def __init__(self, data, model, sess):
"""
:param data:
:param model: expect an initialized and trained model
:return:
"""
roots_size = [tree_util.size_of_tree(root) for root in data]
data = helper.sort_by(data, roots_size)
probs, labels = model.predict_and_label(data, sess)
labels = get_prediction(labels)
predictions = get_prediction(probs)
self.acc = get_accuracy(labels, predictions)
if len(data) < 1500:
print(model.accuracy(data, sess))
self.TP, self.FP, self.TN, self.FN = get_confusion_matrix(labels, predictions)
self.precision = self.TP / (self.TP + self.FP)
self.recall = self.TP / (self.TP + self.FN)
self.F1 = 2 * (self.precision * self.recall) / (self.precision + self.recall)
self.TPR_list, self.FPR_list = get_roc_values(labels, probs)
self.auc = metrics.auc(self.FPR_list, self.TPR_list)
def build_feed_dict(self, roots, sort=True, train=False):
if sort:
roots_size = [tree_util.size_of_tree(root) for root in roots]
roots = helper.sort_by(roots, roots_size)
roots_size = [tree_util.size_of_tree(root) for root in roots]
roots_list, permutation = helper.greedy_bin_packing(roots, roots_size, np.max(roots_size))
node_list_list = []
root_indices = []
internal_nodes_array = []
lstm_prev_list = []
for i, roots in enumerate(roots_list):
node_list = []
root_index = 0
leaf_index = 0
lstm_prev = [0]
lstm_prev_count = 0
for root in roots:
tree_util.depth_first_traverse(root, node_list, lambda node, node_list: node_list.append(node))
leaf_count = tree_util.leafs_in_tree(root)
root_index += leaf_count
root_indices.append([i, root_index])
for j in range(0,leaf_count):
leaf_index += 1
internal_nodes_array.append([i, leaf_index])
leaf_count = tree_util.leafs_in_tree(root)
for x in range(0, leaf_count):
if x == 0:
lstm_prev.append(0)
else:
lstm_prev.append(lstm_prev_count)
lstm_prev_count += 1
node_list_list.append(node_list)
lstm_prev_list.append(lstm_prev)
feed_dict = {
self.dropout_rate: FLAGS.dropout_prob if train else 0,
self.lstm_prev_array: helper.lists_pad(lstm_prev_list, 0),
self.leaf_word_array: helper.lists_pad(
[[0] + [self.word_embed.get_idx(node.value) for node in node_list if node.is_leaf]
for node_list in node_list_list]
,0),
self.loss_array: root_indices if self.use_root_loss else internal_nodes_array,
self.root_array: root_indices,
self.label_array: helper.lists_pad([
[[0, 0]] + [node.label for node in node_list if node.is_leaf]
for node_list in node_list_list], [0, 0])
}
return feed_dict, permutation
from utils import data_util, tree_util, helper
_data_util = data_util.DataUtil()
data = _data_util.get_data()
roots_size = [tree_util.size_of_tree(root) for root in data.train_trees]
roots = helper.sort_by(data.train_trees, roots_size)
for root in roots[-5:]:
print(root.label)
print(root.to_sentence())
print()
def build_feed_dict(self, roots, sort=True, train=False):
if sort:
roots_size = [tree_util.size_of_tree(root) for root in roots]
roots = helper.sort_by(roots, roots_size)
roots_size = [tree_util.size_of_tree(root) for root in roots]
roots_list, permutation = helper.greedy_bin_packing(
roots, roots_size, np.max(roots_size))
node_list_list = []
node_to_index_list = []
root_indices = []
lstm_idx_list = []
internal_nodes_array = []
for i, roots in enumerate(roots_list):
node_list = []
lstm_idx = [0]
root_index = 0
start = 0
for root in roots:
tree_util.depth_first_traverse(
root, node_list,
lambda node, node_list: node_list.append(node))
_, start = tree_util.get_preceding_lstm_index(
root, start, start, lstm_idx)
root_index += tree_util.size_of_tree(root)
root_indices.append([i, root_index])
node_list_list.append(node_list)
node_to_index = helper.reverse_dict(node_list)
node_to_index_list.append(node_to_index)
lstm_idx_list.append(lstm_idx)
for node in node_list:
if not node.is_leaf:
internal_nodes_array.append([i, node_to_index[node] + 1])
internal_nodes_array = internal_nodes_array if len(
internal_nodes_array) > 0 else [[0, 0]]
feed_dict = {
self.dropout_rate:
FLAGS.dropout_prob if train else 0,
self.leaf_word_array:
helper.lists_pad([[0] + [
self.word_embed.get_idx(node.value)
for node in node_list if node.is_leaf
] for node_list in node_list_list], 0),
self.lstm_index_array:
helper.lists_pad(lstm_idx_list, 0),
self.loss_array:
root_indices if self.use_root_loss else internal_nodes_array,
self.root_array:
root_indices,
self.is_leaf_array:
helper.lists_pad(
[[0] + helper.to_int([node.is_leaf for node in node_list])
for node_list in node_list_list], 0),
self.word_index_array:
helper.lists_pad(
[[0] +
