small update in docs

examples/FasterRCNN/train.py

@@ -103,7 +103,7 @@ class Model(ModelDesc):
             boxes_on_featuremap = proposal_boxes * (1.0 / config.ANCHOR_STRIDE)
 
         roi_resized = roi_align(featuremap, boxes_on_featuremap, 14)
-        feature_fastrcnn = resnet_conv5(roi_resized, config.RESNET_NUM_BLOCK[-1])    # nxc
+        feature_fastrcnn = resnet_conv5(roi_resized, config.RESNET_NUM_BLOCK[-1])    # nxcx7x7
         fastrcnn_label_logits, fastrcnn_box_logits = fastrcnn_head('fastrcnn', feature_fastrcnn, config.NUM_CLASS)
 
         if is_training:
@@ -133,8 +133,7 @@ class Model(ModelDesc):
                 fastrcnn_label_loss, fastrcnn_box_loss,
                 wd_cost], 'total_cost')
 
-            for k in self.cost, wd_cost:
-                add_moving_summary(k)
+            add_moving_summary(self.cost, wd_cost)
         else:
             label_probs = tf.nn.softmax(fastrcnn_label_logits, name='fastrcnn_all_probs')  # #proposal x #Class
             anchors = tf.tile(tf.expand_dims(proposal_boxes, 1), [1, config.NUM_CLASS - 1, 1])   # #proposal x #Cat x 4

examples/README.md

 
 # tensorpack examples
 
-Training examples with __reproducible__ and meaningful performance.
+Training examples with __reproducible__ performance.
+
+__Reproducible performance is important__. Usually deep learning code is easy to write,
+but hard to know the correctness -- wrong code will usually still converge.
+Without a setting and performance comparable to someone else, you don't know if an implementation is correct or not.
 
 ## Getting Started:
 + [An illustrative mnist example with explanation of the framework](mnist-convnet.py)
@@ -13,7 +17,7 @@ Training examples with __reproducible__ and meaningful performance.
 | Name | Performance |
 | ---  | --- |
 |	Train [ResNet](ResNet) and [ShuffleNet](ShuffleNet) on ImageNet		| reproduce paper	|
-|	[Train ResNet50-Faster-RCNN on COCO](FasterRCNN)				|	reproduce paper		|
+|	[Train Faster-RCNN on COCO](FasterRCNN)				|	reproduce paper		|
 | [DoReFa-Net: training binary / low-bitwidth CNN on ImageNet](DoReFa-Net) | reproduce paper |
 | [Generative Adversarial Network(GAN) variants](GAN), including DCGAN, InfoGAN, <br/> Conditional GAN, WGAN, BEGAN, DiscoGAN, Image to Image, CycleGAN | visually reproduce |
 | [Inception-BN and InceptionV3](Inception) | reproduce reference code |
@@ -44,7 +48,3 @@ Training examples with __reproducible__ and meaningful performance.
 Example needs to satisfy one of the following:
 + Reproduce performance of a published or well-known paper.
 + Illustrate a new way of using the library that is currently not covered.
-
-__Performance is important__. Usually deep learning code is easy to write,
-but hard to know the correctness -- thanks to SGD things will usually still converge when you've made mistakes.
-Without a setting and performance comparable to someone else, you don't know if an implementation is correct or not.