提交 7ff689f5 编写于 作者: C caoying03

Merge branch 'develop' into add_sequence_slice_layer

......@@ -24,4 +24,5 @@ cmake-build-*
python/paddle/v2/framework/core.so
CMakeFiles
cmake_install.cmake
paddle/.timestamp
python/paddlepaddle.egg-info/
......@@ -28,7 +28,7 @@ RUN apt-get update && \
wget unzip unrar tar xz-utils bzip2 gzip coreutils ntp \
curl sed grep graphviz libjpeg-dev zlib1g-dev \
python-matplotlib gcc-4.8 g++-4.8 \
automake locales clang-format-3.8 swig doxygen cmake \
automake locales clang-format swig doxygen cmake \
liblapack-dev liblapacke-dev libboost-dev \
clang-3.8 llvm-3.8 libclang-3.8-dev \
net-tools && \
......
hash: 1b9b07408ca7fac27a374dc2ccd2433e4bff090484008a037df967284949a582
updated: 2017-08-03T21:46:51.744995189Z
updated: 2017-08-07T23:37:48.867469328Z
imports:
- name: github.com/beorn7/perks
version: 4c0e84591b9aa9e6dcfdf3e020114cd81f89d5f9
......@@ -10,7 +10,7 @@ imports:
- name: github.com/cockroachdb/cmux
version: 112f0506e7743d64a6eb8fedbcff13d9979bbf92
- name: github.com/coreos/etcd
version: c31bec0f29facff13f7c3e3d948e55dd6689ed42
version: d0d1a87aa96ae14914751d42264262cb69eda170
subpackages:
- alarm
- auth
......@@ -24,6 +24,7 @@ imports:
- error
- etcdserver
- etcdserver/api
- etcdserver/api/etcdhttp
- etcdserver/api/v2http
- etcdserver/api/v2http/httptypes
- etcdserver/api/v3client
......@@ -210,11 +211,6 @@ testImports:
version: 04cdfd42973bb9c8589fd6a731800cf222fde1a9
subpackages:
- spew
- name: github.com/docker/docker
version: b6d164e6c46d8115b146e4c3ac93784e9ef8b49e
subpackages:
- pkg/ioutils
- pkg/longpath
- name: github.com/pmezard/go-difflib
version: d8ed2627bdf02c080bf22230dbb337003b7aba2d
subpackages:
......
package master_test
import (
"io/ioutil"
"net/url"
"os"
"strings"
"testing"
"time"
"github.com/PaddlePaddle/Paddle/go/master"
"github.com/coreos/etcd/clientv3"
"github.com/coreos/etcd/embed"
"github.com/docker/docker/pkg/ioutils"
"github.com/stretchr/testify/assert"
)
func TestNewServiceWithEtcd(t *testing.T) {
// setup an embed etcd server
etcdDir, err := ioutils.TempDir("", "")
etcdDir, err := ioutil.TempDir("", "")
if err != nil {
t.Fatal(err)
}
cfg := embed.NewConfig()
lpurl, _ := url.Parse("http://localhost:0")
lcurl, _ := url.Parse("http://localhost:0")
cfg.LPUrls = []url.URL{*lpurl}
cfg.LCUrls = []url.URL{*lcurl}
cfg.Dir = etcdDir
e, err := embed.StartEtcd(cfg)
if err != nil {
......@@ -30,15 +36,13 @@ func TestNewServiceWithEtcd(t *testing.T) {
t.Fatal(err)
}
}()
select {
case <-e.Server.ReadyNotify():
t.Log("Server is ready!")
case <-time.After(60 * time.Second):
e.Server.Stop() // trigger a shutdown
t.Fatal("Server took too long to start!")
}
ep := []string{"127.0.0.1:2379"}
<-e.Server.ReadyNotify()
port := strings.Split(e.Clients[0].Addr().String(), ":")[1]
endpoint := "127.0.0.1:" + port
ep := []string{endpoint}
masterAddr := "127.0.0.1:3306"
store, err := master.NewEtcdClient(ep, masterAddr, master.DefaultLockPath, master.DefaultAddrPath, master.DefaultStatePath, 30)
if err != nil {
......
......@@ -90,8 +90,12 @@ func cArrayToSlice(p unsafe.Pointer, len int) []byte {
type selector bool
func (s selector) Select() bool {
return bool(s)
func (s selector) Select() (bool, error) {
return bool(s), nil
}
func (s selector) Done() error {
return nil
}
type lister []client.Server
......@@ -114,11 +118,10 @@ func paddle_new_pserver_client(addrs *C.char, selected int) C.paddle_pserver_cli
}
//export paddle_new_etcd_pserver_client
func paddle_new_etcd_pserver_client(etcdEndpoints *C.char, selected int) C.paddle_pserver_client {
// TODO(Longfei: use etcd lock to decide which trainer to initialize the parameters)
func paddle_new_etcd_pserver_client(etcdEndpoints *C.char) C.paddle_pserver_client {
addr := C.GoString(etcdEndpoints)
etcdClient := client.NewEtcd(addr)
c := client.NewClient(etcdClient, etcdClient.Desired(), selector(selected != 0))
c := client.NewClient(etcdClient, etcdClient.Desired(), etcdClient)
return add(c)
}
......@@ -136,7 +139,12 @@ func paddle_pserver_client_release(client C.paddle_pserver_client) {
//export paddle_begin_init_params
func paddle_begin_init_params(client C.paddle_pserver_client) C.int {
c := get(client)
if selected := c.BeginInitParams(); selected {
selected, err := c.BeginInitParams()
if err != nil {
panic(err)
}
if selected {
return 1
}
return 0
......
......@@ -27,9 +27,13 @@ import (
// TODO(helin): add RPC call retry logic
// Selector selects if the client should initialize parameter servers.
// Selector selects if the client should initialize parameters and
// reports the initialization process done.
type Selector interface {
Select() bool
// Select selects if the client should initialize parameter servers.
Select() (bool, error)
// Done indicates the initialization process is done.
Done() error
}
// Server is the identification of a parameter Server.
......@@ -115,7 +119,7 @@ func (c *Client) monitorPservers(l Lister, pserverNum int) {
// servers. Other trainers will be blocked until the initialization is
// done, and they need to get the initialized parameters from
// parameter servers using GetParams.
func (c *Client) BeginInitParams() bool {
func (c *Client) BeginInitParams() (bool, error) {
return c.sel.Select()
}
......
......@@ -124,8 +124,12 @@ func initEtcdClient() {
type selector bool
func (s selector) Select() bool {
return bool(s)
func (s selector) Select() (bool, error) {
return bool(s), nil
}
func (s selector) Done() error {
return nil
}
type lister []client.Server
......@@ -135,7 +139,11 @@ func (l lister) List() []client.Server {
}
func testClient(t *testing.T, c *client.Client) {
selected := c.BeginInitParams()
selected, err := c.BeginInitParams()
if err != nil {
t.Fatal(err)
}
if !selected {
t.Fatal("should be selected.")
}
......
......@@ -16,53 +16,60 @@ package client
import (
"context"
"errors"
"fmt"
"strconv"
"strings"
"time"
"github.com/PaddlePaddle/Paddle/go/pserver"
"github.com/coreos/etcd/clientv3"
"github.com/coreos/etcd/clientv3/concurrency"
log "github.com/sirupsen/logrus"
)
const (
defaultEtcdTimeout time.Duration = 5 * time.Second
initLockPath = "/init_ps/lock"
initDonePath = "/init_ps/done"
initDoneVal = "1"
)
// EtcdClient is used by pserver client that is a part of trainer process.
// Etcd is used by pserver client that is a part of trainer process.
// TODO:
// 1. add watcher to watch the change state of pservers)
// 1. add etcd lock)
type EtcdClient struct {
// 1. add watcher to watch the change state of pservers.
type Etcd struct {
client *clientv3.Client
timeout time.Duration
endpoints []string
lock *concurrency.Mutex
}
// Desired read ps desired number from etcd.
func (p *EtcdClient) Desired() int {
func (e *Etcd) Desired() int {
var psDesired int
for {
ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
resp, err := p.client.Get(ctx, pserver.PsDesired)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
resp, err := e.client.Get(ctx, pserver.PsDesired)
cancel()
if err != nil {
log.Errorf("Get ps dresire number failed! recnnectiong..., %v", err)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
kvs := resp.Kvs
if len(kvs) == 0 {
log.Infoln("Waiting for ps desired registered ...")
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
psDesired, err = strconv.Atoi(string(resp.Kvs[0].Value))
if err != nil {
log.Errorf("psDesired %d invalid %v", psDesired, err)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
......@@ -73,26 +80,26 @@ func (p *EtcdClient) Desired() int {
}
// List return the pserver list read from etcd.
func (p *EtcdClient) List() []Server {
psDesired := p.Desired()
func (e *Etcd) List() []Server {
psDesired := e.Desired()
servers := make([]Server, psDesired)
for {
for i := 0; i < psDesired; i++ {
ctx, cancel := context.WithTimeout(context.Background(), p.timeout)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
psKey := pserver.PsPath + strconv.Itoa(i)
log.Debugf("checking %s", psKey)
resp, err := p.client.Get(ctx, psKey)
resp, err := e.client.Get(ctx, psKey)
cancel()
if err != nil {
log.Infof("Get psKey= %s error, %v", psKey, err)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
kvs := resp.Kvs
if len(kvs) == 0 {
log.Infof("Waiting for ps addr registered ...")
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
......@@ -100,7 +107,7 @@ func (p *EtcdClient) List() []Server {
// TODO(Longfei) check the ps address
if psAddr == "" {
log.Infof("Get psKey = %s, psAddr is empty", psKey)
time.Sleep(p.timeout)
time.Sleep(e.timeout)
continue
}
log.Debugf("got value (%s) for key: %s", psAddr, psKey)
......@@ -113,7 +120,7 @@ func (p *EtcdClient) List() []Server {
}
// NewEtcd create a etcd client to return the state of pserver on etcd.
func NewEtcd(endpoints string) *EtcdClient {
func NewEtcd(endpoints string) *Etcd {
ep := strings.Split(endpoints, ",")
var cli *clientv3.Client
var err error
......@@ -130,10 +137,118 @@ func NewEtcd(endpoints string) *EtcdClient {
break
}
log.Infof("Connected to etcd: %s\n", endpoints)
client := &EtcdClient{
client := &Etcd{
client: cli,
timeout: defaultEtcdTimeout,
endpoints: ep,
}
return client
}
// Select indicates if the current trainer is selected to initialize
// the pserver parameters.
func (e *Etcd) Select() (bool, error) {
sess, err := concurrency.NewSession(e.client, concurrency.WithTTL(5))
if err != nil {
return false, err
}
lock := concurrency.NewMutex(sess, initLockPath)
log.Infof("Trying to acquire lock at %s.", initLockPath)
// Do not use timeout context here, since we don't know how
// long does it take for other trainers to initialize the
// parameters.
err = lock.Lock(context.Background())
if err != nil {
return false, err
}
log.Infof("Successfully acquired lock at %s.", initLockPath)
get := clientv3.OpGet(initDonePath)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
tresp, err := e.client.Txn(ctx).If(lock.IsOwner()).Then(get).Commit()
cancel()
if err != nil {
return false, err
}
if !tresp.Succeeded {
return false, errors.New("no longer the owner of the lock")
}
resp := tresp.Responses[0].GetResponseRange()
if len(resp.Kvs) == 0 {
// Key value not set, select current trainer.
e.lock = lock
log.Infoln("Trainer selected.")
return true, nil
}
if string(resp.Kvs[0].Value) == initDoneVal {
log.Infoln("Initialization is already done.")
ctx, cancel = context.WithTimeout(context.Background(), e.timeout)
err = lock.Unlock(ctx)
cancel()
if err != nil {
log.Errorln(err)
}
return false, nil
}
return false, fmt.Errorf("key %s have unexpected value: %v", initDonePath, resp.Kvs[0].Value)
}
// Done indicates the parameter initialization process is done.
func (e *Etcd) Done() error {
if e.lock == nil {
return errors.New("lock is nil, Done called unexpectedly")
}
put := clientv3.OpPut(initDonePath, initDoneVal)
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
tresp, err := e.client.Txn(ctx).If(e.lock.IsOwner()).Then(put).Commit()
cancel()
if err != nil {
return err
}
if !tresp.Succeeded {
return errors.New("no longer the owner of the lock")
}
ctx, cancel = context.WithTimeout(context.Background(), e.timeout)
err = e.lock.Unlock(ctx)
cancel()
if err != nil {
log.Errorln(err)
} else {
e.lock = nil
}
return nil
}
// Close closes the etcd client.
func (e *Etcd) Close() error {
var err error
if e.lock != nil {
ctx, cancel := context.WithTimeout(context.Background(), e.timeout)
err = e.lock.Unlock(ctx)
cancel()
if err == nil {
e.lock = nil
}
}
cErr := e.client.Close()
if cErr != nil {
if err != nil {
log.Errorln(cErr)
return err
}
return cErr
}
return err
}
package client_test
import (
"io/ioutil"
"net/url"
"os"
"strings"
"sync"
"testing"
"github.com/PaddlePaddle/Paddle/go/pserver/client"
"github.com/coreos/etcd/embed"
)
func TestSelector(t *testing.T) {
etcdDir, err := ioutil.TempDir("", "")
if err != nil {
t.Fatal(err)
}
cfg := embed.NewConfig()
lpurl, _ := url.Parse("http://localhost:0")
lcurl, _ := url.Parse("http://localhost:0")
cfg.LPUrls = []url.URL{*lpurl}
cfg.LCUrls = []url.URL{*lcurl}
cfg.Dir = etcdDir
e, err := embed.StartEtcd(cfg)
if err != nil {
t.Fatal(err)
}
defer func() {
e.Close()
if err := os.RemoveAll(etcdDir); err != nil {
t.Fatal(err)
}
}()
<-e.Server.ReadyNotify()
port := strings.Split(e.Clients[0].Addr().String(), ":")[1]
endpoint := "127.0.0.1:" + port
var mu sync.Mutex
selectedCount := 0
var wg sync.WaitGroup
selectAndDone := func(c *client.Etcd) {
defer wg.Done()
selected, err := c.Select()
if err != nil {
panic(err)
}
if selected {
mu.Lock()
selectedCount++
mu.Unlock()
err = c.Done()
if err != nil {
t.Fatal(err)
}
}
}
c0 := client.NewEtcd(endpoint)
c1 := client.NewEtcd(endpoint)
c2 := client.NewEtcd(endpoint)
c3 := client.NewEtcd(endpoint)
wg.Add(3)
go selectAndDone(c0)
go selectAndDone(c1)
go selectAndDone(c2)
wg.Wait()
// simulate trainer crashed and restarted after the
// initialization process.
