def testDeterministicDropoutTest(self):
x = tf.ones([4, 6], dtype=tf.float32)
x = py_utils.DeterministicDropout(x, keep_prob=0.7, seeds=[1234, 5678])
with self.session() as sess:
x_val = sess.run(x)
# pyformat: disable
self.assertAllClose(
[[1.0 / 0.7, 0.0000000, 0.0000000, 0.0000000, 1.0 / 0.7, 1.0 / 0.7],
[1.0 / 0.7, 1.0 / 0.7, 1.0 / 0.7, 1.0 / 0.7, 1.0 / 0.7, 1.0 / 0.7],
[1.0 / 0.7, 0.0000000, 0.0000000, 1.0 / 0.7, 1.0 / 0.7, 0.0000000],
[1.0 / 0.7, 0.0000000, 0.0000000, 1.0 / 0.7, 1.0 / 0.7, 1.0 / 0.7]],
x_val)
# pyformat: enable
self.assertAllClose(22.85714, np.sum(x_val))
self.assertEqual(x_val.dtype, np.float32)
def _GetWeight(self, theta):
p = self.params
filter_w = theta.w
# First normalize filter_w over the temporal dimension here.
filter_w = tf.nn.softmax(filter_w / p.temperature, axis=0)
# Add dropconnect on the weights for regularization.
if p.dropconnect_prob > 0.0 and not self.do_eval:
if p.deterministic_dropout:
filter_w = py_utils.DeterministicDropout(
filter_w, 1.0 - p.dropconnect_prob,
py_utils.GenerateStepSeedPair(p))
else:
filter_w = tf.nn.dropout(
filter_w, rate=p.dropconnect_prob, seed=p.random_seed)
# Tie the parameters of every subsequent number of weight_tiling_factor
# channels.
filter_w = tf.tile(filter_w, [1, 1, p.weight_tiling_factor, 1])
return filter_w