[self.word_embed.get_idx(node.value) for node in node_list]
for node_list in node_list_list],
self.word_embed.get_idx("ZERO")),
self.left_child_array:
helper.lists_pad([[0] + helper.add_one([
node_to_index[node.left_child]
if node.left_child is not None else -1 for node in node_list
])
for node_list, node_to_index in zip(
node_list_list, node_to_index_list)], 0),
self.right_child_array:
helper.lists_pad([[0] + helper.add_one([
node_to_index[node.right_child]
if node.right_child is not None else -1 for node in node_list
])
for node_list, node_to_index in zip(
node_list_list, node_to_index_list)], 0),
self.label_array:
helper.lists_pad([[[0, 0]] + [node.label for node in node_list]
for node_list in node_list_list], [0, 0])
}
return feed_dict, permutation
def get_data():
if not os.path.exists(directories.CLASSIFIER_DATA_DIR):
os.mkdir(directories.CLASSIFIER_DATA_DIR)
if not os.path.exists(directories.CLASSIFIER_DATA(FLAGS.model_name)):
os.mkdir(directories.CLASSIFIER_DATA(FLAGS.model_name))
if os.path.exists(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_train.npy'):
x_train = np.load(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_train.npy')
y_train = np.load(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'y_train.npy')
x_val = np.load(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_val.npy')
y_val = np.load(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'y_val.npy')
x_test = np.load(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_test.npy')
y_test = np.load(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'y_test.npy')
else:
_data_util = data_util.DataUtil()
data = _data_util.get_data()
roots_size = [
tree_util.size_of_tree(root) for root in data.train_trees
]
data.train_trees = helper.sort_by(data.train_trees, roots_size)
roots_size = [tree_util.size_of_tree(root) for root in data.val_trees]
data.val_trees = helper.sort_by(data.val_trees, roots_size)
roots_size = [tree_util.size_of_tree(root) for root in data.test_trees]
data.test_trees = helper.sort_by(data.test_trees, roots_size)
if FLAGS.use_gpu:
config = None
else:
config = tf.ConfigProto(device_count={'GPU': 0})
if FLAGS.word_embed_model == constants.WORD2VEC:
word_embeddings = Word2Vec(mode=FLAGS.word_embed_mode,
dimensions=FLAGS.word_embedding_size)
elif FLAGS.word_embed_model == constants.FASTTEXT:
word_embeddings = FastText(mode=FLAGS.word_embed_mode,
dimensions=FLAGS.word_embedding_size)
else: # FLAGS.word_embed_model == constants.GLOVE
word_embeddings = GloVe(mode=FLAGS.word_embed_mode,
dimensions=FLAGS.word_embedding_size)
g_tree = tf.Graph()
with g_tree.as_default():
model = None
if FLAGS.model == constants.DEEP_RNN:
model = deepRNN(data, word_embeddings, FLAGS.model_name)
elif FLAGS.model == constants.BATCH_TREE_RNN:
model = treeRNN_batch(data, word_embeddings, FLAGS.model_name)
elif FLAGS.model == constants.NEERBEK_TREE_RNN:
model = treeRNN_neerbek(data, word_embeddings,
FLAGS.model_name)
elif FLAGS.model == constants.TREE_LSTM:
model = treeLSTM(data, word_embeddings, FLAGS.model_name)
elif FLAGS.model == constants.TRACKER_TREE_RNN:
model = treeRNN_tracker(data, word_embeddings,
FLAGS.model_name)
elif FLAGS.model == constants.TRACKER_TREE_LSTM:
model = treeLSTM_tracker(data, word_embeddings,
FLAGS.model_name)
elif FLAGS.model == constants.LSTM:
model = LSTM(data, word_embeddings, FLAGS.model_name)
with tf.Session(config=tf.ConfigProto(
device_count={'GPU': 0})) as sess:
saver = tf.train.Saver()
model.load_best(sess, saver, "validation")
x_train = np.array(
model.get_representation(data.train_trees, sess))
y_train = np.array(get_labels(data.train_trees))
x_val = np.array(model.get_representation(
data.val_trees, sess))
y_val = np.array(get_labels(data.val_trees))
x_test = np.array(
model.get_representation(data.test_trees, sess))
y_test = np.array(get_labels(data.test_trees))
np.save(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_train',
x_train)
np.save(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'y_train',
y_train)
np.save(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_val', x_val)
np.save(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'y_val', y_val)
np.save(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'x_test',
x_test)
np.save(
directories.CLASSIFIER_DATA(FLAGS.model_name) + 'y_test',
y_test)
return {
'train': (x_train, y_train),
'val': (x_val, y_val),
'test': (x_test, y_test)
}
def select_data(self, data, cut_off, cluster_predictions=None):
roots_size = [tree_util.size_of_tree(root) for root in data]
