fix rgb/bgr in ResNet script

examples/ResNet/README.md

@@ -8,14 +8,15 @@ Models can be [downloaded here](https://goo.gl/6XjK9V).
 
 | Model              | Top 5 Error | Top 1 Error |
 |:-------------------|-------------|------------:|
-| ResNet 18          |      10.67% |      29.50% |
-| ResNet 34          |      8.66%  |      26.45% |
-| ResNet 50          |      7.13%  |      24.12% |
-| ResNet 101         |      6.54%  |      22.89% |
+| ResNet18           |     10.55%  |      29.73% |
+| ResNet34           |     8.51%   |      26.50% |
+| ResNet50           |     7.24%   |      23.91% |
+| ResNet50-SE        |     6.42%   |      22.94% |
+| ResNet101          |     6.26%   |      22.53% |
 
 To train, just run:
 ```bash
-./imagenet-resnet.py --data /path/to/original/ILSVRC --gpu 0,1,2,3 -d 18
+./imagenet-resnet.py --data /path/to/original/ILSVRC --gpu 0,1,2,3 -d 50
 ```
 You should be able to see good GPU utilization (around 95%), if your data is fast enough.
 See the [tutorial](http://tensorpack.readthedocs.io/en/latest/tutorial/efficient-dataflow.html) on how to speed up your data.

examples/ResNet/imagenet-resnet.py

@@ -45,9 +45,7 @@ class Model(ModelDesc):
 
     def _build_graph(self, inputs):
         image, label = inputs
-        # It should actually use bgr=True here, but for compatibility with
-        # pretrained models, we keep the wrong version.
-        image = image_preprocess(image, bgr=False)
+        image = image_preprocess(image, bgr=True)
 
         if self.data_format == 'NCHW':
             image = tf.transpose(image, [0, 3, 1, 2])

tensorpack/libinfo.py

@@ -25,4 +25,7 @@ os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'  # issue#9339
 os.environ['TF_AUTOTUNE_THRESHOLD'] = '3'   # use more warm-up
 os.environ['TF_AVGPOOL_USE_CUDNN'] = '1'   # issue#8566
 
+import tensorflow as tf  # noqa
+assert int(tf.__version__.split('.')[0]) >= 1, "TF>=1.0 is required!"
+
 __version__ = '0.3.0'