must be 2.");
PADDLE_ENFORCE(outputs[0]->dims().size() == 1,
"label's dimension must be 1.");
outputs[0]->Resize(framework::make_ddim({inputs[0]->dims()[0]}));
outputs[0]->Resize({inputs[0]->dims()[0]});
}
};
class OnehotCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
class OnehotCrossEntropyOpMaker : public OpProtoAndCheckerMaker {
public:
OnehotCrossEntropyOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
OnehotCrossEntropyOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of OnehotCrossEntropyOp");
AddInput("label", "The second input of OnehotCrossEntropyOp");
AddOutput("Y", "The output of OnehotCrossEntropyOp");
......@@ -59,9 +55,7 @@ OnehotCrossEntropy Operator.
} // namespace paddle
REGISTER_OP(onehot_cross_entropy,
paddle::operators::OnehotCrossEntropyOp,
paddle::operators::OnehotCrossEntropyOpMaker);
REGISTER_OP_CPU_KERNEL(
onehot_cross_entropy,
paddle::operators::OnehotCrossEntropyOpKernel<::paddle::platform::CPUPlace,
float>);
ops::OnehotCrossEntropyOp,
ops::OnehotCrossEntropyOpMaker);
REGISTER_OP_CPU_KERNEL(onehot_cross_entropy,
ops::OnehotCrossEntropyOpKernel<ops::CPUPlace, float>);
#include "paddle/operators/cross_entropy_op.h"
#include "paddle/framework/op_registry.h"
REGISTER_OP_GPU_KERNEL(onehot_cross_entropy,
paddle::operators::OnehotCrossEntropyOpKernel<
::paddle::platform::GPUPlace, float>);
\ No newline at end of file
ops::OnehotCrossEntropyOpKernel<ops::GPUPlace, float>);
\ No newline at end of file
......@@ -13,23 +13,21 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "glog/logging.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class OnehotCrossEntropyOpKernel : public framework::OpKernel {
class OnehotCrossEntropyOpKernel : public OpKernel {
public:
constexpr T LOG_THRESHOLD() const { return static_cast<T>(1e-20); }
void Compute(const framework::KernelContext& context) const override {
auto X = context.Input(0)->Get<framework::Tensor>();
void Compute(const KernelContext& context) const override {
auto X = context.Input(0)->Get<Tensor>();
const T* X_data = X.data<T>();
const int* label_data =
context.Input(1)->Get<framework::Tensor>().data<int>();
auto* Y = context.Output(0)->GetMutable<framework::Tensor>();
const int* label_data = context.Input(1)->Get<Tensor>().data<int>();
auto* Y = context.Output(0)->GetMutable<Tensor>();
Y->mutable_data<T>(context.GetPlace());
......
......@@ -12,41 +12,38 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/framework/net.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "type_alias.h"
namespace paddle {
namespace operators {
class FullyConnectedOp : public framework::PlainNet {
class FullyConnectedOp : public NetOp {
public:
void Init() override {
AddOp(framework::OpRegistry::CreateOp("mul",
{
Input("X"), Input("W"),
},
{Output("before_act")},
{}));
AddOp(OpRegistry::CreateOp("mul",
{
Input("X"), Input("W"),
},
{Output("before_act")},
{}));
auto b = Input("b");
if (b != framework::OperatorBase::EMPTY_VAR_NAME()) {
AddOp(framework::OpRegistry::CreateOp("rowwise_add",
{Output("before_act"), Input("b")},
{Output("before_act")},
{}));
if (b != EMPTY_VAR_NAME()) {
AddOp(OpRegistry::CreateOp("rowwise_add",
{Output("before_act"), Input("b")},
{Output("before_act")},
{}));
}
auto activation = GetAttr<std::string>("activation");
AddOp(framework::OpRegistry::CreateOp(
AddOp(OpRegistry::CreateOp(
activation, {Output("before_act")}, {Output("Y")}, {}));
CompleteAddOp(false);
}
};
class FullyConnectedOpMaker : public framework::OpProtoAndCheckerMaker {
class FullyConnectedOpMaker : public OpProtoAndCheckerMaker {
public:
FullyConnectedOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
FullyConnectedOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "the input of fc operator");
AddInput("W", "the weight of fc operator");
......@@ -71,6 +68,4 @@ USE_OP(rowwise_add);
USE_OP(sigmoid);
USE_OP(softmax);
REGISTER_OP(fc,
paddle::operators::FullyConnectedOp,
paddle::operators::FullyConnectedOpMaker);
REGISTER_OP(fc, ops::FullyConnectedOp, ops::FullyConnectedOpMaker);
......@@ -13,17 +13,14 @@
limitations under the License. */
#include "paddle/operators/mul_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor.h"
namespace paddle {
namespace operators {
class MulOp : public framework::OperatorWithKernel {
class MulOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 2, "The mul op must take two inputs");
auto dim0 = inputs[0]->dims();
auto dim1 = inputs[1]->dims();
......@@ -37,10 +34,10 @@ protected:
}
};
class MulOpMaker : public framework::OpProtoAndCheckerMaker {
class MulOpMaker : public OpProtoAndCheckerMaker {
public:
MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of mul op");
AddInput("Y", "The second input of mul op");
AddOutput("Out", "The output of mul op");
......@@ -52,11 +49,10 @@ The equation is: Out = X * Y
}
};
class MulOpGrad : public framework::OperatorWithKernel {
class MulOpGrad : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {}
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {}
std::string DebugString() const override {
LOG(INFO) << "MulGrad";
return "";
......@@ -66,8 +62,7 @@ protected:
} // namespace operators
} // namespace paddle
REGISTER_OP(mul, paddle::operators::MulOp, paddle::operators::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, paddle::operators::MulOpGrad);
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(
mul, paddle::operators::MulKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<ops::CPUPlace, float>);
......@@ -13,8 +13,5 @@
limitations under the License. */
#include "paddle/operators/mul_op.h"
#include "paddle/framework/op_registry.h"
REGISTER_OP_GPU_KERNEL(mul,
paddle::operators::MulKernel<paddle::platform
::GPUPlace, float>);
\ No newline at end of file
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);
\ No newline at end of file
......@@ -14,30 +14,27 @@
#pragma once
#include "glog/logging.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class MulKernel : public framework::OpKernel {
class MulKernel : public OpKernel {
public:
void Compute(const framework::KernelContext& context) const override {
void Compute(const KernelContext& context) const override {
Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
{Eigen::IndexPair<Eigen::DenseIndex>(1, 0)}};
auto input0 = context.Input(0)->Get<framework::Tensor>();
auto input1 = context.Input(1)->Get<framework::Tensor>();
auto* output = context.Output(0)->GetMutable<framework::Tensor>();
auto input0 = context.Input(0)->Get<Tensor>();
auto input1 = context.Input(1)->Get<Tensor>();
auto* output = context.Output(0)->GetMutable<Tensor>();
output->mutable_data<T>(context.GetPlace());
framework::EigenMatrix<T>::From(*output).device(
*(context.GetEigenDevice<Place>())) =
framework::EigenMatrix<T>::From(input0).contract(
framework::EigenMatrix<T>::From(input1), dim_pair);
EigenMatrix<T>::From(*output).device(*(context.GetEigenDevice<Place>())) =
EigenMatrix<T>::From(input0).contract(EigenMatrix<T>::From(input1),
dim_pair);
}
};
} // namespace operators
......
