def get_random_binary_tree(min_depth,
max_depth,
num_vals,
child_prob=0.7,
_cur_depth=0):
root = tree_rnn.BinaryNode(int(random.random() * num_vals))
if max_depth <= 1:
return root
# left child
if _cur_depth < min_depth or random.random() < child_prob:
left_child = get_random_binary_tree(min_depth,
max_depth - 1,
num_vals,
child_prob=child_prob,
_cur_depth=_cur_depth + 1)
root.add_left(left_child)
# right child
if _cur_depth < min_depth or random.random() < child_prob:
right_child = get_random_binary_tree(min_depth,
max_depth - 1,
num_vals,
child_prob=child_prob,
_cur_depth=_cur_depth + 1)
root.add_right(right_child)
return root
def test_tree_rnn_var_degree():
model = DummyBinaryRNN(10, 2, 2, 1, degree=2)
emb = model.embeddings.get_value()
root = tree_rnn.BinaryNode(0)
c1 = tree_rnn.BinaryNode(1)
cc1 = tree_rnn.BinaryNode(2)
ccc1 = tree_rnn.BinaryNode(3)
cc1.add_left(ccc1)
c1.add_right(cc1)
root.add_left(c1)
root_emb = model.evaluate(root)
expected = emb[0] + (emb[1] + (emb[2] + emb[3])**2)
assert_array_almost_equal(expected, root_emb)
cccc1 = tree_rnn.BinaryNode(5)
cccc2 = tree_rnn.BinaryNode(6)
ccc1.add_left(cccc1)
ccc1.add_right(cccc2)
root_emb = model.evaluate(root)
expected = emb[0] + (emb[1] + (emb[2] + (emb[3] + emb[5] + emb[6]**2))**2)
assert_array_almost_equal(expected, root_emb)
# check step works without error
model.train_step(root, np.array([0]).astype(theano.config.floatX))
def test():
# very simple for now... just checks compilation and training step
# TODO: better tests
root = tree_rnn.BinaryNode(0)
c1 = tree_rnn.BinaryNode(1)
cc1 = tree_rnn.BinaryNode(2)
ccc1 = tree_rnn.BinaryNode(3)
cccc1 = tree_rnn.BinaryNode(5)
cccc2 = tree_rnn.BinaryNode(6)
ccc1.add_left(cccc1)
ccc1.add_right(cccc2)
cc1.add_left(ccc1)
c1.add_right(cc1)
root.add_left(c1)
# check child sum
model = tree_gru.ChildSumTreeGRU(10, 20, 30, 1)
model.train_step(root, np.array([0]).astype(theano.config.floatX))
# check n-ary
model = tree_gru.NaryTreeGRU(10, 20, 30, 1)
model.train_step(root, np.array([0]).astype(theano.config.floatX))