update docs

.github/ISSUE_TEMPLATE.md

-Potential Bugs/Feature Requests/Usage Questions Only:
+An issue has to be one of the following:
+1. Unexpected Problems / Potential Bugs
+2. Feature Requests
+3. Usage Questions
 
 Any unexpected problems: __PLEASE ALWAYS INCLUDE__:
 1. What you did:
@@ -18,17 +21,17 @@ Any unexpected problems: __PLEASE ALWAYS INCLUDE__:
 5. About efficiency, PLEASE first read http://tensorpack.readthedocs.io/en/latest/tutorial/performance-tuning.html
 
 Feature Requests:
-+ Improve an existing feature, or add a new feature.
 + You can implement a lot of features by extending tensorpack
 	(See http://tensorpack.readthedocs.io/en/latest/tutorial/index.html#extend-tensorpack).
 	It does not have to be added to tensorpack unless you have a good reason.
-+ We don't take feature requests for examples.
++ We don't take feature requests for implementing new techniques.
+  If you don't know how, ask it as a usage question.
 
 Usage Questions:
 
 + Read the [tutorials](http://tensorpack.readthedocs.io/en/latest/tutorial/index.html#user-tutorials) first.
 + We answer "HOW to do X in tensorpack" for a specific well-defined X.
   We don't answer general machine learning questions,
-  such as "how to improve my model" or "I don't understand the paper".
+  such as "what networks to use" or "I don't understand the paper".
 
 You can also use gitter (https://gitter.im/tensorpack/users) for more casual discussions.

docs/tutorial/efficient-dataflow.md

@@ -230,8 +230,9 @@ Let me summarize what this DataFlow does:
 3. The main process takes data from the pipe, makes batches.
 
 The two DataFlow mentioned in this tutorial (both random read and sequential read) can run at a speed of 1k ~ 2.5k images per second if you have good CPUs, RAM, disks.
+With fewer augmentations, it can reach 5k images/s.
 As a reference, tensorpack can train ResNet-18 at 1.2k images/s on 4 old TitanX.
-A DGX-1 (8 P100s) can train ResNet-50 at 1.7k images/s according to the [official benchmark](https://www.tensorflow.org/performance/benchmarks).
+8 P100s can train ResNet-50 at 1.7k images/s according to the [official benchmark](https://www.tensorflow.org/performance/benchmarks).
 So DataFlow will not be a serious bottleneck if configured properly.
 
 ## Distributed DataFlow

examples/FasterRCNN/README.md

@@ -53,10 +53,10 @@ MaskRCNN results contain both bbox and segm mAP.
 
 |Backbone|`FASTRCNN_BATCH`|resolution |schedule|mAP (bbox/segm)|Time         |
 |   -    |    -           |    -      |   -    |   -           |   -         |
-|R-50    |64              |(600, 1024)|280k    |33.0           |22h on 8 P100|
+|R-50    |64              |(600, 1024)|280k    |33.1           |18h on 8 V100|
 |R-50    |512             |(800, 1333)|280k    |35.6           |55h on 8 P100|
 |R-50    |512             |(800, 1333)|360k    |36.7           |49h on 8 V100|
-|R-50    |256             |(800, 1333)|280k    |36.9/32.3      |39h on 8 P100|
+|R-50    |256             |(800, 1333)|280k    |36.8/32.1      |39h on 8 P100|
 |R-50    |512							|(800, 1333)|360k    |37.7/33.0      |72h on 8 P100|
 |R-101   |512             |(800, 1333)|280k    |40.1/34.4      |70h on 8 P100|
 

tensorpack/tfutils/tower.py

@@ -84,6 +84,7 @@ class TowerContext(object):
     def get_collection_in_tower(self, key):
         """
         Get items from this collection that are added in the current tower.
+        These items may or may not start with the same prefix as the tower.
         """
         return self._collection_guard.get_collection_in_tower(key)
 
@@ -336,6 +337,7 @@ class TowerTensorHandle(object):
     def get_collection(self, name):
         """
         Get items from a collection that are added in this tower.
+        These items may or may not start with the same prefix as the tower.
 
         Args:
             name (str): the name of the collection

tensorpack/train/tower.py

@@ -78,11 +78,25 @@ class TowerTrainer(Trainer):
         Returns a callable predictor built under ``TowerContext(is_training=False)``.
 
         Args:
-            input_names (list), output_names(list): list of names
+            input_names (list): list of input names, matching the inputs declared for the trainer.
+            output_names(list): list of tensor names without the tower prefix.
             device (int): build the predictor on device '/gpu:{device}' or use -1 for '/cpu:0'.
 
         Returns:
             an :class:`OnlinePredictor`.
+
+        Examples:
+
+        .. code-block:: none
+
+            # in the graph:
+            interesting_tensor = tf.identity(x, name='fun')
+            # in _setup_graph callback method:
+            self._predictor = self.trainer.get_predictor(['input1'], ['fun'])
+            # After session is initialized (see Tutorials - Write a Callback), can use it by:
+            outputs = self._predictor(inputs)
+
+        The CycleGAN example and DQN example have more concrete use of this method.
         """
         assert self.tower_func is not None, "Must set tower_func on the trainer to use get_predictor()!"
         tower_name = 'tower-pred-{}'.format(device) if device >= 0 else 'tower-pred-cpu'