......@@ -13,15 +13,13 @@
limitations under the License. */
#include "paddle/operators/rowwise_add_op.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
class RowWiseAddOp : public framework::OperatorWithKernel {
class RowWiseAddOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 2UL, "Two inputs is needed by rowwise add");
auto dim0 = inputs[0]->dims();
auto dim1 = inputs[1]->dims();
......@@ -34,11 +32,10 @@ protected:
}
};
class RowWiseAddOpMaker : public framework::OpProtoAndCheckerMaker {
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The left input of row-wise add op, must be matrix");
AddInput("b", "The right input of row-wise add op, must be vector");
AddOutput("Out", "The output of row-wise add op");
......@@ -53,9 +50,6 @@ for i in xrange(X.shape[0]):
} // namespace operators
} // namespace paddle
REGISTER_OP(rowwise_add,
paddle::operators::RowWiseAddOp,
paddle::operators::RowWiseAddOpMaker);
REGISTER_OP_CPU_KERNEL(
rowwise_add,
paddle::operators::RowWiseAddKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP(rowwise_add, ops::RowWiseAddOp, ops::RowWiseAddOpMaker);
REGISTER_OP_CPU_KERNEL(rowwise_add,
ops::RowWiseAddKernel<ops::CPUPlace, float>);
#include "paddle/framework/op_registry.h"
#include "paddle/operators/rowwise_add_op.h"
REGISTER_OP_GPU_KERNEL(
rowwise_add,
paddle::operators::RowWiseAddKernel<paddle::platform ::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(rowwise_add,
ops::RowWiseAddKernel<ops::GPUPlace, float>);
......@@ -13,25 +13,23 @@
limitations under the License. */
#pragma once
#include "glog/logging.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class RowWiseAddKernel : public framework::OpKernel {
class RowWiseAddKernel : public OpKernel {
public:
void Compute(const framework::KernelContext& context) const override {
auto in0 = context.Input(0)->Get<framework::Tensor>();
auto in1 = context.Input(1)->Get<framework::Tensor>();
auto* out = context.Output(0)->GetMutable<framework::Tensor>();
void Compute(const KernelContext& context) const override {
auto in0 = context.Input(0)->Get<Tensor>();
auto in1 = context.Input(1)->Get<Tensor>();
auto* out = context.Output(0)->GetMutable<Tensor>();
out->mutable_data<T>(context.GetPlace());
auto input = framework::EigenMatrix<T>::From(in0);
auto bias = framework::EigenVector<T>::From(in1);
auto output = framework::EigenMatrix<T>::From(*out);
auto input = EigenMatrix<T>::From(in0);
auto bias = EigenVector<T>::From(in1);
auto output = EigenMatrix<T>::From(*out);
const int bias_size = bias.dimension(0);
const int rest_size = input.size() / bias_size;
......
