update docs and new model

.github/ISSUE_TEMPLATE.md

@@ -2,3 +2,6 @@ An issue has to be one of the following:
 - Unexpected Problems / Potential Bugs
 - Feature Requests
 - Questions on Using/Understanding Tensorpack
+
+To post an issue, please click "New Issue", choose your category, and read
+instructions there.

docs/tutorial/dataflow.md

@@ -13,6 +13,13 @@ A datapoint is a **list** of Python objects which are called the `components` of
 that yields datapoints (lists) of two components:
 a numpy array of shape (64, 28, 28), and an array of shape (64,).
 
+As you saw,
+DataFlow is __independent__ of TensorFlow since it produces any python objects
+(usually numpy arrays).
+To `import tensorpack.dataflow`, you don't even have to install TensorFlow.
+You can simply use DataFlow as a data processing pipeline and plug it into any other frameworks.
+
+
 ### Composition of DataFlow
 One good thing about having a standard interface is to be able to provide
 the greatest code reusability.
@@ -65,8 +72,3 @@ generator = df.get_data()
 for dp in generator:
 	# dp is now a list. do whatever
 ```
-
-DataFlow is __independent__ of both tensorpack and TensorFlow.
-To `import tensorpack.dataflow`, you don't even have to install TensorFlow.
-You can simply use it as a data processing pipeline and plug it into any other frameworks.
-

examples/FasterRCNN/README.md

@@ -72,7 +72,7 @@ prediction will need to be run with the corresponding training configs.
 ## Results
 
 These models are trained with different configurations on trainval35k and evaluated on minival using mAP@IoU=0.50:0.95.
-Performance in [Detectron](https://github.com/facebookresearch/Detectron/) can be roughly reproduced, some are better but some are worse.
+Performance in [Detectron](https://github.com/facebookresearch/Detectron/) can be roughly reproduced.
 MaskRCNN results contain both box and mask mAP.
 
  | Backbone | mAP<br/>(box;mask)                                                                                                            | Detectron mAP <sup>[1](#ft1)</sup><br/> (box;mask) | Time on 8 V100s | Configurations <br/> (click to expand)                                                                                                                                                                                                                 |
@@ -87,12 +87,14 @@ MaskRCNN results contain both box and mask mAP.
  | R50-FPN  | 40.0;36.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R50FPN-MaskRCNN-StandardGN.npz)                         | 40.3;35.7                                          | 40h             | <details><summary>+GN</summary>`MODE_FPN=True`<br/>`FPN.NORM=GN BACKBONE.NORM=GN`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head`<br/>`FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head`                                                               |
  | R101-C4  | 41.4;35.2 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101C4-MaskRCNN-Standard.npz)                           |                                                    | 60h             | <details><summary>standard</summary>`BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]` </details>                                                                                                                                                                  |
  | R101-FPN | 40.4;36.6 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-Standard.npz)                          | 40.9;36.4                                          | 38h             | <details><summary>standard</summary>`MODE_FPN=True`<br/>`BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]` </details>                                                                                                                                              |
- | R101-FPN | 41.1;36.6 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-BetterParams.npz) | 40.9;36.4                                          | 38h             | <details><summary>better params</summary>`MODE_FPN=True`<br/>`BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]`<br/>`TEST.RESULT_SCORE_THRESH=1e-4`<br/>`FRCNN.BBOX_REG_WEIGHTS=[20,20,10,10]` </details>    |
+ | R101-FPN | 46.5;40.1 [:arrow_down:](http://models.tensorpack.com/FasterRCNN/COCO-R101FPN-MaskRCNN-BetterParams.npz) <sup>[3](#ft3)</sup> |                                                    | 73h             | <details><summary>+++</summary>`MODE_FPN=True FPN.CASCADE=True`<br/>`BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]`<br/>`TEST.RESULT_SCORE_THRESH=1e-4`<br/>`PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800]`<br/>`TRAIN.LR_SCHEDULE=[420000,500000,540000]` </details> |
 
  <a id="ft1">1</a>: Here we comapre models that have identical training & inference cost between the two implementation. However their numbers are different due to many small implementation details.
 
  <a id="ft2">2</a>: Numbers taken from [Group Normalization](https://arxiv.org/abs/1803.08494)
 
+ <a id="ft3">3</a>: Our mAP is __10+ point__ better than the official model in [matterport/Mask_RCNN](https://github.com/matterport/Mask_RCNN/releases) with the same R101-FPN backbone.
+
 ## Notes
 
 [NOTES.md](NOTES.md) has some notes about implementation details & speed.

examples/FasterRCNN/train.py

@@ -441,7 +441,7 @@ class EvalCallback(Callback):
             logger.info("[EvalCallback] Will evaluate every {} epochs".format(interval))
 
     def _eval(self):
-        if cfg.TRAINER == 'replicated' or cfg.TRAIN.NUM_GPUS == 1:
+        if cfg.TRAINER == 'replicated':
             with ThreadPoolExecutor(max_workers=self.num_predictor, thread_name_prefix='EvalWorker') as executor, \
                     tqdm.tqdm(total=sum([df.size() for df in self.dataflows])) as pbar:
                 futures = []

tensorpack/callbacks/trigger.py

@@ -86,11 +86,13 @@ class PeriodicRunHooks(ProxyCallback):
 
 class EnableCallbackIf(ProxyCallback):
     """
-    Enable the ``{before,after}_epoch``, ``{before,after}_run``,
+    Disable the ``{before,after}_epoch``, ``{before,after}_run``,
     ``trigger_{epoch,step}``
-    methods of a callback, only when some condition satisfies.
+    methods of a callback, unless some condition satisfies.
     The other methods are unaffected.
 
+    A more accurate name for this callback should be "DisableCallbackUnless", but that's too ugly.
+
     Note:
         If you use ``{before,after}_run``,
         ``pred`` will be evaluated only in ``before_run``.
@@ -101,6 +103,7 @@ class EnableCallbackIf(ProxyCallback):
         Args:
             callback (Callback):
             pred (self -> bool): a callable predicate. Has to be a pure function.
+                The callback is disabled unless this predicate returns True.
         """
         self._pred = pred
         super(EnableCallbackIf, self).__init__(callback)

tensorpack/utils/compatible_serialize.py

-#!/usr/bin/env python
-
 import os
 from .serialize import loads_msgpack, loads_pyarrow, dumps_msgpack, dumps_pyarrow
 

tensorpack/utils/serialize.py

@@ -2,8 +2,6 @@
 # File: serialize.py
 
 import os
-import pyarrow as pa
-
 from .develop import create_dummy_func
 
 __all__ = ['loads', 'dumps']
@@ -46,6 +44,16 @@ def loads_pyarrow(buf):
     return pa.deserialize(buf)
 
 
+try:
+    # import pyarrow has a lot of side effect: https://github.com/apache/arrow/pull/2329
+    # So we need an option to disable it.
+    if os.environ.get('TENSORPACK_SERIALIZE', 'pyarrow') == 'pyarrow':
+        import pyarrow as pa
+except ImportError:
+    pa = None
+    dumps_pyarrow = create_dummy_func('dumps_pyarrow', ['pyarrow'])  # noqa
+    loads_pyarrow = create_dummy_func('loads_pyarrow', ['pyarrow'])  # noqa
+
 try:
     import msgpack
     import msgpack_numpy