diff --git a/.gitignore b/.gitignore index c84b2fc8c79d6e2c9c83e2b830ab176295846fd0..9622ab78e0e0556ec2b4cc974fee93ff680d54d2 100644 --- a/.gitignore +++ b/.gitignore @@ -24,4 +24,5 @@ cmake-build-* python/paddle/v2/framework/core.so CMakeFiles cmake_install.cmake - +paddle/.timestamp +python/paddlepaddle.egg-info/ diff --git a/Dockerfile b/Dockerfile index 06a3d8930769bca2599a7afedb3683b2207cb302..8ac123bf9c0f24b47b741611f3b80213c61b82e9 100644 --- a/Dockerfile +++ b/Dockerfile @@ -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 && \ diff --git a/go/glide.lock b/go/glide.lock index be1fb24d772a6524cb798c6169c23ff03e9fed7b..1ecdd217520e0a62b546b4c7048a25f4316d3f37 100644 --- a/go/glide.lock +++ b/go/glide.lock @@ -1,5 +1,5 @@ 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: diff --git a/go/master/service_test.go b/go/master/service_test.go index 5f91910ecc8cf32289e71e2e41e8b283acc115e6..2d00c22d6feb7177da5c19c557fd16d7925ef6d1 100644 --- a/go/master/service_test.go +++ b/go/master/service_test.go @@ -1,24 +1,30 @@ 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 { diff --git a/go/pserver/client/c/cclient.go b/go/pserver/client/c/cclient.go index 14ad0774550f6e5a5d8610d6007904cd2820432c..a49cd01522b8b49a74f21fcb97e9eeb1fbb2d272 100644 --- a/go/pserver/client/c/cclient.go +++ b/go/pserver/client/c/cclient.go @@ -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 diff --git a/go/pserver/client/client.go b/go/pserver/client/client.go index 15adda4735b022c16cb22715fb690b3740e58b76..20d91e77034e1a0c6825bc401175e6dc1afec52f 100644 --- a/go/pserver/client/client.go +++ b/go/pserver/client/client.go @@ -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() } diff --git a/go/pserver/client/client_test.go b/go/pserver/client/client_test.go index 1243ebd6836550d58144b5033e2755ae8594e948..c3d88e926d7cb5f3027be26a270bee6f2db65f31 100644 --- a/go/pserver/client/client_test.go +++ b/go/pserver/client/client_test.go @@ -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.") } diff --git a/go/pserver/client/etcd_client.go b/go/pserver/client/etcd_client.go index 977ae5af37e2b7d647ae16af9c4403f916b0216d..f9071caaa8f5ac32d426b1d4344a30262202b96d 100644 --- a/go/pserver/client/etcd_client.go +++ b/go/pserver/client/etcd_client.go @@ -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 +} diff --git a/go/pserver/client/etcd_client_test.go b/go/pserver/client/etcd_client_test.go new file mode 100644 index 0000000000000000000000000000000000000000..08742433e7a266fbd39e34f4b92ac4cc4caeb0fb --- /dev/null +++ b/go/pserver/client/etcd_client_test.go @@ -0,0 +1,106 @@ +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) + } +} diff --git a/paddle/framework/CMakeLists.txt b/paddle/framework/CMakeLists.txt index 33e6baf818a728d7bf50ba110274d60000dcc22e..f6df89369c52797f7269c41f635756582fadbc47 100644 --- a/paddle/framework/CMakeLists.txt +++ b/paddle/framework/CMakeLists.txt @@ -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) @@ -47,5 +50,6 @@ cc_library(paddle_pybind SHARED cross_entropy_op recurrent_op uniform_random_op + gaussian_random_op fill_zeros_like_op) endif(WITH_PYTHON) diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc index 47983110fa618e89d455a311af2112fc0ff2b9ae..437a44a8aafa650d654a1a77c60613abe07679fe 100644 --- a/paddle/framework/backward.cc +++ b/paddle/framework/backward.cc @@ -133,8 +133,9 @@ std::shared_ptr BackwardRecursive( std::shared_ptr 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 @@ -167,7 +168,7 @@ std::shared_ptr Backward( std::unordered_set 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); diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc index 6d5835bd2236118b6aff95743c4319faceb05d89..1677a3ed4c85ef293f0aadc64a4caa809cbd6ced 100644 --- a/paddle/framework/backward_test.cc +++ b/paddle/framework/backward_test.cc @@ -171,10 +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) { @@ -182,7 +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"}); @@ -250,18 +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(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) { @@ -313,15 +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) { @@ -331,10 +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"); diff --git a/paddle/framework/details/lod_tensor.cc b/paddle/framework/details/lod_tensor.cc new file mode 100644 index 0000000000000000000000000000000000000000..9ad3979e5b511517f75d2d43004f97ee1576953b --- /dev/null +++ b/paddle/framework/details/lod_tensor.cc @@ -0,0 +1,62 @@ +/* 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 + +namespace paddle { +namespace framework { +namespace details { + +using LOD = LODTensor::LOD; + +std::shared_ptr SliceLOD(const LOD &lod, size_t level_begin, + size_t level_end) { + auto new_lod = std::make_shared(); + 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 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(); + 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 diff --git a/paddle/framework/details/lod_tensor.h b/paddle/framework/details/lod_tensor.h new file mode 100644 index 0000000000000000000000000000000000000000..9a6a6cd2ea41f02db991bdc0a2b917433dafed99 --- /dev/null +++ b/paddle/framework/details/lod_tensor.h @@ -0,0 +1,46 @@ +/* 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 + +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 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 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 diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc index cf7143eba4460e5619188b82ffe23db11a04a236..f1ebbae52f13d9c0fc9408aec8c4160575ad59c0 100644 --- a/paddle/framework/grad_op_builder_test.cc +++ b/paddle/framework/grad_op_builder_test.cc @@ -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({"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( - {"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({"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({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({"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( - {"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( - {"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( - {"in3_1" + f::kGradVarSuffix, "in3_2" + f::kGradVarSuffix})); + {f::GradVarName("in3_1"), f::GradVarName("in3_2")})); } diff --git a/paddle/framework/lod_tensor.cc b/paddle/framework/lod_tensor.cc new file mode 100644 index 0000000000000000000000000000000000000000..70045dbf7afd0935e4df852b2f0e3ecd163a9316 --- /dev/null +++ b/paddle/framework/lod_tensor.cc @@ -0,0 +1,51 @@ +/* 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 + +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 diff --git a/paddle/framework/lod_tensor.h b/paddle/framework/lod_tensor.h new file mode 100644 index 0000000000000000000000000000000000000000..4933479b109694312e99595dc8ad6db70259efa6 --- /dev/null +++ b/paddle/framework/lod_tensor.h @@ -0,0 +1,145 @@ +/* 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 +#if (!PADDLE_ONLY_CPU) +#include +#include +#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; +#else + using Level = thrust::device_vector; +#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 LOD; + + LODTensor() {} + LODTensor(const std::shared_ptr &tensor, + const std::shared_ptr &lod) { + Reset(tensor, lod); + } + + void Reset(const std::shared_ptr &tensor, + const std::shared_ptr &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 + 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 + 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(*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() { return tensor_; } + Tensor *raw_tensor() { return tensor_.get(); } + + private: + std::shared_ptr lod_start_pos_; + std::shared_ptr tensor_; +}; + +} // namespace