......@@ -13,17 +13,14 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/sgd_op.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/tensor.h"
namespace paddle {
namespace operators {
class SGDOp : public framework::OperatorWithKernel {
class SGDOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 2, "Input size of SGDOp must be two");
PADDLE_ENFORCE(outputs.size() == 1, "Output size of SGDOp must be one");
PADDLE_ENFORCE(inputs[0] != nullptr, "inputs[0] mast be set");
......@@ -35,10 +32,10 @@ protected:
}
};
class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
class SGDOpMaker : public OpProtoAndCheckerMaker {
public:
SGDOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
SGDOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("param", "input parameter");
AddInput("grad", "input gradient");
AddOutput("param_out", "output parameter");
......@@ -55,7 +52,5 @@ param_out = param - learning_rate * grad;
} // namespace operators
} // namespace paddle
REGISTER_OP(sgd, paddle::operators::SGDOp, paddle::operators::SGDOpMaker);
typedef paddle::operators::SGDOpKernel<::paddle::platform::CPUPlace, float>
SGDOpKernel_CPU_float;
REGISTER_OP_CPU_KERNEL(sgd, SGDOpKernel_CPU_float);
REGISTER_OP(sgd, ops::SGDOp, ops::SGDOpMaker);
REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel<ops::CPUPlace, float>);
#include "paddle/operators/sgd_op.h"
#include "paddle/framework/op_registry.h"
typedef paddle::operators::SGDOpKernel<::paddle::platform::GPUPlace, float> SGDOpKernel_GPU_float;
REGISTER_OP_GPU_KERNEL(sgd, SGDOpKernel_GPU_float);
\ No newline at end of file
REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);
\ No newline at end of file
......@@ -13,28 +13,24 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "glog/logging.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class SGDOpKernel : public framework::OpKernel {
class SGDOpKernel : public OpKernel {
public:
void Compute(const framework::KernelContext& ctx) const override {
auto param = ctx.Input("param")->Get<framework::Tensor>();
auto grad = ctx.Input("grad")->Get<framework::Tensor>();
auto* param_out = ctx.Output(0)->GetMutable<framework::Tensor>();
void Compute(const KernelContext& ctx) const override {
auto param = ctx.Input("param")->Get<Tensor>();
auto grad = ctx.Input("grad")->Get<Tensor>();
auto* param_out = ctx.Output(0)->GetMutable<Tensor>();
float lr = ctx.op_.GetAttr<float>("learning_rate");
param_out->mutable_data<T>(ctx.GetPlace());
framework::EigenVector<T>::Flatten(*param_out)
.device(*(ctx.GetEigenDevice<Place>())) =
framework::EigenVector<T>::Flatten(param) -
lr * framework::EigenVector<T>::Flatten(grad);
EigenVector<T>::Flatten(*param_out).device(*(ctx.GetEigenDevice<Place>())) =
EigenVector<T>::Flatten(param) - lr * EigenVector<T>::Flatten(grad);
}
};
......
......@@ -13,37 +13,33 @@
limitations under the License. */
#include "paddle/operators/sigmoid_op.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
class SigmoidOp : public framework::OperatorWithKernel {
class SigmoidOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 1, "Sigmoid Op only have one input");
PADDLE_ENFORCE(outputs.size() == 1, "Sigmoid Op only have one output");
outputs[0]->Resize(inputs[0]->dims());
}
};
class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
class SigmoidOpMaker : public OpProtoAndCheckerMaker {
public:
SigmoidOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "sigmoid input");
AddOutput("Y", "sigmoid output");
AddComment("Sigmoid function");
}
};
class SigmoidOpGrad : public framework::OperatorWithKernel {
class SigmoidOpGrad : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {}
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {}
std::string DebugString() const override {
LOG(INFO) << "SigmoidGrad";
return "";
......@@ -53,11 +49,7 @@ protected:
} // namespace operators
} // namespace paddle
REGISTER_OP(sigmoid,
paddle::operators::SigmoidOp,
paddle::operators::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, paddle::operators::SigmoidOpGrad);
REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad);
REGISTER_OP_CPU_KERNEL(
sigmoid,
paddle::operators::SigmoidKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::CPUPlace, float>);
#include "paddle/operators/sigmoid_op.h"
#include "paddle/framework/op_registry.h"
REGISTER_OP_GPU_KERNEL(
sigmoid, paddle::operators::SigmoidKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::GPUPlace, float>);
......@@ -14,25 +14,23 @@
#pragma once
#include "glog/logging.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class SigmoidKernel : public framework::OpKernel {
class SigmoidKernel : public OpKernel {
public:
void Compute(const framework::KernelContext& context) const override {
auto input = context.Input(0)->Get<framework::Tensor>();
auto* output = context.Output(0)->GetMutable<framework::Tensor>();
void Compute(const KernelContext& context) const override {
auto input = context.Input(0)->Get<Tensor>();
auto* output = context.Output(0)->GetMutable<Tensor>();
output->mutable_data<T>(context.GetPlace());
framework::EigenVector<T>::Flatten(*output).device(
EigenVector<T>::Flatten(*output).device(
*(context.GetEigenDevice<Place>())) =
1.0 / (1.0 + (-1.0 * framework::EigenVector<T>::Flatten(input)).exp());
1.0 / (1.0 + (-1.0 * EigenVector<T>::Flatten(input)).exp());
}
};
} // namespace operators
......
......@@ -12,16 +12,14 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/softmax_op.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
class SoftmaxOp : public framework::OperatorWithKernel {
class SoftmaxOp : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {
PADDLE_ENFORCE(inputs.size() == 1, "Only one input is need for softmax");
PADDLE_ENFORCE(inputs[0]->dims().size() == 2,
"The input of softmax op must be matrix");
......@@ -31,10 +29,9 @@ protected:
}
};
class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
class SoftmaxOpMaker : public OpProtoAndCheckerMaker {
public:
SoftmaxOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
SoftmaxOpMaker(OpProto *proto, OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "input of softmax");
AddOutput("Y", "output of softmax");
......@@ -42,11 +39,10 @@ public:
}
};
class SoftmaxOpGrad : public framework::OperatorWithKernel {
class SoftmaxOpGrad : public OperatorWithKernel {
protected:
void InferShape(
const std::vector<const framework::Tensor *> &inputs,
const std::vector<framework::Tensor *> &outputs) const override {}
void InferShape(const std::vector<const Tensor *> &inputs,
const std::vector<Tensor *> &outputs) const override {}
std::string DebugString() const override {
LOG(INFO) << "SoftmaxOpGrad";
return "";
......@@ -56,9 +52,6 @@ protected:
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker);
REGISTER_GRADIENT_OP(softmax, softmax_grad, paddle::operators::SoftmaxOpGrad);
REGISTER_OP_CPU_KERNEL(softmax,
ops::SoftmaxKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel<ops::CPUPlace, float>);
#include "paddle/framework/op_registry.h"
#include "paddle/operators/softmax_op.h"
REGISTER_OP_GPU_KERNEL(
softmax, paddle::operators::SoftmaxKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(softmax, ops::SoftmaxKernel<ops::GPUPlace, float>);
......@@ -14,23 +14,21 @@
#pragma once
#include "glog/logging.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
template <typename Place, typename T>
class SoftmaxKernel : public framework::OpKernel {
class SoftmaxKernel : public OpKernel {
public:
void Compute(const framework::KernelContext& context) const override {
auto input = context.Input(0)->Get<framework::Tensor>();
auto* output = context.Output(0)->GetMutable<framework::Tensor>();
void Compute(const KernelContext& context) const override {
auto input = context.Input(0)->Get<Tensor>();
auto* output = context.Output(0)->GetMutable<Tensor>();
output->mutable_data<T>(context.GetPlace());
auto logits = framework::EigenMatrix<T>::From(input);
auto softmax = framework::EigenMatrix<T>::From(*output);
auto logits = EigenMatrix<T>::From(input);
auto softmax = EigenMatrix<T>::From(*output);
const int kBatchDim = 0;
const int kClassDim = 1;
......
