def expand_sentence(text_p: Union[str, Example],
word_indices: List[int],
query=None,
additional_mask_indices: List[int] = None,
schedule_idx=[-1]) -> List[Example]:
if additional_mask_indices is None:
additional_mask_indices = []
if isinstance(text_p, Example):
text = text_p.sentence
else:
text = text_p
text = Sentence(text)
word_indices = list(word_indices)
word_indices = [
wi for wi in word_indices
if not all([s in ":,;.*" for s in text.words[wi]])
]
if len(word_indices) == 0:
return []
original_words = {i: text.words[i] for i in word_indices}
max_words = query.consider_max_words if query is not None else Query(
None).consider_max_words
masked_sentence = Sentence(
text.get_with_masked(word_indices + additional_mask_indices))
predictions = masked_sentence.calc_mask_predictions(max_words)
result = []
for word_idx in word_indices:
if not predictions[word_idx]:
continue
sentences = []
for predicted_token, score in predictions[word_idx]:
new_sen = text.replace_word(word_idx, predicted_token)
sentences.append(new_sen)
classification = calc_sentiment_batch(sentences)
for i, (predicted_token, score) in enumerate(predictions[word_idx]):
if original_words[word_idx] != predicted_token:
if isinstance(text_p, str):
e = Example(sentences[i],
classification[i], [(word_idx, score)],
pred_ind=[i],
sched_ind=[schedule_idx],
sent_ind=[0])
else:
e = Example(sentences[i],
classification[i],
text_p.changes + [(word_idx, score)],
pred_ind=text_p.prediction_indices + [i],
sched_ind=text_p.schedule_indices +
[schedule_idx],
sent_ind=text_p.sentence_indices + [0])
result.append(e)
return result
def generate_example_table():
result = ""
text = Sentence(ex)
for query in [
Query(wanted_cls=y_prime,
c=0.2,
num_needed=5,
mini_beam_search=False,
allow_splitting=False,
consider_top_k=10),
Query(wanted_cls=y_prime,
c=0.2,
num_needed=5,
mini_beam_search=True,
allow_splitting=False,
consider_top_k=10),
Query(wanted_cls=y_prime,
c=0.2,
num_needed=5,
mini_beam_search=False,
allow_splitting=True,
consider_top_k=10),
Query(wanted_cls=y_prime,
c=0.2,
num_needed=5,
mini_beam_search=True,
allow_splitting=True,
consider_top_k=10),
]:
q_result = generate_counterfactuals(ex, query)
print(q_result)
ex_df = []
for change_group in q_result.examples:
for e in change_group:
se = Sentence(e.sentence)
d = e.changed_word_distances()
cwi = e.changed_word_indices()
entry = {
"Original":
', '.join([text.words[wi] for wi in cwi]),
"Counterfactual":
f"{', '.join([se.words[wi] for wi in cwi])}",
"Klassifikation":
f"{e.cls[1]:.2f}",
"Distanz":
f"{sum([d_i ** 2 for d_i in d]) + COST_PER_ADDITIONAL_WORD * len(d):.2f}"
}
ex_df.append(entry)
ex_df = pd.DataFrame(ex_df)
# result += f"\n\n\nOriginale Klassifikation: {text.calc_sentiment()[1]:.2f} \\\\ \n"
# result += f"\nMBS={query.mini_beam_search}, ST={query.allow_splitting}, MAX\\_WORDS={query.consider_max_words} \\\\ \n"
result += "\n\n"
result += ex_df.to_latex(
index=False,
caption=
f"{query.alg()} (Originale Klassifikation: {text.calc_sentiment()[1]:.2f})"
)
return result
def get_sentence_word_mapping(text: str) -> List[Tuple[int, int]]:
# can be longer than original, because Sentence(..) limited to 512 tokens
tok_sen = nltk_tokenizer.sent_tokenize(text)
original = model_config.tokenizer.clean_up_tokenization(" ".join(
Sentence(text.lower().strip()).words))
sentence = Sentence(original)
word_sentence_map = []
last_start = 0
for xp in tok_sen:
cxp = clean_for_comparison(xp.lower())
for i in range(last_start, len(sentence.words) + 1):
ccomp = clean_for_comparison("".join(sentence.words[last_start:i]))
if ccomp == cxp:
word_sentence_map.append((last_start, i - 1))
last_start = i
break
if last_start != len(sentence.words):
word_sentence_map.append((last_start, len(sentence.words)))
# merge small sentences (<6) into neighbors
while True:
dists = [(stop - start) for (start, stop) in word_sentence_map]
if all([d > MIN_SEN_LEN for d in dists]):
return word_sentence_map
else:
# if True not in [d > MIN_SEN_LEN for d in dists]:
# return word_sentence_map
sen_idx = [d > MIN_SEN_LEN for d in dists].index(False)
# calc left side sen len
if sen_idx - 1 < 0:
left_len = None
else:
o_start, o_stop = word_sentence_map[sen_idx - 1]
left_len = o_stop - o_start
# calc right side sen len
if sen_idx + 1 >= len(word_sentence_map):
right_len = None