wg.Add(1)
go selectAndDone(c3)
wg.Wait()
mu.Lock()
if selectedCount != 1 {
t.Fatal("selected count wrong:", selectedCount)
}
mu.Unlock()
err = c0.Close()
if err != nil {
t.Fatal(err)
}
err = c1.Close()
if err != nil {
t.Fatal(err)
}
err = c2.Close()
if err != nil {
t.Fatal(err)
}
err = c3.Close()
if err != nil {
t.Fatal(err)
}
}
......@@ -7,6 +7,9 @@ cc_library(tensor SRCS tensor.cc DEPS ddim place paddle_memory device_context)
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
cc_library(lod_tensor SRCS lod_tensor.cc details/lod_tensor.cc DEPS ddim place tensor)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor)
cc_test(variable_test SRCS variable_test.cc)
cc_library(scope SRCS scope.cc)
......@@ -45,6 +48,8 @@ cc_library(paddle_pybind SHARED
add_op
mean_op
cross_entropy_op
fill_zeros_like_op
recurrent_op)
recurrent_op
uniform_random_op
gaussian_random_op
fill_zeros_like_op)
endif(WITH_PYTHON)
......@@ -13,6 +13,7 @@
limitations under the License. */
#include "paddle/framework/backward.h"
#include <list>
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
......@@ -132,8 +133,9 @@ std::shared_ptr<OperatorBase> BackwardRecursive(
std::shared_ptr<OperatorBase> grad_op = OpRegistry::CreateGradOp(forwardOp);
for (std::string& grad_input : grad_op->inputs_) {
if (no_grad_names.count(grad_input)) {
std::string prefix =
grad_input.substr(0, grad_input.size() - kGradVarSuffix.size());
// +1 for \0
std::string prefix = grad_input.substr(
0, grad_input.size() - sizeof(kGradVarSuffix) / sizeof(char) + 1);
grad_input = prefix + kZeroVarSuffix;
// If part of input gradient of that operator is not calculated, fill
......@@ -166,7 +168,7 @@ std::shared_ptr<OperatorBase> Backward(
std::unordered_set<std::string> no_grad_names;
no_grad_names.reserve(no_grad_vars.size());
no_grad_names.insert(kEmptyVarName + kGradVarSuffix);
no_grad_names.insert(std::string(kEmptyVarName) + kGradVarSuffix);
for (auto& name : no_grad_vars) {
no_grad_names.insert(name + kGradVarSuffix);
......
......@@ -17,16 +17,21 @@
#include <gtest/gtest.h>
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace framework {
using OperatorBase = framework::OperatorBase;
using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
using OpProto = framework::OpProto;
using OpAttrChecker = framework::OpAttrChecker;
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
class EmptyOp : public OperatorBase {
public:
void InferShape(const Scope &scope) const override {}
void Run(const Scope &scope,
const platform::DeviceContext &dev_ctx) const override {}
void Run(const Scope &scope, const DeviceContext &dev_ctx) const override {}
};
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
......@@ -71,7 +76,7 @@ class NoGradOpMaker : public OpProtoAndCheckerMaker {
}
};
class FcOp : public ops::NetOp {
class FcOp : public operators::NetOp {
public:
void Init() override {
AddOp(OpRegistry::CreateOp("mul", {Input("X"), Input("W")},
......@@ -143,6 +148,7 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
} // namespace paddle
namespace f = paddle::framework;
namespace ops = paddle::operators;
using EnforceNotMet = paddle::platform::EnforceNotMet;
REGISTER_OP(rowwise_add, f::EmptyOp, f::RowWiseAddOpMaker);
REGISTER_GRADIENT_OP(rowwise_add, rowwise_add_grad, f::EmptyOp);
......@@ -165,10 +171,10 @@ TEST(Backward, simple_op_grad) {
ASSERT_EQ(4UL, gop->inputs_.size());
ASSERT_EQ(f::kEmptyVarName, gop->inputs_[0]);
ASSERT_EQ("rowwise_add_grad", gop->type_);
ASSERT_EQ("X" + f::kGradVarSuffix, gop->outputs_[0]);
ASSERT_EQ("b" + f::kGradVarSuffix, gop->outputs_[1]);
ASSERT_EQ(f::GradVarName("X"), gop->outputs_[0]);
ASSERT_EQ(f::GradVarName("b"), gop->outputs_[1]);
ASSERT_EQ("X" + f::kGradVarSuffix, gop->Output("X" + f::kGradVarSuffix));
ASSERT_EQ(f::GradVarName("X"), gop->Output(f::GradVarName("X")));
}
TEST(Backward, simple_op_not_need_grad) {
......@@ -176,7 +182,7 @@ TEST(Backward, simple_op_not_need_grad) {
ASSERT_NE(fwd, nullptr);
auto gop = f::Backward(*fwd, {"X"});
ASSERT_EQ(std::find(gop->outputs_.begin(), gop->outputs_.end(),
"X" + f::kGradVarSuffix),
f::GradVarName("X")),
gop->outputs_.end());
auto no_input_gop = f::Backward(*fwd, {"X", "b"});
......@@ -244,18 +250,18 @@ TEST(Backward, net_input_of_network_not_need_grad) {
all_output.erase(f::kEmptyVarName);
for (auto &out : {"W1", "b1", "hidden0", "W2", "b2"}) {
ASSERT_NE(all_output.find(out + f::kGradVarSuffix), all_output.end());
ASSERT_NE(all_output.find(f::GradVarName(out)), all_output.end());
}
// Not Generated X
ASSERT_EQ(all_output.find("X" + f::kGradVarSuffix), all_output.end());
ASSERT_EQ(all_output.find(f::GradVarName("X")), all_output.end());
ASSERT_EQ(2UL, bwd_net->ops_.size());
ASSERT_TRUE(bwd_net->ops_[1]->IsNetOp());
auto first_fc_grad = static_cast<ops::NetOp *>(bwd_net->ops_[1].get());
ASSERT_EQ(3UL, first_fc_grad->ops_.size());
ASSERT_EQ(f::kEmptyVarName,
first_fc_grad->ops_[2]->Output("A" + f::kGradVarSuffix));
first_fc_grad->ops_[2]->Output(f::GradVarName("A")));
}
TEST(Backward, net_shared_weight) {
......@@ -307,15 +313,15 @@ TEST(Backward, op_part_of_output_are_not_need) {
ASSERT_EQ(1UL, fill_zero.inputs_.size());
ASSERT_EQ("Z", fill_zero.inputs_[0]);
ASSERT_EQ(1UL, fill_zero.outputs_.size());
ASSERT_EQ("Z" + f::kZeroVarSuffix, fill_zero.outputs_[0]);
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix, fill_zero.outputs_[0]);
auto &d_many_out = *net->ops_[1];
ASSERT_EQ("many_output_op_grad", d_many_out.type_);
ASSERT_EQ(1UL + 2UL + 2UL, d_many_out.inputs_.size()); // I/O/OG
ASSERT_EQ("Z" + f::kZeroVarSuffix, d_many_out.Input("z" + f::kGradVarSuffix));
ASSERT_EQ("Y" + f::kGradVarSuffix, d_many_out.Input("y" + f::kGradVarSuffix));
ASSERT_EQ("X" + f::kGradVarSuffix,
d_many_out.Output("x" + f::kGradVarSuffix));
ASSERT_EQ(std::string("Z") + f::kZeroVarSuffix,
d_many_out.Input(f::GradVarName("z")));
ASSERT_EQ(f::GradVarName("Y"), d_many_out.Input(f::GradVarName("y")));
ASSERT_EQ(f::GradVarName("X"), d_many_out.Output(f::GradVarName("x")));
}
TEST(Backward, op_part_of_input_are_not_need) {
......@@ -325,10 +331,9 @@ TEST(Backward, op_part_of_input_are_not_need) {
ASSERT_EQ(grad_mul.type_, "mul_grad");
ASSERT_EQ(grad_mul.inputs_.size(), 2UL + 1UL + 1UL);
ASSERT_EQ(grad_mul.outputs_.size(), 2UL);
ASSERT_EQ(grad_mul.Output("A" + f::kGradVarSuffix), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output("B" + f::kGradVarSuffix), "b" + f::kGradVarSuffix);
ASSERT_EQ(grad_mul.Input("Out" + f::kGradVarSuffix),
"out" + f::kGradVarSuffix);
ASSERT_EQ(grad_mul.Output(f::GradVarName("A")), f::kEmptyVarName);
ASSERT_EQ(grad_mul.Output(f::GradVarName("B")), f::GradVarName("b"));
ASSERT_EQ(grad_mul.Input(f::GradVarName("Out")), f::GradVarName("out"));
ASSERT_EQ(grad_mul.Input("A"), "a");
ASSERT_EQ(grad_mul.Input("B"), "b");
ASSERT_EQ(grad_mul.Input("Out"), "out");
......
/* 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. */
#include "paddle/framework/lod_tensor.h"
#include <memory>
namespace paddle {
namespace framework {
namespace details {
using LOD = LODTensor::LOD;
std::shared_ptr<LOD> SliceLOD(const LOD &lod, size_t level_begin,
size_t level_end) {
auto new_lod = std::make_shared<LOD>();
new_lod->reserve(level_end - level_begin);
for (size_t i = level_begin; i < level_end; i++) {
new_lod->emplace_back(lod[i]);
}
return new_lod;
}
std::shared_ptr<LOD> SliceLOD(const LOD &lod, size_t level, size_t elem_begin,
size_t elem_end, bool tensor_shared) {
// slice the lod.
auto new_lod = std::make_shared<LOD>();
new_lod->reserve(lod.size() - level);
auto start = lod.at(level)[elem_begin];
auto end = lod.at(level)[elem_end];
for (auto it = lod.begin() + level; it != lod.end(); it++) {
auto it_begin = std::find(it->begin(), it->end(), start);
auto it_end = std::find(it_begin, it->end(), end);
PADDLE_ENFORCE(it_begin != it->end(), "error in parsing lod info");
PADDLE_ENFORCE(it_end != it->end(), "error in parsing lod info");
new_lod->emplace_back(it_begin, it_end + 1);
if (!tensor_shared) {
// reset offset if tensor is copyed and sliced.
std::transform(new_lod->back().begin(), new_lod->back().end(),
new_lod->back().begin(),
[start](int v) { return v - start; });
PADDLE_ENFORCE(new_lod->back().front() == 0, "error in slice LOD");
}
}
return new_lod;
}
} // namespace details
} // namespace framework
} // namespace paddle
/* 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 <memory>
namespace paddle {
namespace framework {
namespace details {
/*
* Slice levels from LOD.
*
* @lod: LOD to slice.
* @level_begin: level to begin slice.
* @level_end: level to end slice.
*/
std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
size_t level_begin, size_t level_end);
/*
* Slice elements from a level of LOD.
*
* @lod: LOD to slice.
* @level: which level to slice.
* @elem_begin: element's index to begin slice.
* @elem_end: element's index to end slice.
*/
std::shared_ptr<LODTensor::LOD> SliceLOD(const LODTensor::LOD &lod,
size_t level, size_t elem_begin,
size_t elem_end, bool tensor_shared);
} // namespace details
} // namespace framework
} // namespace paddle
......@@ -83,21 +83,19 @@ TEST(GradOpBuilder, MutiInOut) {
EXPECT_EQ(grad_test_op->Input("Out1"), "out1");
EXPECT_EQ(grad_test_op->Inputs("Out2_mult"),
std::vector<std::string>({"out2_1", "out2_2"}));
EXPECT_EQ(grad_test_op->Input("Out1" + f::kGradVarSuffix),
"out1" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Inputs("Out2_mult" + f::kGradVarSuffix),
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out1")),
f::GradVarName("out1"));
EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out2_mult")),
std::vector<std::string>(
{"out2_1" + f::kGradVarSuffix, "out2_2" + f::kGradVarSuffix}));
{f::GradVarName("out2_1"), f::GradVarName("out2_2")}));
ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
"in1" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
std::vector<std::string>({"in2_1" + f::kGradVarSuffix,
"in2_2" + f::kGradVarSuffix,
"in2_3" + f::kGradVarSuffix}));
EXPECT_EQ(grad_test_op->Output("In3" + f::kGradVarSuffix),
"in3" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>({f::GradVarName("in2_1"),
f::GradVarName("in2_2"),
f::GradVarName("in2_3")}));
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In3")), f::GradVarName("in3"));
}
TEST(GradOpBuilder, IOIgnoredInGradient) {
......@@ -119,19 +117,18 @@ TEST(GradOpBuilder, IOIgnoredInGradient) {
EXPECT_EQ(grad_test_op->Inputs("Out1_mult"),
std::vector<std::string>({"out1_1", "out1_2"}));
EXPECT_EQ(grad_test_op->Input("Out2"), f::kEmptyVarName);
EXPECT_EQ(grad_test_op->Inputs("Out1_mult" + f::kGradVarSuffix),
EXPECT_EQ(grad_test_op->Inputs(f::GradVarName("Out1_mult")),
std::vector<std::string>(
{"out1_1" + f::kGradVarSuffix, "out1_2" + f::kGradVarSuffix}));
EXPECT_EQ(grad_test_op->Input("Out2" + f::kGradVarSuffix),
"out2" + f::kGradVarSuffix);
{f::GradVarName("out1_1"), f::GradVarName("out1_2")}));
EXPECT_EQ(grad_test_op->Input(f::GradVarName("Out2")),
f::GradVarName("out2"));
ASSERT_EQ(grad_test_op->outputs_.size(), 5UL);
EXPECT_EQ(grad_test_op->Output("In1" + f::kGradVarSuffix),
"in1" + f::kGradVarSuffix);
EXPECT_EQ(grad_test_op->Outputs("In2_mult" + f::kGradVarSuffix),
EXPECT_EQ(grad_test_op->Output(f::GradVarName("In1")), f::GradVarName("in1"));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In2_mult")),
std::vector<std::string>(
{"in2_1" + f::kGradVarSuffix, "in2_2" + f::kGradVarSuffix}));
EXPECT_EQ(grad_test_op->Outputs("In3_mult" + f::kGradVarSuffix),
{f::GradVarName("in2_1"), f::GradVarName("in2_2")}));
EXPECT_EQ(grad_test_op->Outputs(f::GradVarName("In3_mult")),
std::vector<std::string>(
{"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix}));
{f::GradVarName("in3_1"), f::GradVarName("in3_2")}));
}
/* 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. */
#include "paddle/framework/lod_tensor.h"
#include <glog/logging.h>
namespace paddle {
namespace framework {
LODTensor LODTensor::SliceShared(size_t level_begin, size_t level_end) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
// slice levels just need to update LOD info, each level will contains the
// whole tensor_, so no need to modify tensor_.
return LODTensor(tensor_, new_lod);
}
LODTensor LODTensor::SliceShared(size_t level, size_t elem_begin,
size_t elem_end) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
PADDLE_ENFORCE(elem_begin < NumElements(level),
"element begin [%d] out of range [%d]", elem_begin,
NumElements(level));
PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
"element end [%d] out of range [%d]", elem_end,
NumElements(level));
auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
true /*tensor_shared*/);
// slice elements just need to update LOD info, because offsets are not
// changed, so the original tensor_ can be reused.
return LODTensor(tensor_, new_lod);
}
} // namespace framework
} // namespace paddle
/* 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 <memory>
#if (!PADDLE_ONLY_CPU)
#include <thrust/device_vector.h>
#include <thrust/host_vector.h>
#endif
#include "paddle/framework/ddim.h"
#include "paddle/framework/tensor.h"
#include "paddle/platform/enforce.h"
namespace paddle {
namespace framework {
/*
* LODTensor (Level of details Tensor)
* see https://en.wikipedia.org/wiki/Level_of_details for reference.