/* 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. */
#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/net.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using OpKernel = framework::OpKernel;
using KernelContext = framework::KernelContext;
template <typename T,
int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename T,
int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename T,
size_t D,
int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
using Tensor = framework::Tensor;
using OperatorWithKernel = framework::OperatorWithKernel;
using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
using OpProto = framework::OpProto;
using OpAttrChecker = framework::OpAttrChecker;
using CPUPlace = platform::CPUPlace;
using GPUPlace = platform::GPUPlace;
using NetOp = framework::NetOp;
using OpRegistry = framework::OpRegistry;
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
......@@ -20,12 +20,96 @@ Eigen::DefaultDevice* DeviceContext::get_eigen_device<Eigen::DefaultDevice>()
return reinterpret_cast<const CPUDeviceContext*>(this)->eigen_device();
}
CPUDeviceContext::CPUDeviceContext() {
eigen_device_.reset(new Eigen::DefaultDevice());
}
CPUDeviceContext::CPUDeviceContext(CPUPlace place) {
eigen_device_.reset(new Eigen::DefaultDevice());
}
Eigen::DefaultDevice* CPUDeviceContext::eigen_device() const {
return eigen_device_.get();
}
Place CPUDeviceContext::GetPlace() const { return CPUPlace(); }
#ifndef PADDLE_ONLY_CPU
template <>
Eigen::GpuDevice* DeviceContext::get_eigen_device<Eigen::GpuDevice>() const {
return reinterpret_cast<const CUDADeviceContext*>(this)->eigen_device();
}
#endif
CUDADeviceContext::CUDADeviceContext(GPUPlace place) : place_(place) {
SetDeviceId(place_.device);
PADDLE_ENFORCE(cudaStreamCreate(&stream_));
eigen_stream_.reset(new Eigen::CudaStreamDevice(&stream_));
eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
}
CUDADeviceContext::~CUDADeviceContext() {
SetDeviceId(place_.device);
Wait();
if (cublas_handle_) {
PADDLE_ENFORCE(dynload::cublasDestroy(cublas_handle_));
}
if (cudnn_handle_) {
PADDLE_ENFORCE(dynload::cudnnDestroy(cudnn_handle_));
}
if (curand_generator_) {
PADDLE_ENFORCE(dynload::curandDestroyGenerator(curand_generator_));
}
eigen_stream_.reset();
eigen_device_.reset();
PADDLE_ENFORCE(cudaStreamDestroy(stream_));
}
Place CUDADeviceContext::GetPlace() const { return place_; }
cudaStream_t CUDADeviceContext::stream() const { return stream_; }
void CUDADeviceContext::Wait() const {
PADDLE_ENFORCE(cudaStreamSynchronize(stream_));
}
Eigen::GpuDevice* CUDADeviceContext::eigen_device() const {
return eigen_device_.get();
}
cublasHandle_t CUDADeviceContext::cublas_handle() {
if (!cublas_handle_) {
SetDeviceId(place_.device);
PADDLE_ENFORCE(dynload::cublasCreate(&cublas_handle_));
PADDLE_ENFORCE(dynload::cublasSetStream(cublas_handle_, stream_));
}
return cublas_handle_;
}
cudnnHandle_t CUDADeviceContext::cudnn_handle() {
if (!cudnn_handle_) {
SetDeviceId(place_.device);
PADDLE_ENFORCE(dynload::cudnnCreate(&cudnn_handle_));
PADDLE_ENFORCE(dynload::cudnnSetStream(cudnn_handle_, stream_));
}
return cudnn_handle_;
}
curandGenerator_t CUDADeviceContext::curand_generator() {
if (!curand_generator_) {
SetDeviceId(place_.device);
PADDLE_ENFORCE(dynload::curandCreateGenerator(&curand_generator_,
CURAND_RNG_PSEUDO_DEFAULT));
PADDLE_ENFORCE(
dynload::curandSetPseudoRandomGeneratorSeed(curand_generator_, seed_));
PADDLE_ENFORCE(dynload::curandSetStream(curand_generator_, stream_));
}
return curand_generator_;
}
#endif // PADDLE_ONLY_CPU
} // namespace platform
} // namespace paddle
......@@ -39,14 +39,13 @@ class DeviceContext {
class CPUDeviceContext : public DeviceContext {
public:
CPUDeviceContext() { eigen_device_.reset(new Eigen::DefaultDevice()); }
CPUDeviceContext();
CPUDeviceContext(CPUPlace);
virtual ~CPUDeviceContext() {}
Eigen::DefaultDevice* eigen_device() const { return eigen_device_.get(); }
Eigen::DefaultDevice* eigen_device() const;
Place GetPlace() const override {
Place retv = CPUPlace();
return retv;
}
Place GetPlace() const override;
private:
std::unique_ptr<Eigen::DefaultDevice> eigen_device_;
......@@ -54,119 +53,51 @@ class CPUDeviceContext : public DeviceContext {
#ifndef PADDLE_ONLY_CPU
class GPUPlaceGuard {
class CUDADeviceContext : public DeviceContext {
public:
explicit GPUPlaceGuard(GPUPlace new_place) : previous_(GetCurrentDeviceId()) {
if (previous_ != new_place) {