else:
o_start, o_stop = word_sentence_map[sen_idx + 1]
right_len = o_stop - o_start
if right_len is None and left_len is None:
return word_sentence_map
elif left_len is None or (right_len is not None and left_len <
right_len): # merge with right
new_entry = (word_sentence_map[sen_idx][0],
word_sentence_map[sen_idx + 1][1])
word_sentence_map[sen_idx:sen_idx + 2] = [new_entry]
elif right_len is None or (right_len is not None and right_len <=
left_len): # merge with left
new_entry = (word_sentence_map[sen_idx - 1][0],
word_sentence_map[sen_idx][1])
word_sentence_map[sen_idx - 1:sen_idx + 1] = [new_entry]
def example_as_latex_string(idx):
example_with_ans = data[idx]
res_obj: Result
res_obj = example_with_ans[1]
if len(res_obj.examples) == 0:
return ""
cf_ex = [
escape_latex(w)
for w in Sentence(res_obj.examples[0][0].sentence).words
]
origi = [
escape_latex(w)
for w in Sentence(res_obj.stats.original_sentence).words
]
assert len(cf_ex) == len(origi)
i = 0
result_str = f"""
\\begin{{figure}}[h]
\\begin{{center}}
\\begin{{tabular}}{{|l|c|c|}}
\\multicolumn{{3}}{{c}}{{Variante: {res_obj.query.alg()}}} \\\\
\\hline
{{}} & Counterfactual Example & Original \\\\
\\hline
Perplexity & {res_obj.examples[0][0].calc_perplexity():.2f} & {Sentence(res_obj.stats.original_sentence).calc_perplexity():.2f} \\\\
Polarität & {res_obj.examples[0][0].cls[1]:.2f} & {res_obj.stats.original_classification[1]:.2f} \\\\
\\hline
\\end{{tabular}}
\\end{{center}}
\n
"""
while i != len(cf_ex):
if cf_ex[i] != origi[i]:
combine_from, combine_to = i, i + 1
while origi[combine_to] != cf_ex[combine_to]:
combine_to += 1
left_part = ' '.join(origi[combine_from:combine_to])
right_part = ' '.join(cf_ex[combine_from:combine_to])
change = "\\mbox{[" + tex_color(
left_part, "blue", False) + " \\to{} " + tex_color(
right_part, "blue", True) + "]} "
result_str += change
i += (combine_to - combine_from)
else:
result_str += cf_ex[i] + " "
i += 1
result_str += f"\n\\caption{{ Beispiel {idx // 4} aus der {DATASET} Evaluation ({res_obj.query.alg()}) }}\n\\label{{{idx // 4}_{res_obj.query.alg()}}}\n\\end{{figure}} "
return result_str + "\n\n\n"
def calc_sentence_edit_schedule(query, sw_map, text):
if USE_GRADIENTS_FOR_SENTENCE_RELEVANCE:
gradients = text.calc_gradients(query.wanted_cls)
sent_grad = defaultdict(list)
for i in range(len(text.words)):
idx = [a <= i <= b for (a, b) in sw_map].index(True)
sent_grad[idx].append(gradients[i])
gradients_per_sentence = [(si, np.linalg.norm(g))
for (si, g) in sent_grad.items()]
edit_sentence_order = [
y[0] for y in sorted(
gradients_per_sentence, key=lambda x: x[1], reverse=True)
]
else:
# use distance to wanted classification for relevance
dist_to_wanted_cls = []
for start, stop in sw_map:
sub = model_config.tokenizer.clean_up_tokenization(" ".join(
text.words[start:stop + 1]))
cls = Sentence(sub).calc_sentiment()
dst = mse_loss(torch.tensor(cls),
torch.tensor(query.wanted_cls, dtype=torch.float32))
dist_to_wanted_cls.append(dst)
edit_sentence_order = np.argsort(-np.array(dist_to_wanted_cls))
return edit_sentence_order
def default(self, obj):
if isinstance(obj, np.integer):
return int(obj)
elif isinstance(obj, np.floating):
return float(obj)
elif isinstance(obj, np.ndarray):
return obj.tolist()
elif isinstance(obj, Example):
ex_dict = dict(obj.__dict__)
ex_dict["sentence"] = Sentence(clean(obj.sentence)).words
ex_dict["perplexity"] = Sentence(obj.sentence).calc_perplexity()
del ex_dict["prediction_indices"]
del ex_dict["schedule_indices"]
del ex_dict["sentence_indices"]
del ex_dict["changes"]
return ex_dict
else:
return super(CustomEncoder, self).default(obj)
def info(self):
result_str = ""
sen = Sentence(self.stats.original_sentence)
result_str += f"pp={sen.calc_perplexity():.2f}, {len(sen.words)} words, y={np.round(self.stats.original_classification, 2)}\n"
result_str += f"Duration: {self.stats.total_duration} | {self.stats.find_matching_words_duration} searching words | {self.stats.merging_duration} merging.\n"
result_str += f"{self.total_valid_examples()} examples, {len(self.rest)} in rest, found {len(self.examples)} of {self.query.num_needed} groups with different indices.\n"
for e in self.simple_results():
result_str += "\t" + e.info() + "\n"
result_str += "Successful!\n" if self.success else "Query not fullfilled!"
