Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into scope-impl

CMakeLists.txt

@@ -27,6 +27,7 @@ if(NOT CMAKE_CROSSCOMPILING)
 endif(NOT CMAKE_CROSSCOMPILING)
 find_package(Git REQUIRED)
 find_package(Threads REQUIRED)
+find_package(Boost QUIET)
 
 include(simd)
 
@@ -92,6 +93,7 @@ include(external/openblas)  # download, build, install openblas
 include(external/swig)      # download, build, install swig
 include(external/warpctc)   # download, build, install warpctc
 include(external/any)       # download libn::any
+include(external/eigen)     # download eigen3
 
 include(generic)            # simplify cmake module
 include(package)            # set paddle packages
@@ -109,6 +111,7 @@ include_directories("${PROJ_ROOT}")
 include_directories("${PROJ_ROOT}/paddle/cuda/include")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/proto")
 include_directories("${CMAKE_CURRENT_BINARY_DIR}/go/pserver/cclient")
+include_directories(${Boost_INCLUDE_DIRS})
 
 set(EXTERNAL_LIBS
     ${GFLAGS_LIBRARIES}

cmake/external/eigen.cmake

+INCLUDE(ExternalProject)
+
+SET(EIGEN_SOURCE_DIR ${THIRD_PARTY_PATH}/eigen3)
+
+INCLUDE_DIRECTORIES(${EIGEN_SOURCE_DIR}/src/eigen3)
+
+ExternalProject_Add(
+    eigen3
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    URL            "https://bitbucket.org/eigen/eigen/get/3.3.4.tar.gz"
+    URL_MD5        "1a47e78efe365a97de0c022d127607c3"
+    PREFIX          ${EIGEN_SOURCE_DIR}
+    UPDATE_COMMAND  ""
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND     ""
+    INSTALL_COMMAND   ""
+    TEST_COMMAND      ""
+)
+
+LIST(APPEND external_project_dependencies eigen3)

doc/getstarted/concepts/use_concepts_cn.rst

@@ -111,7 +111,7 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
     # define training dataset reader
     def train_reader():
         train_x = np.array([[1, 1], [1, 2], [3, 4], [5, 2]])
-        train_y = np.array([-2, -3, -7, -7])
+        train_y = np.array([[-2], [-3], [-7], [-7]])
         def reader():
             for i in xrange(train_y.shape[0]):
                 yield train_x[i], train_y[i]

go/cmd/pserver/pserver.go

@@ -30,7 +30,13 @@ func main() {
 	log.SetLevel(level)
 
 	timeout := time.Second * time.Duration((*etcdTimeout))
-	s, err := pserver.NewService(*etcdEndpoint, *numPservers, timeout)
+	e := pserver.NewEtcdClient(*etcdEndpoint, *numPservers, timeout)
+	idx, err := e.Register()
+	if err != nil {
+		panic(err)
+	}
+
+	s, err := pserver.NewService(idx)
 	if err != nil {
 		panic(err)
 	}

go/master/c/client.go

@@ -13,10 +13,13 @@ typedef int paddle_master_client;
 import "C"
 
 import (
+	"strings"
 	"sync"
+	"time"
 	"unsafe"
 
 	"github.com/PaddlePaddle/Paddle/go/master"
+	"github.com/coreos/etcd/clientv3"
 	log "github.com/sirupsen/logrus"
 )
 
@@ -48,16 +51,33 @@ func remove(client C.paddle_master_client) *master.Client {
 	return h
 }
 
-type addresser string
-
-func (a addresser) Address() string {
-	return string(a)
+//export paddle_new_etcd_master_client
+func paddle_new_etcd_master_client(etcdEndpoints *C.char, timeout int, bufSize int) C.paddle_master_client {
+	p := C.GoString(etcdEndpoints)
+	cli, err := clientv3.New(clientv3.Config{
+		Endpoints:   strings.Split(p, ","),
+		DialTimeout: time.Second * time.Duration(timeout),
+	})
+	if err != nil {
+		panic(err)
+	}
+	ch := make(chan string, 1)
+	a, err := master.GetKey(cli, master.DefaultAddrPath, timeout)
+	if err != nil {
+		panic(err)
+	}
+	ch <- a
+	go master.WatchKey(cli, master.DefaultAddrPath, ch)
+	c := master.NewClient(ch, bufSize)
+	return add(c)
 }
 
 //export paddle_new_master_client
 func paddle_new_master_client(addr *C.char, bufSize int) C.paddle_master_client {
 	a := C.GoString(addr)
-	c := master.NewClient(addresser(a), bufSize)
+	ch := make(chan string, 1)
+	ch <- a
+	c := master.NewClient(ch, bufSize)
 	return add(c)
 }
 

go/master/client.go

@@ -2,18 +2,12 @@ package master
 
 import (
 	"os"
-	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/connection"
 	"github.com/PaddlePaddle/recordio"
 	log "github.com/sirupsen/logrus"
 )
 
-// Addresser provide the address of the master server.
-type Addresser interface {
-	Address() string
-}
-
 // Client is the client of the master server.
 type Client struct {
 	conn *connection.Conn
@@ -24,11 +18,11 @@ type Client struct {
 //
 // bufSize is the record buffer size. NextRecord will read from this
 // buffer.
-func NewClient(addr Addresser, bufSize int) *Client {
+func NewClient(addrCh <-chan string, bufSize int) *Client {
 	c := &Client{}
 	c.conn = connection.New()
 	c.ch = make(chan []byte, bufSize)
-	go c.monitorMaster(addr)
+	go c.monitorMaster(addrCh)
 	go c.getRecords()
 	return c
 }
@@ -72,12 +66,10 @@ func (c *Client) getRecords() {
 	}
 }
 
-func (c *Client) monitorMaster(addr Addresser) {
+func (c *Client) monitorMaster(addrCh <-chan string) {
 	lastMaster := ""
-	monitor := func() {
-		// get the lastest address of the master server,
+	for curMaster := range addrCh {
 		// connect to the new address once address changed.
-		curMaster := addr.Address()
 		if curMaster != lastMaster {
 			if curMaster == "" {
 				err := c.conn.Close()
@@ -94,18 +86,10 @@ func (c *Client) monitorMaster(addr Addresser) {
 					// to retry next time.
 					curMaster = lastMaster
 				}
-
 			}
 		}
-
 		lastMaster = curMaster
 	}
-
-	monitor()
-	ticker := time.NewTicker(10 * time.Second)
-	for _ = range ticker.C {
-		monitor()
-	}
 }
 
 // SetDataset set dataset for the master server to dispatch.

go/master/client_internal_test.go

@@ -26,12 +26,6 @@ func init() {
 	log.SetLevel(log.ErrorLevel)
 }
 
-type TestAddresser string
-
-func (a TestAddresser) Address() string {
-	return string(a)
-}
-
 func TestGetFinishTask(t *testing.T) {
 	const path = "/tmp/master_client_test_0"
 
@@ -45,7 +39,6 @@ func TestGetFinishTask(t *testing.T) {
 	if err != nil {
 		panic(err)
 	}
-
 	go func(l net.Listener) {
 		s, err := NewService(&InMemStore{}, chunkPerTask, time.Second, 1)
 		if err != nil {
@@ -82,9 +75,11 @@ func TestGetFinishTask(t *testing.T) {
 	// Manually intialize client to avoid calling c.getRecords()
 	c := &Client{}
 	c.conn = connection.New()
-	go c.monitorMaster(TestAddresser(fmt.Sprintf(":%d", p)))
+	addr := fmt.Sprintf(":%d", p)
+	ch := make(chan string, 1)
+	ch <- addr
+	go c.monitorMaster(ch)
 	c.SetDataset([]string{path})
-
 	checkOnePass := func(i int) {
 		var tasks []Task
 		for idx := 0; idx < totalTask; idx++ {

go/master/client_test.go

@@ -20,7 +20,6 @@ func TestNextRecord(t *testing.T) {
 		path  = "/tmp/master_client_TestFull"
 		total = 50
 	)
-
 	l, err := net.Listen("tcp", ":0")
 	if err != nil {
 		panic(err)
@@ -31,7 +30,6 @@ func TestNextRecord(t *testing.T) {
 	if err != nil {
 		panic(err)
 	}
-
 	go func(l net.Listener) {
 		s, err := master.NewService(&master.InMemStore{}, 10, time.Second, 1)
 		if err != nil {
@@ -63,10 +61,10 @@ func TestNextRecord(t *testing.T) {
 	}
 	w.Close()
 	f.Close()
-
-	c := master.NewClient(master.TestAddresser(fmt.Sprintf(":%d", p)), 10)
+	curAddr := make(chan string, 1)
+	curAddr <- fmt.Sprintf(":%d", p)
+	c := master.NewClient(curAddr, 10)
 	c.SetDataset([]string{path})
-
 	for pass := 0; pass < 50; pass++ {
 		received := make(map[byte]bool)
 		for i := 0; i < total; i++ {

go/master/etcd_client.go

@@ -18,8 +18,8 @@ const (
 	DefaultAddrPath = "/master/addr"
 )
 
-// EtcdClient is the etcd client that master uses for fault tolerance
-// and service registry.
+// EtcdClient is the etcd client that the master uses for fault
+// tolerance and service registry.
 type EtcdClient struct {
 	lockPath  string
 	statePath string
@@ -142,3 +142,31 @@ func (e *EtcdClient) Load() ([]byte, error) {
 	state := kvs[0].Value
 	return state, nil
 }
+
+// GetKey gets the value by the specify key.
+func GetKey(c *clientv3.Client, key string, timeout int) (string, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*time.Duration(timeout))
+	resp, err := c.Get(ctx, key)
+	cancel()
+	if err != nil {
+		return "", err
+	}
+	kvs := resp.Kvs
+	if len(kvs) == 0 {
+		return "", nil
+	}
+	v := kvs[0].Value
+	return string(v), nil
+}
+
+// WatchKey watches the specify key and send to valChan if there is some event.
+func WatchKey(c *clientv3.Client, key string, valChan chan<- string) {
+	rch := c.Watch(context.Background(), key)
+	for wresp := range rch {
+		for _, ev := range wresp.Events {
+			// if received event is DELETE, the value will be an empty string
+			log.Infof("received event %s, %q : %q\n", ev.Type, ev.Kv.Key, ev.Kv.Value)
+			valChan <- string(ev.Kv.Value)
+		}
+	}
+}

go/pserver/client.go

 package pserver
 
 import (
+	"errors"
 	"hash/fnv"
 	"sort"
 	"time"
@@ -123,6 +124,9 @@ func (c *Client) FinishInitParams() error {
 // SendGrads sends gradients to parameter servers for updating
 // parameters.
 func (c *Client) SendGrads(grads []Gradient) error {
+	if len(grads) == 0 {
+		return errors.New("no gradient received")
+	}
 	errCh := make(chan error, len(grads))
 	for _, g := range grads {
 		go func(g Gradient) {

go/pserver/client_test.go

@@ -7,7 +7,6 @@ import (
 	"strconv"
 	"strings"
 	"testing"
-	"time"
 