paddle::platform::SetDeviceId(new_place.device);
}
}
explicit CUDADeviceContext(GPUPlace);
virtual ~CUDADeviceContext();
~GPUPlaceGuard() { paddle::platform::SetDeviceId(previous_.device); }
/*! \brief Wait for all operations completion in the stream. */
void Wait() const;
private:
GPUPlace previous_;
};
/*! \brief Return CUDA stream in the device context. */
cudaStream_t stream() const;
class CUDADeviceContext : public DeviceContext {
public:
explicit CUDADeviceContext(const GPUPlace gpu_place) : gpu_place_(gpu_place) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(cudaStreamCreate(&stream_), "cudaStreamCreate failed");
eigen_stream_.reset(new Eigen::CudaStreamDevice(&stream_));
eigen_device_.reset(new Eigen::GpuDevice(eigen_stream_.get()));
}
Place GetPlace() const override {
Place retv = GPUPlace();
return retv;
}
void Wait() {
PADDLE_ENFORCE(cudaStreamSynchronize(stream_),
"cudaStreamSynchronize failed");
}
cudaStream_t stream() { return stream_; }
Eigen::GpuDevice* eigen_device() const { return eigen_device_.get(); }
cublasHandle_t cublas_handle() {
if (!blas_handle_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::cublasCreate(&blas_handle_),
"cublasCreate failed");
PADDLE_ENFORCE(
paddle::platform::dynload::cublasSetStream(blas_handle_, stream_),
"cublasSetStream failed");
}
return blas_handle_;
}
cudnnHandle_t cudnn_handle() {
if (!dnn_handle_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::cudnnCreate(&dnn_handle_),
"cudnnCreate failed");
PADDLE_ENFORCE(
paddle::platform::dynload::cudnnSetStream(dnn_handle_, stream_),
"cudnnSetStream failed");
}
return dnn_handle_;
}
curandGenerator_t curand_generator() {
if (!rand_generator_) {
GPUPlaceGuard guard(gpu_place_);
PADDLE_ENFORCE(paddle::platform::dynload::curandCreateGenerator(
&rand_generator_, CURAND_RNG_PSEUDO_DEFAULT),
"curandCreateGenerator failed");
PADDLE_ENFORCE(
paddle::platform::dynload::curandSetPseudoRandomGeneratorSeed(
rand_generator_, random_seed_),
"curandSetPseudoRandomGeneratorSeed failed");
PADDLE_ENFORCE(
paddle::platform::dynload::curandSetStream(rand_generator_, stream_),
"curandSetStream failed");
}
return rand_generator_;
}
~CUDADeviceContext() {
Wait();
if (blas_handle_) {
PADDLE_ENFORCE(paddle::platform::dynload::cublasDestroy(blas_handle_),
"cublasDestroy failed");
}
if (dnn_handle_) {
PADDLE_ENFORCE(paddle::platform::dynload::cudnnDestroy(dnn_handle_),
"cudnnDestroy failed");
}
if (rand_generator_) {
PADDLE_ENFORCE(
paddle::platform::dynload::curandDestroyGenerator(rand_generator_),
"curandDestroyGenerator failed");
}
eigen_stream_.reset();
eigen_device_.reset();
PADDLE_ENFORCE(cudaStreamDestroy(stream_), "cudaStreamDestroy failed");
}
/*! \brief Return place in the device context. */
Place GetPlace() const override;
/*! \brief Return eigen device in the device context. */
Eigen::GpuDevice* eigen_device() const;
// clang-format off
/*! \brief Return cublas handle in the device context. */
cublasHandle_t cublas_handle ();
/*! \brief Return cudnn handle in the device context. */
cudnnHandle_t cudnn_handle ();
/*! \brief Return curand handle in the device context. */
curandGenerator_t curand_generator();
// clang-format on
private:
GPUPlace gpu_place_;
cudaStream_t stream_;
GPUPlace place_;
std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
private:
std::unique_ptr<Eigen::GpuDevice> eigen_device_;
std::unique_ptr<Eigen::CudaStreamDevice> eigen_stream_;
cublasHandle_t blas_handle_{nullptr};
private:
uint64_t seed_;
cudnnHandle_t dnn_handle_{nullptr};
cudaStream_t stream_;
int random_seed_;
curandGenerator_t rand_generator_{nullptr};
// clang-format off
cudnnHandle_t cudnn_handle_ = nullptr;
cublasHandle_t cublas_handle_ = nullptr;
curandGenerator_t curand_generator_ = nullptr;
// clang-format on
};
#endif
......
......@@ -36,6 +36,21 @@ limitations under the License. */
namespace paddle {
namespace platform {
struct EnforceNotMet : public std::exception {
std::exception_ptr exp_;
std::string err_str_;
EnforceNotMet(std::exception_ptr e, const char* f, int l) : exp_(e) {
try {
std::rethrow_exception(exp_);
} catch (const std::exception& exp) {
err_str_ = string::Sprintf("%s at [%s:%d]", exp.what(), f, l);
}
}
const char* what() const noexcept { return err_str_.c_str(); }
};
// Because most enforce conditions would evaluate to true, we can use
// __builtin_expect to instruct the C++ compiler to generate code that
// always forces branch prediction of true.
......@@ -43,18 +58,11 @@ namespace platform {
// For more details, please check https://stackoverflow.com/a/43870188/724872.
#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
template <typename T>
inline void throw_on_error(T e) {
throw_on_error(e, "");
}
template <typename... Args>
inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
int stat, const Args&... args) {
if (UNLIKELY(!(stat))) {
throw std::runtime_error(
string::Sprintf(args...) +
string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
throw std::runtime_error(string::Sprintf(args...));
}
}
......@@ -64,12 +72,8 @@ template <typename... Args>
inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
cudaError_t e, const Args&... args) {
if (UNLIKELY(e)) {
// clang-format off
throw thrust::system_error(
e, thrust::cuda_category(),
string::Sprintf(args...) +
string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
// clang-format on
throw thrust::system_error(e, thrust::cuda_category(),
string::Sprintf(args...));
}
}
......@@ -77,12 +81,8 @@ template <typename... Args>
inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
curandStatus_t stat, const Args&... args) {
if (stat != CURAND_STATUS_SUCCESS) {
// clang-format off
throw thrust::system_error(
cudaErrorLaunchFailure, thrust::cuda_category(),
string::Sprintf(args...) +
string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
// clang-format on
throw thrust::system_error(cudaErrorLaunchFailure, thrust::cuda_category(),
string::Sprintf(args...));
}
}
......@@ -92,12 +92,8 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
if (stat == CUDNN_STATUS_SUCCESS) {
return;
} else {
// clang-format off
throw std::runtime_error(
platform::dynload::cudnnGetErrorString(stat) +
string::Sprintf(args...) +
string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
// clang-format on
throw std::runtime_error(platform::dynload::cudnnGetErrorString(stat) +
string::Sprintf(args...));
}
}
......@@ -126,22 +122,32 @@ inline typename std::enable_if<sizeof...(Args) != 0, void>::type throw_on_error(
} else if (stat == CUBLAS_STATUS_LICENSE_ERROR) {
err = "CUBLAS: license error, ";
}
throw std::runtime_error(err + string::Sprintf(args...) +
string::Sprintf(" at [%s:%s];", __FILE__, __LINE__));
throw std::runtime_error(err + string::Sprintf(args...));
}
#endif // PADDLE_ONLY_CPU
#define PADDLE_THROW(...) \
do { \
throw std::runtime_error( \
string::Sprintf(__VA_ARGS__) + \
string::Sprintf(" at [%s:%s];", __FILE__, __LINE__)); \
template <typename T>
inline void throw_on_error(T e) {
throw_on_error(e, "");
}
#define PADDLE_THROW(...) \
do { \
throw ::paddle::platform::EnforceNotMet( \
std::make_exception_ptr( \
std::runtime_error(string::Sprintf(__VA_ARGS__))), \
__FILE__, __LINE__); \
} while (0)
#define PADDLE_ENFORCE(...) \
do { \
::paddle::platform::throw_on_error(__VA_ARGS__); \
#define PADDLE_ENFORCE(...) \
do { \
try { \
::paddle::platform::throw_on_error(__VA_ARGS__); \
} catch (...) { \
throw ::paddle::platform::EnforceNotMet(std::current_exception(), \
__FILE__, __LINE__); \
} \
} while (0)
} // namespace platform
......
......@@ -23,7 +23,7 @@ TEST(ENFORCE, FAILED) {
bool in_catch = false;
try {
PADDLE_ENFORCE(false, "Enforce is not ok %d at all", 123);
} catch (const std::runtime_error& error) {
} catch (paddle::platform::EnforceNotMet error) {
// your error handling code here
in_catch = true;
std::string msg = "Enforce is not ok 123 at all";
......
......@@ -146,22 +146,22 @@ All parameter, weight, gradient are variables in Paddle.
});
ExposeOperator(operator_base);
using PlainNetPtr = std::shared_ptr<pd::PlainNet>;
py::class_<pd::PlainNet, PlainNetPtr> net(m, "Net");
py::class_<pd::NetOp, std::shared_ptr<pd::NetOp>> net(m, "Net");
net.def_static("create",
[]() -> std::shared_ptr<pd::PlainNet> {
auto retv = std::make_shared<pd::PlainNet>();
[]() -> std::shared_ptr<pd::NetOp> {
auto retv = std::make_shared<pd::NetOp>();
retv->type_ = "plain_net";
return retv;
})
.def("add_op", &pd::PlainNet::AddOp)
.def("add_op", &pd::NetOp::AddOp)
.def("add_op",
[](PlainNetPtr& self, const PlainNetPtr& net) -> void {
self->AddOp(std::static_pointer_cast<pd::OperatorBase>(net));
[](pd::NetOp& self, const std::shared_ptr<pd::NetOp>& net) -> void {
self.AddOp(std::static_pointer_cast<pd::OperatorBase>(net));
})
.def("complete_add_op", &pd::PlainNet::CompleteAddOp)
.def("complete_add_op", [](PlainNetPtr& self) { self->CompleteAddOp(); });
.def("complete_add_op", &pd::NetOp::CompleteAddOp)
.def("complete_add_op",
[](std::shared_ptr<pd::NetOp>& self) { self->CompleteAddOp(); });
ExposeOperator(net);
m.def("unique_integer", UniqueIntegerGenerator);
......
......@@ -76,7 +76,11 @@ void NewRemoteParameterUpdater::init(
sgdConfigV2->set_decay(paramConfig.decay_rate());
optimizeConfigV2.set_lr_policy(paddle::OptimizerConfig::Const);
auto constlr = optimizeConfigV2.mutable_const_lr();
constlr->set_learning_rate(paramConfig.learning_rate());
if (paramConfig.has_learning_rate()) {
constlr->set_learning_rate(paramConfig.learning_rate());
} else {
constlr->set_learning_rate(trainerConfig_.learning_rate());
}
if (trainerConfig_.algorithm() == "sgd") {
optimizeConfigV2.set_optimizer(paddle::OptimizerConfig::SGD);
// FIXME: config all algorithms
......