return result_str
def __init__(self, sentence: str, classification, changes, pred_ind, sched_ind, sent_ind):
self.cls = classification
self.sentence = sentence
self.changes: List[Tuple[WordIdx, float]] = changes
self.prediction_indices = pred_ind # n-th word in list of alternative words
self.schedule_indices = sched_ind # 1st highest gradient word, 2nd hgw, ...
self.sentence_indices = sent_ind # 1st sentence,... only != 0 if splitting text
self.perplexity = None
for (i, _) in changes:
assert i < len(Sentence(self.sentence).words)
assert len(self.schedule_indices) == len(self.prediction_indices) == len(self.changes)
def extract_colors(r, per_sentence):
text = r.stats.original_sentence
cmap = matplotlib.colors.LinearSegmentedColormap.from_list(
"", ["white", "blue"])
text_s = Sentence(text)
if not per_sentence:
word_gradients = text_s.calc_gradients(r.query.wanted_cls)
wgn = np.interp(word_gradients,
(np.min(word_gradients), np.max(word_gradients)),
(0., 1.))
fg, bg = [], []
for ind in range(len(wgn)):
ctpl = cmap(wgn[ind])[:3]
tc = twofivefive(text_color(ctpl))
ctpl = twofivefive(ctpl)
fg.append(str(tc)[1:-1])
bg.append(str(ctpl)[1:-1])
return fg, bg
else:
sw_map = get_sentence_word_mapping(text)
edit_sentence_order = calc_sentence_edit_schedule(
r.query, sw_map, text_s)
fg, bg = [], []
for enm_si, si in enumerate(edit_sentence_order):
start, stop = sw_map[si]
sub = model_config.tokenizer.clean_up_tokenization(" ".join(
text_s.words[start:stop + 1]))
subtext = Sentence(sub)
word_gradients = np.array(
subtext.calc_gradients(r.query.wanted_cls))
word_gradients /= np.linalg.norm(word_gradients)
wgn = np.interp(word_gradients,
(np.min(word_gradients), np.max(word_gradients)),
(0., 1.))
for ind in range(len(wgn)):
ctpl = cmap(wgn[ind])[:3]
tc = twofivefive(text_color(ctpl))
ctpl = twofivefive(ctpl)
fg.append(str(tc)[1:-1])
bg.append(str(ctpl)[1:-1])
return fg, bg
def merge_all_changes_into_one(grouped_items, original) -> Example:
# Merge all available -> check
merge_all_changes_into_this_one = list(original.words)
merge_changes = []
pred_indices = []
sched_indices = []
sent_indices = []
for changes_a, la in grouped_items:
if len(changes_a) == 1 and len(la) > 0:
e: Example = la[0]
s = Sentence(e)
merge_all_changes_into_this_one[changes_a[0]] = s.words[
changes_a[0]]
merge_changes += e.changes
pred_indices.append(e.prediction_indices)
sched_indices.append(e.schedule_indices)
sent_indices.append(e.sentence_indices)
merge_all_changes_into_this_one = " ".join(merge_all_changes_into_this_one)
all_sentiment = calc_sentiment_batch([merge_all_changes_into_this_one])
return Example(merge_all_changes_into_this_one, all_sentiment[0],
merge_changes, pred_indices, sched_indices, sent_indices)
def __repr__(self):
# DEBUG HELP
left_right_window = 6
sen = Sentence(self.sentence)
changed_indices = self.changed_word_indices()
sp = [w if i not in changed_indices else f"#{w}#" for i, w in enumerate(sen.words)]
relevant_parts = ' '.join(sp)
if len(relevant_parts) > 160:
relevant_parts = []
for i, (w_id, score) in enumerate(self.changes):
if i != 0:
relevant_parts.append("[...]")
