def testJacobianNumerically(self):
x = tf.random_normal([_BATCH_SIZE, 2])
h1 = tf.contrib.layers.fully_connected(x, 20)
h2 = tf.contrib.layers.fully_connected(h1, 20)
f = tf.contrib.layers.fully_connected(h2, 10)
j_tensor = jacobian_conditioning.compute_jacobian(xs=x, fx=f)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
x_np = sess.run(x)
jacobian = sess.run(j_tensor, feed_dict={x: x_np})
# Test 10 random elements.
for _ in range(10):
# Pick a random element of jacobian to test.
batch_idx = np.random.randint(_BATCH_SIZE)
x_idx = np.random.randint(2)
f_idx = np.random.randint(10)
# Test with finite differences.
epsilon = 1e-4
x_plus = x_np.copy()
x_plus[batch_idx, x_idx] += epsilon
f_plus = sess.run(f, feed_dict={x: x_plus})[batch_idx, f_idx]
x_minus = x_np.copy()
x_minus[batch_idx, x_idx] -= epsilon
f_minus = sess.run(f, feed_dict={x: x_minus})[batch_idx, f_idx]
self.assertAllClose(jacobian[batch_idx, f_idx, x_idx],
(f_plus - f_minus) / (2. * epsilon),
rtol=1e-3,
atol=1e-3)
def testJacobianSimpleCase(self):
x = tf.random_normal([_BATCH_SIZE, 2])
W = tf.constant([[2., -1.], [1.5, 1.]]) # pylint: disable=invalid-name
f = tf.matmul(x, W)
j_tensor = jacobian_conditioning.compute_jacobian(xs=x, fx=f)
with tf.Session() as sess:
jacobian = sess.run(j_tensor)
# Transpose of W in 'expected' is expected because in vector notation
# f = W^T * x.
expected = tf.tile([[[2, 1.5], [-1, 1]]], [_BATCH_SIZE, 1, 1])
self.assertAllClose(jacobian, expected)
def testJacobianAgainstSlowVersion(self):
x = tf.random_normal([_BATCH_SIZE, 2])
h1 = tf.contrib.layers.fully_connected(x, 20)
h2 = tf.contrib.layers.fully_connected(h1, 20)
f = tf.contrib.layers.fully_connected(h2, 10)
j_slow_tensor = slow_jacobian(xs=x, fx=f)
j_fast_tensor = jacobian_conditioning.compute_jacobian(xs=x, fx=f)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
j_fast, j_slow = sess.run([j_fast_tensor, j_slow_tensor])
self.assertAllClose(j_fast, j_slow)
def ComputeGeneratorConditionNumber(sess, gan):
"""Computes the generator condition number.
Computes the Jacobian of the generator in session, then postprocesses to get
the condition number.
Args:
sess: tf.Session object.
gan: AbstractGAN object, that is already present in the current tf.Graph.
Returns:
A list of length gan.batch_size. Each element is the condition number
computed at a single z sample within a minibatch.
"""
shape = gan.fake_images.get_shape().as_list()
flat_generator_output = tf.reshape(
gan.fake_images, [gan.batch_size, np.prod(shape[1:])])
tf_jacobian = conditioning_lib.compute_jacobian(xs=gan.z,
fx=flat_generator_output)
z_sample = gan.z_generator(gan.batch_size, gan.z_dim)
np_jacobian = sess.run(tf_jacobian, feed_dict={gan.z: z_sample})
result_dict = conditioning_lib.analyze_jacobian(np_jacobian)
return result_dict["metric_tensor"]["log_condition_number"]