I'm training a small neural network using Tensorflow 1.10. The training process goes well and I get expected results, but it works weirdly in the validation or in the testing process. I'm checking out my code multiple times, but can't find the reason. Why I'm getting different results in the training and validation process.
In training process you should control some variable updates. As you know batch normalization contains some not trainable weights, which getting changed during training process. These variables are moving_mean and moving_variance. To force updating these variables you should put this code in your training process.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer(learning_rate=L_RATE).minimize(loss=loss)