 	"github.com/PaddlePaddle/Paddle/go/pserver"
 )
@@ -31,7 +30,7 @@ func init() {
 		port[i] = p
 
 		go func(l net.Listener) {
-			s, err := pserver.NewService("", time.Second*5)
+			s, err := pserver.NewService(0)
 			if err != nil {
 				panic(err)
 			}

go/pserver/etcd_client.go

+package pserver
+
+import (
+	"context"
+	"errors"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/PaddlePaddle/Paddle/go/utils/networkhelper"
+	"github.com/coreos/etcd/clientv3"
+	"github.com/coreos/etcd/clientv3/concurrency"
+	log "github.com/sirupsen/logrus"
+)
+
+// EtcdClient is the etcd client that the pserver uses for fault
+// tolerance, service registry and coordination.
+type EtcdClient struct {
+	numPservers   int
+	etcdEndpoints string
+	etcdClient    *clientv3.Client
+	// etcdTimeout is also used as retry intervals.
+	etcdTimeout time.Duration
+	// FIXME: ensure GetExternalIP gets the correct ip for trainers to connect.
+	externalIP string
+	// desired number of pservers in the job.
+	// assume desired will not change during one training job.
+	desired int
+}
+
+// NewEtcdClient creates an EtcdClient
+func NewEtcdClient(endpoints string, numPservers int, timeout time.Duration) *EtcdClient {
+	return &EtcdClient{
+		etcdTimeout:   timeout,
+		numPservers:   numPservers,
+		etcdEndpoints: endpoints,
+	}
+}
+
+// Register registers the pserver on etcd
+//
+// Register returns the index of the current pserver.
+func (e *EtcdClient) Register() (int, error) {
+
+	var err error
+	e.externalIP, err = networkhelper.GetExternalIP()
+	if err != nil {
+		return 0, err
+	}
+
+	// initialize connection to etcd.
+	ep := strings.Split(e.etcdEndpoints, ",")
+	for {
+		cli, err := clientv3.New(clientv3.Config{
+			Endpoints:   ep,
+			DialTimeout: e.etcdTimeout,
+		})
+		if err != nil {
+			log.Errorf("connect to etcd error: %v", err)
+			time.Sleep(e.etcdTimeout)
+			continue
+		}
+		e.etcdClient = cli
+		log.Debugf("inited client to %s", e.etcdEndpoints)
+		break
+	}
+	// init /ps_desired using transaction, for multiple pservers may want to write
+	// it at the same time.
+	for {
+		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+		_, err := e.initDesiredPsercers(ctx, e.numPservers)
+		cancel()
+		if err != nil {
+			log.Warn(err)
+			time.Sleep(e.etcdTimeout)
+			continue
+		}
+		break
+	}
+	// TODO: when implementing extending or reducing pservers, /ps_desired is
+	// changed, then we need to watch /ps_desired node for events. For now, just
+	// write once when init and read from it.
+	// wait and set s.desired init value
+	for {
+		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+		resp, err := e.etcdClient.Get(ctx, PsDesired)
+		cancel()
+		if err != nil {
+			log.Errorf("getting %s error: %v", PsDesired, err)
+			time.Sleep(e.etcdTimeout)
+			continue
+		}
+		if len(resp.Kvs) != 0 {
+			e.desired, err = strconv.Atoi(string(resp.Kvs[0].Value))
+			if err != nil {
+				log.Errorf("value of %s invalid %v\n", PsDesired, err)
+				time.Sleep(e.etcdTimeout)
+				// NOTE: wait util ps_desired value change
+				continue
+			}
+			break
+		}
+	}
+
+	var pserverIdx int
+	// try register pserver node on etcd
+	for {
+		ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+		var err error
+		pserverIdx, err = e.registerPserverEtcd(ctx)
+		cancel()
+		if err != nil {
+			log.Warn(err)
+			time.Sleep(e.etcdTimeout)
+			continue
+		}
+		break
+	}
+
+	return pserverIdx, nil
+}
+
+func (e *EtcdClient) initDesiredPsercers(ctx context.Context, numPservers int) (*clientv3.TxnResponse, error) {
+	return concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
+		dsStr := c.Get(PsDesired)
+		if dsStr == "" {
+			c.Put(PsDesired, strconv.Itoa(numPservers))
+		}
+		return nil
+	}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
+}
+
+// registerPserverEtcd registers pserver node on etcd using transaction.
+func (e *EtcdClient) registerPserverEtcd(ctx context.Context) (int, error) {
+	var idx int
+	_, err := concurrency.NewSTM(e.etcdClient, func(c concurrency.STM) error {
+		registered := false
+		for i := 0; i < e.desired; i++ {
+			psKey := "/ps/" + strconv.Itoa(i)
+			log.Debugf("checking %s", psKey)
+			ps := c.Get(psKey)
+			log.Debugf("got value (%s) for key: %s", ps, psKey)
+
+			if ps == "" {
+				resp, err := e.etcdClient.Grant(context.TODO(), 5)
+				if err != nil {
+					log.Fatal(err)
+				}
+				// find the first id and write info
+				c.Put(psKey, e.externalIP, clientv3.WithLease(resp.ID))
+				log.Debugf("set pserver node %s with value %s", psKey, e.externalIP)
+				ch, kaerr := e.etcdClient.KeepAlive(context.TODO(), resp.ID)
+				if kaerr != nil {
+					log.Errorf("keepalive etcd node error: %v", kaerr)
+					return kaerr
+				}
+
+				// Eat the keep alive message so etcd
+				// will not expire the lease.
+				go func(ch <-chan *clientv3.LeaseKeepAliveResponse) {
+					ka := <-ch
+					log.Debugf("keepalive: %d\n", ka.TTL)
+				}(ch)
+				log.Debug("register finished")
+				idx = i
+				registered = true
+				break
+			}
+		}
+		if registered == true {
+			return nil
+		}
+		return errors.New("not registerd, may due to already have enough pservers")
+	}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
+
+	if err != nil {
+		return 0, err
+	}
+
+	return idx, nil
+}

go/pserver/service.go

 package pserver
 
 import (
-	"context"
 	"errors"
 	"fmt"
-	"strconv"
-	"strings"
 	"sync"
-	"time"
-
-	"github.com/PaddlePaddle/Paddle/go/utils/networkhelper"
-	"github.com/coreos/etcd/clientv3"
-	"github.com/coreos/etcd/clientv3/concurrency"
-	log "github.com/sirupsen/logrus"
 )
 
 // ElementType is the type of elements of a Parameter.
@@ -55,160 +46,25 @@ type Gradient Parameter
 // Service is the RPC service for pserver.
 type Service struct {
 	initialized chan struct{}
+	idx         int
 
 	mu       sync.Mutex
 	opt      *optimizer
 	paramMap map[string]Parameter
-
-	etcdEndpoints string
-	etcdClient    *clientv3.Client
-	// etcdTimeout is also used as retry intervals.
-	etcdTimeout time.Duration
-	// desired number of pservers in the job.
-	// assume desired will not change during one training job.
-	desired int
-	// FIXME: ensure GetExternalIP gets the correct ip for trainers to connect.
-	externalIP string
 }
 
 // NewService creates a new service, will bypass etcd registration if no
 // endpoints specified.
-func NewService(endpoints string, numPservers int, timeout time.Duration) (*Service, error) {
-	s := &Service{opt: newOptimizer(sgd, 0.005)}
+func NewService(idx int) (*Service, error) {
+	s := &Service{
+		idx: idx,
+		opt: newOptimizer(sgd, 0.005),
+	}
 	s.paramMap = make(map[string]Parameter)
 	s.initialized = make(chan struct{})
-	s.etcdEndpoints = endpoints
-	s.etcdTimeout = timeout
-
-	var err error
-	s.externalIP, err = networkhelper.GetExternalIP()
-	if err != nil {
-		return nil, err
-	}
-
-	if endpoints != "" {
-		// initialize connection to etcd, try
-		ep := strings.Split(s.etcdEndpoints, ",")
-		for {
-			cli, err := clientv3.New(clientv3.Config{
-				Endpoints:   ep,
-				DialTimeout: s.etcdTimeout,
-			})
-			if err != nil {
-				log.Errorf("connect to etcd error: %v", err)
-				time.Sleep(s.etcdTimeout)
-				continue
-			}
-			s.etcdClient = cli
-			log.Debugf("inited client to %s", s.etcdEndpoints)
-			break
-		}
-		// init /ps_desired using transaction, for multiple pservers may want to write
-		// it at the same time.
-		for {
-			ctx, cancel := context.WithTimeout(context.Background(), time.Second)
-			_, err := s.initDesiredPsercers(ctx, numPservers)
-			cancel()
-			if err != nil {
-				log.Warn(err)
-				time.Sleep(s.etcdTimeout)
-				continue
-			}
-			break
-		}
-		// TODO: when implementing extending or reducing pservers, /ps_desired is
-		// changed, then we need to watch /ps_desired node for events. For now, just
-		// write once when init and read from it.
-		// wait and set s.desired init value
-		for {
-			ctx, cancel := context.WithTimeout(context.Background(), time.Second)
-			resp, err := s.etcdClient.Get(ctx, PsDesired)
-			cancel()
-			if err != nil {
-				log.Errorf("getting %s error: %v", PsDesired, err)
-				time.Sleep(s.etcdTimeout)
-				continue
-			}
-			if len(resp.Kvs) != 0 {
-				s.desired, err = strconv.Atoi(string(resp.Kvs[0].Value))
-				if err != nil {
-					log.Errorf("value of %s invalid %v\n", PsDesired, err)
-					time.Sleep(s.etcdTimeout)
-					// NOTE: wait util ps_desired value change
-					continue
-				}
-				break
-			}
-		}
-		// try register pserver node on etcd
-		for {
-			ctx, cancel := context.WithTimeout(context.Background(), time.Second)
-			_, err := s.registerPserverEtcd(ctx)
-			cancel()
-			if err != nil {
-				log.Warn(err)
-				time.Sleep(s.etcdTimeout)
-				continue
-			}
-			break
-		}
-	} // if endpoints != ""
-	// Bypass etcd registration if no endpoints specified
 	return s, nil
 }
 