*/
class LODTensor {
public:
// Level save offsets of each unit.
#ifdef PADDLE_ONLY_CPU
using Level = std::vector<size_t>;
#else
using Level = thrust::device_vector<size_t>;
#endif
// LOD stores offsets of each level of units, the largest units level first,
// then the smaller units level. Each Level stores the offsets of units in
// Tesor.
typedef std::vector<Level> LOD;
LODTensor() {}
LODTensor(const std::shared_ptr<Tensor> &tensor,
const std::shared_ptr<LOD> &lod) {
Reset(tensor, lod);
}
void Reset(const std::shared_ptr<Tensor> &tensor,
const std::shared_ptr<LOD> &lod) {
tensor_ = tensor;
lod_start_pos_ = lod;
}
/*
* Get a element from LOD.
*/
size_t lod_element(size_t level, size_t elem) const {
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
PADDLE_ENFORCE(elem < NumElements(level),
"element begin [%d] out of range [%d]", elem,
NumElements(level));
return (*lod_start_pos_)[level][elem];
}
/*
* Number of LODTensor's levels, each level has units of data, for example,
* in the sentence's view, article, paragraph, sentence are 3 levels.
*/
size_t NumLevels() const {
return lod_start_pos_ ? lod_start_pos_->size() : 0UL;
}
/*
* Number of elements in a level.
*/
size_t NumElements(size_t level = 0) const {
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
// the last offset is the end of last element
return lod_start_pos_->at(level).size() - 1;
}
/*
* Slice of levels[level_begin:level_end], with tensor copied.
*/
template <typename T>
LODTensor SliceCopied(size_t level_begin, size_t level_end,
const platform::Place &dst_place) const;
/*
* Slice of levels[level_begin:level_end], with tensor shared.
*/
LODTensor SliceShared(size_t level_begin, size_t level_end) const;
/*
* Slice of elements of a level, [elem_begin: elem_end], with tensor copied.
* @note: low performance in slice lod_start_pos_.
*/
template <typename T>
LODTensor SliceCopied(size_t level, size_t elem_begin, size_t elem_end,
const platform::Place &dst_place) const;
/*
* Slice of elements of a level, [elem_begin: elem_end], with tensor shared.
* @note: low performance in slice lod_start_pos_.
*/
LODTensor SliceShared(size_t level, size_t elem_begin, size_t elem_end) const;
/*
* Copy other's lod_start_pos_, to share LOD info.
* @note: the LOD info should not be changed.
*/
void ShareLOD(const LODTensor &other) {
lod_start_pos_ = other.lod_start_pos_;
}
/*
* Copy other's lod_start_pos_'s content, free to mutate.
*/
void CopyLOD(const LODTensor &other) {
lod_start_pos_ = std::make_shared<LOD>(*other.lod_start_pos_);
}
/*
* Determine whether LODTensor has a valid LOD info.
*/
bool HasLOD() const { return bool(lod_start_pos_); }
LOD *lod() const { return lod_start_pos_.get(); }
std::shared_ptr<Tensor> &tensor() { return tensor_; }
Tensor *raw_tensor() { return tensor_.get(); }
private:
std::shared_ptr<LOD> lod_start_pos_;
std::shared_ptr<Tensor> tensor_;
};
} // namespace framework
} // namespace paddle
#include "paddle/framework/lod_tensor_impl.h"
/* 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/details/lod_tensor.h"
namespace paddle {
namespace framework {
template <typename T>
LODTensor LODTensor::SliceCopied(size_t level_begin, size_t level_end,
const platform::Place &dst_place) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
auto new_lod = details::SliceLOD(*lod_start_pos_, level_begin, level_end);
auto new_tensor = std::make_shared<Tensor>();
new_tensor->CopyFrom<T>(*tensor_, dst_place);
return LODTensor(new_tensor, new_lod);
}
template <typename T>
LODTensor LODTensor::SliceCopied(size_t level, size_t elem_begin,
size_t elem_end,
const platform::Place &dst_place) const {
PADDLE_ENFORCE(HasLOD(), "has no LOD info, can't be sliced.");
PADDLE_ENFORCE(level < NumLevels(), "level [%d] out of range [%d]", level,
NumLevels());
PADDLE_ENFORCE(elem_begin < NumElements(level),
"element begin [%d] out of range [%d]", elem_begin,
NumElements(level));
PADDLE_ENFORCE(elem_end < NumElements(level) + 1,
"element end [%d] out of range [%d]", elem_end,
NumElements(level));
auto new_lod = details::SliceLOD(*lod_start_pos_, level, elem_begin, elem_end,
false /*tensor_shared*/);
auto start_idx = new_lod->front().front();
auto end_idx = new_lod->front().back() - 1 /*the next element's start*/;
auto sliced_tensor = tensor_->Slice<T>(start_idx, end_idx);
auto new_tensor = std::make_shared<Tensor>();
new_tensor->CopyFrom<T>(sliced_tensor, dst_place);
return LODTensor(new_tensor, new_lod);
}
} // namespace framework
} // namespace paddle
/*
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.
*/
#include "paddle/framework/lod_tensor.h"
#include <glog/logging.h>
#include <gtest/gtest.h>
#include <memory>
namespace paddle {
namespace framework {
class LODTensorTester : public ::testing::Test {
public:
virtual void SetUp() override {
lod_tensor.reset(new LODTensor);
// tensor's batch_size: 30
// 3 levels
// 0 10 20
// 0 5 10 15 20
// 0 2 5 7 10 12 15 20
auto lod = std::make_shared<LODTensor::LOD>();
lod->push_back(std::vector<size_t>{0, 10, 20});
lod->push_back(std::vector<size_t>{0, 5, 10, 15, 20});
lod->push_back(std::vector<size_t>{0, 2, 5, 7, 10, 12, 15, 17, 20});
auto tensor = std::make_shared<Tensor>();
tensor->Resize({20 /*batch size*/, 128 /*dim*/});
// malloc memory
tensor->mutable_data<float>(place);
lod_tensor->Reset(tensor, lod);
}
protected:
std::unique_ptr<LODTensor> lod_tensor;
platform::CPUPlace place;
};
TEST_F(LODTensorTester, NumLevels) { ASSERT_EQ(lod_tensor->NumLevels(), 3UL); }
TEST_F(LODTensorTester, NumElements) {
ASSERT_EQ(lod_tensor->NumElements(0), 2UL);
ASSERT_EQ(lod_tensor->NumElements(1), 4UL);
ASSERT_EQ(lod_tensor->NumElements(2), 8UL);
}
TEST_F(LODTensorTester, SliceShared_Level) {
// slice 1 level
for (size_t level = 0; level < 3UL; ++level) {
auto new_lod_tensor = lod_tensor->SliceShared(level, level + 1);
ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
}
// slice 2 level
for (size_t level = 0; level < 2UL; ++level) {
auto new_lod_tensor = lod_tensor->SliceShared(level, level + 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
ASSERT_EQ(new_lod_tensor.NumElements(1),
lod_tensor->NumElements(level + 1));
ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
}
}
TEST_F(LODTensorTester, SliceCopied_Level) {
// slice 1 level
for (size_t level = 0; level < 3UL; ++level) {
auto new_lod_tensor =
lod_tensor->SliceCopied<float>(level, level + 1, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 1UL);
ASSERT_EQ(new_lod_tensor.NumElements(0UL), lod_tensor->NumElements(level));
// ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
// TODO(superjom) add tensor comparation here.
}
// slice 2 level
for (size_t level = 0; level < 2UL; ++level) {
auto new_lod_tensor =
lod_tensor->SliceCopied<float>(level, level + 2, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), lod_tensor->NumElements(level));
ASSERT_EQ(new_lod_tensor.NumElements(1),
lod_tensor->NumElements(level + 1));
// ASSERT_EQ(new_lod_tensor.tensor(), lod_tensor->tensor());
// TODO(superjom) add tensor comparation here.
}
}
TEST_F(LODTensorTester, SliceShared_Element) {
size_t level = 0;
auto new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
level = 1;
new_lod_tensor = lod_tensor->SliceShared(level, 0, 2);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
}
TEST_F(LODTensorTester, SliceCopied_Element) {
size_t level = 0;
auto new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 3UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.NumElements(2), 8UL);
ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
level = 1;
new_lod_tensor = lod_tensor->SliceCopied<float>(level, 0, 2, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_NE(new_lod_tensor.raw_tensor(), lod_tensor->raw_tensor());
level = 1;
// LOD is
// 0 5 10
// 0 2 5 7 10
new_lod_tensor = lod_tensor->SliceCopied<float>(level, 1, 3, place);
ASSERT_EQ(new_lod_tensor.NumLevels(), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(0), 2UL);
ASSERT_EQ(new_lod_tensor.NumElements(1), 4UL);
ASSERT_EQ(new_lod_tensor.lod_element(0, 0), 0UL);
ASSERT_EQ(new_lod_tensor.lod_element(0, 1), 5UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 0), 0UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 1), 2UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 2), 5UL);
ASSERT_EQ(new_lod_tensor.lod_element(1, 3), 7UL);
// TODO(superjom) compare the content of these tensors
}
TEST_F(LODTensorTester, ShareLOD) {
LODTensor new_lod_tensor;
new_lod_tensor.ShareLOD(*lod_tensor);
ASSERT_EQ(new_lod_tensor.lod(), lod_tensor->lod());
}
TEST_F(LODTensorTester, CopyLOD) {
LODTensor new_lod_tensor;
new_lod_tensor.CopyLOD(*lod_tensor);
ASSERT_NE(new_lod_tensor.lod(), lod_tensor->lod());
}
} // namespace framework
} // namespace paddle
......@@ -260,12 +260,6 @@ class OpRegistry {
return CreateOp(op_desc.type(), inputs, outputs, attrs);
}
static bool SupportGPU(const std::string& op_type) {
OperatorWithKernel::OpKernelKey key;
key.place_ = platform::GPUPlace();
return OperatorWithKernel::AllOpKernels().at(op_type).count(key) != 0;
}
static std::shared_ptr<OperatorBase> CreateGradOp(const OperatorBase& op) {
PADDLE_ENFORCE(!op.IsNetOp(),
"Use framework::Backward to get backward ops");
......
......@@ -33,19 +33,19 @@ namespace paddle {
namespace framework {
/// If a variable is a empty variable, that name will be used.
const std::string kEmptyVarName = "@EMPTY@";
constexpr char kEmptyVarName[] = "@EMPTY@";
/// If a variable is a temporary variable, that name will be set in Python,
/// but it will be convert to a unique name in scope after OpCreator.
const std::string kTempVarName = "@TEMP@";
constexpr char kTempVarName[] = "@TEMP@";
/// If a variable's name has a certain suffix, it means that the
/// variable is the gradient of another varibale.
/// e.g. Variable "x@GRAD" is the gradient of varibale "x".
const std::string kGradVarSuffix = "@GRAD";
constexpr char kGradVarSuffix[] = "@GRAD";
/// Variables with this suffix are supposed to be filled up with zeros.
const std::string kZeroVarSuffix = "@ZERO";
constexpr char kZeroVarSuffix[] = "@ZERO";
inline std::string GradVarName(const std::string& var_name) {
return var_name + kGradVarSuffix;
......@@ -88,6 +88,8 @@ class OperatorBase {
virtual bool IsNetOp() const { return false; }
virtual bool SupportGPU() const { return false; }
/// rename inputs outputs name
void Rename(const std::string& old_name, const std::string& new_name);
......@@ -118,10 +120,10 @@ class OperatorBase {
std::shared_ptr<std::unordered_map<std::string, int>> in_out_idxs_;
};
class OperatorContext {
class InferShapeContext {
public:
OperatorContext(const OperatorBase* op, const Scope& scope)
: op_(*op), scope_(scope) {}
InferShapeContext(const OperatorBase& op, const Scope& scope)
: op_(op), scope_(scope) {}
size_t InputSize() const { return op_.inputs_.size(); }
......@@ -232,12 +234,6 @@ class OperatorContext {
const Scope& scope_;
};
class InferShapeContext : public OperatorContext {
public:
InferShapeContext(const OperatorBase* op, const Scope& scope)
: OperatorContext(op, scope) {}
};
template <typename T>
struct EigenDeviceConverter;
......@@ -253,11 +249,11 @@ struct EigenDeviceConverter<platform::GPUPlace> {
};
#endif
class ExecutionContext : public OperatorContext {
class ExecutionContext : public InferShapeContext {
public:
ExecutionContext(const OperatorBase* op, const Scope& scope,
ExecutionContext(const OperatorBase& op, const Scope& scope,
const platform::DeviceContext* device_context)
: OperatorContext(op, scope), device_context_(device_context) {}
: InferShapeContext(op, scope), device_context_(device_context) {}
template <typename PlaceType,
typename DeviceType =
......@@ -308,14 +304,14 @@ class OperatorWithKernel : public OperatorBase {
using OpKernelMap =
std::unordered_map<OpKernelKey, std::unique_ptr<OpKernel>, OpKernelHash>;
void InferShape(const Scope& scope) const {
InferShape(InferShapeContext(this, scope));
void InferShape(const Scope& scope) const override {
InferShape(InferShapeContext(*this, scope));
}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const final {
auto& opKernel = AllOpKernels().at(type_).at(OpKernelKey(dev_ctx));
opKernel->Compute(ExecutionContext(this, scope, &dev_ctx));
opKernel->Compute(ExecutionContext(*this, scope, &dev_ctx));
}
static std::unordered_map<std::string /* op_type */, OpKernelMap>&
......@@ -324,6 +320,12 @@ class OperatorWithKernel : public OperatorBase {
return g_all_op_kernels;
}
bool SupportGPU() const override {
OperatorWithKernel::OpKernelKey key;
key.place_ = platform::GPUPlace();
return OperatorWithKernel::AllOpKernels().at(type_).count(key) != 0;
}
protected:
virtual void InferShape(const InferShapeContext& ctx) const = 0;
};
......
......@@ -18,13 +18,11 @@ limitations under the License. */
#include "paddle/framework/backward.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/scope.h"
#include "paddle/framework/tensor_py.h"
#include "paddle/operators/net_op.h"
#include "paddle/operators/type_alias.h"
#include "paddle/platform/enforce.h"
#include "paddle/platform/place.h"
#include "paddle/string/to_string.h"
#include "pybind11/numpy.h"
#include "pybind11/pybind11.h"
#include "pybind11/stl.h"
......@@ -42,8 +40,14 @@ USE_OP(softmax);
USE_OP(rowwise_add);
USE_OP(fill_zeros_like);
USE_OP_WITHOUT_KERNEL(recurrent_op);
USE_OP(gaussian_random);
USE_OP(uniform_random);
namespace paddle {
namespace framework {
using Tensor = framework::Tensor;
template <typename ClassType>
void ExposeOperator(ClassType &m) {
m.def("infer_shape", &ClassType::type::InferShape)
......@@ -56,6 +60,26 @@ void ExposeOperator(ClassType &m) {
[](const typename ClassType::type &op) -> std::vector<std::string> {
return op.outputs_;
})
.def("inputs",
[](const typename ClassType::type &op) -> std::vector<std::string> {
return op.inputs_;
})
.def("support_gpu", &ClassType::type::SupportGPU)
.def("temp_outputs",
[](const typename ClassType::type &op) -> std::vector<std::string> {
auto iter = op.attrs_.find("temporary_index");
std::vector<std::string> ret;
if (iter == op.attrs_.end()) {
return ret;
} else {
auto tmp_idx = boost::get<std::vector<int>>(iter->second);
ret.reserve(tmp_idx.size());
for (auto &index : tmp_idx) {
ret.push_back(op.outputs_.at(index));
}
return ret;
}
})
.def("__str__", &ClassType::type::DebugString);
}
......@@ -129,8 +153,8 @@ All parameter, weight, gradient are variables in Paddle.
[](Variable &self) -> Tensor * { return self.GetMutable<Tensor>(); },
py::return_value_policy::reference)
.def("get_net",
[](Variable &self) -> ops::NetOp * {
return self.GetMutable<ops::NetOp>();
[](Variable &self) -> operators::NetOp * {
return self.GetMutable<operators::NetOp>();
},
py::return_value_policy::reference);
......@@ -184,9 +208,13 @@ All parameter, weight, gradient are variables in Paddle.
});
// clang-format on
py::class_<paddle::platform::GPUPlace>(m, "GPUPlace").def(py::init<int>());
py::class_<platform::GPUPlace>(m, "GPUPlace")
.def(py::init<int>())
.def("__str__", string::to_string<const platform::GPUPlace &>);
py::class_<paddle::platform::CPUPlace>(m, "CPUPlace").def(py::init<>());
py::class_<paddle::platform::CPUPlace>(m, "CPUPlace")
.def(py::init<>())
.def("__str__", string::to_string<const platform::CPUPlace &>);
py::class_<OperatorBase, std::shared_ptr<OperatorBase>> operator_base(
m, "Operator");
......@@ -201,8 +229,6 @@ All parameter, weight, gradient are variables in Paddle.
return OpRegistry::CreateOp(desc);
});
operator_base.def_static("support_gpu", &OpRegistry::SupportGPU);
operator_base.def("backward",
[](const OperatorBase &forwardOp,
const std::unordered_set<std::string> &no_grad_vars) {
......@@ -211,23 +237,24 @@ All parameter, weight, gradient are variables in Paddle.
ExposeOperator(operator_base);
py::class_<ops::NetOp, std::shared_ptr<ops::NetOp>> net(m, "Net");
py::class_<operators::NetOp, std::shared_ptr<operators::NetOp>> net(m, "Net");
net.def_static("create",
[]() -> std::shared_ptr<ops::NetOp> {
auto retv = std::make_shared<ops::NetOp>();
[]() -> std::shared_ptr<operators::NetOp> {
auto retv = std::make_shared<operators::NetOp>();
retv->type_ = "plain_net";
return retv;
})
.def("add_op", &ops::NetOp::AddOp)
.def(
"add_op",
[](ops::NetOp &self, const std::shared_ptr<ops::NetOp> &net) -> void {
.def("add_op", &operators::NetOp::AddOp)
.def("add_op",
[](operators::NetOp &self,
const std::shared_ptr<operators::NetOp> &net) -> void {
self.AddOp(std::static_pointer_cast<OperatorBase>(net));
})
.def("complete_add_op", &ops::NetOp::CompleteAddOp)
.def("complete_add_op",
[](std::shared_ptr<ops::NetOp> &self) { self->CompleteAddOp(); });
.def("complete_add_op", &operators::NetOp::CompleteAddOp)
.def("complete_add_op", [](std::shared_ptr<operators::NetOp> &self) {
self->CompleteAddOp();
});
ExposeOperator(net);
......
......@@ -18,6 +18,8 @@ limitations under the License. */
#include <cstring>
#include <memory>
#include <typeindex>
#include <vector>
#include "paddle/framework/ddim.h"
#include "paddle/memory/memory.h"
#include "paddle/platform/device_context.h"
......
......@@ -19,7 +19,7 @@ TEST(Tensor, Dims) {
using namespace paddle::framework;
using namespace paddle::platform;
Tensor tt;
tt.Resize(make_ddim({2, 3, 4}));
tt.Resize({2, 3, 4});
DDim dims = tt.dims();
ASSERT_EQ(arity(dims), 3);
for (int i = 0; i < 3; ++i) {
......
......@@ -93,8 +93,8 @@ TEST(Arguments, Matrix) {
MatrixPtr matrix = Matrix::create(100, 200);
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 2U);
EXPECT_EQ(arg.shape()[0], 100);
EXPECT_EQ(arg.shape()[1], 200);
EXPECT_EQ(arg.shape()[0], 100U);
EXPECT_EQ(arg.shape()[1], 200U);
EXPECT_EQ(arg.data(), matrix->getData());
EXPECT_EQ(arg.matrix<DEVICE_TYPE_CPU>().getHeight(), matrix->getHeight());
......@@ -112,8 +112,8 @@ TEST(Arguments, Matrix) {
TEST(Arguments, Vector) {
VectorPtr vector = Vector::create(100, false);
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 1);
EXPECT_EQ(arg.shape()[0], 100);
EXPECT_EQ(arg.shape().ndims(), 1U);
EXPECT_EQ(arg.shape()[0], 100U);
EXPECT_EQ(arg.data(), vector->getData());
CpuVector inVector = arg.vector<real, DEVICE_TYPE_CPU>();
......@@ -131,9 +131,9 @@ TEST(Arguments, Vector) {
TEST(Arguments, CpuSparseMatrix) {
CpuSparseMatrix sparse(200, 300, 50);
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 2);
EXPECT_EQ(arg.shape()[0], 200);
EXPECT_EQ(arg.shape()[1], 300);
EXPECT_EQ(arg.shape().ndims(), 2U);
EXPECT_EQ(arg.shape()[0], 200U);
EXPECT_EQ(arg.shape()[1], 300U);
EXPECT_EQ(arg.data(), sparse.getData());
// CHECK_EQ(arg.sparse().nnz(), 50);
// CHECK_EQ(arg.sparse().dataFormat(), SPARSE_CSR_FORMAT);
......@@ -152,10 +152,10 @@ TEST(Arguments, CpuSparseMatrix) {
TEST(Arguments, BufferArg) {
BufferArg arg(nullptr, VALUE_TYPE_FLOAT, {1, 2, 3});
CheckBufferArg check = [=](const BufferArg& arg) {
EXPECT_EQ(arg.shape().ndims(), 3);
EXPECT_EQ(arg.shape()[0], 1);
EXPECT_EQ(arg.shape()[1], 2);
EXPECT_EQ(arg.shape()[2], 3);
EXPECT_EQ(arg.shape().ndims(), 3U);
EXPECT_EQ(arg.shape()[0], 1U);
EXPECT_EQ(arg.shape()[1], 2U);
EXPECT_EQ(arg.shape()[2], 3U);
};
BufferArgs argments;
......
......@@ -44,7 +44,7 @@ TEST(TensorShape, GetAndSet) {
EXPECT_EQ(t.ndims(), 3U);
EXPECT_EQ(t.getElements(), 6U);
EXPECT_EQ(t[1], 2);
EXPECT_EQ(t[1], 2U);
t.setDim(1, 100);
EXPECT_EQ(t.getElements(), 300U);
EXPECT_EQ(t[1], 100U);
......
......@@ -98,10 +98,12 @@ void KmaxSeqScoreLayer::forward(PassType passType) {
}
// TODO(caoying)
// Here selSubSeqIdx is automatically converted from real to int
// This is very dangerous if user fill this matrix himself, invalid data may
// occur. The selected indices should be stored in
// CpuSparseMatrix with SparseValueType set to NO_VALUE.
// In PaddlePaddle, the currently available matrixes all a have real-typed
// data field, but the selected indices information are actually int-typed
// (with -1 as a special token). Storing indices information in real-typed
// Matrix leads to converting real to int. This is very dangerous if a user
// fills this matrix himself, invalid data may occur.
// The selected indices should be stored in an int-typed matrix.
Matrix::resizeOrCreate(
output_.value,
input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(),
......
......@@ -32,10 +32,13 @@ public:
private:
// TODO(caoying)
// Here selSubSeqIdx is automatically converted from real to int
// This is very dangerous if user fill this matrix himself, invalid data
// may occur. The selected indices should be stored in CpuSparseMatrix
// with SparseValueType set to NO_VALUE.
// In PaddlePaddle, the currently available matrixes all a have real-typed
// data field, but the selected indices information are actually int-typed
// (with -1 as a special token). Storing indices information in real-typed
// Matrix leads to converting real to int. This is very dangerous if a user
// fills this matrix himself, invalid data may occur.
// The selected indices should be stored in an int-typed matrix.
MatrixPtr startIdsOnCpu_;
MatrixPtr endIdsOnCpu_;
......
......@@ -59,6 +59,13 @@ private:
const std::vector<std::vector<int>>& inputSeqInfo);
// if the second input of this layer is on GPU memory, copy it to CPU memory.
// TODO(caoying)
// In PaddlePaddle, the currently available matrixes all a have real-typed
// data field, but the selected indices information are actually int-typed
// (with -1 as a special token). Storing indices information in real-typed
// Matrix leads to converting real to int. This is very dangerous if a user
// fills this matrix himself, invalid data may occur.
// The selected indices should be stored in an int-typed matrix.
MatrixPtr selIdsCpu_;
// reorganized sequenceStartPositions and subSequenceStartPositions
......@@ -95,12 +102,7 @@ void SubNestedSequenceLayer::calSelectedRows(
for (size_t i = 0; i < seqNum; ++i) {
for (size_t j = 0; j < beamSize; ++j) {
if (selectedIndices->getElement(i, j) == -1.) break;
// TODO(caoying)
// Here selSubSeqIdx is automatically converted from real to int
// This is very dangerous if user fill this matrix himself, invalid data
// may occur. The selected indices should be stored in
// CpuSparseMatrix with SparseValueType set to NO_VALUE.
int selSubSeqIdx = selectedIndices->getElement(i, j);
size_t selSubSeqIdx = selectedIndices->getElement(i, j);
CHECK_GT(inputSeqInfoVec_[i].size() - 1, selSubSeqIdx);
size_t subSeqLen = inputSeqInfoVec_[i][selSubSeqIdx + 1] -
......@@ -139,7 +141,7 @@ void SubNestedSequenceLayer::forward(PassType passType) {
CHECK(inputSeq.hasSubseq()) << "The first input of SubNestSequence layer "
<< "must be a nested sequence.";
const MatrixPtr selectedIndices = getInputValue(1);
CHECK_EQ(inputSeq.getNumSequences(), selectedIndices->getHeight());
CHECK_EQ(size_t(inputSeq.getNumSequences()), selectedIndices->getHeight());
if (dynamic_cast<GpuMatrix*>(selectedIndices.get())) {
/*
......
......@@ -88,7 +88,7 @@ void checkLayerOut(vector<vector<int>> groundTruth,
TEST(Layer, kmaxSeqScoreLayer) {
const size_t maxBeamSize = 100;
int beamSize = 1 + (rand() % maxBeamSize);
size_t beamSize = 1 + (rand() % maxBeamSize);
vector<int> seqStartPosition;
vector<int> subSeqStartPosition;
......
......@@ -45,16 +45,15 @@ cc_library(net_op SRCS net_op.cc DEPS op_registry)
cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
op_library(add_op SRCS add_op.cc add_op.cu)
cc_test(add_op_test SRCS add_op_test.cc DEPS add_op)
op_library(mean_op SRCS mean_op.cc mean_op.cu)
cc_test(mean_op_test SRCS mean_op_test.cc DEPS mean_op)
op_library(mul_op SRCS mul_op.cc mul_op.cu)
op_library(rowwise_add_op SRCS rowwise_add_op.cu rowwise_add_op.cc)
op_library(sigmoid_op SRCS sigmoid_op.cc sigmoid_op.cu)
op_library(softmax_op SRCS softmax_op.cc softmax_op.cu)
op_library(gaussian_random_op SRCS gaussian_random_op.cc gaussian_random_op.cu)
op_library(cross_entropy_op SRCS cross_entropy_op.cc cross_entropy_op.cu)
op_library(fill_zeros_like_op SRCS fill_zeros_like_op.cc fill_zeros_like_op.cu)
......@@ -66,3 +65,5 @@ op_library(fc_op
op_library(recurrent_op SRCS recurrent_op.cc rnn/recurrent_op_utils.cc
DEPS op_desc tensor op_registry operator net_op)
cc_test(recurrent_op_test SRCS recurrent_op_test.cc DEPS recurrent_op gtest mul_op add_op)
op_library(uniform_random_op
SRCS uniform_random_op.cc uniform_random_op.cu)
......@@ -17,9 +17,9 @@ limitations under the License. */
namespace paddle {
namespace operators {
class AddOp : public OperatorWithKernel {
class AddOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.InputSize(), 2);
PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1);
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "Inputs of AddOp must all be set");
......@@ -31,9 +31,9 @@ class AddOp : public OperatorWithKernel {
}
};
class AddOpMaker : public OpProtoAndCheckerMaker {
class AddOpMaker : public framework::OpProtoAndCheckerMaker {
public:
AddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of add op");
AddInput("Y", "The second input of add op");
......@@ -46,14 +46,17 @@ The equation is: Out = X + Y
}
};
class AddOpGrad : public OperatorWithKernel {
class AddOpGrad : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {}
void InferShape(const framework::InferShapeContext &ctx) const override {}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
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>);
REGISTER_OP_CPU_KERNEL(add_two,
ops::AddKernel<paddle::platform::CPUPlace, float>);
......@@ -16,4 +16,6 @@
#include "paddle/framework/op_registry.h"
#include "paddle/operators/add_op.h"
REGISTER_OP_GPU_KERNEL(add_two, ops::AddKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(add_two,
ops::AddKernel<paddle::platform::GPUPlace, float>);
......@@ -13,15 +13,21 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename Place, typename T>
class AddKernel : public OpKernel {
class AddKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
auto input0 = context.Input<Tensor>(0);
auto input1 = context.Input<Tensor>(1);
auto output = context.Output<Tensor>(0);
......
