more pickability in dataflow

tensorpack · Dec 28, 2019 · 4768b26 · 4768b26
1 parent 1648ddd
commit 4768b26
Show file tree

Hide file tree

Showing 2 changed files with 25 additions and 14 deletions.
diff --git a/examples/FasterRCNN/convert_d2/README.md b/examples/FasterRCNN/convert_d2/README.md
@@ -9,6 +9,7 @@ It produces a corresponding tensorpack configs, as well as a tensorpack-compatib
 
 It currently supports ResNet{50,101}-{C4,FPN}-{Faster,Mask,Cascade} R-CNN models in
 [detectron2 model zoo](https://github.com/facebookresearch/detectron2/blob/master/MODEL_ZOO.md).
+You may add new architectures in `../modeling` to support more models.
 
 ### Usage:
 
@@ -27,12 +28,17 @@ $ ./predict.py --evaluate out.json --load R50FPN-d2-converted.npz  --config DATA
 $ ./predict.py --output-pb out.pb --load R50FPN-d2-converted.npz  --config DATA.BASEDIR=~/data/coco 'MODE_MASK=True' 'MODE_FPN=True' 'BACKBONE.STRIDE_1X1=True' 'PREPROC.PIXEL_MEAN=[123.675,116.28,103.53]' 'PREPROC.PIXEL_STD=[1.0,1.0,1.0]'
 ```
 
-Note: this script does not support arbitrary detectron2 config.
-When run against an unsupported config, it may fail silently and produce
-erroneous models.
+Note:
 
-For models in detectron2 model zoo, there is a small incompatibility:
+1. This script does not support arbitrary detectron2 config.
+   When run against an unsupported config, it may fail silently and produce
+   erroneous models.
 
-* `POOLER_SAMPLING_RATIO=0` in RoIAlign: there is no equivalence in TensorFlow.
-	Our RoIAlign only implements `POOLER_SAMPLING_RATIO=2`.
-	The results are quite similar, and the final AP may be different by <0.5.
+2. The above steps produces a TensorFlow's pb file without any inference-time optimization (such as fusion).
+   Tensorpack is a training framework so it does not provide any such tools.
+	 It's up to the user to further optimize the final graph.
+
+3. There could be a small incompatibility for converted models.
+	 For the implementation of RoIAlign,  there is no equivalence of `POOLER_SAMPLING_RATIO=0` in tensorpack or TensorFlow.
+	 Our RoIAlign only implements `POOLER_SAMPLING_RATIO=2`.
+	 The results are quite similar, and the final AP may be different by <0.5.
diff --git a/tensorpack/dataflow/raw.py b/tensorpack/dataflow/raw.py
@@ -103,26 +103,31 @@ def __init__(self, gen):
         Args:
             gen: iterable, or a callable that returns an iterable
         """
-        if not callable(gen):
-            self._gen = lambda: gen
-        else:
-            self._gen = gen
+        self._gen = gen
 
     def __iter__(self):
-        yield from self._gen()
+        if not callable(self._gen):
+            yield from self._gen
+        else:
+            yield from self._gen()
 
 
 class DataFromIterable(DataFlow):
     """ Wrap an iterable of datapoints to a DataFlow"""
     def __init__(self, iterable):
         """
         Args:
-            iterable: an iterable object with length
+            iterable: an iterable object
         """
         self._itr = iterable
-        self._len = len(iterable)
+        try:
+            self._len = len(iterable)
+        except Exception:
+            self._len = None
 
     def __len__(self):
+        if self._len is None:
+            raise NotImplementedError
         return self._len
 
     def __iter__(self):