-func (s *Service) initDesiredPsercers(ctx context.Context, numPservers int) (*clientv3.TxnResponse, error) {
-	return concurrency.NewSTM(s.etcdClient, func(c concurrency.STM) error {
-		dsStr := c.Get(PsDesired)
-		if dsStr == "" {
-			c.Put(PsDesired, strconv.Itoa(numPservers))
-		}
-		return nil
-	}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
-}
-
-// registerPserverEtcd registers pserver node on etcd using transaction.
-func (s *Service) registerPserverEtcd(ctx context.Context) (*clientv3.TxnResponse, error) {
-	return concurrency.NewSTM(s.etcdClient, func(c concurrency.STM) error {
-		registered := false
-		for i := 0; i < s.desired; i++ {
-			psKey := "/ps/" + strconv.Itoa(i)
-			log.Debugf("checking %s", psKey)
-			ps := c.Get(psKey)
-			log.Debugf("got value (%s) for key: %s", ps, psKey)
-
-			if ps == "" {
-				resp, err := s.etcdClient.Grant(context.TODO(), 5)
-				if err != nil {
-					log.Fatal(err)
-				}
-				// find the first id and write info
-				c.Put(psKey, s.externalIP, clientv3.WithLease(resp.ID))
-				log.Debugf("set pserver node %s with value %s", psKey, s.externalIP)
-				ch, kaerr := s.etcdClient.KeepAlive(context.TODO(), resp.ID)
-				if kaerr != nil {
-					log.Errorf("keepalive etcd node error: %v", kaerr)
-					return kaerr
-				}
-
-				// Eat the keep alive message so etcd
-				// will not expire the lease.
-				go func(ch <-chan *clientv3.LeaseKeepAliveResponse) {
-					ka := <-ch
-					log.Debugf("keepalive: %d\n", ka.TTL)
-				}(ch)
-				log.Debug("register finished")
-				registered = true
-				break
-			}
-		}
-		if registered == true {
-			return nil
-		}
-		return errors.New("not registerd, may due to already have enough pservers")
-	}, concurrency.WithAbortContext(ctx), concurrency.WithIsolation(concurrency.RepeatableReads))
-}
-
 // InitParam initializes a parameter.
 func (s *Service) InitParam(paramWithConfigs ParameterWithConfig, dummy *int) error {
 	select {

go/pserver/service_test.go

@@ -10,7 +10,7 @@ import (
 )
 
 func TestFull(t *testing.T) {
-	s, err := pserver.NewService("", time.Second*5)
+	s, err := pserver.NewService(0)
 	if err != nil {
 		t.Error(err)
 	}
@@ -75,7 +75,7 @@ func TestFull(t *testing.T) {
 }
 
 func TestMultipleInit(t *testing.T) {
-	s, err := pserver.NewService("", time.Second*5)
+	s, err := pserver.NewService(0)
 	if err != nil {
 		t.Error(err)
 	}
@@ -91,7 +91,7 @@ func TestMultipleInit(t *testing.T) {
 }
 
 func TestUninitialized(t *testing.T) {
-	s, err := pserver.NewService("", time.Second*5)
+	s, err := pserver.NewService(0)
 	err = s.SendGrad(pserver.Gradient{}, nil)
 	if err.Error() != pserver.Uninitialized {
 		t.FailNow()
@@ -99,7 +99,7 @@ func TestUninitialized(t *testing.T) {
 }
 
 func TestBlockUntilInitialized(t *testing.T) {
-	s, err := pserver.NewService("", time.Second*5)
+	s, err := pserver.NewService(0)
 	if err != nil {
 		t.Error(err)
 	}

paddle/CMakeLists.txt

@@ -9,17 +9,10 @@ add_subdirectory(pserver)
 add_subdirectory(trainer)
 add_subdirectory(scripts)
 add_subdirectory(optimizer)
-add_subdirectory(strings)
-
-# Do not build go directory until go cmake is working smoothly.
-# if(CMAKE_Go_COMPILER)
-#   add_subdirectory(go)
-# endif()
-
-find_package(Boost QUIET)
+add_subdirectory(string)
 
 if(Boost_FOUND)
-  include_directories(${Boost_INCLUDE_DIRS})
+  add_subdirectory(memory)
   add_subdirectory(platform)
   add_subdirectory(framework)
 endif()

paddle/framework/CMakeLists.txt

 # ddim lib
 cc_library(ddim SRCS ddim.cc)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
-
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
-
 cc_test(variable_test SRCS variable_test.cc)
-
 cc_test(scope_test SRCS scope_test.cc)
+cc_test(enforce_test SRCS enforce_test.cc)

paddle/framework/enforce.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/string/printf.h>
+#include <exception>
+#include <sstream>
+
+namespace paddle {
+namespace framework {
+
+/**
+ * @brief Enforce exception. Inherits std::exception
+ *
+ * All enforce condition not met, will throw an EnforceNotMet exception.
+ */
+class EnforceNotMet : public std::exception {
+ public:
+  EnforceNotMet(const std::string& msg, const char* file, int fileline) {
+    std::ostringstream sout;
+    sout << msg << " at [" << file << ":" << fileline << "];";
+    all_msg_ = sout.str();
+  }
+
+  const char* what() const noexcept override { return all_msg_.c_str(); }
+
+ private:
+  std::string all_msg_;
+};
+
+// From https://stackoverflow.com/questions/30130930/
+// __buildin_expect is in C++ 11 standard. Since the condition which enforced
+// should be true in most situation, it will make the compiler generate faster
+// code by adding `UNLIKELY` macro.
+#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
+
+/**
+ * @brief Throw a EnforceNotMet exception, automatically filled __FILE__ &
+ * __LINE__
+ *
+ * This macro take __VA_ARGS__, user can pass any type if that type can
+ * serialize to std::ostream
+ */
+#define PADDLE_THROW(...)                                            \
+  do {                                                               \
+    throw ::paddle::framework::EnforceNotMet(                        \
+        ::paddle::string::Sprintf(__VA_ARGS__), __FILE__, __LINE__); \
+  } while (0)
+
+/**
+ * @brief Enforce a condition, otherwise throw an EnforceNotMet
+ */
+#define PADDLE_ENFORCE(condition, ...) \
+  do {                                 \
+    if (UNLIKELY(!(condition))) {      \
+      PADDLE_THROW(__VA_ARGS__);       \
+    }                                  \
+  } while (0)
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/enforce_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. */
+
+#include <gtest/gtest.h>
+#include <paddle/framework/enforce.h>
+
+TEST(ENFORCE, OK) {
+  PADDLE_ENFORCE(true, "Enforce is ok %d now %f", 123, 0.345);
+  size_t val = 1;
+  const size_t limit = 10;
+  PADDLE_ENFORCE(val < limit, "Enforce is OK too");
+}
+
+TEST(ENFORCE, FAILED) {
+  bool in_catch = false;
+  try {
+    PADDLE_ENFORCE(false, "Enforce is not ok %d at all", 123);
+  } catch (paddle::framework::EnforceNotMet err) {
+    in_catch = true;
+    std::string msg = "Enforce is not ok 123 at all";
+    const char* what = err.what();
+    for (size_t i = 0; i < msg.length(); ++i) {
+      ASSERT_EQ(what[i], msg[i]);
+    }
+  }
+  ASSERT_TRUE(in_catch);
+}
\ No newline at end of file

paddle/gserver/gradientmachines/MultiGradientMachine.cpp

@@ -601,7 +601,7 @@ void TrainerThread::backward() {
 
 void TrainerThread::backwardCallback(Parameter* para) {
   // CPU parameters are merged in the end
-  if (!para->useGpu()) return;
+  if (!para->useGpu() || para->isStatic()) return;
 
   int paramId = para->getID();
   if (multiMachine_->getNumThreads() == 1) {

paddle/memory/.clang-format

+---
+Language:        Cpp
+BasedOnStyle:  Google
+Standard:  Cpp11 
+...

paddle/memory/CMakeLists.txt

+add_subdirectory(detail)

paddle/memory/README.md

@@ -97,6 +97,7 @@ class BuddyAllocator {
   struct Block {
     size_t size;
     Block* left, right;
+    size_t index; // allocator id
   };
   ...
 };

paddle/memory/detail/CMakeLists.txt

+if(${WITH_GPU})
+  nv_library(system_allocator SRCS system_allocator.cc DEPS gflags)
+  nv_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags)
+else(${WITH_GPU})
+  cc_library(system_allocator SRCS system_allocator.cc DEPS gflags)
+  cc_test(system_allocator_test SRCS system_allocator_test.cc DEPS system_allocator gflags)
+endif(${WITH_GPU})

paddle/memory/detail/buddy_allocator.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 "paddle/memory/detail/buddy_allocator.h"
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+BuddyAllocator::BuddyAllocator(size_t pool_size, size_t max_pools,
+                               SystemAllocator* system_allocator)
+    : pool_size_(pool_size),
+      max_pools_(max_pools),
+      system_allocator_(system_allocator) {
+  PADDLE_ASSERT(pool_size > 0);
+  PADDLE_ASSERT(max_pools > 0);
+  PADDLE_ASSERT(system_allocator != nullptr);
+}
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/buddy_allocator.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/memory/detail/system_allocator.h"
+
+#include <mutex>
+#include <vector>
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+class BuddyAllocator {
+ public:
+  BuddyAllocator(size_t pool_size, size_t max_pools,
+                 SystemAllocator* system_allocator);
+  ~BuddyAllocator();
+
+  void* Alloc(size_t size);
+  void Free(void*);
+  size_t Used();
+
+ private:
+  struct Block {
+    size_t size_;
+    Block* left_;   // left buddy
+    Block* right_;  // right buddy
+  };
+
+  // Initially, there is only one pool.  If a Alloc founds not enough
+  // memory from that pool, and there has not been max_num_pools_,
+  // create a new pool by calling system_allocator_.Alloc(pool_size_).
+  std::vector<void*> pools_;
+
+  size_t pool_size_;      // the size of each pool;
+  size_t max_num_pools_;  // the size of all pools;
+
+  SystemAllocator* system_allocator_;
+
+  std::mutex mutex_;
+
+  // Disable copy and assignment.
+  BuddyAllocator(const BuddyAllocator&) = delete;
+  BuddyAllocator& operator=(const BuddyAllocator&) = delete;
+};
+
+BuddyAllocator<CPUAllocator>* GetCPUBuddyAllocator() {
+  static BuddyAllocator<CPUAllocator>* a = nullptr;
+  if (a == nullptr) {
+    a = new BuddyAllocator<CPUAllocator>();
+  }
+  return a;
+}
+
+#ifndef PADDLE_ONLY_CPU  // The following code are for CUDA.
+
+BuddyAllocator<GPUAllocator>* GetGPUBuddyAllocator(int gpu_id) {
+  static BuddyAllocator<GPUAllocator>** as = NULL;
+  if (as == NULL) {
+    int gpu_num = platform::GetDeviceCount();
+    as = new BuddyAllocator<GPUAllocator>*[gpu_num];
+    for (int gpu = 0; gpu < gpu_num; gpu++) {
+      as[gpu] = new BuddyAllocator<GPUAllocator>();
+    }
+  }
+  return as[gpu_id];
+}
+
+#endif  // PADDLE_ONLY_CPU
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/system_allocator.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. */
+
+#include "paddle/memory/detail/system_allocator.h"
+
+#include <stdlib.h>    // for malloc and free
+#include <sys/mman.h>  // for mlock and munlock
+
+#include "gflags/gflags.h"
+#include "paddle/platform/assert.h"
+#include "paddle/platform/cuda.h"
+
+// If use_pinned_memory is true, CPUAllocator calls mlock, which
+// returns pinned and locked memory as staging areas for data exchange
+// between host and device.  Allocates too much would reduce the amount
+// of memory available to the system for paging.  So, by default, we
+// should set false to use_pinned_memory.
+DEFINE_bool(use_pinned_memory, false,
+            "If set, allocate cpu/gpu pinned memory.");
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+void* CPUAllocator::Alloc(size_t size) {
+  // According to http://www.cplusplus.com/reference/cstdlib/malloc/,
+  // malloc might not return nullptr if size is zero, but the returned
+  // pointer shall not be dereferenced -- so we make it nullptr.
+  if (size <= 0) return nullptr;
+
+  void* p = malloc(size);
+  if (p != nullptr && FLAGS_use_pinned_memory) {
+    mlock(p, size);
+  }
+  return p;
+}
+
+void CPUAllocator::Free(void* p, size_t size) {
+  if (p != nullptr && FLAGS_use_pinned_memory) {
+    munlock(p, size);
+  }
+  free(p);
+}
+
+#ifndef PADDLE_ONLY_CPU
+
+void* GPUAllocator::Alloc(size_t size) {
+  // CUDA documentation doesn't explain if cudaMalloc returns nullptr
+  // if size is 0.  We just make sure it does.
+  if (size <= 0) {
+    return nullptr;
+  }
+
+  void* p = 0;
+  cudaError_t result =
+      FLAGS_use_pinned_memory ? cudaMallocHost(&p, size) : cudaMalloc(&p, size);
+  if (result != cudaSuccess) {
+    cudaGetLastError();  // clear error if there is any.
+  }
+  return result == cudaSuccess ? p : nullptr;
+}
+
+void GPUAllocator::Free(void* p, size_t size) {
+  // Purposefully allow cudaErrorCudartUnloading, because
+  // that is returned if you ever call cudaFree after the
+  // driver has already shutdown. This happens only if the
+  // process is terminating, in which case we don't care if
+  // cudaFree succeeds.
+  cudaError_t err = FLAGS_use_pinned_memory ? cudaFreeHost(p) : cudaFree(p);
+  if (err != cudaErrorCudartUnloading) {
+    platform::throw_on_error(err, "cudaFree{Host} failed");
+  }
+}
+
+#endif  // PADDLE_ONLY_CPU
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/system_allocator.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 <stddef.h>  // for size_t
+
+namespace paddle {
+namespace memory {
+namespace detail {
+
+// SystemAllocator is the parent class of CPUAllocator and
+// GPUAllocator.  A BuddyAllocator object uses a SystemAllocator*
+// pointing to the underlying system allocator.  An alternative to
+// this class hierarchy is to pass a system allocator class to
+// BuddyAllocator as a template parameter.  This approach makes
+// BuddyAllocator a class template, and it's very complicated
+// algorithm would make the buddy_allocator.h messy.
+class SystemAllocator {
+ public:
+  virtual ~SystemAllocator() {}
+  virtual void* Alloc(size_t size) = 0;
+  virtual void Free(void* p, size_t size) = 0;
+};
+
+class CPUAllocator : public SystemAllocator {
+ public:
+  virtual void* Alloc(size_t size);
+  virtual void Free(void* p, size_t size);
+};
+
+#ifndef PADDLE_ONLY_CPU
+class GPUAllocator : public SystemAllocator {
+ public:
+  virtual void* Alloc(size_t size);
+  virtual void Free(void* p, size_t size);
+};
+#endif  // PADDLE_ONLY_CPU
+
+}  // namespace detail
+}  // namespace memory
+}  // namespace paddle

paddle/memory/detail/system_allocator_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. */
+
+#include "paddle/memory/detail/system_allocator.h"
+
+#include <memory>
+#include <vector>
+
+#include "gflags/gflags.h"
+#include "gtest/gtest.h"
+
+DECLARE_bool(use_pinned_memory);
+
+void TestAllocator(paddle::memory::detail::SystemAllocator& a, size_t size) {
+  bool freed = false;
+  {
+    void* p = a.Alloc(size);
+    if (size > 0) {
+      EXPECT_NE(p, nullptr);
+    } else {
+      EXPECT_EQ(p, nullptr);
+    }
+
+    int* i = static_cast<int*>(p);
+    std::shared_ptr<int> ptr(i, [&](void* p) {
+      freed = true;
+      a.Free(p, size);
+    });
+  }
+  EXPECT_TRUE(freed);
+}
+
+TEST(CPUAllocator, NoLockMem) {
+  FLAGS_use_pinned_memory = false;
+  paddle::memory::detail::CPUAllocator a;
+  TestAllocator(a, 2048);
+  TestAllocator(a, 0);
+}
+
+TEST(CPUAllocator, LockMem) {
+  FLAGS_use_pinned_memory = true;
+  paddle::memory::detail::CPUAllocator a;
+  TestAllocator(a, 2048);
+  TestAllocator(a, 0);
+}
+
+#ifndef PADDLE_ONLY_CPU
+TEST(GPUAllocator, NoStaging) {
+  FLAGS_use_pinned_memory = false;
+  paddle::memory::detail::GPUAllocator a;
+  TestAllocator(a, 2048);
+  TestAllocator(a, 0);
+}
+TEST(GPUAllocator, Staging) {
+  FLAGS_use_pinned_memory = true;
+  paddle::memory::detail::GPUAllocator a;
+  TestAllocator(a, 2048);
+  TestAllocator(a, 0);
+}
+#endif  // PADDLE_ONLY_CPU

paddle/memory/memory.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. */
+
+#include "paddle/memory/memory.h"
+#include "paddle/memory/detail/buddy_allocator.h"
+#include "paddle/memory/detail/system_allocator.h"
+#include "paddle/platform/assert.h"
+
+#include <boost/variant.hpp>
+
+namespace paddle {
+namespace memory {
+
+void* Alloc(platform::Place pl, size_t size) {
+#ifndef PADDLE_ONLY_CPU
+  if (paddle::platform::is_gpu_place(pl)) {
+    size_t gpu_id = boost::get<platform::GPUPlace>(pl).device;
+    return detail::GetGPUBuddyAllocator(gpu_id)->Alloc(size);
+  }
+#endif  // PADDLE_ONLY_CPU
+  PADDLE_ASSERT(paddle::platform::is_cpu_place(pl));
+  return detail::GetCPUBuddyAllocator()->Alloc(size);
+}
+
+void Free(paddle::platform::Place pl, void* p) {
+#ifndef PADDLE_ONLY_CPU
+  if (paddle::platform::is_gpu_place(pl)) {
+    size_t gpu_id = boost::get<platform::GPUPlace>(pl).device;
+    detail::GetGPUBuddyAllocator(gpu_id)->Free(p);
+  }
+#endif  // PADDLE_ONLY_CPU
+  PADDLE_ASSERT(paddle::platform::is_cpu_place(pl));
+  detail::GetCPUBuddyAllocator()->Free(p);
+}
+
+size_t Used(paddle::platform::Place pl) {
+#ifndef PADDLE_ONLY_CPU
+  if (paddle::platform::is_gpu_place(pl)) {
+    size_t gpu_id = boost::get<platform::GPUPlace>(pl).device;
+    return detail::GetGPUBuddyAllocator(gpu_id)->Used();
+  }
+#endif  // PADDLE_ONLY_CPU
+  PADDLE_ASSERT(paddle::platform::is_cpu_place(pl));
+  return detail::GetCPUBuddyAllocator()->Used();
+}
+
+}  // namespace memory
+}  // namespace paddle

paddle/platform/must_check.h → paddle/memory/memory.h

-/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+/* 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.
@@ -10,17 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-/**
- * __must_check macro. It make the function's return value must be used,
- * otherwise it will raise a compile warning. And also Paddle treat all compile
- * warnings as errors.
- */
-#ifdef __GNUC__
-#if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) >= 30400
-#define __must_check __attribute__((warn_unused_result))
-#else
-#define __must_check
-#endif
-#else
-#define __must_check
-#endif
+
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace memory {
+
+void* Alloc(paddle::platform::Place, size_t);
+void Free(paddle::platform::Place, void*);
+size_t Used(paddle::platform::Place);
+
+}  // namespace memory
+}  // namespace paddle

paddle/platform/CMakeLists.txt

@@ -2,4 +2,3 @@ nv_test(cuda_test SRCS cuda_test.cu)
 
 cc_library(place SRCS place.cc)
 cc_test(place_test SRCS place_test.cc DEPS place glog gflags)
-cc_test(must_check_test SRCS must_check_test.cc)

paddle/platform/cuda.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
+
+#ifndef PADDLE_ONLY_CPU
+
+#include <thrust/system/cuda/error.h>
+#include <thrust/system_error.h>
+
+namespace paddle {
+namespace platform {
+
+inline void throw_on_error(cudaError_t e, const char* message) {
+  if (e) {
+    throw thrust::system_error(e, thrust::cuda_category(), message);
+  }
+}
+
+int GetDeviceCount(void) {
+  int count;
+  throw_on_error(cudaGetDeviceCount(&count), "cudaGetDeviceCount failed");
+  return count;
+}
+
+}  // namespace platform
+}  // namespace paddle
+
+#endif  // PADDLE_ONLY_CPU

paddle/platform/must_check_test.cc

-#include <gtest/gtest.h>
-#include <paddle/platform/must_check.h>
-
-int __must_check SomeFunctionMustCheck() { return 0; }
-
-TEST(MustCheck, all) {
-  //  This line should not be compiled, because the
-  //  return value of SomeFunctionMustCheck marked as __must_check
-  //  SomeFunctionMustCheck();
-}
\ No newline at end of file

paddle/platform/place.cc

@@ -8,8 +8,8 @@ namespace detail {
 class PlacePrinter : public boost::static_visitor<> {
  public:
   PlacePrinter(std::ostream &os) : os_(os) {}
-  void operator()(const CpuPlace &) { os_ << "CpuPlace"; }
-  void operator()(const GpuPlace &p) { os_ << "GpuPlace(" << p.device << ")"; }
+  void operator()(const CPUPlace &) { os_ << "CPUPlace"; }
+  void operator()(const GPUPlace &p) { os_ << "GPUPlace(" << p.device << ")"; }
 
  private:
   std::ostream &os_;
@@ -22,14 +22,14 @@ static Place the_default_place;
 void set_place(const Place &place) { the_default_place = place; }
 const Place &get_place() { return the_default_place; }
 
-const GpuPlace default_gpu() { return GpuPlace(0); }
-const CpuPlace default_cpu() { return CpuPlace(); }
+const GPUPlace default_gpu() { return GPUPlace(0); }
+const CPUPlace default_cpu() { return CPUPlace(); }
 
 bool is_gpu_place(const Place &p) {
-  return boost::apply_visitor(IsGpuPlace(), p);
+  return boost::apply_visitor(IsGPUPlace(), p);
 }
 bool is_cpu_place(const Place &p) {
-  return !boost::apply_visitor(IsGpuPlace(), p);
+  return !boost::apply_visitor(IsGPUPlace(), p);
 }
 
 bool places_are_same_class(const Place &p1, const Place &p2) {

paddle/platform/place.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 <boost/variant.hpp>
 #include <iostream>
 
 namespace paddle {
 namespace platform {
 
-struct CpuPlace {
+struct CPUPlace {
   // WORKAROUND: for some reason, omitting this constructor
   // causes errors with boost 1.59 and OSX
-  CpuPlace() {}
+  CPUPlace() {}
 
   // needed for variant equality comparison
-  inline bool operator==(const CpuPlace &) const { return true; }
-  inline bool operator!=(const CpuPlace &) const { return false; }
+  inline bool operator==(const CPUPlace &) const { return true; }
+  inline bool operator!=(const CPUPlace &) const { return false; }
 };
 
-struct GpuPlace {
-  GpuPlace() : GpuPlace(0) {}
-  GpuPlace(int d) : device(d) {}
+struct GPUPlace {
+  GPUPlace() : GPUPlace(0) {}
+  GPUPlace(int d) : device(d) {}
 
   // needed for variant equality comparison
-  inline bool operator==(const GpuPlace &o) const { return device == o.device; }
-  inline bool operator!=(const GpuPlace &o) const { return !(*this == o); }
+  inline bool operator==(const GPUPlace &o) const { return device == o.device; }
+  inline bool operator!=(const GPUPlace &o) const { return !(*this == o); }
 
   int device;
 };
 