word = sp[w_id]
left = sp[max(0, w_id - left_right_window):w_id]
right = sp[w_id + 1:w_id + left_right_window]
relevant_parts.extend(left + [word] + right)
relevant_parts = ' '.join(relevant_parts)
return f"PRED_IDX={self.prediction_indices}, " \
f"cls={np.round(self.cls, 2)}, " \
f"idxs={changed_indices} " \
f"dist={np.round(list(dict(self.changes).values()), 2)}, {relevant_parts}"
def one_mask():
max_words = 20
result = defaultdict(list)
result_dist_diff_only = defaultdict(list)
result_wdiff = defaultdict(list)
for enm in range(len(dataset)):
print(f"\nSentence {enm}: ", end="")
data = dataset[enm]
x, y = data["sentence"], data["label"]
x = model_config.tokenizer.clean_up_tokenization(x)
y = [0, 1] if y == 1 else [1, 0]
y_prime = [1 - y[0], 1 - y[1]]
s = Sentence(x)
word_gradients = s.calc_gradients(y_prime)
sorted_highest = np.argsort(word_gradients)[::-1]
for observed_idx in sorted_highest[:10]:
# observed_idx = sorted_highest[0]
print(f"{observed_idx},", end="")
sdir = 1 if len(s.words) - observed_idx > observed_idx else -1
alt_s = Sentence(s.get_with_masked([observed_idx]))
original_answer = alt_s.calc_mask_predictions()[observed_idx]
if len(original_answer) != 0:
for mask_distance in range(1, max_words):
if observed_idx + mask_distance * sdir < 0 or observed_idx + mask_distance * sdir >= len(
alt_s.words):
continue
new_sen = Sentence(
alt_s.get_with_masked([
observed_idx + mask_distance * sdir, observed_idx
]))
alt_sen_pred = new_sen.calc_mask_predictions(
)[observed_idx]
avg_distance, avg_word_diff, dist_diff_only = find_differences(
original_answer, alt_sen_pred)
# print(f"Mask offset {mask_distance}: dist={avg_distance:.3f} word_dist={avg_word_diff:.3f}")
result[mask_distance].append(avg_distance)
result_wdiff[mask_distance].append(avg_word_diff)
result_dist_diff_only[mask_distance].append(dist_diff_only)
if enm % 50 == 0 or enm == len(dataset) - 1:
fig = plt.figure(figsize=(11, 8))
plt.title(
"Relation Bewertung der Wörter zur Nähe des nächsten [MASK]-Token"
)
plt.xlabel("Entfernung zum zusätzlichen [MASK]-Token")
plt.xlim(0, max_words)
plt.ylim(0., 0.65)
plt.ylabel("Veränderung der Bewertung")
idx, mean, std = list(
zip(*[(md, np.mean(lst), np.std(lst))
for (md, lst) in result_wdiff.items()]))
mean = np.array(mean)
std = np.array(std)
plt.plot(idx, mean, color='r', label="Wort-Unterschiede")
plt.fill_between(idx, mean - std, mean + std, color='r', alpha=.2)
idx, mean, std = list(
zip(*[(md, np.mean(lst), np.std(lst))
for (md, lst) in result_dist_diff_only.items()]))
mean = np.array(mean)
std = np.array(std)
plt.plot(idx, mean, color='green', label="Distanz-Unterschiede")
plt.fill_between(idx,
mean - std,
mean + std,
color='green',
alpha=.2)
plt.xticks(idx)
plt.legend()
plt.savefig(f'{root}saved_plots/all/_besser_{enm}.png')
# plt.show()
plt.close(fig)
def two_mask():
max_words = 15
result = defaultdict(list)
result_dist_diff_only = defaultdict(list)
result_wdiff = defaultdict(list)
for enm in range(len(dataset)):
print(f"\nSentence {enm}: ", end="")
data = dataset[enm]
x, y = data["sentence"], data["label"]
x = model_config.tokenizer.clean_up_tokenization(x)
y = [0, 1] if y == 1 else [1, 0]
y_prime = [1 - y[0], 1 - y[1]]
s = Sentence(x)
word_gradients = s.calc_gradients(y_prime)
sorted_highest = np.argsort(word_gradients)[::-1]
for observed_idx in sorted_highest[:10]:
print(f"{observed_idx},", end="")
alt_s = Sentence(s.get_with_masked([observed_idx]))
original_answer = alt_s.calc_mask_predictions()[observed_idx]
if len(original_answer) != 0:
for mask_distance1 in range(-max_words, max_words + 1):
for mask_distance2 in range(-max_words, max_words + 1):
if not (0 <= observed_idx + mask_distance1 < len(
alt_s.words)):
continue
if not (0 <= observed_idx + mask_distance2 < len(
alt_s.words)):
continue
new_sen = Sentence(
alt_s.get_with_masked(
[observed_idx + mask_distance1, observed_idx]))
new_sen = Sentence(
new_sen.get_with_masked(
[observed_idx + mask_distance2, observed_idx]))
alt_sen_pred = new_sen.calc_mask_predictions(
)[observed_idx]
avg_distance, avg_word_diff, dist_diff_only = find_differences(
original_answer, alt_sen_pred)
