Simple batch normalization prediction fix for Keras 1.x weights

BN_test_time.py

import keras.backend as K

# Load the weights
# If there are more than 1 BN, you might wanna have a proper layer
moving_mean = f['batchnormalization_1_running_mean']
moving_std = f['batchnormalization_1_running_std']
beta = f['batchnormalization_1_beta']
gamma = f['batchnormalization_2_beta']

# BE CAREFUL! Keras 1 stores std
# , while still taking var in the backend function.
moving_variance = moving_std ** 2

def BN_test_time(x):
    return K.batch_normalization(
                    x,
                    moving_mean,
                    moving_variance,
                    beta,
                    gamma,
                    epsilon=self.epsilon)

def BN_output_shape(input_shape):
    return input_shape

# in your model...
model.add(Lambda(BN_test_time,
         output_shape=BN_output_shape))