-struct IsGpuPlace : public boost::static_visitor<bool> {
-  bool operator()(const CpuPlace &) const { return false; }
-  bool operator()(const GpuPlace &gpu) const { return true; }
+struct IsGPUPlace : public boost::static_visitor<bool> {
+  bool operator()(const CPUPlace &) const { return false; }
+  bool operator()(const GPUPlace &gpu) const { return true; }
 };
 
-typedef boost::variant<GpuPlace, CpuPlace> Place;
+typedef boost::variant<GPUPlace, CPUPlace> Place;
 
 void set_place(const Place &);
 const Place &get_place();
 
-const GpuPlace default_gpu();
-const CpuPlace default_cpu();
+const GPUPlace default_gpu();
+const CPUPlace default_cpu();
 
 bool is_gpu_place(const Place &);
 bool is_cpu_place(const Place &);

paddle/platform/place_test.cc

@@ -3,8 +3,8 @@
 #include "gtest/gtest.h"
 
 TEST(Place, Equality) {
-  paddle::platform::CpuPlace cpu;
-  paddle::platform::GpuPlace g0(0), g1(1), gg0(0);
+  paddle::platform::CPUPlace cpu;
+  paddle::platform::GPUPlace g0(0), g1(1), gg0(0);
 
   EXPECT_EQ(cpu, cpu);
   EXPECT_EQ(g0, g0);
@@ -22,19 +22,19 @@ TEST(Place, Default) {
   EXPECT_TRUE(paddle::platform::is_gpu_place(paddle::platform::default_gpu()));
   EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::default_cpu()));
 
-  paddle::platform::set_place(paddle::platform::CpuPlace());
+  paddle::platform::set_place(paddle::platform::CPUPlace());
   EXPECT_TRUE(paddle::platform::is_cpu_place(paddle::platform::get_place()));
 }
 
 TEST(Place, Print) {
   {
     std::stringstream ss;
-    ss << paddle::platform::GpuPlace(1);
-    EXPECT_EQ("GpuPlace(1)", ss.str());
+    ss << paddle::platform::GPUPlace(1);
+    EXPECT_EQ("GPUPlace(1)", ss.str());
   }
   {
     std::stringstream ss;
-    ss << paddle::platform::CpuPlace();
-    EXPECT_EQ("CpuPlace", ss.str());
+    ss << paddle::platform::CPUPlace();
+    EXPECT_EQ("CPUPlace", ss.str());
   }
 }

paddle/py_paddle/dataprovider_converter.py

@@ -109,6 +109,10 @@ class DenseScanner(IScanner):
             if len(self.__shape__) > 3:
                 raise ValueError(
                     "The dimension of input cannot be greater than 3.")
+            if len(self.__shape__) == 0:
+                raise ValueError(
+                    "The input should be a vector, please check your input data."
+                )
             self.__dim__ = reduce(lambda x, y: x * y, self.__shape__)
             if len(self.__shape__) == 1 and self.__dim__ != self.input_type.dim:
                 raise ValueError(
@@ -140,7 +144,7 @@ class DenseScanner(IScanner):
         if len(self.__shape__) > 1:
             # The last-two dimenstions are the frame height and width.
             # For example, the layout is CHW for 3-D feature of image.
-            # The H and W are the fram height and width.
+            # The H and W are the frame height and width.
             h, w = self.__shape__[-2:]
             argument.setSlotFrameHeight(self.pos, h)
             argument.setSlotFrameWidth(self.pos, w)

paddle/string/CMakeLists.txt

+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)

paddle/strings/stringpiece.cc → paddle/string/piece.cc

@@ -14,7 +14,7 @@
   limitations under the License.
 */
 
-#include "paddle/strings/stringpiece.h"
+#include "paddle/string/piece.h"
 
 #include <string.h>
 
@@ -23,29 +23,25 @@
 #include <stdexcept>
 
 namespace paddle {
+namespace string {
 
-StringPiece::StringPiece() : data_(NULL), size_(0) {}
+Piece::Piece() : data_(NULL), size_(0) {}
 
-StringPiece::StringPiece(const char* d, size_t n) : data_(d), size_(n) {
+Piece::Piece(const char* d, size_t n) : data_(d), size_(n) {
   if (d == NULL && n != 0)
-    throw std::invalid_argument(
-        "StringPiece requires len to be 0 for NULL data");
+    throw std::invalid_argument("Piece requires len to be 0 for NULL data");
 }
 
-StringPiece::StringPiece(const char* s) : data_(s) {
-  size_ = (s == NULL) ? 0 : strlen(s);
-}
+Piece::Piece(const char* s) : data_(s) { size_ = (s == NULL) ? 0 : strlen(s); }
 
-StringPiece::StringPiece(const std::string& s)
-    : data_(s.data()), size_(s.size()) {}
+Piece::Piece(const std::string& s) : data_(s.data()), size_(s.size()) {}
 
-char StringPiece::operator[](size_t n) const {
-  if (n >= len())
-    throw std::invalid_argument("index out of StringPiece length");
+char Piece::operator[](size_t n) const {
+  if (n >= len()) throw std::invalid_argument("index out of Piece length");
   return data_[n];
 }
 
-int Compare(StringPiece a, StringPiece b) {
+int Compare(Piece a, Piece b) {
   const size_t min_len = (a.len() < b.len()) ? a.len() : b.len();
   int r = memcmp(a.data(), b.data(), min_len);
   if (r == 0) {
@@ -57,85 +53,86 @@ int Compare(StringPiece a, StringPiece b) {
   return r;
 }
 
-bool operator==(StringPiece x, StringPiece y) {
+bool operator==(Piece x, Piece y) {
   return ((x.len() == y.len()) &&
           (x.data() == y.data() || memcmp(x.data(), y.data(), x.len()) == 0));
 }
 
-bool operator!=(StringPiece x, StringPiece y) { return !(x == y); }
+bool operator!=(Piece x, Piece y) { return !(x == y); }
 
-bool operator<(StringPiece x, StringPiece y) { return Compare(x, y) < 0; }
-bool operator>(StringPiece x, StringPiece y) { return Compare(x, y) > 0; }
+bool operator<(Piece x, Piece y) { return Compare(x, y) < 0; }
+bool operator>(Piece x, Piece y) { return Compare(x, y) > 0; }
 
-bool operator<=(StringPiece x, StringPiece y) { return Compare(x, y) <= 0; }
-bool operator>=(StringPiece x, StringPiece y) { return Compare(x, y) >= 0; }
+bool operator<=(Piece x, Piece y) { return Compare(x, y) <= 0; }
+bool operator>=(Piece x, Piece y) { return Compare(x, y) >= 0; }
 
-bool HasPrefix(StringPiece s, StringPiece x) {
+bool HasPrefix(Piece s, Piece x) {
   return ((s.len() >= x.len()) && (memcmp(s.data(), x.data(), x.len()) == 0));
 }
 
-bool HasSuffix(StringPiece s, StringPiece x) {
+bool HasSuffix(Piece s, Piece x) {
   return ((s.len() >= x.len()) &&
           (memcmp(s.data() + (s.len() - x.len()), x.data(), x.len()) == 0));
 }
 
-StringPiece SkipPrefix(StringPiece s, size_t n) {
+Piece SkipPrefix(Piece s, size_t n) {
   if (n > s.len())
-    throw std::invalid_argument("Skip distance larger than StringPiece length");
-  return StringPiece(s.data() + n, s.len() - n);
+    throw std::invalid_argument("Skip distance larger than Piece length");
+  return Piece(s.data() + n, s.len() - n);
 }
 
-StringPiece SkipSuffix(StringPiece s, size_t n) {
+Piece SkipSuffix(Piece s, size_t n) {
   if (n > s.len())
-    throw std::invalid_argument("Skip distance larger than StringPiece length");
-  return StringPiece(s.data(), s.len() - n);
+    throw std::invalid_argument("Skip distance larger than Piece length");
+  return Piece(s.data(), s.len() - n);
 }
 
-StringPiece TrimPrefix(StringPiece s, StringPiece x) {
+Piece TrimPrefix(Piece s, Piece x) {
   return HasPrefix(s, x) ? SkipPrefix(s, x.len()) : s;
 }
 
-StringPiece TrimSuffix(StringPiece s, StringPiece x) {
+Piece TrimSuffix(Piece s, Piece x) {
   return HasSuffix(s, x) ? SkipSuffix(s, x.len()) : s;
 }
 
-bool Contains(StringPiece s, StringPiece sub) {
+bool Contains(Piece s, Piece sub) {
   return std::search(s.begin(), s.end(), sub.begin(), sub.end()) != s.end();
 }
 
-size_t Index(StringPiece s, StringPiece sub) {
+size_t Index(Piece s, Piece sub) {
   auto e = std::search(s.begin(), s.end(), sub.begin(), sub.end());
-  return e != s.end() ? e - s.data() : StringPiece::npos;
+  return e != s.end() ? e - s.data() : Piece::npos;
 }
 
-size_t Find(StringPiece s, char c, size_t pos) {
+size_t Find(Piece s, char c, size_t pos) {
   if (pos >= s.len()) {
-    return StringPiece::npos;
+    return Piece::npos;
   }
   const char* result =
       reinterpret_cast<const char*>(memchr(s.data() + pos, c, s.len() - pos));
-  return result != nullptr ? result - s.data() : StringPiece::npos;
+  return result != nullptr ? result - s.data() : Piece::npos;
 }
 
-size_t RFind(StringPiece s, char c, size_t pos) {
-  if (s.len() == 0) return StringPiece::npos;
+size_t RFind(Piece s, char c, size_t pos) {
+  if (s.len() == 0) return Piece::npos;
   for (const char* p = s.data() + std::min(pos, s.len() - 1); p >= s.data();
        p--) {
     if (*p == c) {
       return p - s.data();
     }
   }
-  return StringPiece::npos;
+  return Piece::npos;
 }
 
-StringPiece SubStr(StringPiece s, size_t pos, size_t n) {
+Piece SubStr(Piece s, size_t pos, size_t n) {
   if (pos > s.len()) pos = s.len();
   if (n > s.len() - pos) n = s.len() - pos;
-  return StringPiece(s.data() + pos, n);
+  return Piece(s.data() + pos, n);
 }
 
-std::ostream& operator<<(std::ostream& o, StringPiece piece) {
+std::ostream& operator<<(std::ostream& o, Piece piece) {
   return o << piece.ToString();
 }
 
+}  // namespace string
 }  // namespace paddle

paddle/strings/stringpiece.h → paddle/string/piece.h

@@ -20,33 +20,34 @@
 #include <string>
 
 namespace paddle {
+namespace string {
 
-// StringPiece points into a std::string object but doesn't own the
+// Piece points into a std::string object but doesn't own the
 // string.  It is for efficient access to strings.  Like Go's string
-// type.  Not that StringPiece doesn't mutate the underlying string,
+// type.  Not that Piece doesn't mutate the underlying string,
 // so it is thread-safe given that the underlying string doesn't
-// change.  Because StringPiece contains a little data members, and
+// change.  Because Piece contains a little data members, and
 // its syntax is simple as it doesn't own/manage the string, it is
-// cheap to construct StringPieces and pass them around.
-class StringPiece {
+// cheap to construct Pieces and pass them around.
+class Piece {
 public:
   static const size_t npos = static_cast<size_t>(-1);
 
