Fix batch normalization during evaluation

This commit solves the bug observed in all previous versions of this
code in which validation loss/accuracy are approximately/exactly
constant for every epoch and all validation predictions are of the
same class.

The problem was due to incorrect implementation of batch normalization
in two ways. First, tensorflow does not automatically collect the update
ops for updating the moving_mean and moving_variance. This is now being
done by using slim.learning.create_train_op() instead of native
tf.train.Optimizer().minimize() to create the train op. Second, the
decay parameter has been decreased from the default 0.999 to 0.95, as
with too high of a value batch_norm takes too long to converge on a
small dataset.

For more information, see
tensorflow/tensorflow#1122.

models/mobilenet.py

@@ -59,7 +59,8 @@ def mobilenet_base(scope, inputs, conv_defs, is_training=True, reuse=None):
     with tf.variable_scope(scope, inputs, reuse=reuse):
         with slim.arg_scope([slim.conv2d, slim.separable_conv2d],
                 padding='SAME'):
-            with slim.arg_scope([slim.batch_norm], is_training=is_training):
+            with slim.arg_scope([slim.batch_norm], is_training=is_training, 
+                    decay=0.95):
                 net = inputs
                 for i, conv_def in enumerate(conv_defs):
                     end_point_base = 'Conv2d_%d' % i
@@ -124,7 +125,7 @@ def mobilenet_head(inputs, dropout_keep_prob=0.9, num_classes=2,
         is_training=True):
     # Define the network
     net, end_points = mobilenet_base("MobileNetHead", inputs, HEAD_CONV_DEFS, 
-            is_training)
+            is_training=is_training)
 
     with tf.variable_scope('Logits'):
         net = slim.avg_pool2d(net, [15, 15], padding='VALID', 

scripts/train.py

@@ -12,6 +12,7 @@
 os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
 
 import tensorflow as tf
+slim = tf.contrib.slim
 from tables import *
 import numpy as np
 
@@ -164,18 +165,21 @@ def load_val_data(index):
         filter_summ_op = tf.summary.image('filter',tf.slice(tf.transpose(kernel, perm=[3, 0, 1, 2]),begin=[0,0,0,0],size=[96,11,11,1]),max_outputs=IMAGE_VIZ_MAX_OUTPUTS)
         activations_summ_op = tf.summary.image('activations',tf.slice(activations,begin=[0,0,0,0],size=[TRAIN_BATCH_SIZE,58,58,1]),max_outputs=IMAGE_VIZ_MAX_OUTPUTS)
 
-    #train op
+    # Define the train op
     if args.optimizer == 'adadelta':
-        train_op = tf.train.AdadeltaOptimizer(learning_rate=variable_learning_rate).minimize(loss)
+        optimizer = tf.train.AdadeltaOptimizer(
+                learning_rate=variable_learning_rate)
     elif args.optimizer == 'adam':
-        train_op = tf.train.AdamOptimizer(learning_rate=variable_learning_rate,
-        beta1=0.9,
-        beta2=0.999,
-        epsilon=0.1,
-        use_locking=False,
-        name='Adam').minimize(loss)
+        optimizer = tf.train.AdamOptimizer(
+                learning_rate=variable_learning_rate,
+                beta1=0.9,
+                beta2=0.999,
+                epsilon=0.1,
+                use_locking=False,
+                name='Adam')
     else:
-        train_op = tf.train.GradientDescentOptimizer(variable_learning_rate).minimize(loss)
+        optimizer = tf.train.GradientDescentOptimizer(variable_learning_rate)
+    train_op = slim.learning.create_train_op(loss, optimizer)
 
     #for embeddings visualization
     if embedding: