docs update; use abstract sockets for a3c

docs/tutorial/efficient-dataflow.md

@@ -251,11 +251,11 @@ send_dataflow_zmq(df, 'tcp://1.2.3.4:8877')
 ```python
 # Training Machine, process 1-10:
 df = MyLargeData()
-send_dataflow_zmq(df, 'ipc:///tmp/ipc-socket')
+send_dataflow_zmq(df, 'ipc://@my-socket')
 ```
 ```python
 # Training Machine, training process
-df = RemoteDataZMQ('ipc:///tmp/ipc-socket', 'tcp://0.0.0.0:8877')
+df = RemoteDataZMQ('ipc://@my-socket', 'tcp://0.0.0.0:8877')
 TestDataSpeed(df).start()
 ```
 

examples/A3C-Gym/simulator.py

@@ -174,7 +174,7 @@ if __name__ == '__main__':
             client.memory = []
             client.state = 0
 
-    name = 'ipc://whatever'
+    name = 'ipc://@whatever'
     procs = [NaiveSimulator(k, name) for k in range(10)]
     [k.start() for k in procs]
 

examples/A3C-Gym/train-atari.py

@@ -233,9 +233,8 @@ def train():
 
     # setup simulator processes
     name_base = str(uuid.uuid1())[:6]
-    PIPE_DIR = os.environ.get('TENSORPACK_PIPEDIR', '.').rstrip('/')
-    namec2s = 'ipc://{}/sim-c2s-{}'.format(PIPE_DIR, name_base)
-    names2c = 'ipc://{}/sim-s2c-{}'.format(PIPE_DIR, name_base)
+    namec2s = 'ipc://@sim-c2s-{}'.format(name_base)
+    names2c = 'ipc://@sim-s2c-{}'.format(name_base)
     procs = [MySimulatorWorker(k, namec2s, names2c) for k in range(SIMULATOR_PROC)]
     ensure_proc_terminate(procs)
     start_proc_mask_signal(procs)

examples/FasterRCNN/README.md

@@ -63,7 +63,8 @@ MaskRCNN results contain both bbox and segm mAP.
 The two 360k models have identical configurations with
 `R50-C4-2x` configuration in
 [Detectron Model Zoo](https://github.com/facebookresearch/Detectron/blob/master/MODEL_ZOO.md#end-to-end-faster--mask-r-cnn-baselines).
-They get the __same performance__ with the official models, and are about 14% slower than the official implementation, due to the lack of specialized ops.
+They get the __same performance__ with the official models, and are about 14% slower than the official implementation,
+probably due to the lack of specialized ops in TensorFlow.
 
 ## Notes
 

tensorpack/dataflow/remote.py

@@ -168,7 +168,7 @@ if __name__ == '__main__':
     args = parser.parse_args()
 
     # tcp addr like "tcp://127.0.0.1:8877"
-    # ipc addr like "ipc:///tmp/ipc-test"
+    # ipc addr like "ipc://@ipc-test"
     if args.task == 'send':
         # use random=True to make it slow and cpu-consuming
         ds = FakeData([(128, 244, 244, 3)], 1000, random=True)

tensorpack/tfutils/summary.py

@@ -216,6 +216,9 @@ def add_moving_summary(*args, **kwargs):
     # allow ctx to be none
     if ctx is not None and not ctx.is_main_training_tower:
         return []
+    if tf.get_variable_scope().reuse is True:
+        logger.warn("add_moving_summary() called under reuse=True scope, ignored.")
+        return []
 
     if not isinstance(args[0], list):
         v = args