Spaces:
Runtime error
Runtime error
Updates
Browse files
app.py
CHANGED
@@ -5,79 +5,72 @@ from model2vec import StaticModel
|
|
5 |
from reach import Reach
|
6 |
from difflib import ndiff
|
7 |
|
8 |
-
# Load the model
|
9 |
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
10 |
|
11 |
-
# Default
|
12 |
-
|
13 |
-
|
14 |
-
default_dataset2_name = "sst2"
|
15 |
-
default_dataset2_split = "validation"
|
16 |
default_text_column = "sentence"
|
17 |
default_threshold = 0.9
|
18 |
|
19 |
-
# Load the default datasets at startup
|
20 |
-
ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
21 |
-
ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
22 |
-
|
23 |
def batch_iterable(iterable, batch_size):
|
24 |
-
"""
|
25 |
for i in range(0, len(iterable), batch_size):
|
26 |
yield iterable[i:i + batch_size]
|
27 |
|
28 |
-
def compute_embeddings(texts, batch_size, progress, desc
|
|
|
29 |
embeddings = []
|
30 |
total_batches = (len(texts) + batch_size - 1) // batch_size
|
31 |
for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
32 |
-
|
33 |
-
embeddings.append(batch_embeddings)
|
34 |
progress((i + 1) / total_batches, desc=desc)
|
35 |
return np.concatenate(embeddings, axis=0)
|
36 |
|
37 |
-
def
|
38 |
-
|
39 |
-
|
|
|
40 |
batch_size: int = 1024,
|
41 |
progress=None
|
42 |
-
)
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
progress.tqdm(results, desc="Processing duplicates", total=
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
70 |
-
|
71 |
-
for sim_idx in similar_indices:
|
72 |
-
if sim_idx in deduplicated_indices:
|
73 |
-
deduplicated_indices.remove(sim_idx)
|
74 |
-
duplicate_to_original_mapping[sim_idx] = i
|
75 |
-
|
76 |
-
return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
77 |
|
78 |
def display_word_differences(x: str, y: str) -> str:
|
|
|
79 |
diff = ndiff(x.split(), y.split())
|
80 |
-
return " ".join(
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
def perform_deduplication(
|
83 |
deduplication_type,
|
@@ -91,208 +84,86 @@ def perform_deduplication(
|
|
91 |
progress=gr.Progress(track_tqdm=True),
|
92 |
):
|
93 |
try:
|
94 |
-
# Convert threshold to float
|
95 |
threshold = float(threshold)
|
96 |
|
97 |
-
#
|
98 |
-
|
|
|
|
|
|
|
99 |
|
100 |
if deduplication_type == "Single dataset":
|
101 |
-
#
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
dataset1_name == default_dataset1_name
|
106 |
-
and dataset1_split == default_dataset1_split
|
107 |
-
):
|
108 |
-
ds = ds_default1
|
109 |
-
else:
|
110 |
-
ds = load_dataset(dataset1_name, split=dataset1_split)
|
111 |
-
|
112 |
-
# Extract texts
|
113 |
-
status = "Extracting texts from Dataset 1..."
|
114 |
-
yield status, ""
|
115 |
-
texts = [example[dataset1_text_column] for example in ds]
|
116 |
-
|
117 |
-
# Compute embeddings
|
118 |
-
status = "Computing embeddings for Dataset 1..."
|
119 |
-
yield status, ""
|
120 |
-
embedding_matrix = compute_embeddings(
|
121 |
-
texts,
|
122 |
-
batch_size=64,
|
123 |
-
progress=progress,
|
124 |
-
desc="Computing embeddings for Dataset 1",
|
125 |
-
)
|
126 |
-
|
127 |
-
# Deduplicate
|
128 |
-
status = "Deduplicating embeddings..."
|
129 |
-
yield status, ""
|
130 |
-
deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
131 |
-
embedding_matrix, threshold, progress=progress
|
132 |
)
|
133 |
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
result_text = f"**Total documents:** {num_total}\n"
|
140 |
-
result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
141 |
-
result_text += (
|
142 |
-
f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
143 |
)
|
144 |
|
145 |
-
# Show deduplicated examples
|
146 |
if num_duplicates > 0:
|
147 |
-
result_text += "**
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
|
|
157 |
else:
|
158 |
result_text += "No duplicates found."
|
159 |
|
160 |
-
|
161 |
-
status = "Deduplication completed."
