def test_metrics_equivalence():
import numpy as np
a = np.random.random((10, 16, 64, 64, 3))
b = np.random.random((10, 16, 64, 64, 3))
metrics = [
mean_squared_error, peak_signal_to_noise_ratio, structural_similarity,
vgg_cosine_distance
]
metrics_np = [
mean_squared_error_np, peak_signal_to_noise_ratio_np,
structural_similarity_np, vgg_cosine_distance_np
]
sess = tf.Session()
# this initializes the vgg network for the pure tf (i.e. non-np) metrics
with tf.variable_scope('vgg'):
vgg_network.vgg16(tf.placeholder(tf.float32, shape=[None] * 4))
vgg_network.vgg_assign_from_values_fn(var_name_prefix='vgg/')(sess)
for keep_axis in (None, 0, 1, (0, 1)):
for metric, metric_np in zip(metrics, metrics_np):
m = metric(a, b, keep_axis=keep_axis)
m_np = metric_np(a, b, keep_axis=keep_axis)
assert np.allclose(sess.run(m), m_np, atol=1e-7)
print('The test metrics returned the same values.')
def initialize_graph(self):
self._image_placeholder = tf.placeholder(dtype=tf.float32)
with tf.variable_scope('vgg', reuse=tf.AUTO_REUSE):
_, self._feature_op = _with_flat_batch(vgg_network.vgg16)(
self._image_placeholder)
self._assign_from_values_fn = vgg_network.vgg_assign_from_values_fn(
var_name_prefix='vgg/')