   // We provide non-explicit singleton constructors so users can
-  // pass in a "const char*" or a "string" wherever a "StringPiece"
+  // pass in a "const char*" or a "string" wherever a "Piece"
   // is expected.  These contructors ensure that if data_ is NULL,
   // size_ is 0.
-  StringPiece();
-  StringPiece(const char* d, size_t n);
-  StringPiece(const char* d);
-  StringPiece(const std::string& s);
+  Piece();
+  Piece(const char* d, size_t n);
+  Piece(const char* d);
+  Piece(const std::string& s);
 
   const char* data() const { return data_; }
   size_t len() const { return size_; }
 
   char operator[](size_t n) const;
 
-  // StringPiece doesn't own the string, so both iterator and const
+  // Piece doesn't own the string, so both iterator and const
   // iterator are const char* indeed.
   typedef const char* const_iterator;
   typedef const char* iterator;
@@ -63,43 +64,44 @@ private:
   // Intentionally copyable
 };
 
-int Compare(StringPiece a, StringPiece b);
+int Compare(Piece a, Piece b);
 
-bool operator==(StringPiece x, StringPiece y);
-bool operator!=(StringPiece x, StringPiece y);
-bool operator<(StringPiece x, StringPiece y);
-bool operator>(StringPiece x, StringPiece y);
-bool operator<=(StringPiece x, StringPiece y);
-bool operator>=(StringPiece x, StringPiece y);
+bool operator==(Piece x, Piece y);
+bool operator!=(Piece x, Piece y);
+bool operator<(Piece x, Piece y);
+bool operator>(Piece x, Piece y);
+bool operator<=(Piece x, Piece y);
+bool operator>=(Piece x, Piece y);
 
-bool HasPrefix(StringPiece s, StringPiece prefix);
-bool HasSuffix(StringPiece s, StringPiece suffix);
+bool HasPrefix(Piece s, Piece prefix);
+bool HasSuffix(Piece s, Piece suffix);
 
-StringPiece SkipPrefix(StringPiece s, size_t n);
-StringPiece SkipSuffix(StringPiece s, size_t n);
+Piece SkipPrefix(Piece s, size_t n);
+Piece SkipSuffix(Piece s, size_t n);
 
 // Skip the prefix (or suffix) if it matches with the string.
-StringPiece TrimPrefix(StringPiece s, StringPiece prefix);
-StringPiece TrimSuffix(StringPiece s, StringPiece suffix);
+Piece TrimPrefix(Piece s, Piece prefix);
+Piece TrimSuffix(Piece s, Piece suffix);
 
 // Returns if s contains sub.  Any s except for empty s contains an
 // empty sub.
-bool Contains(StringPiece s, StringPiece sub);
+bool Contains(Piece s, Piece sub);
 
 // Return the first occurrence of sub in s, or npos.  If both s and
 // sub is empty, it returns npos; otherwise, if only sub is empty, it
 // returns 0.
-size_t Index(StringPiece s, StringPiece sub);
+size_t Index(Piece s, Piece sub);
 
 // Return the first occurrence of c in s[pos:end], or npos.
-size_t Find(StringPiece s, char c, size_t pos);
+size_t Find(Piece s, char c, size_t pos);
 
 // Search range is [0..pos] inclusive.  If pos == npos, search everything.
-size_t RFind(StringPiece s, char c, size_t pos);
+size_t RFind(Piece s, char c, size_t pos);
 
-StringPiece SubStr(StringPiece s, size_t pos, size_t n);
+Piece SubStr(Piece s, size_t pos, size_t n);
 
-// allow StringPiece to be logged
-std::ostream& operator<<(std::ostream& o, StringPiece piece);
+// allow Piece to be logged
+std::ostream& operator<<(std::ostream& o, Piece piece);
 
+}  // namespace string
 }  // namespace paddle

paddle/strings/stringpiece_test.cc → paddle/string/piece_test.cc

@@ -14,7 +14,7 @@
   limitations under the License.
 */
 
-#include "paddle/strings/stringpiece.h"
+#include "paddle/string/piece.h"
 
 #include <sstream>
 
@@ -22,42 +22,44 @@
 
 TEST(StringPiece, Construct) {
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ(NULL, s.data());
     EXPECT_EQ(0U, s.len());
   }
-  { EXPECT_THROW(paddle::StringPiece s(NULL, 10000U), std::invalid_argument); }
   {
-    paddle::StringPiece s(NULL);
+    EXPECT_THROW(paddle::string::Piece s(NULL, 10000U), std::invalid_argument);
+  }
+  {
+    paddle::string::Piece s(NULL);
     EXPECT_EQ(0U, s.len());
   }
   {
     std::string a;
     EXPECT_EQ(0U, a.size());
-    paddle::StringPiece s(a);
+    paddle::string::Piece s(a);
     EXPECT_EQ(0U, s.len());
   }
 }
 
 TEST(StringPiece, CopyAndAssign) {
-  paddle::StringPiece empty;
+  paddle::string::Piece empty;
   EXPECT_EQ(0U, empty.len());
 
-  paddle::StringPiece a("hello");
-  paddle::StringPiece b = a;
+  paddle::string::Piece a("hello");
+  paddle::string::Piece b = a;
   EXPECT_EQ(b.len(), strlen("hello"));
   EXPECT_EQ(a, b);
 
   std::string storage("hello");
-  paddle::StringPiece c(storage);
+  paddle::string::Piece c(storage);
   EXPECT_EQ(a, c);
   EXPECT_NE(a.data(), c.data());
 }
 
 TEST(StringPiece, Compare) {
   {
-    paddle::StringPiece a("hello");
-    paddle::StringPiece b("world");
+    paddle::string::Piece a("hello");
+    paddle::string::Piece b("world");
     EXPECT_TRUE(a != b);
     EXPECT_FALSE(a == b);
     EXPECT_TRUE(a < b);
@@ -68,7 +70,7 @@ TEST(StringPiece, Compare) {
     EXPECT_GT(Compare(b, a), 0);
   }
   {
-    paddle::StringPiece a, b;
+    paddle::string::Piece a, b;
     EXPECT_TRUE(a == b);
     EXPECT_FALSE(a != b);
     EXPECT_FALSE(a < b);
@@ -82,31 +84,31 @@ TEST(StringPiece, Compare) {
 
 TEST(StringPiece, ToString) {
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ(std::string(""), s.ToString());
   }
   {
-    paddle::StringPiece s(NULL);
+    paddle::string::Piece s(NULL);
     EXPECT_EQ(std::string(""), s.ToString());
   }
   {
-    paddle::StringPiece s("hello");
+    paddle::string::Piece s("hello");
     EXPECT_EQ(std::string("hello"), s.ToString());
   }
 }
 
 TEST(StringPiece, HasPrefixSuffix) {
-  using paddle::HasPrefix;
-  using paddle::HasSuffix;
+  using paddle::string::HasPrefix;
+  using paddle::string::HasSuffix;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_FALSE(HasPrefix(s, "something"));
     EXPECT_TRUE(HasPrefix(s, ""));
     EXPECT_FALSE(HasSuffix(s, "something"));
     EXPECT_TRUE(HasSuffix(s, ""));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_TRUE(HasPrefix(s, ""));
     EXPECT_TRUE(HasPrefix(s, "a"));
     EXPECT_TRUE(HasPrefix(s, "ap"));
@@ -120,10 +122,10 @@ TEST(StringPiece, HasPrefixSuffix) {
 }
 
 TEST(StringPiece, SkipPrefixSuffix) {
-  using paddle::SkipPrefix;
-  using paddle::SkipSuffix;
+  using paddle::string::SkipPrefix;
+  using paddle::string::SkipSuffix;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ("", SkipPrefix(s, 0));
     EXPECT_THROW(SkipPrefix(s, 1), std::invalid_argument);
 
@@ -131,7 +133,7 @@ TEST(StringPiece, SkipPrefixSuffix) {
     EXPECT_THROW(SkipSuffix(s, 1), std::invalid_argument);
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_EQ("app", SkipPrefix(s, 0));
     EXPECT_EQ("pp", SkipPrefix(s, 1));
     EXPECT_EQ("p", SkipPrefix(s, 2));
@@ -147,10 +149,10 @@ TEST(StringPiece, SkipPrefixSuffix) {
 }
 
 TEST(StringPiece, TrimPrefixSuffix) {
-  using paddle::TrimPrefix;
-  using paddle::TrimSuffix;
+  using paddle::string::TrimPrefix;
+  using paddle::string::TrimSuffix;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ("", TrimPrefix(s, ""));
     EXPECT_EQ("", TrimPrefix(s, "something"));
 
@@ -158,7 +160,7 @@ TEST(StringPiece, TrimPrefixSuffix) {
     EXPECT_EQ("", TrimSuffix(s, "something"));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_EQ("app", TrimPrefix(s, ""));
     EXPECT_EQ("pp", TrimPrefix(s, "a"));
     EXPECT_EQ("p", TrimPrefix(s, "ap"));
@@ -174,14 +176,14 @@ TEST(StringPiece, TrimPrefixSuffix) {
 }
 
 TEST(StringPiece, Contains) {
-  using paddle::Contains;
+  using paddle::string::Contains;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_FALSE(Contains(s, ""));
     EXPECT_FALSE(Contains(s, "something"));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_TRUE(Contains(s, ""));
     EXPECT_TRUE(Contains(s, "a"));
     EXPECT_TRUE(Contains(s, "p"));
@@ -193,15 +195,15 @@ TEST(StringPiece, Contains) {
 }
 
 TEST(StringPiece, Index) {
-  using paddle::Index;
-  auto npos = paddle::StringPiece::npos;
+  using paddle::string::Index;
+  auto npos = paddle::string::Piece::npos;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ(npos, Index(s, ""));
     EXPECT_EQ(npos, Index(s, "something"));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_EQ(0U, Index(s, ""));
     EXPECT_EQ(0U, Index(s, "a"));
     EXPECT_EQ(1U, Index(s, "p"));
@@ -213,14 +215,14 @@ TEST(StringPiece, Index) {
 }
 
 TEST(StringPiece, Find) {
-  using paddle::Find;
-  auto npos = paddle::StringPiece::npos;
+  using paddle::string::Find;
+  auto npos = paddle::string::Piece::npos;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ(npos, Find(s, 'a', 0U));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_EQ(0U, Find(s, 'a', 0U));
     EXPECT_EQ(1U, Find(s, 'p', 0U));
     EXPECT_EQ(1U, Find(s, 'p', 1U));
@@ -230,14 +232,14 @@ TEST(StringPiece, Find) {
 }
 
 TEST(StringPiece, RFind) {
-  using paddle::RFind;
-  auto npos = paddle::StringPiece::npos;
+  using paddle::string::RFind;
+  auto npos = paddle::string::Piece::npos;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ(npos, RFind(s, 'a', 0U));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_EQ(2U, RFind(s, 'p', 2U));
     EXPECT_EQ(0U, RFind(s, 'a', 2U));
     EXPECT_EQ(1U, RFind(s, 'p', 1U));
@@ -247,15 +249,15 @@ TEST(StringPiece, RFind) {
 }
 