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <gtest/gtest.h>
#define private public
#include <paddle/framework/op_registry.h>
USE_OP(add_two);
// USE_OP(add_two_grad);
TEST(AddOp, GetOpProto) {
auto& protos = paddle::framework::OpRegistry::protos();
auto it = protos.find("add_two");
ASSERT_NE(it, protos.end());
auto& op_creators = paddle::framework::OpRegistry::op_creators();
auto it1 = op_creators.find("add_two_grad");
ASSERT_NE(it1, op_creators.end());
}
......@@ -17,9 +17,9 @@ limitations under the License. */
namespace paddle {
namespace operators {
class OnehotCrossEntropyOp : public OperatorWithKernel {
class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.InputSize(), 2,
"Input size of OnehotCrossEntropyOp must be two");
PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1,
......@@ -37,9 +37,9 @@ class OnehotCrossEntropyOp : public OperatorWithKernel {
}
};
class OnehotCrossEntropyGradientOp : public OperatorWithKernel {
class OnehotCrossEntropyGradientOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
auto X_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto X = ctx.Input<Tensor>("X");
......@@ -48,9 +48,10 @@ class OnehotCrossEntropyGradientOp : public OperatorWithKernel {
}
};
class OnehotCrossEntropyOpMaker : public OpProtoAndCheckerMaker {
class OnehotCrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
public:
OnehotCrossEntropyOpMaker(OpProto *proto, OpAttrChecker *op_checker)
OnehotCrossEntropyOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of OnehotCrossEntropyOp");
AddInput("label", "The second input of OnehotCrossEntropyOp");
......@@ -66,11 +67,14 @@ OnehotCrossEntropy Operator.
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(onehot_cross_entropy, ops::OnehotCrossEntropyOp,
ops::OnehotCrossEntropyOpMaker);
REGISTER_OP_CPU_KERNEL(onehot_cross_entropy,
ops::OnehotCrossEntropyOpKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
onehot_cross_entropy,
ops::OnehotCrossEntropyOpKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(onehot_cross_entropy, onehot_cross_entropy_grad,
ops::OnehotCrossEntropyGradientOp);
REGISTER_OP_CPU_KERNEL(
onehot_cross_entropy_grad,
ops::OnehotCrossEntropyGradientOpKernel<ops::CPUPlace, float>);
ops::OnehotCrossEntropyGradientOpKernel<paddle::platform::CPUPlace, float>);
......@@ -14,3 +14,8 @@
#define EIGEN_USE_GPU
#include "paddle/operators/cross_entropy_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
onehot_cross_entropy,
ops::OnehotCrossEntropyOpKernel<paddle::platform::GPUPlace, float>);
......@@ -13,11 +13,13 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T>
T tolerable_value(T x) {
static_assert(std::is_floating_point<T>::value,
......@@ -38,9 +40,9 @@ T tolerable_value(T x) {
}
template <typename Place, typename T>
class OnehotCrossEntropyOpKernel : public OpKernel {
class OnehotCrossEntropyOpKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& ctx) const override {
void Compute(const framework::ExecutionContext& ctx) const override {
auto X = ctx.Input<Tensor>("X");
const T* Xdata = X->data<T>();
const int* label_data = ctx.Input<Tensor>(1)->data<int>();
......@@ -61,9 +63,9 @@ class OnehotCrossEntropyOpKernel : public OpKernel {
};
template <typename Place, typename T>
class OnehotCrossEntropyGradientOpKernel : public OpKernel {
class OnehotCrossEntropyGradientOpKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& ctx) const override {
void Compute(const framework::ExecutionContext& ctx) const override {
auto X = ctx.Input<Tensor>("X");
auto dX = ctx.Output<Tensor>(framework::GradVarName("X"));
auto dY = ctx.Input<Tensor>(framework::GradVarName("Y"));
......
......@@ -12,11 +12,16 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include "type_alias.h"
#include "paddle/operators/net_op.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using OpRegistry = framework::OpRegistry;
class FullyConnectedOp : public NetOp {
public:
void Init() override {
......@@ -39,9 +44,10 @@ class FullyConnectedOp : public NetOp {
}
};
class FullyConnectedOpMaker : public OpProtoAndCheckerMaker {
class FullyConnectedOpMaker : public framework::OpProtoAndCheckerMaker {
public:
FullyConnectedOpMaker(OpProto *proto, OpAttrChecker *op_checker)
FullyConnectedOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "the input of fc operator");
AddInput("W", "the weight of fc operator");
......@@ -66,4 +72,5 @@ USE_OP(rowwise_add);
USE_OP(sigmoid);
USE_OP(softmax);
namespace ops = paddle::operators;
REGISTER_OP(fc, ops::FullyConnectedOp, ops::FullyConnectedOpMaker);
......@@ -50,8 +50,8 @@ The output will have the same size with input.
} // namespace operators
} // namespace paddle
REGISTER_OP(fill_zeros_like, paddle::operators::FillZerosLikeOp,
paddle::operators::FillZerosLikeOpMaker);
namespace ops = paddle::operators;
REGISTER_OP(fill_zeros_like, ops::FillZerosLikeOp, ops::FillZerosLikeOpMaker);
REGISTER_OP_CPU_KERNEL(
fill_zeros_like,
paddle::operators::FillZerosLikeKernel<paddle::platform::CPUPlace, float>);
ops::FillZerosLikeKernel<paddle::platform::CPUPlace, float>);
......@@ -16,6 +16,7 @@
#include "paddle/framework/op_registry.h"
#include "paddle/operators/fill_zeros_like_op.h"
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
fill_zeros_like,
paddle::operators::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
ops::FillZerosLikeKernel<paddle::platform::GPUPlace, float>);
......@@ -13,7 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
......
/* 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. */
#include <random>
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
template <typename T>
class GaussianRandomKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
float mean = context.op_.GetAttr<float>("mean");
float std = context.op_.GetAttr<float>("std");
auto* tensor = context.Output<framework::Tensor>(0);
T* data = tensor->mutable_data<T>(context.GetPlace());
// TODO(dzh): attribute does not support unsigned int.
// And we need a global random seed configuration.
int seed = context.op_.GetAttr<int>("seed");
if (seed == 0) {
seed = std::random_device()();
}
std::mt19937 g(seed);
std::normal_distribution<T> distribution(mean, std);
ssize_t size = framework::product(tensor->dims());
for (int i = 0; i < size; ++i) {
data[i] = distribution(g);
}
}
};
class GaussianRandomOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext& context) const override {
auto* tensor = context.Output<framework::Tensor>(0);
auto dims = GetAttr<std::vector<int>>("dims");
PADDLE_ENFORCE(dims.size() > 0UL,
"dims can be one int or array. dims must be set.");
tensor->Resize(framework::make_ddim(dims));
}
};
class GaussianRandomOpMaker : public framework::OpProtoAndCheckerMaker {
public:
GaussianRandomOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
AddOutput("Out", "output matrix of random op");
AddComment(R"DOC(
GaussianRandom operator.
Use to initialize tensor with gaussian random generator.
)DOC");
AddAttr<std::vector<int>>("dims", "The dimension of random tensor.");
AddAttr<float>("mean", "mean value of random.").SetDefault(.0f);
AddAttr<float>("std", "minimum value of random value.").SetDefault(1.0f);
AddAttr<int>("seed",
"Random seed of generator."
"0 means use system wide seed")
.SetDefault(0);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(gaussian_random, ops::GaussianRandomOp, ops::GaussianRandomOpMaker);
REGISTER_OP_CPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);
......@@ -12,44 +12,41 @@
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <memory>
#include <random>
#include "paddle/platform/dynload/curand.h"
#include "paddle/platform/gpu_info.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
namespace paddle {
namespace operators {
using OpKernel = framework::OpKernel;
using OperatorBase = framework::OperatorBase;
using InferShapeContext = framework::InferShapeContext;
using ExecutionContext = framework::ExecutionContext;
using Variable = framework::Variable;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
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 Scope = framework::Scope;
using OperatorWithKernel = framework::OperatorWithKernel;
using OperatorBase = framework::OperatorBase;
using OpProtoAndCheckerMaker = framework::OpProtoAndCheckerMaker;
using OpProto = framework::OpProto;
using OpAttrChecker = framework::OpAttrChecker;
using CPUPlace = platform::CPUPlace;
using GPUPlace = platform::GPUPlace;
using OpRegistry = framework::OpRegistry;
template <typename T>
class GaussianRandomKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
float mean = context.op_.GetAttr<float>("mean");
float std = context.op_.GetAttr<float>("std");
auto* tensor = context.Output<framework::Tensor>(0);
T* data = tensor->mutable_data<T>(context.GetPlace());
int seed = context.op_.GetAttr<int>("seed");
if (seed == 0) {
seed = std::random_device()();
}
curandGenerator_t g;
PADDLE_ENFORCE(platform::dynload::curandCreateGenerator(
&g, CURAND_RNG_PSEUDO_DEFAULT));
PADDLE_ENFORCE(
platform::dynload::curandSetPseudoRandomGeneratorSeed(g, seed));
curandGenerateNormal(g, data, framework::product(tensor->dims()), mean,
std);
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(gaussian_random, ops::GaussianRandomKernel<float>);
\ No newline at end of file
......@@ -17,9 +17,9 @@ limitations under the License. */
namespace paddle {
namespace operators {
class MeanOp : public OperatorWithKernel {
class MeanOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.InputSize(), 1, "Input size of AddOp must be one");
PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of AddOp must be one");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "input should be set");
......@@ -28,9 +28,9 @@ class MeanOp : public OperatorWithKernel {
}
};
class MeanOpMaker : public OpProtoAndCheckerMaker {
class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
public:
MeanOpMaker(OpProto *proto, OpAttrChecker *op_checker)
MeanOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The input of mean op");
AddOutput("Out", "The output of mean op").IgnoreGradient();
......@@ -38,10 +38,10 @@ class MeanOpMaker : public OpProtoAndCheckerMaker {
}
};
class MeanGradOp : public OperatorWithKernel {
class MeanGradOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
ctx.Output<Tensor>("X" + framework::kGradVarSuffix)
void InferShape(const framework::InferShapeContext &ctx) const override {
ctx.Output<Tensor>(framework::GradVarName("X"))
->Resize(ctx.Input<Tensor>("X")->dims());
}
};
......@@ -49,7 +49,10 @@ class MeanGradOp : public OperatorWithKernel {
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(mean, ops::MeanOp, ops::MeanOpMaker);
REGISTER_OP_CPU_KERNEL(mean, ops::MeanKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mean,
ops::MeanKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(mean, mean_grad, ops::MeanGradOp);
REGISTER_OP_CPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mean_grad,
ops::MeanGradKernel<paddle::platform::CPUPlace, float>);
......@@ -16,5 +16,8 @@
#include "paddle/operators/mean_op.h"
REGISTER_OP_GPU_KERNEL(mean, ops::MeanKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mean_grad, ops::MeanGradKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(mean,
ops::MeanKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(mean_grad,
ops::MeanGradKernel<paddle::platform::GPUPlace, float>);
......@@ -13,15 +13,24 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenScalar = framework::EigenScalar<T, MajorType, IndexType>;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename Place, typename T>
class MeanKernel : public OpKernel {
class MeanKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
auto input = context.Input<Tensor>(0);
auto output = context.Output<Tensor>(0);
......@@ -36,13 +45,13 @@ class MeanKernel : public OpKernel {
};
template <typename Place, typename T>
class MeanGradKernel : public OpKernel {
class MeanGradKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
auto OG = context.Input<Tensor>("Out" + framework::kGradVarSuffix);
void Compute(const framework::ExecutionContext& context) const override {
auto OG = context.Input<Tensor>(framework::GradVarName("Out"));
PADDLE_ENFORCE(framework::product(OG->dims()) == 1,
"Mean Gradient should be scalar");
auto IG = context.Output<Tensor>("X" + framework::kGradVarSuffix);
auto IG = context.Output<Tensor>(framework::GradVarName("X"));
IG->mutable_data<T>(context.GetPlace());
T ig_size = (T)framework::product(IG->dims());
......
/* 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. */
#include <gtest/gtest.h>
#include <paddle/framework/op_registry.h>
USE_OP(mean);
TEST(MeanOp, GetOpProto) {
auto& protos = paddle::framework::OpRegistry::protos();
auto it = protos.find("mean");
ASSERT_NE(it, protos.end());
}
......@@ -17,9 +17,9 @@
namespace paddle {
namespace operators {
class MulOp : public OperatorWithKernel {
class MulOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE(ctx.InputSize() == 2, "The mul op must take two inputs");
auto dim0 = ctx.Input<Tensor>(0)->dims();
auto dim1 = ctx.Input<Tensor>(1)->dims();
......@@ -37,9 +37,9 @@ class MulOp : public OperatorWithKernel {
}
};
class MulOpMaker : public OpProtoAndCheckerMaker {
class MulOpMaker : public framework::OpProtoAndCheckerMaker {
public:
MulOpMaker(OpProto *proto, OpAttrChecker *op_checker)
MulOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "The first input of mul op");
AddInput("Y", "The second input of mul op");
......@@ -52,9 +52,9 @@ The equation is: Out = X * Y
}
};
class MulOpGrad : public OperatorWithKernel {
class MulOpGrad : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {}
void InferShape(const framework::InferShapeContext &ctx) const override {}
std::string DebugString() const override {
LOG(INFO) << "MulGrad";
return "";
......@@ -64,7 +64,8 @@ class MulOpGrad : public OperatorWithKernel {
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker);
REGISTER_GRADIENT_OP(mul, mul_grad, ops::MulOpGrad);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
......@@ -15,4 +15,6 @@
#define EIGEN_USE_GPU
#include "paddle/operators/mul_op.h"
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(mul, ops::MulKernel<paddle::platform::GPUPlace, float>);
......@@ -13,16 +13,21 @@
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename Place, typename T>
class MulKernel : public OpKernel {
class MulKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
Eigen::array<Eigen::IndexPair<Eigen::DenseIndex>, 1> dim_pair = {
{Eigen::IndexPair<Eigen::DenseIndex>(1, 0)}};
......@@ -40,5 +45,6 @@ class MulKernel : public OpKernel {
Z.device(place) = X.contract(Y, dim_pair);
}
};
} // namespace operators
} // namespace paddle
......@@ -15,7 +15,6 @@
*/
#include "paddle/operators/net_op.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
......