|
162 |
-
yield status, result_text
|
163 |
-
|
164 |
-
elif deduplication_type == "Cross-dataset":
|
165 |
-
# Similar code for cross-dataset deduplication
|
166 |
-
# Load Dataset 1
|
167 |
-
status = "Loading Dataset 1..."
|
168 |
-
yield status, ""
|
169 |
-
if (
|
170 |
-
dataset1_name == default_dataset1_name
|
171 |
-
and dataset1_split == default_dataset1_split
|
172 |
-
):
|
173 |
-
ds1 = ds_default1
|
174 |
-
else:
|
175 |
-
ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
176 |
-
|
177 |
-
# Load Dataset 2
|
178 |
-
status = "Loading Dataset 2..."
|
179 |
-
yield status, ""
|
180 |
-
if (
|
181 |
-
dataset2_name == default_dataset2_name
|
182 |
-
and dataset2_split == default_dataset2_split
|
183 |
-
):
|
184 |
-
ds2 = ds_default2
|
185 |
-
else:
|
186 |
-
ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
187 |
-
|
188 |
-
# Extract texts from Dataset 1
|
189 |
-
status = "Extracting texts from Dataset 1..."
|
190 |
-
yield status, ""
|
191 |
-
texts1 = [example[dataset1_text_column] for example in ds1]
|
192 |
-
|
193 |
-
# Extract texts from Dataset 2
|
194 |
-
status = "Extracting texts from Dataset 2..."
|
195 |
-
yield status, ""
|
196 |
-
texts2 = [example[dataset2_text_column] for example in ds2]
|
197 |
-
|
198 |
-
# Compute embeddings for Dataset 1
|
199 |
-
status = "Computing embeddings for Dataset 1..."
|
200 |
-
yield status, ""
|
201 |
-
embedding_matrix1 = compute_embeddings(
|
202 |
-
texts1,
|
203 |
-
batch_size=64,
|
204 |
-
progress=progress,
|
205 |
-
desc="Computing embeddings for Dataset 1",
|
206 |
-
)
|
207 |
|
208 |
-
|
209 |
-
|
210 |
-
yield
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
|
|
|
|
|
|
|
216 |
)
|
217 |
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
)
|
224 |
|
225 |
-
num_duplicates = len(duplicate_indices_in_ds2)
|
226 |
-
num_total_ds2 = len(texts2)
|
227 |
-
num_unique_ds2 = num_total_ds2 - num_duplicates
|
228 |
-
|
229 |
-
result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
230 |
-
result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
231 |
-
result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
232 |
-
|
233 |
-
# Show deduplicated examples
|
234 |
if num_duplicates > 0:
|
235 |
-
result_text += "**
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
else:
|
247 |
result_text += "No duplicates found."
|
248 |
|
249 |
-
|
250 |
-
status = "Deduplication completed."
|
251 |
-
yield status, result_text
|
252 |
|
253 |
except Exception as e:
|
254 |
yield f"An error occurred: {e}", ""
|
255 |
raise e
|
256 |
|
257 |
-
def deduplicate_across_datasets(
|
258 |
-
embedding_matrix_1: np.ndarray,
|
259 |
-
embedding_matrix_2: np.ndarray,
|
260 |
-
threshold: float,
|
261 |
-
batch_size: int = 1024,
|
262 |
-
progress=None
|
263 |
-
) -> tuple[list[int], dict[int, int]]:
|
264 |
-
# Building the index from Dataset 1
|
265 |
-
progress(0, desc="Building search index from Dataset 1...")
|
266 |
-
reach = Reach(
|
267 |
-
vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))]
|
268 |
-
)
|
269 |
-
|
270 |
-
duplicate_indices_in_test = []
|
271 |
-
duplicate_to_original_mapping = {}
|
272 |
-
|
273 |
-
# Finding nearest neighbors between datasets
|
274 |
-
progress(0, desc="Finding nearest neighbors between datasets...")