result[(mask_distance1,
mask_distance2)].append(avg_distance)
result_wdiff[(mask_distance1,
mask_distance2)].append(avg_word_diff)
result_dist_diff_only[(
mask_distance1,
mask_distance2)].append(dist_diff_only)
if enm % 2 == 0 or enm == len(dataset) - 1:
all_variants = [(result, "result"), (result_wdiff, "wdiff"),
(result_dist_diff_only, "ddiff")]
with open('used_data.pickle', 'wb') as handle:
pickle.dump(all_variants, handle)
for res, name in all_variants:
data = [(k, np.mean(v)) for k, v in res.items()]
matrix = np.zeros(shape=(2 * max_words + 1, 2 * max_words + 1))
for (i, j), m in data:
matrix[(i + max_words), j + max_words] = m
plt.figure(figsize=(15, 12))
ax = sns.heatmap(
np.flip(matrix, axis=0),
linewidth=0.0,
xticklabels=list(range(-max_words, max_words + 1)),
yticklabels=list(reversed(range(-max_words,
max_words + 1))))
ax.set_title(
"Durchschnittliche Veränderung der Wörter bei 2 MASK-Tokens"
)
plt.savefig(f'{root}saved_plots/2d/{name}_{enm}.pdf')
plt.close()
def clean(s):
return model_config.tokenizer.clean_up_tokenization(' '.join(
Sentence(s).words)).replace(" - ", "-").lower()
import gensim.downloader as api
import pandas as pd
from gensim.models.keyedvectors import Word2VecKeyedVectors
from config import model_config
from search_utils.Sentence import Sentence
wv = api.load('word2vec-google-news-300')
wv: Word2VecKeyedVectors
s = """my thoughts were focused on the characters ."""
s = model_config.tokenizer.clean_up_tokenization(s)
sen_s = Sentence(s)
df_wv = dict()
df = dict()
for i in range(len(sen_s.words)):
word = sen_s.words[i]
if word in ".:,?!-(){}[]/\\|&%":
continue
new_s = sen_s.get_with_masked(i)
bert_preds = Sentence(new_s).calc_mask_predictions()[i]
wv_preds = wv.most_similar(word, topn=15, restrict_vocab=200_000)
df_wv[word] = list(list(zip(*wv_preds))[0])
df[word] = list(zip(*bert_preds))[0][:15]
df = pd.DataFrame(df)
df_wv = pd.DataFrame(df_wv)
def _gen_cf_ex(text: str, query) -> Result:
"""
Parameters
----------
text: str
query : search_helper.classes.Query.Query
Returns
-------
"""
gc.collect()
torch.cuda.empty_cache()
stats = Statistics(original_sentence=text)
stats.total_duration.resume()
stats.query = query
text = model_config.tokenizer.clean_up_tokenization(text)
text = Sentence(text)
original_cls = text.calc_sentiment()
stats.original_classification = original_cls
examples = []
schedule = text.calc_edit_schedule(query)
for schedule_idx, (edit_strategy, word_indices) in enumerate(schedule):
assert isinstance(schedule_idx, int)
with stats.find_matching_words_duration:
if edit_strategy == Edit.NEIGHBOURS:
batch = expand_sentence_neighbour(text.text, query,
word_indices, schedule_idx)
else:
word_indices, mask_indices = word_indices
batch = expand_sentence(text.text, word_indices, query,
mask_indices, schedule_idx)
# filtering only 'relevant' makes the found words more extreme
# relevant_batch = [b for b in batch if abs(original_cls[cls_idx] - b.cls[cls_idx]) > MIN_SENTIMENT_CHANGE]
relevant_batch = batch
debug(
f"{len(examples)} examples total | {len(batch)} new for {len(word_indices)} words with {schedule_idx} highest gradient"
)
stats.tried_examples += len(batch)
examples.extend(relevant_batch)
with stats.merging_duration:
num_per_group = max(4 -
schedule_idx, 1) if schedule_idx < 10 else -1
merged_examples = generate_merged_examples(text, examples, query,
num_per_group)
examples.extend(merged_examples)
stats.tried_examples += len(merged_examples)
results: Result = examples_are_sufficient(examples, query)
if results.sufficient():
stats.total_duration.pause()
assert stats.all_timers_stopped()
results.stats = stats
results.query = query
return results
results: Result = examples_are_sufficient(examples, query)
stats.total_duration.pause()
assert stats.all_timers_stopped()
results.stats = stats
results.query = query
return results
def _gen_cf_ex_long(text: str, query) -> Result:
"""
Parameters
----------
text : str
query: search_helper.classes.Query.Query
Returns
-------
Result
"""
gc.collect()
torch.cuda.empty_cache()
stats = Statistics(original_sentence=text)
stats.total_duration.resume()
stats.query = query
