speed up prefetch

examples/cifar10-convnet.py

@@ -97,7 +97,7 @@ def get_data(train_or_test):
     ds = AugmentImageComponent(ds, augmentors)
     ds = BatchData(ds, 128, remainder=not isTrain)
     if isTrain:
-        ds = PrefetchData(ds, 10, 5)
+        ds = PrefetchDataZMQ(ds, 5)
     return ds
 
 def get_config():

tensorpack/dataflow/common.py

@@ -145,7 +145,8 @@ class FakeData(DataFlow):
 
     def get_data(self):
         for _ in range(self._size):
-            yield [self.rng.random_sample(k) for k in self.shapes]
+            yield [self.rng.random_sample(k).astype('float32') for k in self.shapes]
+            #yield [self.rng.random_sample(k) for k in self.shapes]
 
 class MapData(ProxyDataFlow):
     """ Apply map/filter a function on the datapoint"""

tensorpack/dataflow/format.py

@@ -11,7 +11,7 @@ from six.moves import range
 try:
     import h5py
 except ImportError:
-    logger.error("Error in 'import h5py'. HDF5Data won't be imported.")
+    logger.error("Error in 'import h5py'. HDF5Data won't be available.")
     __all__ = []
 else:
     __all__ = ['HDF5Data']

tensorpack/dataflow/prefetch.py

@@ -7,7 +7,6 @@ from threading import Thread
 from six.moves import range
 from six.moves.queue import Queue
 import uuid
-import zmq
 import os
 
 from .base import ProxyDataFlow
@@ -15,7 +14,14 @@ from ..utils.concurrency import ensure_proc_terminate
 from ..utils.serialize import *
 from ..utils import logger
 
-__all__ = ['PrefetchData', 'PrefetchDataZMQ']
+try:
+    import zmq
+except ImportError:
+    logger.error("Error in 'import zmq'. PrefetchDataZMQ won't be available.")
+    __all__ = ['PrefetchData']
+else:
+    __all__ = ['PrefetchData', 'PrefetchDataZMQ']
+
 
 class PrefetchProcess(multiprocessing.Process):
     def __init__(self, ds, queue):
@@ -69,64 +75,57 @@ class PrefetchData(ProxyDataFlow):
         logger.info("Prefetch process exited.")
 
 
-
 class PrefetchProcessZMQ(multiprocessing.Process):
-    def __init__(self, ds, conn_name, qsize=1):
+    def __init__(self, ds, conn_name):
+        """
+        :param ds: a `DataFlow` instance.
+        :param conn_name: the name of the IPC connection
+        """
         super(PrefetchProcessZMQ, self).__init__()
         self.ds = ds
-        self.qsize = qsize
         self.conn_name = conn_name
 
     def run(self):
+        self.ds.reset_state()
         self.context = zmq.Context()
         self.socket = self.context.socket(zmq.PUSH)
-        self.socket.set_hwm(self.qsize)
+        self.socket.set_hwm(1)
         self.socket.connect(self.conn_name)
-        self.id = os.getpid()
-        cnt = 0
         while True:
             for dp in self.ds.get_data():
-                self.socket.send(dumps(dp))
-                cnt += 1
-                print("Proc {} send {}".format(self.id, cnt))
+                self.socket.send(dumps(dp), copy=False)
 
 class PrefetchDataZMQ(ProxyDataFlow):
     """ Work the same as `PrefetchData`, but faster. """
-    def __init__(self, ds, nr_prefetch, nr_proc=1):
+    def __init__(self, ds, nr_proc=1):
+        """
+        :param ds: a `DataFlow` instance.
+        :param nr_proc: number of processes to use. When larger than 1, order
+            of datapoints will be random.
+        """
         super(PrefetchDataZMQ, self).__init__(ds)
-        self.ds = ds
         self._size = ds.size()
         self.nr_proc = nr_proc
 
         self.context = zmq.Context()
         self.socket = self.context.socket(zmq.PULL)
-        name = "ipc://whatever-" + str(uuid.uuid1())[:6]
-        self.socket.bind(name)
-
-        # TODO local queue again? probably don't affect training
-        self.queue = Queue(maxsize=nr_prefetch)
-        def enque():
-            while True:
-                self.queue.put(loads(self.socket.recv(copy=False)))
-        self.th = Thread(target=enque)
-        self.th.daemon = True
-        self.th.start()
-
-        self.procs = [PrefetchProcessZMQ(self.ds, name)
+        self.pipename = "ipc://dataflow-pipe-" + str(uuid.uuid1())[:6]
+        self.socket.set_hwm(5)  # a little bit faster than default, don't know why
+        self.socket.bind(self.pipename)
+
+        self.procs = [PrefetchProcessZMQ(self.ds, self.pipename)
                       for _ in range(self.nr_proc)]
         for x in self.procs:
             x.start()
 
     def get_data(self):
         for _ in range(self._size):
-            dp = self.queue.get()
+            dp = loads(self.socket.recv(copy=False))
             yield dp
-            #print(self.queue.qsize())
 
     def __del__(self):
         logger.info("Prefetch process exiting...")
-        self.queue.close()
+        self.context.destroy(0)
         for x in self.procs:
             x.terminate()
-        self.th.terminate()
         logger.info("Prefetch process exited.")

tensorpack/utils/serialize.py

@@ -3,17 +3,17 @@
 # File: serialize.py
 # Author: Yuxin Wu <ppwwyyxxc@gmail.com>
 
-import msgpack
-import msgpack_numpy
-msgpack_numpy.patch()
-#import dill
+#import msgpack
+#import msgpack_numpy
+#msgpack_numpy.patch()
+import dill
 
 __all__ = ['loads', 'dumps']
 
 def dumps(obj):
-    #return dill.dumps(obj)
-    return msgpack.dumps(obj, use_bin_type=True)
+    return dill.dumps(obj)
+    #return msgpack.dumps(obj, use_bin_type=True)
 
 def loads(buf):
-    #return dill.loads(buf)
-    return msgpack.loads(buf)
+    return dill.loads(buf)
+    #return msgpack.loads(buf)