|
275 |
-
results = reach.nearest_neighbor_threshold(
|
276 |
-
embedding_matrix_2,
|
277 |
-
threshold=threshold,
|
278 |
-
batch_size=batch_size,
|
279 |
-
show_progressbar=False, # Disable internal progress bar
|
280 |
-
)
|
281 |
-
|
282 |
-
total_items = len(embedding_matrix_2)
|
283 |
-
# Processing duplicates with a progress bar
|
284 |
-
for i, similar_items in enumerate(
|
285 |
-
progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)
|
286 |
-
):
|
287 |
-
similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
288 |
-
|
289 |
-
if similar_indices:
|
290 |
-
duplicate_indices_in_test.append(i)
|
291 |
-
duplicate_to_original_mapping[i] = similar_indices[0]
|
292 |
-
|
293 |
-
return duplicate_indices_in_test, duplicate_to_original_mapping
|
294 |
-
|
295 |
-
# Adjust the height of the status_output component using custom CSS
|
296 |
with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
297 |
gr.Markdown("# Semantic Deduplication")
|
298 |
|
@@ -303,38 +174,27 @@ with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
|
303 |
)
|
304 |
|
305 |
with gr.Row():
|
306 |
-
dataset1_name = gr.Textbox(value=
|
307 |
-
dataset1_split = gr.Textbox(value=
|
308 |
dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
309 |
|
310 |
dataset2_inputs = gr.Column(visible=False)
|
311 |
with dataset2_inputs:
|
312 |
gr.Markdown("### Dataset 2")
|
313 |
with gr.Row():
|
314 |
-
dataset2_name = gr.Textbox(value=
|
315 |
-
dataset2_split = gr.Textbox(value=
|
316 |
dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
317 |
|
318 |
-
threshold = gr.Slider(
|
319 |
-
minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold"
|
320 |
-
)
|
321 |
-
|
322 |
compute_button = gr.Button("Compute")
|
323 |
-
|
324 |
-
# Use 'gr.Markdown' with 'elem_id' and custom CSS to adjust height
|
325 |
status_output = gr.Markdown(elem_id="status_output")
|
326 |
result_output = gr.Markdown()
|
327 |
|
328 |
-
|
329 |
-
|
330 |
-
if deduplication_type_value == "Cross-dataset":
|
331 |
-
return gr.update(visible=True)
|
332 |
-
else:
|
333 |
-
return gr.update(visible=False)
|
334 |
|
335 |
-
deduplication_type.change(
|
336 |
-
update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
337 |
-
)
|
338 |
|
339 |
compute_button.click(
|
340 |
fn=perform_deduplication,
|
@@ -353,19 +213,17 @@ with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
|
353 |
|
354 |
demo.launch()
|
355 |
|
356 |
-
|
357 |
# import gradio as gr
|
358 |
# from datasets import load_dataset
|
359 |
# import numpy as np
|
360 |
# from model2vec import StaticModel
|
361 |
# from reach import Reach
|
362 |
# from difflib import ndiff
|
363 |
-
# import tqdm
|
364 |
|
365 |
# # Load the model at startup
|
366 |
# model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
367 |
|
368 |
-
# #
|
369 |
# default_dataset1_name = "sst2"
|
370 |
# default_dataset1_split = "train"
|
371 |
# default_dataset2_name = "sst2"
|
@@ -384,29 +242,42 @@ demo.launch()
|
|
384 |
|
385 |
# def compute_embeddings(texts, batch_size, progress, desc="Computing embeddings"):
|
386 |
# embeddings = []
|
387 |
-
#
|
388 |
-
#
|
|
|
389 |
# embeddings.append(batch_embeddings)
|
|
|
390 |
# return np.concatenate(embeddings, axis=0)
|
391 |
|
392 |
-
# def deduplicate(
|
393 |
-
#
|
394 |
-
#
|
395 |
-
#
|
396 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
397 |
|
398 |
# deduplicated_indices = set(range(len(embedding_matrix)))
|
399 |
# duplicate_to_original_mapping = {}
|
400 |
|
|
|
|
|
401 |
# results = reach.nearest_neighbor_threshold(
|
402 |
# embedding_matrix,
|
403 |
# threshold=threshold,
|
404 |
# batch_size=batch_size,
|
405 |
-
# show_progressbar=False
|
406 |
# )
|
407 |
|
|
|
408 |
# total_items = len(embedding_matrix)
|
409 |
-
# for i, similar_items in enumerate(
|
|
|
|
|
410 |
# if i not in deduplicated_indices:
|
411 |
# continue
|
412 |
|
@@ -419,35 +290,9 @@ demo.launch()
|
|
419 |
|
420 |
# return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
421 |
|
422 |
-
# def deduplicate_across_datasets(embedding_matrix_1: np.ndarray, embedding_matrix_2: np.ndarray, threshold: float, batch_size: int = 1024, progress=None) -> tuple[list[int], dict[int, int]]:
