Use a better timer when available.

README.md

@@ -62,7 +62,7 @@ demonstrating its flexibility for actual research.
 
 Dependencies:
 
-+ Python 2.7 or 3
++ Python 2.7 or 3.3+
 + Python bindings for OpenCV (Optional, but required by a lot of features)
 + TensorFlow >= 1.3.0 (Optional if you only want to use `tensorpack.dataflow` alone as a data processing library)
 ```

examples/FasterRCNN/README.md

@@ -73,14 +73,15 @@ MaskRCNN results contain both box and mask mAP.
  | R50-C4   | 37.8/33.1          | 37.8/32.8                      | 49h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True` </details>                                                                                                                                  |
  | R50-FPN  | 38.2/34.9          | 38.6/34.5<sup>[1](#ft1)</sup>  | 32h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True MODE_FPN=True` </details>                                                                                                                    |
  | R50-FPN  | 38.5/34.8          | 38.6/34.2<sup>[2](#ft2)</sup>  | 34h on 8 V100s | <details><summary>standard+ConvHead</summary>`MODE_MASK=True MODE_FPN=True`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_head` </details>                                                          |
- | R50-FPN  | 39.5/35.2          | 39.5/34.4<sup>[2](#ft2)</sup>  | 34h on 8 V100s | <details><summary>standard+ConvGNHead</summary>`MODE_MASK=True MODE_FPN=True`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head` </details>                                                          |
+ | R50-FPN  | 39.5/35.2          | 39.5/34.4<sup>[2](#ft2)</sup>  | 34h on 8 V100s | <details><summary>standard+ConvGNHead</summary>`MODE_MASK=True MODE_FPN=True`<br/>`FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head` </details>                                                     |
  | R101-C4  | 40.8/35.1          |                                | 63h on 8 V100s | <details><summary>standard</summary>`MODE_MASK=True `<br/>`BACKBONE.RESNET_NUM_BLOCK=[3,4,23,3]` </details>                                                                                      |
  
  <a id="ft1">1</a>: Slightly different configurations.
 
  <a id="ft2">2</a>: Numbers taken from [Group Normalization](https://arxiv.org/abs/1803.08494)
  
- Performance in [Detectron](https://github.com/facebookresearch/Detectron/) can be reproduced.
+ Performance in [Detectron](https://github.com/facebookresearch/Detectron/) can
+ be roughly reproduced, some are better but some are worse, probably due to many tiny implementation details.
  Note that most of these numbers are better than what's in the paper. 
 
 ## Notes

examples/FasterRCNN/model_fpn.py

@@ -218,5 +218,5 @@ def generate_fpn_proposals(
             cfg.RPN.TRAIN_POST_NMS_TOPK if ctx.is_training else cfg.RPN.TEST_POST_NMS_TOPK)
 
     tf.sigmoid(proposal_scores, name='probs')  # for visualization
-    return tf.identity(proposal_boxes, name='boxes'), \
-        tf.identity(proposal_scores, name='scores')
+    return tf.stop_gradient(proposal_boxes, name='boxes'), \
+        tf.stop_gradient(proposal_scores, name='scores')

tensorpack/callbacks/group.py

@@ -4,14 +4,18 @@
 
 import tensorflow as tf
 from contextlib import contextmanager
-import time
+from time import time as timer
 import traceback
+import six
 
 from .base import Callback
 from .hooks import CallbackToHook
 from ..utils import logger
 from ..utils.utils import humanize_time_delta
 
+if six.PY3:
+    from time import perf_counter as timer  # noqa
+
 __all__ = ['Callbacks']
 
 
@@ -26,9 +30,9 @@ class CallbackTimeLogger(object):
 
     @contextmanager
     def timed_callback(self, name):
-        s = time.time()
+        s = timer()
         yield
-        self.add(name, time.time() - s)
+        self.add(name, timer() - s)
 
     def log(self):
 

tensorpack/train/interface.py

@@ -52,7 +52,7 @@ def launch_train_with_config(config, trainer):
     present a simple training interface. It basically does the following
     3 things (and you can easily do them by yourself if you need more control):
 
-    1. Setup the input with automatic prefetching,
+    1. Setup the input with automatic prefetching heuristics,
        from `config.data` or `config.dataflow`.
     2. Call `trainer.setup_graph` with the input as well as `config.model`.
     3. Call `trainer.train` with rest of the attributes of config.

tensorpack/utils/timer.py

@@ -3,14 +3,18 @@
 
 
 from contextlib import contextmanager
-import time
 from collections import defaultdict
 import six
 import atexit
+from time import time as timer
 
 from .stats import StatCounter
 from . import logger
 
+if six.PY3:
+    from time import perf_counter as timer  # noqa
+
+
 __all__ = ['total_timer', 'timed_operation',
            'print_total_timer', 'IterSpeedCounter']
 
@@ -38,10 +42,10 @@ def timed_operation(msg, log_start=False):
     """
     if log_start:
         logger.info('Start {} ...'.format(msg))
-    start = time.time()
+    start = timer()
     yield
     logger.info('{} finished, time:{:.4f}sec.'.format(
-        msg, time.time() - start))
+        msg, timer() - start))
 
 
 _TOTAL_TIMER_DATA = defaultdict(StatCounter)
@@ -50,9 +54,9 @@ _TOTAL_TIMER_DATA = defaultdict(StatCounter)
 @contextmanager
 def total_timer(msg):
     """ A context which add the time spent inside to TotalTimer. """
-    start = time.time()
+    start = timer()
     yield
-    t = time.time() - start
+    t = timer() - start
     _TOTAL_TIMER_DATA[msg].feed(t)
 
 
@@ -96,7 +100,7 @@ class IterSpeedCounter(object):
         self.name = name if name else 'IterSpeed'
 
     def reset(self):
-        self.start = time.time()
+        self.start = timer()
 
     def __call__(self):
         if self.cnt == 0:
@@ -104,6 +108,6 @@ class IterSpeedCounter(object):
         self.cnt += 1
         if self.cnt % self.print_every != 0:
             return
-        t = time.time() - self.start
+        t = timer() - self.start
         logger.info("{}: {:.2f} sec, {} times, {:.3g} sec/time".format(
             self.name, t, self.cnt, t / self.cnt))