fasterrcnn notes

examples/FasterRCNN/README.md

 # Faster-RCNN / Mask-RCNN on COCO
 This example aims to provide a minimal (1.3k lines) multi-GPU implementation of
-Faster-RCNN / Mask-RCNN (without FPN) on COCO.
+Faster-RCNN & Mask-RCNN (with ResNet backbones) on COCO.
 
 ## Dependencies
 + Python 3; TensorFlow >= 1.4.0
@@ -53,14 +53,14 @@ MaskRCNN results contain both bbox and segm mAP.
 
 |Backbone | `FASTRCNN_BATCH` | resolution | mAP (bbox/segm) | Time |
 | - | - | - | - | - |
-| Res50 | 64 | (600, 1024) | 33.0 | 22h on 8 P100 |
-| Res50 | 256 | (600, 1024) | 34.4 | 49h on 8 M40 |
-| Res50 | 512 | (800, 1333) | 35.6 | 55h on 8 P100|
-| Res50 | 256 | (800, 1333) | 36.9/32.3 | 39h on 8 P100|
-| Res101 | 512 | (800, 1333) | 40.1/34.4 | 70h on 8 P100|
+| R50 | 64 | (600, 1024) | 33.0 | 22h on 8 P100 |
+| R50 | 256 | (600, 1024) | 34.4 | 49h on 8 M40 |
+| R50 | 512 | (800, 1333) | 35.6 | 55h on 8 P100|
+| R50 | 256 | (800, 1333) | 36.9/32.3 | 39h on 8 P100|
+| R101 | 512 | (800, 1333) | 40.1/34.4 | 70h on 8 P100|
 
-Note that these models are trained with a larger ROI batch size than the paper,
-and get about 1mAP better performance.
+Note that these models are trained with different ROI batch size and LR schedule.
+The performance is slightly better than the paper.
 
 ## Notes
 

tensorpack/dataflow/parallel.py

@@ -160,7 +160,6 @@ class MultiProcessPrefetchData(ProxyDataFlow):
             self._size = -1
         self.nr_proc = nr_proc
         self.nr_prefetch = nr_prefetch
-        self._guard = DataFlowReentrantGuard()
 
         if nr_proc > 1:
             logger.info("[MultiProcessPrefetchData] Will fork a dataflow more than one times. "
@@ -173,12 +172,11 @@ class MultiProcessPrefetchData(ProxyDataFlow):
         start_proc_mask_signal(self.procs)
 
     def get_data(self):
-        with self._guard:
-            for k in itertools.count():
-                if self._size > 0 and k >= self._size:
-                    break
-                dp = self.queue.get()
-                yield dp
+        for k in itertools.count():
+            if self._size > 0 and k >= self._size:
+                break
+            dp = self.queue.get()
+            yield dp
 
     def reset_state(self):
         # do nothing. all ds are reset once and only once in spawned processes