 TEST(StringPiece, SubStr) {
-  using paddle::SubStr;
+  using paddle::string::SubStr;
   {
-    paddle::StringPiece s;
+    paddle::string::Piece s;
     EXPECT_EQ("", SubStr(s, 0, 0));
     EXPECT_EQ("", SubStr(s, 0, 1));
     EXPECT_EQ("", SubStr(s, 1, 0));
   }
   {
-    paddle::StringPiece s("app");
+    paddle::string::Piece s("app");
     EXPECT_EQ("", SubStr(s, 0, 0));
     EXPECT_EQ("", SubStr(s, 1, 0));
     EXPECT_EQ("", SubStr(s, 2, 0));
@@ -279,15 +281,15 @@ TEST(StringPiece, SubStr) {
 }
 
 TEST(StringPiece, StreamOutput) {
-  using paddle::StringPiece;
+  using paddle::string::Piece;
 
   std::stringstream o;
-  o << StringPiece();
+  o << paddle::string::Piece();
   EXPECT_EQ("", o.str());
 
-  o << StringPiece("hello");
+  o << paddle::string::Piece("hello");
   EXPECT_EQ("hello", o.str());
 
-  o << StringPiece();
+  o << paddle::string::Piece();
   EXPECT_EQ("hello", o.str());
 }

paddle/string/printf.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.
+*/
+
+// Compared with std::stringstream, there are primary purpose of
+// string::Printf:
+//
+// 1. Type-safe printing, with why and how explained in
+//    http://www.drdobbs.com/stringprintf-a-typesafe-printf-family-fo/184401999.
+//    Implementation includes
+//
+//    https://github.com/c42f/tinyformat
+//    boost::format
+//    std::stringstream
+//
+//    std::stringstream is not convenient enough in many cases.  For example:
+//
+//      std::cout << std::setprecision(2) << std::fixed << 1.23456 << "\n";
+//
+//    boost::format is the most convenient one.  We can have
+//
+//      std::cout << format("%2% %1%") % 36 % 77;
+//
+//    or
+//
+//      format fmter("%2% %1%");
+//      fmter % 36; fmter % 77;
+//      std::cout << fmter.c_str();
+//
+//    But the overloading of % might be overkilling and it would be
+//    more efficient if it can write to std::cout directly.
+//
+//    tinyformat has an interface compatible with the C-printf style,
+//    and it can writes to a stream or returns a std::string:
+//
+//      std::cout << tfm::printf(
+//                  "%s, %s %d, %.2d:%.2d\n",
+//                  weekday, month, day, hour, min);
+//
+//    or
+//
+//      tfm::format(std::cout,
+//                  "%s, %s %d, %.2d:%.2d\n",
+//                  weekday, month, day, hour, min);
+//
+// 2. High-performance -- most printed strings are not too long and
+//    doens't need dynamic memory allocation.  Many StringPrintf
+//    implementations doesn't enforce type-safe, but are
+//    high-performance, including
+//
+//    https://developers.google.com/optimization/reference/base/stringprintf/
+//    https://github.com/adobe/chromium/blob/master/base/stringprintf.h
+//    https://github.com/google/protobuf/blob/master/src/google/protobuf/stubs/stringprintf.h
+//
+// According to
+// https://github.com/c42f/tinyformat#compile-time-and-code-bloat,
+// boost::format runs too slow and results in large executable binary
+// files.  So here we port tinyformat.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+#include "paddle/string/tinyformat/tinyformat.h"  // https://github.com/c42f/tinyformat
+
+namespace paddle {
+namespace string {
+
+template <typename... Args>
+void Fprintf(std::ostream& out, const char* fmt, const Args&... args) {
+  tinyformat::vformat(out, fmt, tinyformat::makeFormatList(args...));
+}
+
+template <typename... Args>
+std::string Sprintf(const char* fmt, const Args&... args) {
+  std::ostringstream oss;
+  Fprintf(oss, fmt, args...);
+  return oss.str();
+}
+
+template <typename... Args>
+void Printf(const char* fmt, const Args&... args) {
+  Fprintf(std::cout, fmt, args...);
+}
+
+}  // namespace string
+}  // namespace paddle

paddle/string/printf_test.cc

+#include "paddle/string/printf.h"
+
+#include <string>
+
+#include "gtest/gtest.h"
+
+TEST(StringPrintf, StringPrintf) {
+  std::string weekday = "Wednesday";
+  const char* month = "July";
+  size_t day = 27;
+  long hour = 14;
+  int min = 44;
+  EXPECT_EQ(std::string("Wednesday, July 27, 14:44"),
+            paddle::string::Sprintf(
+                "%s, %s %d, %.2d:%.2d", weekday, month, day, hour, min));
+}

paddle/strings/CMakeLists.txt

-cc_library(stringpiece SRCS stringpiece.cc)
-cc_test(stringpiece_test SRCS stringpiece_test.cc DEPS stringpiece glog gflags)

paddle/utils/Error.h

@@ -19,7 +19,21 @@ limitations under the License. */
 #include <stdio.h>
 #include <memory>
 #include <string>
-#include "paddle/platform/must_check.h"
+
+/**
+ * __must_check macro. It make the function's return value must be used,
+ * otherwise it will raise a compile warning. And also Paddle treat all compile
+ * warnings as errors.
+ */
+#ifdef __GNUC__
+#if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) >= 30400
+#define __must_check __attribute__((warn_unused_result))
+#else
+#define __must_check
+#endif
+#else
+#define __must_check
+#endif
 
 namespace paddle {
 

python/paddle/v2/dataset/__init__.py

@@ -25,8 +25,9 @@ import uci_housing
 import sentiment
 import wmt14
 import mq2007
+import flowers
 
 __all__ = [
     'mnist', 'imikolov', 'imdb', 'cifar', 'movielens', 'conll05', 'sentiment'
-    'uci_housing', 'wmt14', 'mq2007'
+    'uci_housing', 'wmt14', 'mq2007', 'flowers'
 ]

python/paddle/v2/dataset/flowers.py

@@ -13,18 +13,18 @@
 # limitations under the License.
 """
 This module will download dataset from
-http://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html 
+http://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html
 and parse train/test set intopaddle reader creators.
 
-This set contains images of flowers belonging to 102 different categories. 
+This set contains images of flowers belonging to 102 different categories.
 The images were acquired by searching the web and taking pictures. There are a
 minimum of 40 images for each category.
 
 The database was used in:
 
 Nilsback, M-E. and Zisserman, A. Automated flower classification over a large
- number of classes.Proceedings of the Indian Conference on Computer Vision, 
-Graphics and Image Processing (2008) 
+ number of classes.Proceedings of the Indian Conference on Computer Vision,
+Graphics and Image Processing (2008)
 http://www.robots.ox.ac.uk/~vgg/publications/papers/nilsback08.{pdf,ps.gz}.
 
 """
@@ -34,9 +34,9 @@ from common import download
 import tarfile
 import scipy.io as scio
 from paddle.v2.image import *
+from paddle.v2.reader import *
 import os
 import numpy as np
-import paddle.v2 as paddle
 from multiprocessing import cpu_count
 __all__ = ['train', 'test', 'valid']
 
@@ -46,6 +46,12 @@ SETID_URL = 'http://www.robots.ox.ac.uk/~vgg/data/flowers/102/setid.mat'
 DATA_MD5 = '52808999861908f626f3c1f4e79d11fa'
 LABEL_MD5 = 'e0620be6f572b9609742df49c70aed4d'
 SETID_MD5 = 'a5357ecc9cb78c4bef273ce3793fc85c'
+# In official 'readme', tstid is the flag of test data
+# and trnid is the flag of train data. But test data is more than train data.
+# So we exchange the train data and test data.
+TRAIN_FLAG = 'tstid'
+TEST_FLAG = 'trnid'
+VALID_FLAG = 'valid'
 
 
 def default_mapper(sample):
@@ -53,8 +59,8 @@ def default_mapper(sample):
     map image bytes data to type needed by model input layer
     '''
     img, label = sample
-    img = paddle.image.load_image_bytes(img)
-    img = paddle.image.simple_transform(img, 256, 224, True)
+    img = load_image_bytes(img)
+    img = simple_transform(img, 256, 224, True)
     return img.flatten().astype('float32'), label
 
 
@@ -63,22 +69,23 @@ def reader_creator(data_file,
                    setid_file,
                    dataset_name,
                    mapper=default_mapper,
-                   buffered_size=1024):
+                   buffered_size=1024,
+                   use_xmap=True):
     '''
-    1. read images from tar file and 
+    1. read images from tar file and
         merge images into batch files in 102flowers.tgz_batch/
     2. get a reader to read sample from batch file
-    
-    :param data_file: downloaded data file 
+
+    :param data_file: downloaded data file
     :type data_file: string
-    :param label_file: downloaded label file 
+    :param label_file: downloaded label file
     :type label_file: string
     :param setid_file: downloaded setid file containing information
                         about how to split dataset
     :type setid_file: string
     :param dataset_name: data set name (tstid|trnid|valid)
     :type dataset_name: string
-    :param mapper: a function to map image bytes data to type 
+    :param mapper: a function to map image bytes data to type
                     needed by model input layer
     :type mapper: callable
     :param buffered_size: the size of buffer used to process images
@@ -105,15 +112,17 @@ def reader_creator(data_file,
             for sample, label in itertools.izip(data, batch['label']):
                 yield sample, int(label)
 
-    return paddle.reader.xmap_readers(mapper, reader,
-                                      cpu_count(), buffered_size)
+    if use_xmap:
+        return xmap_readers(mapper, reader, cpu_count(), buffered_size)
+    else:
+        return map_readers(mapper, reader)
 
 
-def train(mapper=default_mapper, buffered_size=1024):
+def train(mapper=default_mapper, buffered_size=1024, use_xmap=True):
     '''
-    Create flowers training set reader. 
-    It returns a reader, each sample in the reader is   
-    image pixels in [0, 1] and label in [1, 102] 
+    Create flowers training set reader.
+    It returns a reader, each sample in the reader is
+    image pixels in [0, 1] and label in [1, 102]
     translated from original color image by steps:
     1. resize to 256*256
     2. random crop to 224*224
@@ -128,15 +137,15 @@ def train(mapper=default_mapper, buffered_size=1024):
     return reader_creator(
         download(DATA_URL, 'flowers', DATA_MD5),
         download(LABEL_URL, 'flowers', LABEL_MD5),
-        download(SETID_URL, 'flowers', SETID_MD5), 'trnid', mapper,
-        buffered_size)
+        download(SETID_URL, 'flowers', SETID_MD5), TRAIN_FLAG, mapper,
+        buffered_size, use_xmap)
 
 
-def test(mapper=default_mapper, buffered_size=1024):
+def test(mapper=default_mapper, buffered_size=1024, use_xmap=True):
     '''
-    Create flowers test set reader. 
-    It returns a reader, each sample in the reader is   
-    image pixels in [0, 1] and label in [1, 102] 
+    Create flowers test set reader.
+    It returns a reader, each sample in the reader is
+    image pixels in [0, 1] and label in [1, 102]
     translated from original color image by steps:
     1. resize to 256*256
     2. random crop to 224*224
@@ -151,15 +160,15 @@ def test(mapper=default_mapper, buffered_size=1024):
     return reader_creator(
         download(DATA_URL, 'flowers', DATA_MD5),
         download(LABEL_URL, 'flowers', LABEL_MD5),
-        download(SETID_URL, 'flowers', SETID_MD5), 'tstid', mapper,
-        buffered_size)
+        download(SETID_URL, 'flowers', SETID_MD5), TEST_FLAG, mapper,
+        buffered_size, use_xmap)
 
