def test_actnorm_init_conv(self):
np.random.seed(52321)
images_np = np.random.rand(8, 32, 32, 3)
images = tf.to_float(images_np)
flow = fl.InputLayer(images)
layer = fl.ActnormLayer(scale=np.sqrt(np.pi))
with self.assertRaises(AssertionError):
layer.get_ddi_init_ops()
new_flow = layer(flow, forward=True)
x, logdet, z = new_flow
init_ops = layer.get_ddi_init_ops()
self.assertEqual(z, None)
with self.test_session() as sess:
# initialize network
sess.run(tf.global_variables_initializer())
sess.run(init_ops)
x, logdet = sess.run([x, logdet])
self.assertEqual(x.shape, images_np.shape)
# check var after passing act norm
self.assertAllClose(np.var(x.reshape([-1, 3]), axis=0),
[np.pi] * 3,
atol=0.01)
self.assertAllClose(np.mean(x.reshape([-1, 3]), axis=0), [0.0] * 3,
atol=0.01)
self.forward_inverse(layer, flow)
def test_actnorm_init_conv_iter(self):
np.random.seed(52321)
images_ph = tf.placeholder(tf.float32, shape=[8, 32, 32, 3])
flow = fl.InputLayer(images_ph)
layer = fl.ActnormLayer(scale=np.sqrt(np.pi))
new_flow = layer(flow, forward=True)
x, logdet, z = new_flow
init_ops = layer.get_ddi_init_ops(num_init_iterations=50)
with self.test_session() as sess:
# initialize network
sess.run(tf.global_variables_initializer())
for i in range(200):
sess.run(init_ops,
feed_dict={images_ph: np.random.rand(8, 32, 32, 3)})
for i in range(5):
x_np, logdet_np = sess.run(
[x, logdet],
feed_dict={images_ph: np.random.rand(8, 32, 32, 3)})
self.assertEqual(x.shape, x_np.shape)
self.assertAllClose(np.var(x_np.reshape([-1, 3]), axis=0),
[np.pi] * 3,
atol=0.1)
self.assertAllClose(np.mean(x_np.reshape([-1, 3]), axis=0),
[0.0] * 3,
atol=0.1)
self.forward_inverse(
layer, flow, feed_dict={images_ph: np.random.rand(8, 32, 32, 3)})
def feed_dict_fn():
return {images_ph: np.random.rand(8, 32, 32, 3)}
def post_init_fn(sess):
init_ops = layer.get_ddi_init_ops()
sess.run(init_ops, {images_ph: np.random.rand(8, 32, 32, 3)})
with tf.variable_scope("TestTrain"):
layer = fl.ActnormLayer(scale=np.sqrt(np.pi))
self.try_to_train_identity_layer(layer,
flow,
feed_dict_fn=feed_dict_fn,
post_init_fn=post_init_fn)
def test_actnorm_conv(self):
images_np = np.random.rand(8, 32, 32, 3)
images = tf.to_float(images_np)
flow = fl.InputLayer(images)
layer = fl.ActnormLayer()
new_flow = layer(flow, forward=True)
x, logdet, z = new_flow
self.assertEqual(z, None)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
x, logdet = sess.run([x, logdet])
self.assertEqual(x.shape, images_np.shape)
self.forward_inverse(layer, flow)
def step_flow(
name: str, shift_and_log_scale_fn: Callable[[tf.Tensor], tf.Tensor]
) -> Tuple[fl.ChainLayer, fl.ActnormLayer]:
"""Create single step of the Glow model:
1. actnorm
2. invertible conv
3. affine coupling layer
Returns:
step_layer: a flow layer which perform 1-3 operations
actnorm: a reference of actnorm layer from step 1. This reference can be
used to initialize this layer using data dependent initialization
"""
actnorm = fl.ActnormLayer()
layers = [
actnorm,
fl.InvertibleConv1x1Layer(),
fl.AffineCouplingLayer(shift_and_log_scale_fn=shift_and_log_scale_fn),
]
return fl.ChainLayer(layers, name=name), actnorm