|
423 |
-
# """
|
424 |
-
# Deduplicate embeddings across two datasets and return the indices of duplicates between them.
|
425 |
-
# """
|
426 |
-
# reach = Reach(vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))])
|
427 |
-
|
428 |
-
# duplicate_indices_in_test = []
|
429 |
-
# duplicate_to_original_mapping = {}
|
430 |
-
|
431 |
-
# results = reach.nearest_neighbor_threshold(
|
432 |
-
# embedding_matrix_2,
|
433 |
-
# threshold=threshold,
|
434 |
-
# batch_size=batch_size,
|
435 |
-
# show_progressbar=False
|
436 |
-
# )
|
437 |
-
|
438 |
-
# total_items = len(embedding_matrix_2)
|
439 |
-
# for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)):
|
440 |
-
# similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
441 |
-
|
442 |
-
# if similar_indices:
|
443 |
-
# duplicate_indices_in_test.append(i)
|
444 |
-
# duplicate_to_original_mapping[i] = similar_indices[0]
|
445 |
-
|
446 |
-
# return duplicate_indices_in_test, duplicate_to_original_mapping
|
447 |
-
|
448 |
# def display_word_differences(x: str, y: str) -> str:
|
449 |
# diff = ndiff(x.split(), y.split())
|
450 |
-
# return " ".join([word for word in diff if word.startswith((
|
451 |
|
452 |
# def perform_deduplication(
|
453 |
# deduplication_type,
|
@@ -458,26 +303,61 @@ demo.launch()
|
|
458 |
# dataset2_split="",
|
459 |
# dataset2_text_column="",
|
460 |
# threshold=default_threshold,
|
461 |
-
# progress=gr.Progress(track_tqdm=True)
|
462 |
# ):
|
463 |
# try:
|
|
|
464 |
# threshold = float(threshold)
|
465 |
|
|
|
|
|
|
|
466 |
# if deduplication_type == "Single dataset":
|
467 |
-
#
|
468 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
469 |
|
470 |
-
#
|
471 |
-
#
|
|
|
|
|
472 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
473 |
# num_duplicates = len(duplicate_to_original_mapping)
|
474 |
# num_total = len(texts)
|
475 |
# num_deduplicated = len(deduplicated_indices)
|
476 |
|
477 |
# result_text = f"**Total documents:** {num_total}\n"
|
478 |
# result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
479 |
-
# result_text +=
|
|
|
|
|
480 |
|
|
|
481 |
# if num_duplicates > 0:
|
482 |
# result_text += "**Examples of duplicates found:**\n\n"
|
483 |
# num_examples = min(5, num_duplicates)
|
@@ -492,19 +372,70 @@ demo.launch()
|
|
492 |
# else:
|
493 |
# result_text += "No duplicates found."
|
494 |
|
495 |
-
#
|
|
|
|
|
496 |
|
497 |
# elif deduplication_type == "Cross-dataset":
|
498 |
-
#
|
499 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
500 |
|
|
|
|
|
|
|
501 |
# texts1 = [example[dataset1_text_column] for example in ds1]
|
502 |
-
# texts2 = [example[dataset2_text_column] for example in ds2]
|
503 |
|
504 |
-
#
|
505 |
-
#
|
|
|
|
|
506 |
|
507 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
508 |
|
509 |
# num_duplicates = len(duplicate_indices_in_ds2)
|
510 |
# num_total_ds2 = len(texts2)
|
@@ -514,6 +445,7 @@ demo.launch()
|
|
514 |
# result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
515 |
# result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
516 |
|
|
|
517 |
# if num_duplicates > 0:
|
518 |
# result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
519 |
# num_examples = min(5, num_duplicates)
|
@@ -529,19 +461,60 @@ demo.launch()