framework +} // namespace paddle + +#include "paddle/framework/lod_tensor_impl.h" diff --git a/paddle/framework/lod_tensor_impl.h b/paddle/framework/lod_tensor_impl.h new file mode 100644 index 0000000000000000000000000000000000000000..0eb6469aea3ae25f035751da985b5bebb489d961 --- /dev/null +++ b/paddle/framework/lod_tensor_impl.h @@ -0,0 +1,60 @@ +/* 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 +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(); + new_tensor->CopyFrom(*tensor_, dst_place); + + return LODTensor(new_tensor, new_lod); +} + +template +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(start_idx, end_idx); + auto new_tensor = std::make_shared(); + new_tensor->CopyFrom(sliced_tensor, dst_place); + + return LODTensor(new_tensor, new_lod); +} + +} // namespace framework +} // namespace paddle diff --git a/paddle/framework/lod_tensor_test.cc b/paddle/framework/lod_tensor_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..511716375e81e8fd89b071c940ee97327c268b8b --- /dev/null +++ b/paddle/framework/lod_tensor_test.cc @@ -0,0 +1,165 @@ +/* + 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 +#include +#include + +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(); + lod->push_back(std::vector{0, 10, 20}); + lod->push_back(std::vector{0, 5, 10, 15, 20}); + lod->push_back(std::vector{0, 2, 5, 7, 10, 12, 15, 17, 20}); + + auto tensor = std::make_shared(); + tensor->Resize({20 /*batch size*/, 128 /*dim*/}); + // malloc memory + tensor->mutable_data(place); + + lod_tensor->Reset(tensor, lod); + } + + protected: + std::unique_ptr 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(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(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(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(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(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 diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h index aed244d61acd8efb49b9650151d53e754bd3ab0f..111709c64ac97e8360abaa2e36522e8592270107 100644 --- a/paddle/framework/op_registry.h +++ b/paddle/framework/op_registry.h @@ -416,7 +416,7 @@ class OpKernelRegistrar : public Registrar { static int use_op_itself_##op_type##_ __attribute__((unused)) = \ TouchOpRegistrar_##op_type() -// TODO(jiayi): Most ops' gradient op have not been compeleted. So we use +// TODO(fengjiayi): Most ops' gradient op have not been compeleted. So we use // `NO_GRAD` to disable micro USE_OP_GRADIENT(op_type). Otherwise the code can't // be compiled. `NO_GRAD` should be removed after all gradient ops are // compeleted. @@ -442,7 +442,7 @@ class OpKernelRegistrar : public Registrar { __attribute__((unused)) = \ TouchOpKernelRegistrar_##op_type##_##DEVICE_TYPE() -// TODO(jiayi): The following macros seems ugly, do we have better method? +// TODO(fengjiayi): The following macros seems ugly, do we have better method? #ifdef PADDLE_ONLY_CPU #define USE_OP_KERNEL(op_type) USE_OP_DEVICE_KERNEL(op_type, CPU) diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index c324fa6702de1eabab3f75cbf4e6568c99b60470..8949baf60e80d9802693cb4b28c99bb3c258c79c 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -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; @@ -120,10 +120,10 @@ class OperatorBase { std::shared_ptr> 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(); } @@ -234,12 +234,6 @@ class OperatorContext { const Scope& scope_; }; -class InferShapeContext : public OperatorContext { - public: - InferShapeContext(const OperatorBase* op, const Scope& scope) - : OperatorContext(op, scope) {} -}; - template struct EigenDeviceConverter; @@ -255,11 +249,11 @@ struct EigenDeviceConverter { }; #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 , OpKernelHash>; void InferShape(const Scope& scope) const override { - InferShape(InferShapeContext(this, scope)); + 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& diff --git a/paddle/framework/pybind.cc b/paddle/framework/pybind.cc index 9c618ad90096476ae123c84574ecd20b14cd66ff..c18d38d2f9a146e40a6b3de8c46b453d79c7c11c 100644 --- a/paddle/framework/pybind.cc +++ b/paddle/framework/pybind.cc @@ -22,6 +22,7 @@ limitations under the License. */ #include "paddle/operators/net_op.