......@@ -14,13 +14,7 @@ limitations under the License. */
#pragma once
#include "paddle/framework/op_desc.pb.h"
#include "paddle/framework/op_proto.pb.h"
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/scope.h"
#include "paddle/operators/type_alias.h"
#include "paddle/platform/device_context.h"
namespace paddle {
namespace operators {
......@@ -65,6 +59,15 @@ class NetOp : public framework::OperatorBase {
}
}
bool SupportGPU() const override {
for (auto& op : ops_) {
if (!op->SupportGPU()) {
return false;
}
}
return true;
}
/**
* @brief Add an operator by ptr
*/
......
......@@ -2,31 +2,27 @@
#include <gtest/gtest.h>
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
namespace paddle {
namespace operators {
using Scope = framework::Scope;
using DeviceContext = platform::DeviceContext;
static int infer_shape_cnt = 0;
static int run_cnt = 0;
class TestOp : public OperatorBase {
class TestOp : public framework::OperatorBase {
public:
void InferShape(const framework::Scope& scope) const override {
++infer_shape_cnt;
}
void Run(const framework::Scope& scope,
const paddle::platform::DeviceContext& dev_ctx) const override {
void InferShape(const Scope& scope) const override { ++infer_shape_cnt; }
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {
++run_cnt;
}
};
class EmptyOp : public OperatorBase {
class EmptyOp : public framework::OperatorBase {
public:
void InferShape(const Scope& scope) const override {}
void Run(const Scope& scope,
const platform::DeviceContext& dev_ctx) const override {}
void Run(const Scope& scope, const DeviceContext& dev_ctx) const override {}
};
template <typename T>
......@@ -72,7 +68,7 @@ TEST(OpKernel, all) {
net->Run(scope, dev_ctx);
ASSERT_EQ(2, infer_shape_cnt);
ASSERT_EQ(2, run_cnt);
ASSERT_THROW(net->AddOp(op2), paddle::platform::EnforceNotMet);
ASSERT_THROW(net->AddOp(op2), platform::EnforceNotMet);
}
TEST(NetOp, insert_op) {
......
......@@ -14,17 +14,19 @@
#include "paddle/operators/recurrent_op.h"
#include <glog/logging.h>
#include <cstring>
#include <sstream>
#include "paddle/framework/op_registry.h"
#include "paddle/operators/net_op.h"
#include "paddle/platform/enforce.h"
namespace paddle {
namespace operators {
using Scope = framework::Scope;
using Variable = framework::Variable;
using Tensor = framework::Tensor;
void RecurrentAlgorithm::InferShape(const Scope& scope) const {
seq_len_ = scope.FindVar((arg_->inlinks[0]).external)
->GetMutable<Tensor>()
......@@ -135,10 +137,11 @@ void RecurrentOp::Init() {
alg_.Init(std::move(arg));
}
class RecurrentAlgorithmProtoAndCheckerMaker : public OpProtoAndCheckerMaker {
class RecurrentAlgorithmProtoAndCheckerMaker
: public framework::OpProtoAndCheckerMaker {
public:
RecurrentAlgorithmProtoAndCheckerMaker(OpProto* proto,
OpAttrChecker* op_checker)
RecurrentAlgorithmProtoAndCheckerMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
const auto& name = RecurrentOp::kArgName;
// inputs and outputs stored in proto
......
......@@ -27,6 +27,10 @@ namespace operators {
using framework::make_ddim;
using framework::DDim;
using framework::Tensor;
using framework::Variable;
using framework::Scope;
using framework::OpRegistry;
class RecurrentOpTest : public ::testing::Test {
protected:
......@@ -164,7 +168,7 @@ class RecurrentOpTest : public ::testing::Test {
// father scope
Scope scope_;
std::shared_ptr<OperatorBase> rnn_op_;
std::shared_ptr<framework::OperatorBase> rnn_op_;
};
TEST_F(RecurrentOpTest, Run) {
......
......@@ -18,7 +18,9 @@ namespace paddle {
namespace operators {
namespace rnn {
namespace fmw = paddle::framework;
namespace f = paddle::framework;
using Tensor = framework::Tensor;
void SegmentInputs(const std::vector<Scope*>& step_scopes,
const std::vector<Link>& inlinks, const size_t seq_len,
......@@ -30,10 +32,10 @@ void SegmentInputs(const std::vector<Scope*>& step_scopes,
inlinks[i].external);
Tensor* input = input_var->GetMutable<Tensor>();
fmw::DDim dims = input->dims();
f::DDim dims = input->dims();
PADDLE_ENFORCE(static_cast<size_t>(dims[0]) == seq_len,
"all the inlinks must have same length");
fmw::DDim step_dims = slice_ddim(dims, 1, dims.size());
f::DDim step_dims = slice_ddim(dims, 1, dims.size());
for (size_t j = 0; j < seq_len; j++) {
Tensor* step_input =
step_scopes[j]->NewVar(inlinks[i].internal)->GetMutable<Tensor>();
......@@ -58,11 +60,10 @@ void ConcatOutputs(const std::vector<Scope*>& step_scopes,
auto step_scope_var = step_scopes[0]->FindVar(outlinks[i].internal);
PADDLE_ENFORCE(step_scope_var != nullptr, "%s not in scope",
outlinks[i].internal);
fmw::DDim step_dims =
step_scope_var->template GetMutable<Tensor>()->dims();
f::DDim step_dims = step_scope_var->template GetMutable<Tensor>()->dims();
std::vector<int> dims_vec = vectorize(step_dims);
dims_vec.insert(dims_vec.begin(), seq_len);
output->Resize(fmw::make_ddim(dims_vec));
output->Resize(f::make_ddim(dims_vec));
} else {
output->mutable_data<float>(platform::CPUPlace());
for (size_t j = 0; j < seq_len; j++) {
......@@ -104,7 +105,7 @@ void LinkMemories(const std::vector<Scope*>& scopes,
}
void InitArgument(const ArgumentName& name, Argument* arg,
const OperatorBase& op) {
const framework::OperatorBase& op) {
arg->step_net = op.Input(name.step_net);
arg->step_scopes = op.Output(name.step_scopes);
......
......@@ -17,12 +17,13 @@
#include <string>
#include "paddle/framework/operator.h"
#include "paddle/operators/type_alias.h"
namespace paddle {
namespace operators {
namespace rnn {
using Scope = framework::Scope;
/**
* Memory of a RNN (same as the role of `Momory` in PaddlePaddle).
*
......@@ -86,7 +87,7 @@ void LinkMemories(const std::vector<Scope*>& step_scopes,
const int offset, bool infer_shape_mode);
void InitArgument(const ArgumentName& name, Argument* arg,
const OperatorBase& op);
const framework::OperatorBase& op);
} // namespace rnn
} // namespace operators
......
......@@ -13,12 +13,13 @@
limitations under the License. */
#include "paddle/operators/rowwise_add_op.h"
namespace paddle {
namespace operators {
class RowWiseAddOp : public OperatorWithKernel {
class RowWiseAddOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE(ctx.InputSize() == 2UL,
"Two inputs is needed by rowwise add");
auto dim0 = ctx.Input<Tensor>(0)->dims();
......@@ -32,9 +33,10 @@ class RowWiseAddOp : public OperatorWithKernel {
}
};
class RowWiseAddOpMaker : public OpProtoAndCheckerMaker {
class RowWiseAddOpMaker : public framework::OpProtoAndCheckerMaker {
public:
RowWiseAddOpMaker(OpProto *proto, OpAttrChecker *op_checker)
RowWiseAddOpMaker(framework::OpProto *proto,
framework::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");
......@@ -50,6 +52,7 @@ for i in xrange(X.shape[0]):
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(rowwise_add, ops::RowWiseAddOp, ops::RowWiseAddOpMaker);
REGISTER_OP_CPU_KERNEL(rowwise_add,
ops::RowWiseAddKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
rowwise_add, ops::RowWiseAddKernel<paddle::platform::CPUPlace, float>);
......@@ -15,5 +15,6 @@
#define EIGEN_USE_GPU
#include "paddle/operators/rowwise_add_op.h"
REGISTER_OP_GPU_KERNEL(rowwise_add,
ops::RowWiseAddKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(
rowwise_add, ops::RowWiseAddKernel<paddle::platform::GPUPlace, float>);
......@@ -13,15 +13,24 @@
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
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 Place, typename T>
class RowWiseAddKernel : public OpKernel {
class RowWiseAddKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
auto out = context.Output<Tensor>(0);
out->mutable_data<T>(context.GetPlace());
......
......@@ -17,9 +17,9 @@ limitations under the License. */
namespace paddle {
namespace operators {
class SGDOp : public OperatorWithKernel {
class SGDOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.InputSize(), 2, "Input size of SGDOp must be two");
PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1, "Output size of SGDOp must be one");
PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(0), "inputs[0] mast be set");
......@@ -31,9 +31,9 @@ class SGDOp : public OperatorWithKernel {
}
};
class SGDOpMaker : public OpProtoAndCheckerMaker {
class SGDOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SGDOpMaker(OpProto *proto, OpAttrChecker *op_checker)
SGDOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("param", "input parameter");
AddInput("grad", "input gradient");
......@@ -51,5 +51,7 @@ param_out = param - learning_rate * grad;
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(sgd, ops::SGDOp, ops::SGDOpMaker);
REGISTER_OP_CPU_KERNEL(sgd, ops::SGDOpKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(sgd,
ops::SGDOpKernel<paddle::platform::CPUPlace, float>);
......@@ -15,4 +15,6 @@
#define EIGEN_USE_GPU
#include "paddle/operators/sgd_op.h"
REGISTER_OP_GPU_KERNEL(sgd, ops::SGDOpKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(sgd,
ops::SGDOpKernel<paddle::platform::GPUPlace, float>);
......@@ -13,15 +13,21 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename Place, typename T>
class SGDOpKernel : public OpKernel {
class SGDOpKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& ctx) const override {
void Compute(const framework::ExecutionContext& ctx) const override {
auto param = ctx.Input<Tensor>("param");
auto grad = ctx.Input<Tensor>("grad");
auto param_out = ctx.Output<Tensor>(0);
......
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <gtest/gtest.h>
#include <paddle/framework/op_registry.h>
USE_OP(sgd);
TEST(SGDOp, GetOpProto) {
auto& protos = paddle::framework::OpRegistry::protos();
auto it = protos.find("sgd");
ASSERT_NE(it, protos.end());
}
......@@ -13,21 +13,23 @@
limitations under the License. */
#include "paddle/operators/sigmoid_op.h"
namespace paddle {
namespace operators {
class SigmoidOp : public OperatorWithKernel {
class SigmoidOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE(ctx.InputSize() == 1, "Sigmoid Op only have one input");
PADDLE_ENFORCE(ctx.OutputSize() == 1, "Sigmoid Op only have one output");
ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
}
};
class SigmoidOpMaker : public OpProtoAndCheckerMaker {
class SigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SigmoidOpMaker(OpProto *proto, OpAttrChecker *op_checker)
SigmoidOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "sigmoid input");
AddOutput("Y", "sigmoid output");
......@@ -35,9 +37,9 @@ class SigmoidOpMaker : public OpProtoAndCheckerMaker {
}
};
class SigmoidOpGrad : public OperatorWithKernel {
class SigmoidOpGrad : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
ctx.Output<Tensor>(0)->Resize(ctx.Input<Tensor>(0)->dims());
}
};
......@@ -45,9 +47,11 @@ class SigmoidOpGrad : public OperatorWithKernel {
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(sigmoid, ops::SigmoidOp, ops::SigmoidOpMaker);
REGISTER_GRADIENT_OP(sigmoid, sigmoid_grad, ops::SigmoidOpGrad);
REGISTER_OP_CPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(sigmoid_grad,
ops::SigmoidGradKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(sigmoid,
ops::SigmoidKernel<paddle::platform::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
sigmoid_grad, ops::SigmoidGradKernel<paddle::platform::CPUPlace, float>);
......@@ -15,6 +15,9 @@
#define EIGEN_USE_GPU
#include "paddle/operators/sigmoid_op.h"
REGISTER_OP_GPU_KERNEL(sigmoid, ops::SigmoidKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(sigmoid_grad,
ops::SigmoidGradKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(sigmoid,
ops::SigmoidKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
sigmoid_grad, ops::SigmoidGradKernel<paddle::platform::GPUPlace, float>);
......@@ -13,16 +13,21 @@
limitations under the License. */
#pragma once
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
template <typename Place, typename T>
class SigmoidKernel : public OpKernel {
class SigmoidKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
auto input = context.Input<Tensor>(0);
auto output = context.Output<Tensor>(0);
output->mutable_data<T>(context.GetPlace());
......@@ -37,9 +42,9 @@ class SigmoidKernel : public OpKernel {
};
template <typename Place, typename T>
class SigmoidGradKernel : public OpKernel {
class SigmoidGradKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
auto Y_t = context.Input<Tensor>("Y");
auto dY_t = context.Input<Tensor>(framework::GradVarName("Y"));
auto dX_t = context.Output<Tensor>(framework::GradVarName("X"));
......
......@@ -17,9 +17,9 @@ limitations under the License. */
namespace paddle {
namespace operators {
class SoftmaxOp : public OperatorWithKernel {
class SoftmaxOp : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.InputSize(), 1UL,
"Only one input is need for softmax");
PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims().size(), 2UL,
......@@ -30,9 +30,10 @@ class SoftmaxOp : public OperatorWithKernel {
}
};
class SoftmaxOpMaker : public OpProtoAndCheckerMaker {
class SoftmaxOpMaker : public framework::OpProtoAndCheckerMaker {
public:
SoftmaxOpMaker(OpProto *proto, OpAttrChecker *op_checker)
SoftmaxOpMaker(framework::OpProto *proto,
framework::OpAttrChecker *op_checker)
: OpProtoAndCheckerMaker(proto, op_checker) {
AddInput("X", "input of softmax");
AddOutput("Y", "output of softmax");
......@@ -40,9 +41,9 @@ class SoftmaxOpMaker : public OpProtoAndCheckerMaker {
}
};
class SoftmaxOpGrad : public OperatorWithKernel {
class SoftmaxOpGrad : public framework::OperatorWithKernel {
protected:
void InferShape(const InferShapeContext &ctx) const override {
void InferShape(const framework::InferShapeContext &ctx) const override {
PADDLE_ENFORCE_EQ(ctx.InputSize(), 3UL,
"Input of SoftmaxOpGrad should be 3, X, Y, YG");
PADDLE_ENFORCE_EQ(ctx.OutputSize(), 1UL,
......@@ -61,8 +62,11 @@ class SoftmaxOpGrad : public OperatorWithKernel {
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OP(softmax, ops::SoftmaxOp, ops::SoftmaxOpMaker);
REGISTER_OP_CPU_KERNEL(softmax, ops::SoftmaxKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(softmax,
ops::SoftmaxKernel<paddle::platform::CPUPlace, float>);
REGISTER_GRADIENT_OP(softmax, softmax_grad, ops::SoftmaxOpGrad);
REGISTER_OP_CPU_KERNEL(softmax_grad,
ops::SoftmaxGradKernel<ops::CPUPlace, float>);
REGISTER_OP_CPU_KERNEL(
softmax_grad, ops::SoftmaxGradKernel<paddle::platform::CPUPlace, float>);
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
/* 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.