data = np.array(helper.sort_by(data, roots_size))
t = time()
if cluster_predictions is None:
# Get representations
representations, predictions, labels, permutations = [], [], [], []
batch_size = 500
batches = helper.batches(data, batch_size, perm=False)
pbar = tqdm(
bar_format=
'{percentage:.0f}%|{bar}| Elapsed: {elapsed}, Remaining: {remaining} (batches: {n_fmt}/{total_fmt}) ',
total=len(batches))
for i, batch in enumerate(batches):
feed_dict, permuts = self.model.build_feed_dict(batch,
sort=True)
reps, labs = self.session.run(
[self.model.sentence_representations, self.model.labels],
feed_dict=feed_dict)
representations.extend(reps)
labels.extend(labs)
permutations.extend(list(i * batch_size + np.array(permuts)))
pbar.update(1)
pbar.close()
print()
self.representations = np.array(representations)[permutations]
self.labels = np.array(performance.get_prediction(
np.array(labels)))[permutations]
# Get clusters
try_cluster = True
tries = 10
while try_cluster:
tries -= 1
self.cluster_predictions = self.cluster_model.cluster(
self.representations)
if np.bincount(self.cluster_predictions).max() <= 0.8 * len(
self.representations) or tries >= 0:
try_cluster = False
else:
self.cluster_predictions = cluster_predictions
self.labels = tree_util.get_labels(data)
# Get acc of clusters
cluster_mfo = []
cluster_mfo_labels = []
for i in range(self.num_clusters):
mfo, l = self.mfo(i)
cluster_mfo.append((i, mfo))
cluster_mfo_labels.append((i, l))
# Return data
cluster_mfo.sort(key=lambda el: el[1], reverse=True)
helper._print(f'Cluster MFO scores:')
for (k, mfo), (_, l) in zip(cluster_mfo, cluster_mfo_labels):
helper._print(
f'\tCluster {k}: {mfo}, highest label: {l}, size: {len(self.labels[self.cluster_predictions == k])}/{len(data)}'
)
removed_percent = 0
data_to_use = []
for cluster, acc in cluster_mfo:
new_percent = removed_percent + len(
data[self.cluster_predictions == cluster]) / len(data)
removed_percent = new_percent
if acc < cut_off:
data_to_use.extend(data[self.cluster_predictions == cluster])
helper._print(
f'Done selecting data for training. Overall time used for selection is {int((time() - t)/60)} minutes and {int((time() - t) % 60)} seconds'
)
return data_to_use, self.cluster_predictions
def parse_trees(dataset="small",
type='train',
remove=False): # todo maybe change input param
"""
https://github.com/erickrf/treernn/blob/master/tree.py
:param type: what dataset to use
:return: a list of trees
"""
file = directories.TREES_DIRS[dataset] + '%s.txt' % type
if not os.path.isdir(directories.TREES_DIRS[dataset]):
os.makedirs(directories.TREES_DIRS[dataset])
if not os.path.isfile(file):
if dataset == 'all':
helper._print(f'Creating new {file}...')
with open(file, 'w+') as f:
for l in directories.TREES_ZIP_PATHS:
smaller_tree_file = directories.TREES_DIRS[
l] + '%s.txt' % type
helper._print(f'Merging from {smaller_tree_file}...')
if not os.path.isfile(smaller_tree_file):
helper._print(
f'Extracting {directories.TREES_ZIP_PATHS[l]}...')
with zipfile.ZipFile(directories.TREES_ZIP_PATHS[l],
'r') as zip:
zip.extractall(path=directories.TREES_DIRS[l])
correct_labels(constants.TREE_LABELS[l], l)
with open(smaller_tree_file, 'r+') as sf:
for tree in sf:
f.write(tree)
elif dataset == 'small':
helper._print(
'No small dataset. Try pulling from Git... Or make your own you lazy bastard!'
)
else:
helper._print(
f'Extracting {directories.TREES_ZIP_PATHS[dataset]}...')
with zipfile.ZipFile(directories.TREES_ZIP_PATHS[dataset],
'r') as zip:
zip.extractall(path=directories.TREES_DIRS[dataset])
correct_labels(constants.TREE_LABELS[dataset], dataset)
helper._print("Loading %s trees.." % type)
with open(file, 'r') as fid:
trees = []
lines = fid.readlines()
pbar = tqdm(
bar_format=
'{percentage:.0f}%|{bar}| Elapsed: {elapsed}, Remaining: {remaining} ({n_fmt}/{total_fmt}) ',
total=len(lines))
for i, l in enumerate(lines):
if (i + 1) % 1000 == 0:
pbar.update(1000)
trees.append(parse_tree(l))
pbar.update(len(lines) % 1000)
pbar.close()
print()
sentence_length = [count_leaf(tree) for tree in trees]
sentence_length = np.array(sentence_length)
helper._print("Avg length:", np.average(sentence_length))
trees = np.array(trees)
if remove:
helper._print(
"Shorten then 90 word:",
int(
np.sum(np.array(sentence_length) <= 90) /
len(sentence_length) * 100), "%")
helper._print("Ratio of removed labels:",
ratio_of_labels(trees[np.array(sentence_length) > 90]))
trees = np.array(
helper.sort_by(trees[np.array(sentence_length) <= 90],
sentence_length[np.array(sentence_length) <= 90]))
return trees