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" @@ -39,7 +40,9 @@ USE_OP(softmax); USE_OP(rowwise_add); USE_OP(fill_zeros_like); USE_OP_ITSELF(recurrent_op); +USE_OP(gaussian_random); USE_OP(uniform_random); + namespace paddle { namespace framework { @@ -205,9 +208,13 @@ All parameter, weight, gradient are variables in Paddle. }); // clang-format on - py::class_(m, "GPUPlace").def(py::init()); + py::class_(m, "GPUPlace") + .def(py::init()) + .def("__str__", string::to_string); - py::class_(m, "CPUPlace").def(py::init<>()); + py::class_(m, "CPUPlace") + .def(py::init<>()) + .def("__str__", string::to_string); py::class_> operator_base( m, "Operator"); diff --git a/paddle/framework/tensor.h b/paddle/framework/tensor.h index c44df05e4b0fceed858fbf4f68eddc407a44c894..b57958591fb752132407c35958db0781d0e023f0 100644 --- a/paddle/framework/tensor.h +++ b/paddle/framework/tensor.h @@ -18,6 +18,8 @@ limitations under the License. */ #include #include #include +#include + #include "paddle/framework/ddim.h" #include "paddle/memory/memory.h" #include "paddle/platform/device_context.h" diff --git a/paddle/framework/tensor_test.cc b/paddle/framework/tensor_test.cc index 20276181b974bb5b3d6cb40fb5e6c1295cf1c02f..7db38d5caeebccf710334e854faf785ef0f64063 100644 --- a/paddle/framework/tensor_test.cc +++ b/paddle/framework/tensor_test.cc @@ -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) { diff --git a/paddle/operators/CMakeLists.txt b/paddle/operators/CMakeLists.txt index 9e4026d1c66ccd30ecfc37b3e819241cb85b1a1a..b3399aaf0fb864857ecbf19a7ebeb498b29510f5 100644 --- a/paddle/operators/CMakeLists.txt +++ b/paddle/operators/CMakeLists.txt @@ -41,25 +41,25 @@ function(op_library TARGET) endif() endfunction() +cc_test(gather_test SRCS gather_test.cc DEPS tensor) + 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) op_library(sgd_op SRCS sgd_op.cc sgd_op.cu) -cc_test(sgd_op_test SRCS sgd_op_test.cc DEPS sgd_op) op_library(fc_op SRCS fc_op.cc diff --git a/paddle/operators/add_op_test.cc b/paddle/operators/add_op_test.cc deleted file mode 100644 index bf529defb20d27200a28666278db8607b986e2d5..0000000000000000000000000000000000000000 --- a/paddle/operators/add_op_test.cc +++ /dev/null @@ -1,28 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#define private public -#include "paddle/framework/op_registry.h" - -USE_OP(add_two); - -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()); -} diff --git a/paddle/operators/gather.h b/paddle/operators/gather.h new file mode 100644 index 0000000000000000000000000000000000000000..0c73717d38aca9f3430e66cafc3ecccdd2eec776 --- /dev/null +++ b/paddle/operators/gather.h @@ -0,0 +1,73 @@ +/* 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 +#include + +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +namespace paddle { +namespace operators { + +// Implementation of CPU copy +template +void CPUGather(const T* params, const int* indices, const int slice_size, + const int index_size, T* output) { + const size_t slice_bytes = slice_size * sizeof(T); + + for (size_t i = 0; i < index_size; ++i) { + int index_ = indices[i]; + memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes); + } +} + +// Implementation of GPU copy: +template +void GPUGather(const T* src, const int* index, const int slice_size, + const int index_size, T* output); + +/** + * Return a new tensor from source tensor, gathered according to index + * input[src]: type-T source Tensor + * input[index]: type-int index