stats.original_sentence = text
text = model_config.tokenizer.clean_up_tokenization(text)
text = Sentence(text)
stats.original_classification = text.calc_sentiment()
sw_map = get_sentence_word_mapping(text.text)
edit_sentence_order = calc_sentence_edit_schedule(query, sw_map, text)
examples = []
for sentence_sched_idx, si in enumerate(
edit_sentence_order[:query.consider_max_sentences]):
debug("> subsentence")
start, stop = sw_map[si]
sub = model_config.tokenizer.clean_up_tokenization(" ".join(
text.words[start:stop + 1]))
sub_query = copy(query)
subresult: Result = _gen_cf_ex(sub, sub_query)
if stats.tried_sentences is None:
stats.tried_sentences = 0
stats.tried_sentences += 1
stats.add(subresult.stats)
subexample: Example
best_subexamples = [j[0] for j in subresult.examples]
debug(
f"SUBEXAMPLE SEARCH FOUND {len(best_subexamples)} of {sub_query.num_needed}"
)
if len(best_subexamples) < query.num_needed:
best_subexamples.extend(subresult.rest[:(query.num_needed -
len(best_subexamples))])
debug(f"Added from rest, now {len(best_subexamples)}")
for subexample in best_subexamples:
new_sen = list(text.words)
new_sen[start:stop + 1] = [subexample.sentence]
new_sen = Sentence(
model_config.tokenizer.clean_up_tokenization(
" ".join(new_sen)))
new_cls = new_sen.calc_sentiment()
# print(np.round(new_cls, 3), new_sen.text)
new_changes = [(pos + start, dist)
for (pos, dist) in subexample.changes]
e = Example(new_sen.text,
new_cls,
new_changes,
pred_ind=subexample.prediction_indices,
sched_ind=subexample.schedule_indices,
sent_ind=[sentence_sched_idx])
examples.append(e)
with stats.merging_duration:
debug("> subsentence merge")
merged_examples = generate_merged_examples(text, examples, query,
1)
examples.extend(merged_examples)
stats.tried_examples += len(merged_examples)
debug("< subsentence merge")
results: Result = examples_are_sufficient(examples, query)
if results.sufficient():
stats.total_duration.pause()
assert stats.all_timers_stopped()
results.stats = stats
results.query = query
results.sentence_map = sw_map
return results
debug("< subsentence")
results: Result = examples_are_sufficient(examples, query)
stats.total_duration.pause()
assert stats.all_timers_stopped()
results.stats = stats
results.query = query
results.sentence_map = sw_map
return results
from search_utils.Sentence import Sentence
Sentence("David [MASK] is a spanish tennis player.").calc_mask_predictions(
) # ferrer
Sentence("David [MASK] is a famous musician.").calc_mask_predictions() # bowie
Sentence("David [MASK] is a musician.").calc_mask_predictions() # smith
Sentence("[MASK] is a musician.").calc_mask_predictions() # he
Sentence(
"David [MASK] was the prime minister.").calc_mask_predictions() # cameron
Sentence(
"David [MASK] is the prime minister.").calc_mask_predictions() # cameron
Sentence("It's over Anakin! I have the [MASK] ground!").calc_mask_predictions(
) # high
Sentence("It's over Anakin! I have the high [MASK]!").calc_mask_predictions(
) # ground
Sentence("It's over Anakin! I [MASK] the high ground!").calc_mask_predictions(
) # need
Sentence("Boy, that [MASK] quickly!").calc_mask_predictions(
) # was, happened, moved
Sentence("The man who passes the sentence should swing the [MASK]."
).calc_mask_predictions() # chair, bail, stick, wheel
Sentence("A Lannister always pays his [MASK].").calc_mask_predictions(
) # debts, taxes, way
Sentence("This is the [MASK].").calc_mask_predictions() # end, truth, way
Sentence("That's what I do: I [MASK] and I know things."
).calc_mask_predictions() # see, think, look
def calc_perplexity(self):
if self.perplexity is None:
self.perplexity = Sentence(self.sentence).calc_perplexity()
return self.perplexity
def marked(self):
ps = Sentence(self.sentence).words
bold_indices = list(zip(*self.changes))[0]
for i, wi in enumerate(bold_indices):
ps[wi] = f"#{i + 1}#={ps[wi]}"
return " ".join(ps)
def generate_gradient_highlights():
text = Sentence(ex)
word_gradients = np.array(text.calc_gradients(y_prime))
word_gradients /= np.linalg.norm(word_gradients)
wgn = np.interp(word_gradients,
(np.min(word_gradients), np.max(word_gradients)), (0., 1.))
