update incv3

README.md

@@ -8,6 +8,7 @@ See some interesting [examples](examples) to learn about the framework:
 + [DoReFa-Net: training binary / low bitwidth CNN](examples/DoReFa-Net)
 + [Double-DQN for playing Atari games](examples/Atari2600)
 + [ResNet for Cifar10 classification](examples/ResNet)
++ [IncpetionV3 on ImageNet](examples/Inception/inceptionv3.py)
 + [char-rnn language model](examples/char-rnn)
 
 ## Features:

examples/Inception/inceptionv3.py

@@ -21,7 +21,7 @@ See "Rethinking the Inception Architecture for Computer Vision", arxiv:1512.0056
 
 This config follows the official inceptionv3 setup (https://github.com/tensorflow/models/tree/master/inception/inception)
 with much much fewer lines of code.
-It reaches 73.5% single-crop validation accuracy, same as the official code,
+It reaches 74.5% single-crop validation accuracy, slightly better than the official code,
 and has the same running speed as well.
 """
 

examples/README.md

@@ -9,7 +9,7 @@ Only allow examples with reproducible and meaningful performancce.
 + [Reproduce some reinforcement learning papers](Atari2600)
 + [char-rnn for fun](char-rnn)
 + [DisturbLabel, because I don't believe the paper](DisturbLabel)
-+ [DoReFa-Net, binary / low-bitwidth CNN on ImageNet](DoReFa-Net)
++ [DoReFa-Net: binary / low-bitwidth CNN on ImageNet](DoReFa-Net)
 + [GoogleNet-InceptionV1 with 71% accuracy](Inception/inception-bn.py)
-+ [GoogleNet-InceptionV3 with 73.5% accuracy (same as the official code)](Inception/inceptionv3.py)
-+ [ResNet for Cifar10 with similar accuracy, and for SVHN with state-of-the-art accuracy](ResNet)
++ [GoogleNet-InceptionV3 with 74.5% accuracy (similar to the official code)](Inception/inceptionv3.py)
++ [ResNet for Cifar10 and SVHN](ResNet)

tensorpack/tfutils/symbolic_functions.py

@@ -77,3 +77,8 @@ def rms(x, name=None):
         with tf.name_scope(None):   # name already contains the scope
             return tf.sqrt(tf.reduce_mean(tf.square(x)), name=name)
     return tf.sqrt(tf.reduce_mean(tf.square(x)), name=name)
+
+def get_scalar_var(name, init_value):
+    return tf.get_variable(name, shape=[],
+            initializer=tf.constant_initializer(init_value),
+            trainable=False)