Tensor (1-D) + * return: output tensor + */ +template +void Gather(const platform::Place& place, const paddle::framework::Tensor* src, + const paddle::framework::Tensor* index, + paddle::framework::Tensor* output) { + // check index of shape 1-D + PADDLE_ENFORCE(index->dims().size() == 1); + int index_size = index->dims()[0]; + + auto src_dims = src->dims(); + paddle::framework::DDim output_dims(src_dims); + output_dims[0] = index_size; + + // slice size + int slice_size = 1; + for (size_t i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i]; + + // Gathering + if (platform::is_cpu_place(place)) { + CPUGather(src->data(), index->data(), slice_size, index_size, + output->data()); + } +} + +} // namespace operators +} // namespace paddle diff --git a/paddle/operators/gather_test.cc b/paddle/operators/gather_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..5de748ec461e4b1a34b75b57c9cd7d5bc9326059 --- /dev/null +++ b/paddle/operators/gather_test.cc @@ -0,0 +1,48 @@ +/* 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/operators/gather.h" +#include "paddle/framework/ddim.h" +#include "paddle/framework/tensor.h" +#include "paddle/platform/place.h" + +#include +#include +#include + +TEST(Gather, GatherData) { + using namespace paddle::framework; + using namespace paddle::platform; + using namespace paddle::operators; + + Tensor* src = new Tensor(); + Tensor* index = new Tensor(); + Tensor* output = new Tensor(); + + int* p_src = nullptr; + int* p_index = nullptr; + p_src = src->mutable_data(make_ddim({3, 4}), CPUPlace()); + p_index = index->mutable_data(make_ddim({2}), CPUPlace()); + + for (size_t i = 0; i < 12; ++i) p_src[i] = i; + p_index[0] = 1; + p_index[1] = 0; + + int* p_output = output->mutable_data(make_ddim({2, 4}), CPUPlace()); + + Gather(CPUPlace(), src, index, output); + + for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4); + for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4); +} diff --git a/paddle/operators/gaussian_random_op.cc b/paddle/operators/gaussian_random_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..ef417ae2f06e8a9f10aed80674015e2ee448f4a3 --- /dev/null +++ b/paddle/operators/gaussian_random_op.cc @@ -0,0 +1,82 @@ +/* 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 +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class GaussianRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + float mean = context.op_.GetAttr("mean"); + float std = context.op_.GetAttr("std"); + auto* tensor = context.Output(0); + T* data = tensor->mutable_data(context.GetPlace()); + + // TODO(dzh): attribute does not support unsigned int. + // And we need a global random seed configuration. + int seed = context.op_.GetAttr("seed"); + if (seed == 0) { + seed = std::random_device()(); + } + std::mt19937 g(seed); + std::normal_distribution 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(0); + auto dims = GetAttr>("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>("dims", "The dimension of random tensor."); + AddAttr("mean", "mean value of random.").SetDefault(.0f); + AddAttr("std", "minimum value of random value.").SetDefault(1.0f); + AddAttr("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); diff --git a/paddle/operators/gaussian_random_op.cu b/paddle/operators/gaussian_random_op.cu new file mode 100644 index 0000000000000000000000000000000000000000..54e4ae5d2b255f72582b9826685bfacf6c565fab --- /dev/null +++ b/paddle/operators/gaussian_random_op.cu @@ -0,0 +1,52 @@ +/* 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 +#include +#include "paddle/platform/dynload/curand.h" +#include "paddle/platform/gpu_info.h" + +#include "paddle/framework/op_registry.h" + +namespace paddle { +namespace operators { + +template +class GaussianRandomKernel : public framework::OpKernel { + public: + void Compute(const framework::ExecutionContext& context) const override { + float mean = context.op_.GetAttr("mean"); + float std = context.op_.GetAttr("std"); + auto* tensor = context.Output(0); + T* data = tensor->mutable_data(context.GetPlace()); + + int seed = context.op_.GetAttr("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); \ No newline at end of file diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc index 997b0c514e96467dc9f9027829616c7b16fe43e1..2ea049cb3605f4dedabb992ebc0e8aa276ad5e9a 100644 --- a/paddle/operators/mean_op.cc +++ b/paddle/operators/mean_op.cc @@ -41,7 +41,7 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker { class MeanGradOp : public framework::OperatorWithKernel { protected: void InferShape(const framework::InferShapeContext &ctx) const override { - ctx.Output("X" + framework::kGradVarSuffix) + ctx.Output(framework::GradVarName("X")) ->Resize(ctx.Input("X")->dims()); } }; diff --git a/paddle/operators/mean_op.h b/paddle/operators/mean_op.h index f3db0a29bb234948d180d964fb82057632ec4414..e8595a14faa7c1b03734f814c78f9cbf1819fbb5 100644 --- a/paddle/operators/mean_op.h +++ b/paddle/operators/mean_op.h @@ -48,10 +48,10 @@ template class MeanGradKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto OG = context.Input("Out" + framework::kGradVarSuffix); + auto OG = context.Input(framework::GradVarName("Out")); PADDLE_ENFORCE(framework::product(OG->dims()) == 1, "Mean Gradient should be scalar"); - auto IG = context.Output("X" + framework::kGradVarSuffix); + auto IG = context.Output(framework::GradVarName("X")); IG->mutable_data(context.GetPlace()); T ig_size = (T)framework::product(IG->dims()); diff --git a/paddle/operators/mean_op_test.cc b/paddle/operators/mean_op_test.cc deleted file mode 100644 index 375dcd50e130355c60f82b9d39d1b94fb2c911b0..0000000000000000000000000000000000000000 --- a/paddle/operators/mean_op_test.cc +++ /dev/null @@ -1,25 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include - -#include - -USE_OP(mean); - -TEST(MeanOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("mean"); - ASSERT_NE(it, protos.end()); -} diff --git a/paddle/operators/sgd_op_test.cc b/paddle/operators/sgd_op_test.cc deleted file mode 100644 index b2a5487f12c9e0287257ca5ac89bf46c19799adc..0000000000000000000000000000000000000000 --- a/paddle/operators/sgd_op_test.cc +++ /dev/null @@ -1,22 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include -#include -USE_NO_GRAD_OP(sgd); -TEST(SGDOp, GetOpProto) { - auto& protos = paddle::framework::OpRegistry::protos(); - auto it = protos.find("sgd"); - ASSERT_NE(it, protos.end()); -} diff --git a/paddle/platform/CMakeLists.txt b/paddle/platform/CMakeLists.txt index bd77bb7daa50e0b273f110624ddf6f4b79a3ceab..4154aad15c39119e2f155cb2c7b5177b5aa78022 100644 --- a/paddle/platform/CMakeLists.txt +++ b/paddle/platform/CMakeLists.txt @@ -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) diff --git a/paddle/platform/enforce.h b/paddle/platform/enforce.h index d2adb997de8e36922d5056b20f238a82eee74f8c..337a059fb1494d500be0fd2437e59c863ae1563c 100644 --- a/paddle/platform/enforce.h +++ b/paddle/platform/enforce.h @@ -15,11 +15,12 @@ limitations under the License. */ #pragma once #include -#include #include #include #include #include +#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 diff --git a/paddle/platform/enforce_test.cc b/paddle/platform/enforce_test.cc index 4dfb69754608cb1120baa295072c3d031a4e1a7b..80bdee3d9dfbe38ef707a6ba60cdb7f7b99714de 100644 --- a/paddle/platform/enforce_test.cc +++ b/paddle/platform/enforce_test.cc @@ -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 +#include #include #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; } +}; - ASSERT_TRUE(in_catch); +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); +} + +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 diff --git a/paddle/string/CMakeLists.txt b/paddle/string/CMakeLists.txt index 