"""
\\newcommand{\\reducedstrut}{\\vrule width 0pt height .9\\ht\\strutbox depth .9\\dp\\strutbox\\relax}
\\newcommand{\\mycb}[3]{%
\\begingroup
\\setlength{\\fboxsep}{0pt}%
\\colorbox[rgb]{#1}{ \\strut \\textcolor[rgb]{#2}{#3} }%
\\endgroup
}
"""
result = "" # new command overwritten error
for cmap in [
# matplotlib.colors.LinearSegmentedColormap.from_list("", ["blue", "white", "red"]),
# matplotlib.colors.LinearSegmentedColormap.from_list("", ["white", "forestgreen"]),
# matplotlib.colors.LinearSegmentedColormap.from_list("", ["white", "orangered"]),
# matplotlib.colors.LinearSegmentedColormap.from_list("", ["white", "crimson"]),
matplotlib.colors.LinearSegmentedColormap.from_list(
"", ["white", "blue"]),
# # matplotlib.colors.LinearSegmentedColormap.from_list("", ["white", "red"]),
# # matplotlib.colors.LinearSegmentedColormap.from_list("", ["white", "black"]),
]:
result += f""
for ind, w in enumerate(text.words):
ctpl = cmap(wgn[ind])[:3]
tc = str(text_color(ctpl))[1:-1]
ctpl = [round(v, 3) for v in ctpl]
rgba = str(ctpl)[1:-1]
result += f"\\mycb{{{rgba}}}{{{tc}}}{{{latex_escape(w)}}}\\allowbreak"
result += f"\n\\\\ Top 10: {', '.join(np.array(text.words)[np.argsort(-wgn)][:10])}\\ \n\n\n"
# Sentence-wise calc gradients
sw_map = get_sentence_word_mapping(text.text)
edit_sentence_order = calc_sentence_edit_schedule(
Query(y_prime), sw_map, text)
for enm_si, si in enumerate(edit_sentence_order):
start, stop = sw_map[si]
sub = model_config.tokenizer.clean_up_tokenization(" ".join(
text.words[start:stop + 1]))
subtext = Sentence(sub)
word_gradients = np.array(subtext.calc_gradients(y_prime))
word_gradients /= np.linalg.norm(word_gradients)
wgn = np.interp(word_gradients,
(np.min(word_gradients), np.max(word_gradients)),
(0., 1.))
result += f"{enm_si + 1} Satz (vorher {si + 1}. Satz): "
for ind, w in enumerate(subtext.words):
ctpl = cmap(wgn[ind])[:3]
tc = str(text_color(ctpl))[1:-1]
ctpl = [round(v, 3) for v in ctpl]
rgba = str(ctpl)[1:-1]
result += f"\\mycb{{{rgba}}}{{{tc}}}{{{latex_escape(w)}}}\\allowbreak"
result += "\\\\ \n\n"
return result
dataset, data = list(pickle.load(file).items())[0]
model_config.load(dataset, "gpt2")
json_datapoint = None
result: Result
for another_idx, (ds_idx, result) in enumerate(tqdm(data)):
max_ds_idx = max(ds_idx, max_ds_idx)
if last_ds_idx != ds_idx:
if another_idx != 0:
json_all.append(json_datapoint)
last_ds_idx = ds_idx
fg, bg = extract_colors(result, per_sentence=False)
fg_ps, bg_ps = extract_colors(result, per_sentence=True)
json_datapoint = {
'sentence':
Sentence(clean(result.stats.original_sentence)).words,
'foreground':
fg,
'foreground_per_sen':
fg_ps,
'background':
bg,
'background_per_sen':
bg_ps,
'original_cls':
result.stats.original_classification,
'wanted_cls':
result.query.wanted_cls,
'original_ppl':
np.round(
Sentence(result.stats.original_sentence).
# text = "if you enjoy more thoughtful comedies with interesting conflicted characters ; this one is for you ."
# text = "no place for this story to go but down"
# text = "minority report is exactly what the title indicates , a report ."
# text = "it 's refreshing to see a girl-power movie that does n't feel it has to prove anything ."
# text = "it has all the excitement of eating oatmeal ."
# text = "his healthy sense of satire is light and fun ..."
# text = "Ultimately feels empty and unsatisfying, like swallowing a Communion wafer without the wine."
# text = "the action sequences are fun and reminiscent of combat scenes from the star wars series ."
# text = "with jump cuts , fast editing and lots of pyrotechnics , yu clearly hopes to camouflage how bad his movie is ."
