def testVariableBatchSize(self):
q_tm1 = tf.placeholder(tf.float32, shape=[None, 3])
q_t = tf.placeholder(tf.float32, shape=[None, 3])
a_tm1 = tf.placeholder(tf.int32, shape=[None])
pcont_t = tf.placeholder(tf.float32, shape=[None])
probs_a_t = tf.placeholder(tf.float32, shape=[None, 3])
r_t = tf.placeholder(tf.float32, shape=[None])
sarse = rl.sarse(q_tm1, a_tm1, r_t, pcont_t, q_t, probs_a_t)
# Check static shapes.
self.assertEqual(sarse.loss.get_shape().as_list(), [None])
self.assertEqual(sarse.extra.td_error.get_shape().as_list(), [None])
self.assertEqual(sarse.extra.target.get_shape().as_list(), [None])
# Check runtime shapes.
batch_size = 11
feed_dict = {
q_tm1: np.random.random([batch_size, 3]),
q_t: np.random.random([batch_size, 3]),
a_tm1: np.random.randint(0, 3, [batch_size]),
pcont_t: np.random.random([batch_size]),
r_t: np.random.random(batch_size),
probs_a_t: np.random.uniform(size=[batch_size, 3])
}
with self.test_session() as sess:
loss, td_error, target = sess.run(
[sarse.loss, sarse.extra.td_error, sarse.extra.target],
feed_dict=feed_dict)
self.assertEqual(loss.shape, (batch_size, ))
self.assertEqual(td_error.shape, (batch_size, ))
self.assertEqual(target.shape, (batch_size, ))
def testVariableBatchSize(self):
q_tm1 = tf.placeholder(tf.float32, shape=[None, 3])
q_t = tf.placeholder(tf.float32, shape=[None, 3])
a_tm1 = tf.placeholder(tf.int32, shape=[None])
pcont_t = tf.placeholder(tf.float32, shape=[None])
probs_a_t = tf.placeholder(tf.float32, shape=[None, 3])
r_t = tf.placeholder(tf.float32, shape=[None])
sarse = rl.sarse(q_tm1, a_tm1, r_t, pcont_t, q_t, probs_a_t)
# Check static shapes.
self.assertEqual(sarse.loss.get_shape().as_list(), [None])
self.assertEqual(sarse.extra.td_error.get_shape().as_list(), [None])
self.assertEqual(sarse.extra.target.get_shape().as_list(), [None])
# Check runtime shapes.
batch_size = 11
feed_dict = {
q_tm1: np.random.random([batch_size, 3]),
q_t: np.random.random([batch_size, 3]),
a_tm1: np.random.randint(0, 3, [batch_size]),
pcont_t: np.random.random([batch_size]),
r_t: np.random.random(batch_size),
probs_a_t: np.random.uniform(size=[batch_size, 3])
}
with self.test_session() as sess:
loss, td_error, target = sess.run(
[sarse.loss, sarse.extra.td_error, sarse.extra.target],
feed_dict=feed_dict)
self.assertEqual(loss.shape, (batch_size,))
self.assertEqual(td_error.shape, (batch_size,))
self.assertEqual(target.shape, (batch_size,))
def testIncorrectProbsTensor(self):
probs_a_t = tf.constant([[0.2, 0.5, 0.3], [0.3, 0.5, 0.3]],
dtype=tf.float32)
with self.test_session() as sess:
with self.assertRaisesRegexp(
tf.errors.InvalidArgumentError, "probs_a_t tensor does not sum to 1"):
self.sarse = rl.sarse(self.q_tm1, self.a_tm1, self.r_t, self.pcont_t,
self.q_t, probs_a_t, debug=True)
sess.run(self.sarse.extra.target)
def setUp(self):
super(SarseTest, self).setUp()
self.q_tm1 = tf.constant([[1, 1, 0.5], [1, 1, 3]], dtype=tf.float32)
self.q_t = tf.constant([[1.5, 1, 2], [3, 2, 1]], dtype=tf.float32)
self.a_tm1 = tf.constant([0, 1], dtype=tf.int32)
self.probs_a_t = tf.constant([[0.2, 0.5, 0.3], [0.3, 0.4, 0.3]],
dtype=tf.float32)
self.pcont_t = tf.constant([1, 1], dtype=tf.float32)
self.r_t = tf.constant([4, 1], dtype=tf.float32)
self.sarse = rl.sarse(self.q_tm1, self.a_tm1, self.r_t, self.pcont_t,
self.q_t, self.probs_a_t)
def setUp(self):
super(SarseTest, self).setUp()
self.q_tm1 = tf.constant([[1, 1, 0.5], [1, 1, 3]], dtype=tf.float32)
self.q_t = tf.constant([[1.5, 1, 2], [3, 2, 1]], dtype=tf.float32)
self.a_tm1 = tf.constant([0, 1], dtype=tf.int32)
self.probs_a_t = tf.constant([[0.2, 0.5, 0.3], [0.3, 0.4, 0.3]],
dtype=tf.float32)
self.pcont_t = tf.constant([1, 1], dtype=tf.float32)
self.r_t = tf.constant([4, 1], dtype=tf.float32)
self.sarse = rl.sarse(self.q_tm1, self.a_tm1, self.r_t, self.pcont_t,
self.q_t, self.probs_a_t)
def testIncorrectProbsTensor(self):
probs_a_t = tf.constant([[0.2, 0.5, 0.3], [0.3, 0.5, 0.3]],
dtype=tf.float32)
with self.test_session() as sess:
with self.assertRaisesRegexp(tf.errors.InvalidArgumentError,
"probs_a_t tensor does not sum to 1"):
self.sarse = rl.sarse(self.q_tm1,
self.a_tm1,
self.r_t,
self.pcont_t,
self.q_t,
probs_a_t,
debug=True)
sess.run(self.sarse.extra.target)
def testCompatibilityCheck(self):
probs_a_t = tf.placeholder(tf.float32, [None, 2])
with self.assertRaisesRegexp(
ValueError, "Sarse: Error in rank and/or compatibility check"):
self.sarse = rl.sarse(self.q_tm1, self.a_tm1, self.r_t,
self.pcont_t, self.q_t, probs_a_t)
def testCompatibilityCheck(self):
probs_a_t = tf.placeholder(tf.float32, [None, 2])
with self.assertRaisesRegexp(
ValueError, "Sarse: Error in rank and/or compatibility check"):
self.sarse = rl.sarse(self.q_tm1, self.a_tm1, self.r_t, self.pcont_t,
self.q_t, probs_a_t)