 
-def valid(mapper=default_mapper, buffered_size=1024):
+def valid(mapper=default_mapper, buffered_size=1024, use_xmap=True):
     '''
-    Create flowers validation set reader. 
-    It returns a reader, each sample in the reader is   
-    image pixels in [0, 1] and label in [1, 102] 
+    Create flowers validation set reader.
+    It returns a reader, each sample in the reader is
+    image pixels in [0, 1] and label in [1, 102]
     translated from original color image by steps:
     1. resize to 256*256
     2. random crop to 224*224
@@ -174,8 +183,8 @@ def valid(mapper=default_mapper, buffered_size=1024):
     return reader_creator(
         download(DATA_URL, 'flowers', DATA_MD5),
         download(LABEL_URL, 'flowers', LABEL_MD5),
-        download(SETID_URL, 'flowers', SETID_MD5), 'valid', mapper,
-        buffered_size)
+        download(SETID_URL, 'flowers', SETID_MD5), VALID_FLAG, mapper,
+        buffered_size, use_xmap)
 
 
 def fetch():

python/paddle/v2/dataset/tests/flowers_test.py

@@ -31,13 +31,13 @@ class TestFlowers(unittest.TestCase):
     def test_train(self):
         instances, max_label_value = self.check_reader(
             paddle.v2.dataset.flowers.train())
-        self.assertEqual(instances, 1020)
+        self.assertEqual(instances, 6149)
         self.assertEqual(max_label_value, 102)
 
     def test_test(self):
         instances, max_label_value = self.check_reader(
             paddle.v2.dataset.flowers.test())
-        self.assertEqual(instances, 6149)
+        self.assertEqual(instances, 1020)
         self.assertEqual(max_label_value, 102)
 
     def test_valid(self):

python/paddle/v2/parameters.py

@@ -51,7 +51,7 @@ class Parameters(object):
     def __init__(self):
         self.__param_conf__ = dict()
         self.__gradient_machines__ = []
-        self.__tmp_params__ = []
+        self.__tmp_params__ = dict()
 
     def __append_config__(self, param_conf):
         """
@@ -128,13 +128,10 @@ class Parameters(object):
 
         if len(self.__gradient_machines__) == 0:
             # create new parameter in python numpy.
-            if len(self.__tmp_params__) != 0:
-                ret_list = [
-                    mat for name, mat in self.__tmp_params__ if name == key
-                ]
-                if len(ret_list) == 1:
-                    return ret_list[0]
-            return np.ndarray(shape=shape, dtype=np.float32)
+            if key in self.__tmp_params__:
+                return self.__tmp_params__[key]
+            else:
+                return np.ndarray(shape=shape, dtype=np.float32)
         else:
             for each_gradient_machine in self.__gradient_machines__:
                 param = __get_parameter_in_gradient_machine__(
@@ -187,7 +184,7 @@ class Parameters(object):
                              (shape, value.shape))
 
         if len(self.__gradient_machines__) == 0:
-            self.__tmp_params__.append((key, value))
+            self.__tmp_params__[key] = value
         else:
             for each_gradient_machine in self.__gradient_machines__:
                 __copy_parameter_to_gradient_machine__(each_gradient_machine,
@@ -231,7 +228,7 @@ class Parameters(object):
             raise ValueError("gradient_machine should be api.GradientMachine")
 
         if len(self.__tmp_params__) != 0:
-            for name, val in self.__tmp_params__:
+            for name, val in self.__tmp_params__.iteritems():
                 try:
                     __copy_parameter_to_gradient_machine__(gradient_machine,
                                                            name, val)
@@ -287,6 +284,18 @@ class Parameters(object):
 
     @staticmethod
     def from_tar(f):
+        """
+        Create a `Parameters` object from the given file. And
+        the `Parameters` only contains the parameters in this
+        file. It is adapted the parameters are same in the
+        defined network and the given file. For example, it
+        can be used in the inference.
+
+        :param f: the initialized model file.
+        :type f: tar file
+        :return: A Parameters object.
+        :rtype: Parameters.
+        """
         params = Parameters()
         tar = tarfile.TarFile(fileobj=f, mode='r')
         for finfo in tar:
@@ -302,6 +311,21 @@ class Parameters(object):
             params.deserialize(param_name, f)
         return params
 
+    def init_from_tar(self, f):
+        """
+        Different from `from_tar`, this interface can be used to
+        init partial network parameters from another saved model.
+
+        :param f: the initialized model file.
+        :type f: tar file
+        :return: Nothing.
+        """
+
+        tar_param = Parameters.from_tar(f)
+        for pname in tar_param.names():
+            if pname in self.names():
+                self.set(pname, tar_param.get(pname))
+
 
 def __get_parameter_in_gradient_machine__(gradient_machine, name):
     """

python/paddle/v2/reader/decorator.py

@@ -166,12 +166,12 @@ def buffered(reader, size):
     The buffered data reader will read and save data entries into a
     buffer. Reading from the buffered data reader will proceed as long
     as the buffer is not empty.
-    
+
     :param reader: the data reader to read from.
     :type reader: callable
     :param size: max buffer size.
     :type size: int
-    
+
     :returns: the buffered data reader.
     """
 
@@ -238,7 +238,7 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False):
     :type mapper: callable
     :param reader: the data reader to read from
     :type reader: callable
-    :param process_num: process number to handle original sample 
+    :param process_num: process number to handle original sample
     :type process_num: int
     :param buffer_size: max buffer size
     :type buffer_size: int
@@ -248,9 +248,6 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False):
     :rtype: callable
     """
     end = XmapEndSignal()
-    in_queue = Queue(buffer_size)
-    out_queue = Queue(buffer_size)
-    out_order = [0]
 
     # define a worker to read samples from reader to in_queue
     def read_worker(reader, in_queue):
@@ -266,12 +263,6 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False):
             in_order += 1
         in_queue.put(end)
 
-    # start a read worker in a thread
-    target = order_read_worker if order else read_worker
-    t = Thread(target=target, args=(reader, in_queue))
-    t.daemon = True
-    t.start()
-
     # define a worker to handle samples from in_queue by mapper
     # and put mapped samples into out_queue
     def handle_worker(in_queue, out_queue, mapper):
@@ -298,19 +289,27 @@ def xmap_readers(mapper, reader, process_num, buffer_size, order=False):
         in_queue.put(end)
         out_queue.put(end)
 
-    # start several handle_workers
-    target = order_handle_worker if order else handle_worker
-    args = (in_queue, out_queue, mapper, out_order) if order else (
-        in_queue, out_queue, mapper)
-    workers = []
-    for i in xrange(process_num):
-        worker = Thread(target=target, args=args)
-        worker.daemon = True
-        workers.append(worker)
-    for w in workers:
-        w.start()
-
     def xreader():
+        in_queue = Queue(buffer_size)
+        out_queue = Queue(buffer_size)
+        out_order = [0]
+        # start a read worker in a thread
+        target = order_read_worker if order else read_worker
+        t = Thread(target=target, args=(reader, in_queue))
+        t.daemon = True
+        t.start()
+        # start several handle_workers
+        target = order_handle_worker if order else handle_worker
+        args = (in_queue, out_queue, mapper, out_order) if order else (
+            in_queue, out_queue, mapper)
+        workers = []
+        for i in xrange(process_num):
+            worker = Thread(target=target, args=args)
+            worker.daemon = True
+            workers.append(worker)
+        for w in workers:
+            w.start()
+
         sample = out_queue.get()
         while not isinstance(sample, XmapEndSignal):
             yield sample

python/paddle/v2/reader/tests/decorator_test.py

@@ -132,15 +132,17 @@ class TestXmap(unittest.TestCase):
         for order in orders:
             for tNum in thread_nums:
                 for size in buffered_size:
-                    result = []
-                    for i in paddle.v2.reader.xmap_readers(mapper,
+                    reader = paddle.v2.reader.xmap_readers(mapper,
                                                            reader_creator_10(0),
-                                                           tNum, size, order)():
-                        result.append(i)
-                    if not order:
-                        result.sort()
-                    for idx, e in enumerate(result):
-                        self.assertEqual(e, mapper(idx))
+                                                           tNum, size, order)
+                    for n in xrange(3):
+                        result = []
+                        for i in reader():
+                            result.append(i)
+                        if not order:
+                            result.sort()
+                        for idx, e in enumerate(result):
+                            self.assertEqual(e, mapper(idx))
 
 
 if __name__ == '__main__':

python/paddle/v2/tests/test_parameters.py

@@ -20,14 +20,17 @@ import cStringIO
 import numpy
 
 
-def __rand_param_config__(name):
+def __rand_param_config__(name, psize=None):
     conf = ParameterConfig()
     conf.name = name
     size = 1
-    for i in xrange(2):
-        dim = random.randint(1, 1000)
-        conf.dims.append(dim)
-        size *= dim
+    if psize is None:
+        for i in xrange(2):
+            dim = random.randint(1, 1000)
+            conf.dims.append(dim)
+            size *= dim
+    else:
+        size = psize
     conf.size = size
     assert conf.IsInitialized()
     return conf
@@ -77,6 +80,50 @@ class TestParameters(unittest.TestCase):
         expected = numpy.array([[1, 1], [1, 2], [1, 1]], numpy.float32)
         assert numpy.logical_and.reduce(numpy.reshape(val == expected, 6))
 
+    def test_init_from_tar(self):
+        def get_param(names, size):
+            p = parameters.Parameters()
+            for k, v in zip(names, size):
+                p.__append_config__(__rand_param_config__(k, v))
+            for name in p.names():
+                param = p.get(name)
+                param[:] = numpy.random.uniform(
+                    -1.0, 1.0, size=p.get_shape(name))
+                p.set(name, param)
+            return p
+
+        def get_parames():
+            name1 = ['param_0', 'param_1']
+            size1 = [128, 256]
+            p1 = get_param(name1, size1)
+            file1 = cStringIO.StringIO()
+            p1.to_tar(file1)
+            file1.seek(0)
+
+            name2 = ['param_0', 'param_1', 'param_2']
+            size2 = [128, 256, 288]
+            p2 = get_param(name2, size2)
+            file2 = cStringIO.StringIO()
+            p2.to_tar(file2)
+            file2.seek(0)
+            return p1, file1, p2, file2
+
+        p1, file1, p2, file2 = get_parames()
+        p2.init_from_tar(file1)
+        for name in p1.names():
+            self.assertEqual(p1.get_shape(name), p2.get_shape(name))
+            v1 = p1.get(name)
+            v2 = p2.get(name)
+            self.assertTrue(numpy.isclose(v1, v2).all())
+
+        p1, file1, p2, file2 = get_parames()
+        p1.init_from_tar(file2)
+        for name in p1.names():
+            self.assertEqual(p1.get_shape(name), p2.get_shape(name))
+            v1 = p1.get(name)
+            v2 = p2.get(name)
+            self.assertTrue(numpy.isclose(v1, v2).all())
+
 
 if __name__ == '__main__':
     unittest.main()

python/setup.py.in

@@ -15,7 +15,8 @@ setup_requires=["requests",
                 "protobuf==3.1",
                 "recordio",
                 "matplotlib",
-                "rarfile"]
+                "rarfile",
+                "scipy>=0.19.0"]
 
 if '${CMAKE_SYSTEM_PROCESSOR}' not in ['arm', 'armv7-a', 'aarch64']:
     setup_requires+=["opencv-python"]