Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[FIX] Ontology - remove cache and other fixes #896

Merged
merged 2 commits into from
Oct 5, 2022
Merged

[FIX] Ontology - remove cache and other fixes #896

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1ed60c8
Ontology - remove cache
PrimozGodec Aug 30, 2022
ef2ee05
Ontology - handle unsuccessful embeddings
PrimozGodec Sep 2, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
162 changes: 34 additions & 128 deletions orangecontrib/text/ontology.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,17 +1,13 @@
from typing import List, Set, Dict, Tuple, Optional, Callable
from collections import Counter
from itertools import chain
import os
import pickle

import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

from orangecontrib.text.vectorization.sbert import SBERT
from Orange.misc.environ import cache_dir
from Orange.util import dummy_callback, wrap_callback

EMB_DIM = 384


class Tree:

Expand Down Expand Up @@ -222,103 +218,58 @@ def generate_ontology(
return generation[best, :], roots[best]


def cos_sim(x: np.array, y: np.array) -> float:
dot = np.dot(x, y)
return 0 if np.allclose(dot, 0) else dot / (np.linalg.norm(x) * np. linalg.norm(y))


class EmbeddingStorage:

def __init__(self):
self.cache_dir = os.path.join(cache_dir(), 'ontology')
if not os.path.isdir(self.cache_dir):
os.makedirs(self.cache_dir)
self.similarities = dict()
try:
with open(os.path.join(self.cache_dir, 'sims.pkl'), 'rb') as file:
self.similarities = pickle.load(file)
except IOError:
self.similarities = dict()
self.embeddings = dict()

def save_similarities(self):
with open(os.path.join(self.cache_dir, 'sims.pkl'), 'wb') as file:
pickle.dump(self.similarities, file)

def get_embedding(self, word: str) -> Optional[np.array]:
if word in self.embeddings:
return self.embeddings[word]
try:
emb = np.load(os.path.join(self.cache_dir, f'{word}.npy'))
self.embeddings[word] = emb
return emb
except IOError:
return None

def save_embedding(self, word: str, emb: np.array) -> None:
self.embeddings[word] = emb
np.save(os.path.join(self.cache_dir, f'{word}.npy'), emb)

def clear_storage(self) -> None:
self.similarities = dict()
self.embeddings = dict()
if os.path.isdir(self.cache_dir):
for file in os.listdir(self.cache_dir):
os.remove(os.path.join(self.cache_dir, file))


class OntologyHandler:

def __init__(self):
self.embedder = SBERT()
self.storage = EmbeddingStorage()

def generate(
self,
words: List[str],
callback: Callable = dummy_callback
) -> Dict:
if len(words) == 0:
return {}
) -> Tuple[Dict, int]:
embeddings = self.embedder(words, wrap_callback(callback, end=0.1))
non_none = [(w, e) for w, e in zip(words, embeddings) if e is not None]
skipped = len(words) - len(non_none)
if len(non_none) == 0:
return {}, skipped
words, embeddings = zip(*non_none)
sims = self._get_similarities(embeddings)
callback(0.2)
if len(words) == 1:
return {words[0]: {}}
if len(words) == 2:
return {sorted(words)[0]: {sorted(words)[1]: {}}}
sims = self._get_similarities(
words,
self._get_embeddings(words, wrap_callback(callback, end=0.1)),
wrap_callback(callback, start=0.1, end=0.2)
)
if len(words) == 3:
return {words[0]: {}}, skipped
elif len(words) == 2:
return {sorted(words)[0]: {sorted(words)[1]: {}}}, skipped
elif len(words) == 3:
root = np.argmin(np.sum(sims, axis=1))
rest = sorted([words[i] for i in range(3) if i != root])
return {words[root]: {rest[0]: {}, rest[1]: {}}}
return {words[root]: {rest[0]: {}, rest[1]: {}}}, skipped
ontology, root = generate_ontology(
words,
sims,
callback=wrap_callback(callback, start=0.2)
)
return Tree.from_prufer_sequence(ontology, words, root).to_dict()
return Tree.from_prufer_sequence(ontology, words, root).to_dict(), skipped

def insert(
self,
tree: Dict,
words: List[str],
callback: Callable = dummy_callback
) -> Dict:
) -> Tuple[Dict, int]:
tree = Tree.from_dict(tree)
self._get_embeddings(words, wrap_callback(callback, end=0.3))
ticks = iter(np.linspace(0.3, 0.9, len(words)))
skipped = 0

for word in words:
tick = next(ticks)
tree.adj_list.append(set())
tree.labels.append(word)
sims = self._get_similarities(
tree.labels,
self._get_embeddings(tree.labels, lambda x: callback(tick)),
lambda x: callback(tick)
)
embeddings = self.embedder(tree.labels)
if embeddings[-1] is None:
# the last embedding is for the newly provided word
# if embedding is not successful skip it
skipped += 1
continue
sims = self._get_similarities(embeddings)
idx = len(tree.adj_list) - 1
fitness_function = FitnessFunction(tree.labels, sims).fitness
scores = list()
Expand All @@ -331,65 +282,20 @@ def insert(
best = np.argmax(scores)
tree.adj_list[best].add(idx)
tree.adj_list[idx].add(best)
callback(tick)
callback(next(ticks))

return tree.to_dict()
return tree.to_dict(), skipped

def score(self, tree: Dict, callback: Callable = dummy_callback) -> float:
if not tree:
return 0
tree = Tree.from_dict(tree)
sims = self._get_similarities(
tree.labels,
self._get_embeddings(tree.labels, wrap_callback(callback, end=0.7)),
wrap_callback(callback, start=0.7, end=0.8)
)
embeddings = self.embedder(tree.labels, wrap_callback(callback, end=0.7))
sims = self._get_similarities(embeddings)
callback(0.9)
fitness_function = FitnessFunction(tree.labels, sims).fitness
return fitness_function(tree, tree.root)[0]

def _get_embeddings(
self,
words: List[str],
callback: Callable = dummy_callback
) -> np.array:
embeddings = np.zeros((len(words), EMB_DIM))
missing, missing_idx = list(), list()
ticks = iter(np.linspace(0.0, 0.6, len(words)))
for i, word in enumerate(words):
callback(next(ticks))
emb = self.storage.get_embedding(word)
if emb is None:
missing.append(word)
missing_idx.append(i)
else:
embeddings[i, :] = emb
if len(missing_idx) > 0:
embs = self.embedder(missing, callback=wrap_callback(callback, start=0.6, end=0.9))
if None in embs:
raise RuntimeError("Couldn't obtain embeddings.")
embeddings[missing_idx, :] = np.array(embs)
for i in missing_idx:
self.storage.save_embedding(words[i], embeddings[i, :])

return embeddings

def _get_similarities(
self,
words: List[str],
embeddings: np.array,
callback: Callable = dummy_callback
) -> np.array:
sims = np.zeros((len(words), len(words)))
ticks = iter(np.linspace(0.0, 1.0, int(len(words) * (len(words) - 1) / 2)))
for i in range(len(words)):
for j in range(i + 1, len(words)):
callback(next(ticks))
key = tuple(sorted((words[i], words[j])))
try:
sim = self.storage.similarities[key]
except KeyError:
sim = cos_sim(embeddings[i, :], embeddings[j, :])
self.storage.similarities[key] = sim
sims[i, j] = sim
sims[j, i] = sim
self.storage.save_similarities()
return sims
@staticmethod
def _get_similarities(embeddings: np.array) -> np.array:
return cosine_similarity(embeddings, embeddings)
Loading