......@@ -126,9 +126,11 @@ public:
}
/**
* @brief operator bool, return True if there is something error.
* @brief check this status by glog.
* @note It is a temp method used during cleaning Paddle code. It will be
* removed later.
*/
operator bool() const { return !this->isOK(); }
void check() const { CHECK(this->isOK()) << msg(); }
/**
* @brief isOK return True if there is no error.
......@@ -136,13 +138,6 @@ public:
*/
bool isOK() const { return msg_ == nullptr; }
/**
* @brief check this status by glog.
* @note It is a temp method used during cleaning Paddle code. It will be
* removed later.
*/
void check() const { CHECK(this->isOK()) << msg(); }
private:
std::shared_ptr<std::string> msg_;
};
......
......@@ -18,17 +18,17 @@ limitations under the License. */
TEST(Error, testAll) {
paddle::Error error;
ASSERT_FALSE(error);
ASSERT_TRUE(error.isOK());
error = paddle::Error("I'm the error");
ASSERT_TRUE(error);
ASSERT_FALSE(error.isOK());
ASSERT_STREQ("I'm the error", error.msg());
error = paddle::Error("error2");
ASSERT_TRUE(error);
ASSERT_FALSE(error.isOK());
ASSERT_STREQ("error2", error.msg());
int i = 3;
auto error3 = paddle::Error("error%d", i);
ASSERT_TRUE(error3);
ASSERT_FALSE(error3.isOK());
ASSERT_STREQ("error3", error3.msg());
}
......@@ -2055,8 +2055,7 @@ class BatchNormLayer(LayerBase):
# Automatically select cudnn_batch_norm for GPU and batch_norm for CPU.
# Also based on cudnn version.
use_cudnn = use_gpu and batch_norm_type != "batch_norm" and \
((not parallel_nn) or self.config.device > -1) and \
cudnn_version >= 4007
((not parallel_nn) or self.config.device > -1)
self.layer_type = "cudnn_batch_norm" if use_cudnn else "batch_norm"
super(BatchNormLayer, self).__init__(
name, self.layer_type, 0, inputs=inputs, **xargs)
......
......@@ -272,7 +272,7 @@ class ExtraLayerAttribute(object):
for key in self.attr:
if not hasattr(self, 'can_%s' % key) or \
not getattr(self, 'can_%s' % key):
raise NotImplementedError("Layer %s cannot support %s" %
raise NotImplementedError("Layer %s does not support %s" %
(layer_name, key))
@staticmethod
......
......@@ -865,7 +865,7 @@ def data_layer(name, size, height=None, width=None, layer_attr=None):
@wrap_name_default("embedding")
@wrap_param_attr_default()
@layer_support(ERROR_CLIPPING)
@layer_support(ERROR_CLIPPING, DROPOUT)
def embedding_layer(input, size, name=None, param_attr=None, layer_attr=None):
"""
Define a embedding Layer.
......@@ -1320,7 +1320,7 @@ def pooling_layer(input,
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
@wrap_act_default(param_names=["act", 'state_act'], act=TanhActivation())
@wrap_name_default("lstmemory")
@layer_support(DROPOUT)
@layer_support()
def lstmemory(input,
name=None,
size=None,
......@@ -1429,7 +1429,7 @@ def lstmemory(input,
@wrap_act_default(param_names=['gate_act'], act=SigmoidActivation())
@wrap_act_default(param_names=["act"], act=TanhActivation())
@wrap_name_default("gru")
@layer_support(DROPOUT)
@layer_support()
def grumemory(input,
size=None,
name=None,
......@@ -1793,7 +1793,7 @@ def repeat_layer(input,
@wrap_name_default("seqreshape")
@wrap_act_default(act=IdentityActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support()
@layer_support(ERROR_CLIPPING, DROPOUT)
def seq_reshape_layer(input,
reshape_size,
act=None,
......@@ -2703,7 +2703,7 @@ def img_cmrnorm_layer(input,
default_factory=lambda _: ParamAttr(initial_mean=1.0, initial_std=0.))
@wrap_act_default(act=ReluActivation())
@wrap_name_default("batch_norm")
@layer_support(DROPOUT)
@layer_support(DROPOUT, ERROR_CLIPPING)
def batch_norm_layer(input,
act=None,
name=None,
......@@ -2783,15 +2783,6 @@ def batch_norm_layer(input,
:return: LayerOutput object.
:rtype: LayerOutput
"""
if not isinstance(act, ReluActivation):
logger.log(logging.WARN,
"%s is not recommend for batch normalization's activation, "
"maybe the relu is better" % act.name)
if not isinstance(input.activation, LinearActivation):
logger.log(logging.WARN,
"The activation should be inside batch normalization, the "
"previous layer's activation may be Linear")
if num_channels is None:
if input.num_filters is not None:
......@@ -2861,7 +2852,7 @@ def sum_to_one_norm_layer(input, name=None, layer_attr=None):
@wrap_name_default("addto")
@wrap_act_default(act=LinearActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support(DROPOUT)
@layer_support(DROPOUT, ERROR_CLIPPING)
def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
"""
AddtoLayer.
......@@ -2940,7 +2931,7 @@ def addto_layer(input, act=None, name=None, bias_attr=None, layer_attr=None):
@wrap_act_default(act=IdentityActivation())
@wrap_name_default("concat")
@layer_support()
@layer_support(DROPOUT, ERROR_CLIPPING)
def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
"""
Concat all input vector into one huge vector.
