Using slim to perform batch normalization

Run python mnist_bn.py --phase=train to train. Run python mnist_bn.py --phase=test to test.

It should achieve an accuracy of ~99.3% or higher on test set.

The keys to use batch normalization in slim are:

1. Set proper decay rate for BN layer. Because a BN layer uses EMA (exponential moving average) to approximate the population mean/variance, it takes sometime to warm up, i.e. to get the EMA close to real population mean/variance. The default decay rate is 0.999, which is kind of high for our little cute MNIST dataset and needs ~1000 steps to get a good estimation. In my code, decay is set to 0.95, then it learns the population statistics very quickly. However, a large value of decay does have it own advantage: it gathers information from more mini-batches thus is more stable.

2. Use slim.learning.create_train_op to create train op instead of tf.train.GradientDescentOptimizer(0.1).minimize(loss) or something else!.

I've added accuracy, cross_entropy and batch normalization paramters into summary. Use tensorboard --logdir=/log to explore the learning curve and parameter distributions!