# text = "why make a documentary about these marginal historical figures ?"
# text = "the character of zigzag is not sufficiently developed to support a film constructed around him ."
# text = "watchable up until the point where the situations and the dialogue spin hopelessly out of control"
text = model_config.tokenizer.clean_up_tokenization(text)
s = Sentence(text)
result = _gen_cf_ex(
text,
Query(wanted_cls=[0., 1.],
max_delta=0.4,
num_needed=5,
consider_max_words=500,
consider_top_k=15))
print(result.info())
result = [lst[0] for lst in result.examples]
data = {i: get_scatter_data(i) for i in range(len(s.words))}
colors = ["red", "green", "orange", "magenta", "lawngreen"]
# cmap_scale = cm.ScalarMappable(norm=mpl.colors.Normalize(vmin=0, vmax=len(result)), cmap=cm.gist_rainbow)
fig = plt.figure(figsize=(10, 14))
def generate_merged_examples(original, examples: List[Example], query,
curr_iter_idx):
"""
Parameters
----------
original : Sentence
examples : List of Example
query: search_helper.classes.Query.Query
curr_iter_idx: int
Returns
-------
list of merged examples
"""
generated = []
cs = query.c if isinstance(query.c, list) else [query.c]
for c in cs:
grouped_by_changed_indices = defaultdict(list)
examples = examples_sorted(examples, query.wanted_cls,
c)[:MAX_EXAMPLES_TO_CONSIDER_FOR_MERGING]
for e in examples:
if len(grouped_by_changed_indices[tuple(
e.changed_word_indices())]) <= 5:
grouped_by_changed_indices[tuple(
e.changed_word_indices())].append(e)
grouped_items = grouped_by_changed_indices.items()
grouped_keys = [set(x) for x in grouped_by_changed_indices.keys()]
if len(grouped_items) <= 1: # -> nothing to merge
return []
if curr_iter_idx == -1:
return [merge_all_changes_into_one(grouped_items, original)]
debug(f"MERGING ({len(grouped_keys)}) {grouped_keys}")
for a_idx, (changes_a, la) in enumerate(grouped_items):
for b_idx, (changes_b, lb) in enumerate(grouped_items):
# changes already done or changes overlap or changes too close together
if set(changes_a + changes_b) in grouped_keys \
or len(set(changes_a).intersection(set(changes_b))) > 0 \
or min_distance(changes_a, changes_b) < MIN_DISTANCE_BETWEEN_CHANGED_WORDS:
continue
resulting_edit_size = len(set(changes_a + changes_b))
take_n = max(1, 5 - resulting_edit_size)
# Merge examples
a: Example
b: Example
for a in la[:take_n]:
for b in lb[:take_n]:
sa, sb = Sentence(a), Sentence(b)
new = list(original.words)
for na in a.changed_word_indices():
new[na] = sa.words[na]
for na in b.changed_word_indices():
new[na] = sb.words[na]
generated.append(
(" ".join(new), tuple(a.changes + b.changes),
a.prediction_indices + b.prediction_indices,
a.schedule_indices + b.schedule_indices,
a.sentence_indices + a.sentence_indices))
grouped_keys.append(
set(dict(a.changes +
b.changes).keys())) # can indent 2?
debug(f"MERGE: generated {len(generated)} MERGED (#c={len(cs)}) examples")
if len(generated) == 0:
return []
unzipped = list(zip(*generated))
sentence_list = list(unzipped[0])
if len(sentence_list) == 0:
return []
sentiment_batch = calc_sentiment_batch(sentence_list)
result = []
for i, (sen, changes, pred_idx, sched_idx,
sent_idx) in enumerate(generated):
example = Example(sen,
sentiment_batch[i],
list(changes),
pred_ind=pred_idx,
sched_ind=sched_idx,
sent_ind=sent_idx)
example.schedule_indices = sched_idx
result.append(example)
return result
def expand_sentence_neighbour(text_p: Union[str, Example], query,
word_indices: List[int],
schedule_idx) -> List[Example]:
"""
Parameters
----------
schedule_idx
text_p
word_indices
query : search_helper.classes.Query.Query
Returns
-------
"""
if isinstance(text_p, Example):
text = text_p.sentence
else:
text = text_p
text = Sentence(text)
word_indices = list(word_indices)
word_indices = [
wi for wi in word_indices
if not all([s in ":,;.*" for s in text.words[wi]])
]
if len(word_indices) == 0:
return []
word_indices = word_indices[:MBS_DEPTH]
masked_text = Sentence(text.get_with_masked(word_indices))
initial_example = [Example(masked_text.text, [], [], [], [], [])]
results = []
for word_idx in word_indices:
debug(f"expand neighbours: {word_idx} of {word_indices}")
tmp_results = examples_sorted(results, query.wanted_cls,
query.c)[:MBS_BEAMS]
results = []
for interm_example in (initial_example if word_idx == word_indices[0]
else tmp_results):
intermediate_sen = Sentence(interm_example)
predictions = Sentence(interm_example).calc_mask_predictions(
query.consider_max_words)
if not predictions[word_idx]:
continue
sentences = []
for predicted_token, score in predictions[word_idx]:
new_sen = intermediate_sen.replace_mask(
word_idx, predicted_token)
sentences.append(new_sen)
classification = calc_sentiment_batch(sentences)
for i, (predicted_token,
score) in enumerate(predictions[word_idx]):
results.append(
Example(sentences[i],
classification[i],
interm_example.changes + [(word_idx, score)],
pred_ind=interm_example.prediction_indices + [i],
sched_ind=interm_example.schedule_indices +
[schedule_idx],
sent_ind=[0]))
return examples_sorted(results, query.wanted_cls, query.c)