def prep_tree_srm_arg(relation_list, arg_pos, wbm, max_length, all_left_branching=False):
assert arg_pos == 1 or arg_pos == 2
n_samples = len(relation_list)
w_indices = np.zeros((2 * max_length, n_samples)).astype("int64")
c_mask = np.zeros((max_length, n_samples), dtype=config.floatX)
node_mask = np.zeros((2 * max_length, n_samples), dtype=config.floatX)
# children = np.zeros((max_length, n_samples, 2), dtype='int64')
children = np.zeros((n_samples, max_length, 3), dtype="int64")
for i, relation in enumerate(relation_list):
if all_left_branching:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.find_parse_tree(relation, arg_pos)
if len(parse_tree.leaves()) == 0:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
indices = wbm.index_tokens(parse_tree.leaves(), ignore_OOV=False)
sequence_length = min(max_length, len(indices))
w_indices[:sequence_length, i] = indices[:sequence_length]
ordering_matrix, num_leaves = tree_util.reverse_toposort(parse_tree)
num_nodes = min(2 * max_length, ordering_matrix.shape[0])
print num_leaves, num_nodes
# assert(num_nodes >= num_leaves)
if num_nodes > num_leaves:
num_inner_nodes = num_nodes - num_leaves
children[i, :num_inner_nodes, :] = ordering_matrix[num_leaves:num_nodes, :]
c_mask[:num_inner_nodes, i] = 1.0
node_mask[num_leaves:num_nodes, i] = 1.0
children = np.swapaxes(children, 0, 1)
embedding_series = (
wbm.wm[w_indices.flatten()].reshape([max_length * 2, n_samples, wbm.num_units]).astype(config.floatX)
)
return embedding_series, children, c_mask, node_mask
def prep_tree_srm_arg(relation_list, arg_pos, wbm, max_length,
all_left_branching=False, node_label_alphabet={}):
"""Make the matrices from the data required for the tree model
T = number of time steps
N = number of samples
d = dimensionality of the embedding
k = dimensionality of the
embedding_series: 2T x N x d serrated matrix word embedding for the leaves
children : T x N x 3 children serrated matrix
c_mask : T x N masking matrix for children matrix
node_mask : 2T x N masking matrix for the internal nodes
(for embedding_series) nice for computing mean h or sum h
node_label_tensor : 2T x N x k. This masks embedding_series matrix
"""
assert arg_pos == 1 or arg_pos == 2
n_samples = len(relation_list)
w_indices = np.zeros((2 * max_length, n_samples)).astype('int64')
c_mask = np.zeros((max_length, n_samples), dtype=config.floatX)
node_mask = np.zeros((2 * max_length, n_samples), dtype=config.floatX)
children = np.zeros((n_samples, max_length, 3), dtype='int64')
node_label_tensor = np.zeros((2 * max_length, n_samples, len(node_label_alphabet)), dtype=config.floatX)
for i, relation in enumerate(relation_list):
if all_left_branching:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.find_parse_tree(relation, arg_pos)
if len(parse_tree.leaves()) == 0:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
indices = wbm.index_tokens(parse_tree.leaves(), ignore_OOV=False)
sequence_length = min(max_length, len(indices))
w_indices[:sequence_length, i] = indices[:sequence_length]
ordering_matrix, node_label_list, num_leaves = \
tree_util.reverse_toposort(parse_tree)
num_nodes = min(2 * max_length, ordering_matrix.shape[0])
if num_nodes > num_leaves:
num_inner_nodes = num_nodes - num_leaves
children[i, :num_inner_nodes, :] = ordering_matrix[num_leaves:num_nodes, :]
c_mask[:num_inner_nodes, i] = 1.
node_mask[num_leaves:num_nodes, i] = 1.
if len(node_label_alphabet) > 0:
for t, node_label in enumerate(node_label_list):
if node_label is not None and t < (2 * max_length):
if node_label in node_label_alphabet:
label_index = node_label_alphabet[node_label]
else:
label_index = node_label_alphabet['OTHERS']
node_label_tensor[t, i, label_index] = 1.
children = np.swapaxes(children, 0, 1)
embedding_series = \
wbm.wm[w_indices.flatten()].\
reshape([max_length * 2, n_samples, wbm.num_units]).\
astype(config.floatX)
return embedding_series, children, c_mask, node_mask, node_label_tensor
def prep_tree_arg(relation_list, arg_pos, all_left_branching=False):
parse_trees = []
for i, relation in enumerate(relation_list):
if all_left_branching:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.find_parse_tree(relation, arg_pos)
print parse_tree
if len(parse_tree.leaves()) == 0:
print "use left branching tree because parse is empty"
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.binarize_tree(parse_tree)
print parse_tree
parse_trees.append(parse_tree)
return parse_trees
def prep_tree_arg(relation_list, arg_pos, all_left_branching=False):
parse_trees = []
for i, relation in enumerate(relation_list):
if all_left_branching:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.find_parse_tree(relation, arg_pos)
print parse_tree
if len(parse_tree.leaves()) == 0:
print 'use left branching tree because parse is empty'
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.binarize_tree(parse_tree)
print parse_tree
parse_trees.append(parse_tree)
return parse_trees
def prep_tree_srm_arg(relation_list,
arg_pos,
wbm,
max_length,
all_left_branching=False):
assert arg_pos == 1 or arg_pos == 2
n_samples = len(relation_list)
w_indices = np.zeros((2 * max_length, n_samples)).astype('int64')
c_mask = np.zeros((max_length, n_samples), dtype=config.floatX)
node_mask = np.zeros((2 * max_length, n_samples), dtype=config.floatX)
#children = np.zeros((max_length, n_samples, 2), dtype='int64')
children = np.zeros((n_samples, max_length, 3), dtype='int64')
for i, relation in enumerate(relation_list):
if all_left_branching:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
else:
parse_tree = tree_util.find_parse_tree(relation, arg_pos)
if len(parse_tree.leaves()) == 0:
parse_tree = tree_util.left_branching_tree(relation, arg_pos)
indices = wbm.index_tokens(parse_tree.leaves(), ignore_OOV=False)
sequence_length = min(max_length, len(indices))
w_indices[:sequence_length, i] = indices[:sequence_length]
ordering_matrix, num_leaves = tree_util.reverse_toposort(parse_tree)
num_nodes = min(2 * max_length, ordering_matrix.shape[0])
print num_leaves, num_nodes
#assert(num_nodes >= num_leaves)
if num_nodes > num_leaves:
num_inner_nodes = num_nodes - num_leaves
children[i, :num_inner_nodes, :] = ordering_matrix[
num_leaves:num_nodes, :]
c_mask[:num_inner_nodes, i] = 1.
node_mask[num_leaves:num_nodes, i] = 1.
children = np.swapaxes(children, 0, 1)
embedding_series = \
wbm.wm[w_indices.flatten()].\
reshape([max_length * 2, n_samples, wbm.num_units]).\
astype(config.floatX)
return embedding_series, children, c_mask, node_mask