5becf62672d0c606c98ea1a1a4383df97088ab05..60667b72873f9422aec1807972a81ab680de2e64 100644 --- a/paddle/string/CMakeLists.txt +++ b/paddle/string/CMakeLists.txt @@ -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) diff --git a/paddle/string/to_string.h b/paddle/string/to_string.h new file mode 100644 index 0000000000000000000000000000000000000000..4f478b6a36b23bdba8ef3ddae94b3eadf18716c2 --- /dev/null +++ b/paddle/string/to_string.h @@ -0,0 +1,40 @@ +/* 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 +#include + +namespace paddle { +namespace string { +template +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 diff --git a/paddle/string/to_string_test.cc b/paddle/string/to_string_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..5ff1b007f1875c7b920a08bd13b8d98cdc5138d3 --- /dev/null +++ b/paddle/string/to_string_test.cc @@ -0,0 +1,39 @@ +/* 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 + +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 diff --git a/paddle/trainer/NewRemoteParameterUpdater.cpp b/paddle/trainer/NewRemoteParameterUpdater.cpp index e1558e3fdfbcf296be0ee64202132f53bf901be9..af1dceed0284c70d68b61b9682b0cb23c28043d6 100644 --- a/paddle/trainer/NewRemoteParameterUpdater.cpp +++ b/paddle/trainer/NewRemoteParameterUpdater.cpp @@ -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); diff --git a/python/paddle/v2/framework/tests/CMakeLists.txt b/python/paddle/v2/framework/tests/CMakeLists.txt index 10659caa882fd3d4060f9947413a392c3b681ee8..f6850e06512d196d51e454bc22cfa3cda8bba84a 100644 --- a/python/paddle/v2/framework/tests/CMakeLists.txt +++ b/python/paddle/v2/framework/tests/CMakeLists.txt @@ -21,5 +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) diff --git a/python/paddle/v2/framework/tests/gradient_checker.py b/python/paddle/v2/framework/tests/gradient_checker.py index b73c4869d14a62a951d8e45dafb14b7523355519..aacc5e88feeb65e08093a35ef85837c916cfd39e 100644 --- a/python/paddle/v2/framework/tests/gradient_checker.py +++ b/python/paddle/v2/framework/tests/gradient_checker.py @@ -92,15 +92,27 @@ def get_numeric_gradient(op, class GradientChecker(unittest.TestCase): - def __is_close(self, numeric_grads, scope, max_relative_error): + def assert_is_close(self, numeric_grads, scope, max_relative_error, + msg_prefix): for name in numeric_grads: - op_grad = numpy.array( - scope.find_var(grad_var_name(name)).get_tensor()) - is_close = numpy.allclose( - numeric_grads[name], op_grad, rtol=max_relative_error, atol=100) - if not is_close: - return False - return True + 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, @@ -145,7 +157,8 @@ class GradientChecker(unittest.TestCase): # 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_numeric_gradient(forward_op, input_vars, output_name, + check_name) # get operator gradient according to different device for place in places: @@ -187,15 +200,8 @@ class GradientChecker(unittest.TestCase): backward_op.infer_shape(scope) backward_op.run(scope, ctx) - if isinstance(place, core.CPUPlace): - msg = "CPU kernel gradient is not close to numeric gradient" - else: - if isinstance(place, core.GPUPlace): - msg = "GPU kernel gradient is not close to numeric gradient" - else: - raise ValueError("unknown place " + type(place)) - self.assertTrue( - self.__is_close(numeric_grad, scope, max_relative_error), msg) + self.assert_is_close(numeric_grad, scope, max_relative_error, + "Gradient Check On %s" % str(place)) if __name__ == '__main__': diff --git a/python/paddle/v2/framework/tests/test_gaussian_random_op.py b/python/paddle/v2/framework/tests/test_gaussian_random_op.py new file mode 100644 index 0000000000000000000000000000000000000000..f95ed70b58d611b3233a21d3f2a34c864ae4d1b3 --- /dev/null +++ b/python/paddle/v2/framework/tests/test_gaussian_random_op.py @@ -0,0 +1,36 @@ +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()