......@@ -13,9 +13,11 @@
limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/framework/op_registry.h"
#include "paddle/operators/softmax_op.h"
REGISTER_OP_GPU_KERNEL(softmax, ops::SoftmaxKernel<ops::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(softmax_grad,
ops::SoftmaxGradKernel<ops::GPUPlace, float>);
namespace ops = paddle::operators;
REGISTER_OP_GPU_KERNEL(softmax,
ops::SoftmaxKernel<paddle::platform::GPUPlace, float>);
REGISTER_OP_GPU_KERNEL(
softmax_grad, ops::SoftmaxGradKernel<paddle::platform::GPUPlace, float>);
......@@ -13,19 +13,21 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/framework/ddim.h"
#include "paddle/framework/operator.h"
#include "paddle/framework/tensor.h"
#include "paddle/operators/type_alias.h"
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename Place, typename T>
class SoftmaxKernel : public OpKernel {
class SoftmaxKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
auto input = context.Input<Tensor>("X");
auto output = context.Output<Tensor>("Y");
output->mutable_data<T>(context.GetPlace());
......@@ -62,9 +64,9 @@ class SoftmaxKernel : public OpKernel {
};
template <typename Place, typename T>
class SoftmaxGradKernel : public OpKernel {
class SoftmaxGradKernel : public framework::OpKernel {
public:
void Compute(const ExecutionContext& context) const override {
void Compute(const framework::ExecutionContext& context) const override {
std::shared_ptr<Tensor> scale_ = std::make_shared<Tensor>();
auto Y = context.Input<Tensor>("Y");
......
/* 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. */
#include <random>
#include <type_traits>
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
namespace paddle {
namespace operators {
// It seems that Eigen::Tensor::random in GPU will SEGFAULT.
// Use std::random and thrust::random(thrust is a std library in CUDA) to
// implement uniform random.
template <typename T>
class CPUUniformRandomKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* tensor = context.Output<framework::Tensor>(0);
T* data = tensor->mutable_data<T>(context.GetPlace());
unsigned int seed =
static_cast<unsigned int>(context.op_.GetAttr<int>("seed"));
std::minstd_rand engine;
if (seed == 0) {
seed = std::random_device()();
}
engine.seed(seed);
std::uniform_real_distribution<T> dist(
static_cast<T>(context.op_.GetAttr<float>("min")),
static_cast<T>(context.op_.GetAttr<float>("max")));
for (ssize_t i = 0; i < framework::product(tensor->dims()); ++i) {
data[i] = dist(engine);
}
}
};
class UniformRandomOp : public framework::OperatorWithKernel {
protected:
void InferShape(const framework::InferShapeContext& ctx) const override {
PADDLE_ENFORCE(GetAttr<float>("min") < GetAttr<float>("max"),
"uniform_random's min must less then max");
auto* tensor = ctx.Output<framework::Tensor>(0);
auto dims = GetAttr<std::vector<int>>("dims");
tensor->Resize(framework::make_ddim(dims));
}
};
class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker {
public:
UniformRandomOpMaker(framework::OpProto* proto,
framework::OpAttrChecker* op_checker)
: framework::OpProtoAndCheckerMaker(proto, op_checker) {
AddOutput("Out", "The output tensor of uniform random op");
AddComment(R"DOC(Uniform random operator.
Used to initialize tensor with uniform random generator.
)DOC");
AddAttr<std::vector<int>>("dims", "the dimension of random tensor");
AddAttr<float>("min", "Minimum value of uniform random").SetDefault(-1.0f);
AddAttr<float>("max", "Maximun value of uniform random").SetDefault(1.0f);
AddAttr<int>("seed",
"Random seed of uniform random. "
"0 means generate a seed by system")
.SetDefault(0);
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP(uniform_random, paddle::operators::UniformRandomOp,
paddle::operators::UniformRandomOpMaker);
REGISTER_OP_CPU_KERNEL(uniform_random,
paddle::operators::CPUUniformRandomKernel<float>);
/* 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. */
#include <thrust/device_ptr.h>
#include <thrust/iterator/counting_iterator.h>
#include <thrust/random.h>
#include <thrust/transform.h>
#include "paddle/framework/op_registry.h"
#include "paddle/framework/operator.h"
namespace paddle {
namespace operators {
template <typename T>
struct UniformGenerator {
T min_, max_;
unsigned int seed_;
__host__ __device__ UniformGenerator(T min, T max, int seed)
: min_(min), max_(max), seed_(seed) {}
__host__ __device__ T operator()(const unsigned int n) const {
thrust::minstd_rand rng;
rng.seed(seed_);
thrust::uniform_real_distribution<T> dist(min_, max_);
rng.discard(n);
return dist(rng);
}
};
// It seems that Eigen::Tensor::random in GPU will SEGFAULT.
// Use std::random and thrust::random(thrust is a std library in CUDA) to
// implement uniform random.
template <typename T>
class GPUUniformRandomKernel : public framework::OpKernel {
public:
void Compute(const framework::ExecutionContext& context) const override {
auto* tensor = context.Output<framework::Tensor>(0);
T* data = tensor->mutable_data<T>(context.GetPlace());
unsigned int seed =
static_cast<unsigned int>(context.op_.GetAttr<int>("seed"));
if (seed == 0) {
seed = std::random_device()();
}
T min = static_cast<T>(context.op_.GetAttr<float>("min"));
T max = static_cast<T>(context.op_.GetAttr<float>("max"));
thrust::counting_iterator<unsigned int> index_sequence_begin(0);
ssize_t N = framework::product(tensor->dims());
thrust::transform(index_sequence_begin, index_sequence_begin + N,
thrust::device_ptr<T>(data),
UniformGenerator<T>(min, max, seed));
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_GPU_KERNEL(uniform_random,
paddle::operators::GPUUniformRandomKernel<float>);
......@@ -8,7 +8,7 @@ cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
add_subdirectory(dynload)
cc_test(enforce_test SRCS enforce_test.cc)
cc_test(enforce_test SRCS enforce_test.cc DEPS stringpiece)
IF(WITH_GPU)
set(GPU_CTX_DEPS dynload_cuda dynamic_loader)
......
......@@ -15,11 +15,12 @@ limitations under the License. */
#pragma once
#include <execinfo.h>
#include <paddle/string/printf.h>
#include <iomanip>
#include <sstream>
#include <stdexcept>
#include <string>
#include "paddle/string/printf.h"
#include "paddle/string/to_string.h"
#ifndef PADDLE_ONLY_CPU
......@@ -194,8 +195,8 @@ inline void throw_on_error(T e) {
#define __PADDLE_BINARY_COMPARE(__VAL0, __VAL1, __CMP, __INV_CMP, ...) \
PADDLE_ENFORCE(__VAL0 __CMP __VAL1, \
"enforce %s " #__CMP " %s failed, %s " #__INV_CMP " %s\n%s", \
#__VAL0, #__VAL1, std::to_string(__VAL0), \
std::to_string(__VAL1), \
#__VAL0, #__VAL1, paddle::string::to_string(__VAL0), \
paddle::string::to_string(__VAL1), \
paddle::string::Sprintf("" __VA_ARGS__));
} // namespace platform
......
......@@ -9,10 +9,16 @@ 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. */
#include <array>
#include <iostream>
#include <memory>
#include "gtest/gtest.h"
#include "paddle/platform/enforce.h"
#include "paddle/string/piece.h"
using StringPiece = paddle::string::Piece;
using paddle::string::HasPrefix;
TEST(ENFORCE, OK) {
PADDLE_ENFORCE(true, "Enforce is ok %d now %f", 123, 0.345);
......@@ -22,19 +28,15 @@ TEST(ENFORCE, OK) {
}
TEST(ENFORCE, FAILED) {
bool in_catch = false;
bool caught_exception = false;
try {
PADDLE_ENFORCE(false, "Enforce is not ok %d at all", 123);
} catch (paddle::platform::EnforceNotMet error) {
// your error handling code here
in_catch = true;
std::string msg = "Enforce is not ok 123 at all";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
}
caught_exception = true;
EXPECT_TRUE(
HasPrefix(StringPiece(error.what()), "Enforce is not ok 123 at all"));
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE, NO_ARG_OK) {
......@@ -47,41 +49,27 @@ TEST(ENFORCE, NO_ARG_OK) {
TEST(ENFORCE_EQ, NO_EXTRA_MSG_FAIL) {
int a = 2;
bool in_catch = false;
bool caught_exception = false;
try {
PADDLE_ENFORCE_EQ(a, 1 + 3);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "enforce a == 1 + 3 failed, 2 != 4";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
}
caught_exception = true;
HasPrefix(StringPiece(error.what()), "enforce a == 1 + 3 failed, 2 != 4");
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_EQ, EXTRA_MSG_FAIL) {
int a = 2;
bool in_catch = false;
bool caught_exception = false;
try {
PADDLE_ENFORCE_EQ(a, 1 + 3, "%s size not match", "their");
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg =
"enforce a == 1 + 3 failed, 2 != 4\ntheir size not match";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
caught_exception = true;
HasPrefix(StringPiece(error.what()),
"enforce a == 1 + 3 failed, 2 != 4\ntheir size not match");
}
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_NE, OK) {
......@@ -89,42 +77,32 @@ TEST(ENFORCE_NE, OK) {
PADDLE_ENFORCE_NE(1.0, 2UL);
}
TEST(ENFORCE_NE, FAIL) {
bool in_catch = false;
bool caught_exception = false;
try {
// 2UL here to check data type compatible
PADDLE_ENFORCE_NE(1.0, 1UL);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "enforce 1.0 != 1UL failed, 1.000000 == 1";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
}
caught_exception = true;
EXPECT_TRUE(HasPrefix(StringPiece(error.what()),
"enforce 1.0 != 1UL failed, 1 == 1"))
<< error.what() << " does not have expected prefix";
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_GT, OK) { PADDLE_ENFORCE_GT(2, 1); }
TEST(ENFORCE_GT, FAIL) {
bool in_catch = false;
bool caught_exception = false;
try {
// 2UL here to check data type compatible
PADDLE_ENFORCE_GT(1, 2UL);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "enforce 1 > 2UL failed, 1 <= 2";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
caught_exception = true;
EXPECT_TRUE(
HasPrefix(StringPiece(error.what()), "enforce 1 > 2UL failed, 1 <= 2"));
}
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_GE, OK) {
......@@ -134,21 +112,16 @@ TEST(ENFORCE_GE, OK) {
PADDLE_ENFORCE_GE(3.21, 2UL);
}
TEST(ENFORCE_GE, FAIL) {
bool in_catch = false;
bool caught_exception = false;
try {
PADDLE_ENFORCE_GE(1, 2UL);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "enforce 1 >= 2UL failed, 1 < 2";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
caught_exception = true;
EXPECT_TRUE(
HasPrefix(StringPiece(error.what()), "enforce 1 >= 2UL failed, 1 < 2"));
}
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_LE, OK) {
......@@ -159,21 +132,16 @@ TEST(ENFORCE_LE, OK) {
PADDLE_ENFORCE_LE(2UL, 3.2);
}
TEST(ENFORCE_LE, FAIL) {
bool in_catch = false;
bool caught_exception = false;
try {
PADDLE_ENFORCE_GT(1, 2UL);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "enforce 1 > 2UL failed, 1 <= 2";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
}
caught_exception = true;
EXPECT_TRUE(
HasPrefix(StringPiece(error.what()), "enforce 1 > 2UL failed, 1 <= 2"));
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_LT, OK) {
......@@ -182,21 +150,15 @@ TEST(ENFORCE_LT, OK) {
PADDLE_ENFORCE_LT(2UL, 3);
}
TEST(ENFORCE_LT, FAIL) {
bool in_catch = false;
bool caught_exception = false;
try {
PADDLE_ENFORCE_LT(1UL, 0.12);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "enforce 1UL < 0.12 failed, 1 >= 0.12";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
caught_exception = true;
EXPECT_TRUE(HasPrefix(StringPiece(error.what()),
"enforce 1UL < 0.12 failed, 1 >= 0.12"));
}
}
ASSERT_TRUE(in_catch);
EXPECT_TRUE(caught_exception);
}
TEST(ENFORCE_NOT_NULL, OK) {
......@@ -205,20 +167,50 @@ TEST(ENFORCE_NOT_NULL, OK) {
delete a;
}
TEST(ENFORCE_NOT_NULL, FAIL) {
bool in_catch = false;
int* a{nullptr};
bool caught_exception = false;
try {
int* a = nullptr;
PADDLE_ENFORCE_NOT_NULL(a);
} catch (paddle::platform::EnforceNotMet error) {
in_catch = true;
const std::string msg = "a should not be null";
const char* what = error.what();
for (size_t i = 0; i < msg.length(); ++i) {
ASSERT_EQ(what[i], msg[i]);
caught_exception = true;
EXPECT_TRUE(HasPrefix(StringPiece(error.what()), "a should not be null"));
}
EXPECT_TRUE(caught_exception);
}
struct Dims {
size_t dims_[4];
bool operator==(const Dims& o) const {
for (size_t i = 0; i < 4; ++i) {
if (dims_[i] != o.dims_[i]) return false;
}
return true;
}
};
std::ostream& operator<<(std::ostream& os, const Dims& d) {
for (size_t i = 0; i < 4; ++i) {
if (i == 0) {
os << "[";
}
os << d.dims_[i];
if (i == 4 - 1) {
os << "]";
} else {
os << ", ";
}
}
return os;
}
TEST(ENFORCE_USER_DEFINED_CLASS, EQ) {
Dims a{{1, 2, 3, 4}}, b{{1, 2, 3, 4}};
PADDLE_ENFORCE_EQ(a, b);
}
ASSERT_TRUE(in_catch);
TEST(ENFORCE_USER_DEFINED_CLASS, NE) {
Dims a{{1, 2, 3, 4}}, b{{5, 6, 7, 8}};
ASSERT_THROW(PADDLE_ENFORCE_EQ(a, b), paddle::platform::EnforceNotMet);
}
\ No newline at end of file
cc_library(paddle_pybind SHARED
SRCS pybind.cc
DEPS pybind python backward
fc_op
sgd_op
add_op
mean_op
cross_entropy_op
recurrent_op
fill_zeros_like_op)
......@@ -2,3 +2,4 @@ cc_library(stringpiece SRCS piece.cc)
cc_test(stringpiece_test SRCS piece_test.cc DEPS stringpiece glog gflags)
cc_test(stringprintf_test SRCS printf_test.cc DEPS glog gflags)
cc_test(to_string_test SRCS to_string_test.cc)
/* 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 <sstream>
#include <string>
namespace paddle {
namespace string {
template <typename T>
inline std::string to_string(T v) {
std::ostringstream sout;
sout << v;
return sout.str();
}
// Faster std::string/const char* type
template <>
inline std::string to_string(std::string v) {
return v;
}
template <>
inline std::string to_string(const char* v) {
return std::string(v);
}
} // namespace string
} // namespace paddle
/* 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. */
#include "paddle/string/to_string.h"
#include <gtest/gtest.h>
constexpr char kOutputString[] = "User Defined Output";
class UserDefinedClass {
public:
};
std::ostream& operator<<(std::ostream& s, const UserDefinedClass& ins) {
s << kOutputString;
return s;
}
TEST(to_string, normal) {
using namespace paddle::string;
ASSERT_EQ("10", to_string(10));
ASSERT_EQ("abc", to_string("abc"));
ASSERT_EQ("1.2", to_string(1.2));
}
TEST(to_string, user_defined) {
using namespace paddle::string;
UserDefinedClass instance;
ASSERT_EQ(kOutputString, to_string(instance));
}
\ No newline at end of file
......@@ -50,8 +50,8 @@ void NewRemoteParameterUpdater::init(
// create parameter server client.
