docs change

docs/tutorial/inference.md

@@ -15,6 +15,8 @@ There are two ways to do inference during training.
 	This will further support prefetch & data-parallel inference.
 	More details to come.
 
+In both methods, your tower function will be called again, with `TowerContext.is_training==False`.
+You can build a different graph using this predicate.
 
 ## Inference After Training
 

examples/FasterRCNN/eval.py

@@ -91,7 +91,7 @@ def detect_one_image(img, model_func):
     return results
 
 
-def eval_on_dataflow(df, detect_func):
+def eval_coco(df, detect_func):
     """
     Args:
         df: a DataFlow which produces (image, image_id)

examples/FasterRCNN/train.py

@@ -39,7 +39,7 @@ from viz import (
     draw_predictions, draw_final_outputs)
 from common import print_config
 from eval import (
-    eval_on_dataflow, detect_one_image, print_evaluation_scores, DetectionResult)
+    eval_coco, detect_one_image, print_evaluation_scores, DetectionResult)
 import config
 
 
@@ -280,7 +280,7 @@ def visualize(model_path, nr_visualize=50, output_dir='output'):
 
 def offline_evaluate(pred_func, output_file):
     df = get_eval_dataflow()
-    all_results = eval_on_dataflow(
+    all_results = eval_coco(
         df, lambda img: detect_one_image(img, pred_func))
     with open(output_file, 'w') as f:
         json.dump(all_results, f)
@@ -309,7 +309,7 @@ class EvalCallback(Callback):
         self.epochs_to_eval.add(self.trainer.max_epoch)
 
     def _eval(self):
-        all_results = eval_on_dataflow(self.df, lambda img: detect_one_image(img, self.pred))
+        all_results = eval_coco(self.df, lambda img: detect_one_image(img, self.pred))
         output_file = os.path.join(
             logger.get_logger_dir(), 'outputs{}.json'.format(self.global_step))
         with open(output_file, 'w') as f:

examples/README.md

@@ -3,12 +3,14 @@
 
 Training examples with __reproducible performance__.
 
+__"Reproduce" should always means reproduce performance__.
 Reproducing a method is usually easy, but you don't know whether you've made mistakes, because wrong code will often appear to work.
-Reproducible performance results are what really matters.
+Reproducing __performance__ results is what really matters, and is something that's hardly seen on github.
 See [Unawareness of Deep Learning Mistakes](https://medium.com/@ppwwyyxx/unawareness-of-deep-learning-mistakes-d5b5774da0ba).
 
 
 ## Getting Started:
+These examples don't have meaningful performance numbers. They are supposed to be just demos.
 + [An illustrative mnist example with explanation of the framework](mnist-convnet.py)
 + The same mnist example using [tf-slim](mnist-tfslim.py), and [with weights visualizations](mnist-visualizations.py)
 + A tiny [Cifar ConvNet](cifar-convnet.py) and [SVHN ConvNet](svhn-digit-convnet.py)

tensorpack/input_source/input_source.py

@@ -223,7 +223,7 @@ class QueueInput(FeedfreeInput):
         with self.cached_name_scope():
             # in TF there is no API to get queue capacity, so we can only summary the size
             size = tf.cast(self.queue.size(), tf.float32, name='queue_size')
-        size_ema_op = add_moving_summary(size, collection=None)[0].op
+        size_ema_op = add_moving_summary(size, collection=None, decay=0.5)[0].op
         return RunOp(
             lambda: size_ema_op,
             run_before=False,