......@@ -3024,7 +3015,7 @@ def concat_layer(input, act=None, name=None, layer_attr=None, bias_attr=None):
@wrap_name_default("seqconcat")
@wrap_act_default(act=IdentityActivation())
@wrap_bias_attr_default(has_bias=False)
@layer_support()
@layer_support(DROPOUT, ERROR_CLIPPING)
def seq_concat_layer(a, b, act=None, name=None, layer_attr=None,
bias_attr=None):
"""
......@@ -3177,7 +3168,7 @@ def memory(name,
@wrap_act_default(param_names=['state_act'], act=TanhActivation())
@wrap_act_default(act=TanhActivation())
@wrap_name_default('lstm_step')
@layer_support(ERROR_CLIPPING, DROPOUT)
@layer_support()
def lstm_step_layer(input,
state,
size=None,
......@@ -4480,7 +4471,7 @@ def tensor_layer(a,
@wrap_param_attr_default()
@wrap_bias_attr_default()
@wrap_act_default()
@layer_support()
@layer_support(DROPOUT, ERROR_CLIPPING)
def selective_fc_layer(input,
size,
select=None,
......@@ -5974,7 +5965,7 @@ def crop_layer(input, offset, axis=2, shape=None, name=None, layer_attr=None):
"""
The crop layer crops images by offset and shape. User can set crop shape by
args 'shape' explicitly or by reference input layer.
The example usage is:
.. code-block:: python
......
......@@ -34,6 +34,7 @@ import minibatch
import plot
import image
import model
import paddle.trainer.config_parser as cp
__all__ = [
'optimizer',
......@@ -58,6 +59,8 @@ __all__ = [
'model',
]
cp.begin_parse()
def init(**kwargs):
import py_paddle.swig_paddle as api
......@@ -73,6 +76,11 @@ def init(**kwargs):
for key in args_dict.keys():
args.append('--%s=%s' % (key, str(args_dict[key])))
if 'use_gpu' in kwargs:
cp.g_command_config_args['use_gpu'] = kwargs['use_gpu']
assert 'parallel_nn' not in kwargs, ("currently 'parallel_nn' is not "
"supported in v2 APIs.")
api.initPaddle(*args)
......
......@@ -166,55 +166,37 @@ def cluster_files_reader(files_pattern,
return reader
def convert(output_path,
reader,
num_shards,
name_prefix,
max_lines_to_shuffle=1000):
def convert(output_path, reader, line_count, name_prefix):
import recordio
"""
Convert data from reader to recordio format files.
:param output_path: directory in which output files will be saved.
:param reader: a data reader, from which the convert program will read data instances.
:param num_shards: the number of shards that the dataset will be partitioned into.
:param name_prefix: the name prefix of generated files.
:param max_lines_to_shuffle: the max lines numbers to shuffle before writing.
"""
assert num_shards >= 1
assert max_lines_to_shuffle >= 1
def open_writers():
w = []
for i in range(0, num_shards):
n = "%s/%s-%05d-of-%05d" % (output_path, name_prefix, i,
num_shards - 1)
w.append(recordio.writer(n))
return w
def close_writers(w):
for i in range(0, num_shards):
w[i].close()
assert line_count >= 1
indx_f = 0
def write_data(w, lines):
def write_data(indx_f, lines):
random.shuffle(lines)
for i, d in enumerate(lines):
filename = "%s/%s-%05d" % (output_path, name_prefix, indx_f)
writer = recordio.writer(filename)
for l in lines:
# FIXME(Yancey1989):
# dumps with protocol: pickle.HIGHEST_PROTOCOL
o = pickle.dumps(d)
w[i % num_shards].write(o)
writer.write(cPickle.dumps(l))
writer.close()
w = open_writers()
lines = []
for i, d in enumerate(reader()):
lines.append(d)
if i % max_lines_to_shuffle == 0 and i >= max_lines_to_shuffle:
write_data(w, lines)
if i % line_count == 0 and i >= line_count:
write_data(indx_f, lines)
lines = []
indx_f += 1
continue
write_data(w, lines)
close_writers(w)
write_data(indx_f, lines)
......@@ -242,9 +242,9 @@ def gen_list(querylist):
if not isinstance(querylist, QueryList):
querylist = QueryList(querylist)
querylist._correct_ranking_()
relevance_score_list = [query.relevance_score for query in querylist]
relevance_score_list = [[query.relevance_score] for query in querylist]
feature_vector_list = [query.feature_vector for query in querylist]
yield np.array(relevance_score_list).T, np.array(feature_vector_list)
yield np.array(relevance_score_list), np.array(feature_vector_list)
def query_filter(querylists):
......
......@@ -35,6 +35,13 @@ class Inference(object):
name = param.getName()
assert isinstance(val, api.Vector)
val.copyFromNumpyArray(parameters.get(name).flatten())
# the setValueUpdated function is called in randomize, zeroMem,
# load function in paddle/parameter/Parameter.cpp. But in the
# inference mode, the setValueUpdated is never called, it will
# cause the parameter will not be dispatched
# in MultiGradientMachine for multi-GPU. So setValueUpdated is
# called here, but it's better to call this function in one place.
param.setValueUpdated()
self.__gradient_machine__ = gm
self.__data_types__ = topo.data_type()
......
......@@ -324,6 +324,3 @@ def parse_network(output_layers, extra_layers=None):
def get_layer(name):
return config_base.__layer_map__.get(name)
cp.begin_parse()
......@@ -49,7 +49,6 @@ class client(object):
def set_dataset(self, paths):
holder_type = ctypes.c_char_p * len(paths)
holder = holder_type()
print paths
for idx, path in enumerate(paths):
c_ptr = ctypes.c_char_p(path)
holder[idx] = c_ptr
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册