if (useEtcd_) {
parameterClient_ = paddle_new_etcd_pserver_client(
(char *)pserverSpec_.c_str(), FLAGS_trainer_id == 0);
parameterClient_ =
paddle_new_etcd_pserver_client((char *)pserverSpec_.c_str());
} else {
parameterClient_ = paddle_new_pserver_client((char *)pserverSpec_.c_str(),
FLAGS_trainer_id == 0);
......
......@@ -13,6 +13,7 @@ py_test(test_protobuf SRCS test_protobuf.py)
py_test(test_add_two_op SRCS test_add_two_op.py)
py_test(test_sigmoid_op SRCS test_sigmoid_op.py)
py_test(test_softmax_op SRCS test_softmax_op.py)
py_test(test_cross_entropy_op SRCS test_cross_entropy_op.py)
py_test(test_fill_zeros_like_op SRCS test_fill_zeros_like_op.py)
py_test(gradient_checker SRCS gradient_checker.py)
......@@ -20,4 +21,8 @@ py_test(gradient_checker SRCS gradient_checker.py)
py_test(test_rowwise_add_op SRCS test_rowwise_add_op.py)
py_test(test_default_scope_funcs SRCS test_default_scope_funcs.py)
py_test(test_operator SRCS test_operator.py)
py_test(test_gaussian_random_op SRCS test_gaussian_random_op.py)
py_test(test_uniform_random_op SRCS test_uniform_random_op.py)
import unittest
import numpy
import paddle.v2.framework.core as core
from paddle.v2.framework.op import Operator
import numpy
import unittest
__all__ = ['get_numeric_gradient']
def create_op(op_type):
kwargs = dict()
for in_name in Operator.get_op_input_names(op_type):
kwargs[in_name] = in_name
for out_name in Operator.get_op_output_names(op_type):
kwargs[out_name] = out_name
return Operator(op_type, **kwargs)
def grad_var_name(var_name):
return var_name + "@GRAD"
def get_numeric_gradient(op,
input_values,
output_name,
input_to_check,
delta=1e-2,
delta=0.005,
local_scope=None):
"""
Get Numeric Gradient for an operator's input.
......@@ -76,6 +91,119 @@ def get_numeric_gradient(op,
return gradient_flat.reshape(tensor_to_check.get_dims())
class GradientChecker(unittest.TestCase):
def assert_is_close(self, numeric_grads, scope, max_relative_error,
msg_prefix):
for name in numeric_grads:
b = numpy.array(scope.find_var(grad_var_name(name)).get_tensor())
a = numeric_grads[name]
abs_a = numpy.abs(a)
# if abs_a is nearly zero, then use abs error for a, not relative
# error.
abs_a[abs_a < 1e-3] = 1
diff_mat = numpy.abs(a - b) / abs_a
max_diff = numpy.max(diff_mat)
def err_msg():
offset = numpy.argmax(diff_mat > max_relative_error)
return "%s Variable %s max gradient diff %f over limit %f, the first " \
"error element is %d" % (
msg_prefix, name, max_diff, max_relative_error, offset)
self.assertLessEqual(max_diff, max_relative_error, err_msg())
def check_grad(self,
forward_op,
input_vars,
inputs_to_check,
output_name,
no_grad_set=None,
only_cpu=False,
max_relative_error=0.005):
"""
:param forward_op: used to create backward_op
:param input_vars: numpy value of input variable. The following
computation will use these variables.
:param inputs_to_check: inputs var names that should check gradient.
:param output_name: output name that used to
:param max_relative_error: The relative tolerance parameter.
:param no_grad_set: used when create backward ops
:param only_cpu: only compute and check gradient on cpu kernel.
:return:
"""
if no_grad_set is None:
no_grad_set = set()
tmp_outs = forward_op.temp_outputs()
no_tmp_out = filter(lambda name: name not in tmp_outs,
forward_op.outputs())
if len(no_tmp_out) != 1:
raise ValueError("non temp out_names should be 1")
in_names = forward_op.inputs()
for no_grad in no_grad_set:
if no_grad not in in_names:
raise ValueError("no_grad should be in in_names")
backward_op = core.Operator.backward(forward_op, no_grad_set)
places = [core.CPUPlace()]
if not only_cpu and core.is_compile_gpu() and backward_op.support_gpu():
places.append(core.GPUPlace(0))
numeric_grad = dict()
# get numeric gradient
for check_name in inputs_to_check:
numeric_grad[check_name] = \
get_numeric_gradient(forward_op, input_vars, output_name,
check_name)
# get operator gradient according to different device
for place in places:
scope = core.Scope()
ctx = core.DeviceContext.create(place)
# create input var and set value
for name, value in input_vars.iteritems():
if name not in in_names:
raise ValueError(name + " not in op.inputs_")
var = scope.new_var(name).get_tensor()
var.set_dims(value.shape)
var.set(value, place)
# create output var
for out_name in forward_op.outputs():
scope.new_var(out_name).get_tensor()
# infer the shape of output var and compute/set value of output var
forward_op.infer_shape(scope)
forward_op.run(scope, ctx)
# create output grad var
# set shape as the output var
# set value of this grad to ones
for name in forward_op.outputs():
out_tensor = scope.find_var(name).get_tensor()
grad_tensor = scope.new_var(grad_var_name(name)).get_tensor()
grad_tensor.set_dims(out_tensor.shape())
data = 1.0 * numpy.ones(out_tensor.shape())
grad_tensor.set(data, place)
# create input grad var
for name in backward_op.outputs():
scope.new_var(name).get_tensor()
# infer the shape of input gradient var and compute/set it's value
# with backward op
backward_op.infer_shape(scope)
backward_op.run(scope, ctx)
self.assert_is_close(numeric_grad, scope, max_relative_error,
"Gradient Check On %s" % str(place))
if __name__ == '__main__':
class GetNumericGradientTest(unittest.TestCase):
......@@ -87,4 +215,28 @@ if __name__ == '__main__':
arr = get_numeric_gradient(add_op, {'X': x, "Y": y}, 'Z', 'X')
self.assertAlmostEqual(arr.mean(), 1.0, delta=1e-2)
def test_softmax_op(self):
def stable_softmax(x):
"""Compute the softmax of vector x in a numerically stable way."""
shiftx = x - numpy.max(x)
exps = numpy.exp(shiftx)
return exps / numpy.sum(exps)
def label_softmax_grad(Y, dY):
dX = Y * 0.0
for i in range(Y.shape[0]):
d = numpy.dot(Y[i, :], dY[i, :])
dX[i, :] = Y[i, :] * (dY[i, :] - d)
return dX
softmax_op = Operator("softmax", X="X", Y="Y")
X = numpy.random.random((2, 2)).astype("float32")
Y = numpy.apply_along_axis(stable_softmax, 1, X)
dY = numpy.ones(Y.shape)
dX = label_softmax_grad(Y, dY)
arr = get_numeric_gradient(softmax_op, {"X": X}, 'Y', 'X')
numpy.testing.assert_almost_equal(arr, dX, decimal=1e-2)
unittest.main()
import paddle.v2.framework.core as core
import unittest
import numpy
import paddle.v2.framework.core as core
from paddle.v2.framework.op import Operator
......@@ -24,7 +23,7 @@ class OpTestMeta(type):
scope = core.Scope()
kwargs = dict()
places = [core.CPUPlace()]
if core.is_compile_gpu() and core.Operator.support_gpu(self.type):
if core.is_compile_gpu():
places.append(core.GPUPlace(0))
for place in places:
......@@ -53,6 +52,8 @@ class OpTestMeta(type):
kwargs[attr_name] = self.attrs[attr_name]
op = Operator(self.type, **kwargs)
if isinstance(place, core.GPUPlace) and not op.support_gpu():
return
op.infer_shape(scope)
......
import unittest
import numpy
from op_test_util import OpTestMeta
from gradient_checker import GradientChecker, create_op
class TestSGD(unittest.TestCase):
class TestCrossEntropy(unittest.TestCase):
__metaclass__ = OpTestMeta
def setUp(self):
......@@ -20,7 +21,18 @@ class TestSGD(unittest.TestCase):
self.outputs = {'Y': numpy.array(Y).astype("float32")}
# TODO(superjom) add gradient check
class CrossEntropyGradOpTest(GradientChecker):
def test_softmax_grad(self):
op = create_op("onehot_cross_entropy")
batch_size = 100
class_num = 10
inputs = {
"X": numpy.random.uniform(
0.1, 1.0, [batch_size, class_num]).astype("float32"),
"label": (class_num / 2) * numpy.ones(batch_size).astype("int32")
}
self.check_grad(op, inputs, set("X"), "Y")
if __name__ == "__main__":
unittest.main()
import unittest
import paddle.v2.framework.core as core
from paddle.v2.framework.op import Operator
import numpy
class GaussianRandomTest(unittest.TestCase):
def test_cpu(self):
self.gaussian_random_test(place=core.CPUPlace())
def test_gpu(self):
if core.is_compile_gpu():
self.gaussian_random_test(place=core.GPUPlace(0))
def gaussian_random_test(self, place):
scope = core.Scope()
scope.new_var("Out").get_tensor()
op = Operator(
"gaussian_random",
Out="Out",
dims=[1000, 784],
mean=.0,
std=1.,
seed=10)
op.infer_shape(scope)
context = core.DeviceContext.create(place)
op.run(scope, context)
tensor = numpy.array(scope.find_var("Out").get_tensor())
self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
self.assertAlmostEqual(numpy.std(tensor), 1., delta=0.1)
if __name__ == '__main__':
unittest.main()
import unittest
import numpy as np
import paddle.v2.framework.core as core
from paddle.v2.framework.op import Operator
from gradient_checker import GradientChecker, create_op
from op_test_util import OpTestMeta
......@@ -25,62 +24,11 @@ class TestSoftmaxOp(unittest.TestCase):
}
class TestSoftmaxGradOp(unittest.TestCase):
def test_softmax_grad(self):
op = Operator('softmax', X="X", Y="Y")
backward_op = core.Operator.backward(op, set())
self.assertEqual(backward_op.type(), "softmax_grad")
expected = '''Op(softmax_grad), inputs:(X, Y, Y@GRAD), outputs:(X@GRAD).'''
self.assertEqual(expected, str(backward_op))
batch_size = 3
class_num = 5
# Initialize X and add 1e-2 for numerical stability
Y = np.random.rand(batch_size, class_num).astype(np.float32)
Y = Y + 1e-2
dY = np.random.rand(batch_size, class_num).astype(np.float32)
# Reference implementation of cross entropy with soft labels
def label_softmax_grad(Y, dY):
dX = Y * 0.0
for i in range(batch_size):
d = np.dot(Y[i, :], dY[i, :])
dX[i, :] = Y[i, :] * (dY[i, :] - d)
return dX
expected = label_softmax_grad(Y, dY)
scope = core.Scope()
places = []
places.append(core.CPUPlace())
if core.is_compile_gpu():
places.append(core.GPUPlace(0))
for place in places:
y = scope.new_var("Y")
y_tensor = y.get_tensor()
y_tensor.set_dims([batch_size, class_num])
y_tensor.alloc_float(place)
y_tensor.set(Y, place)
dy = scope.new_var("Y@GRAD")
dy_tensor = dy.get_tensor()
dy_tensor.set_dims([batch_size, class_num])
dy_tensor.alloc_float(place)
dy_tensor.set(dY, place)
x = scope.new_var("X")
dx = scope.new_var("X@GRAD")
tensor = scope.find_var("X@GRAD").get_tensor()
backward_op.infer_shape(scope)
self.assertEqual([batch_size, class_num], tensor.shape())
ctx = core.DeviceContext.create(place)
backward_op.run(scope, ctx)
actual = np.array(tensor)
np.testing.assert_almost_equal(actual, expected, decimal=3)
class SoftmaxGradOpTest(GradientChecker):
def test_softmax(self):
op = create_op("softmax")
inputs = {"X": np.random.uniform(0.1, 1, [10, 10]).astype("float32")}
self.check_grad(op, inputs, set("X"), "Y")
if __name__ == '__main__':
......
import unittest
from paddle.v2.framework.op import Operator
import paddle.v2.framework.core as core
import numpy
class UniformRandomTest(unittest.TestCase):
def test_uniform_random_cpu(self):
self.uniform_random_test(place=core.CPUPlace())
def test_uniform_random_gpu(self):
if core.is_compile_gpu():
self.uniform_random_test(place=core.GPUPlace(0))
def uniform_random_test(self, place):
scope = core.Scope()
scope.new_var("X").get_tensor()
op = Operator(
"uniform_random",
Out="X",
dims=[1000, 784],
min=-5.0,
max=10.0,
seed=10)
op.infer_shape(scope)
ctx = core.DeviceContext.create(place)
op.run(scope, ctx)
tensor = numpy.array(scope.find_var("X").get_tensor())
self.assertAlmostEqual(tensor.mean(), 2.5, delta=0.1)
if __name__ == '__main__':
unittest.main()
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