|
|
529 |
# else:
|
530 |
# result_text += "No duplicates found."
|
531 |
|
532 |
-
#
|
|
|
|
|
533 |
|
534 |
# except Exception as e:
|
535 |
# yield f"An error occurred: {e}", ""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
536 |
|
537 |
-
#
|
538 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
539 |
# gr.Markdown("# Semantic Deduplication")
|
540 |
|
541 |
# deduplication_type = gr.Radio(
|
542 |
# choices=["Single dataset", "Cross-dataset"],
|
543 |
# label="Deduplication Type",
|
544 |
-
# value="Single dataset"
|
545 |
# )
|
546 |
|
547 |
# with gr.Row():
|
@@ -558,17 +531,16 @@ demo.launch()
|
|
558 |
# dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
559 |
|
560 |
# threshold = gr.Slider(
|
561 |
-
# minimum=0.0,
|
562 |
-
# maximum=1.0,
|
563 |
-
# value=default_threshold,
|
564 |
-
# label="Similarity Threshold"
|
565 |
# )
|
566 |
|
567 |
# compute_button = gr.Button("Compute")
|
568 |
|
|
|
569 |
# status_output = gr.Markdown(elem_id="status_output")
|
570 |
-
# result_output = gr.Markdown(
|
571 |
|
|
|
572 |
# def update_visibility(deduplication_type_value):
|
573 |
# if deduplication_type_value == "Cross-dataset":
|
574 |
# return gr.update(visible=True)
|
@@ -576,9 +548,7 @@ demo.launch()
|
|
576 |
# return gr.update(visible=False)
|
577 |
|
578 |
# deduplication_type.change(
|
579 |
-
# update_visibility,
|
580 |
-
# inputs=deduplication_type,
|
581 |
-
# outputs=dataset2_inputs
|
582 |
# )
|
583 |
|
584 |
# compute_button.click(
|
@@ -591,1115 +561,9 @@ demo.launch()
|
|
591 |
# dataset2_name,
|
592 |
# dataset2_split,
|
593 |
# dataset2_text_column,
|
594 |
-
# threshold
|
595 |
# ],
|
596 |
-
# outputs=[status_output, result_output]
|
597 |
# )
|
598 |
|
599 |
# demo.launch()
|
600 |
-
|
601 |
-
# # import gradio as gr
|
602 |
-
# # from datasets import load_dataset
|
603 |
-
# # import numpy as np
|
604 |
-
# # from model2vec import StaticModel
|
605 |
-
# # from reach import Reach
|
606 |
-
# # from difflib import ndiff
|
607 |
-
# # import tqdm
|
608 |
-
|
609 |
-
# # # Load the model at startup
|
610 |
-
# # model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
611 |
-
|
612 |
-
# # # Update default dataset to 'sst2' and set default threshold to 0.9
|
613 |
-
# # default_dataset1_name = "sst2"
|
614 |
-
# # default_dataset1_split = "train"
|
615 |
-
# # default_dataset2_name = "sst2"
|
616 |
-
# # default_dataset2_split = "validation"
|
617 |
-
# # default_text_column = "sentence"
|
618 |
-
# # default_threshold = 0.9
|
619 |
-
|
620 |
-
# # # Load the default datasets at startup
|
621 |
-
# # ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
622 |
-
# # ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
623 |
-
|
624 |
-
# # def batch_iterable(iterable, batch_size):
|
625 |
-
# # """Helper function to create batches from an iterable."""
|
626 |
-
# # for i in range(0, len(iterable), batch_size):
|
627 |
-
# # yield iterable[i:i + batch_size]
|
628 |
-
|
629 |
-
# # def compute_embeddings(texts, batch_size, progress, desc="Computing embeddings"):
|
630 |
-
# # embeddings = []
|
631 |
-
# # for batch in progress.tqdm(batch_iterable(texts, batch_size), total=(len(texts) + batch_size - 1) // batch_size, desc=desc):
|
632 |
-
# # batch_embeddings = model.encode(batch, show_progressbar=False)
|
633 |
-
# # embeddings.append(batch_embeddings)
|
634 |
-
# # return np.concatenate(embeddings, axis=0)
|
635 |
-
|
636 |
-
# # def deduplicate(embedding_matrix: np.ndarray, threshold: float, batch_size: int = 1024, progress=None) -> tuple[np.ndarray, dict[int, int]]:
|
637 |
-
# # """
|
638 |
-
# # Deduplicate embeddings and return the deduplicated indices and a mapping of removed indices to their corresponding original indices.
|
639 |
-
# # """
|
640 |
-
# # # Building the index
|
641 |
-
# # progress(0, desc="Building search index...")
|
642 |
-
# # reach = Reach(vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))])
|
643 |
-
|
644 |
-
# # deduplicated_indices = set(range(len(embedding_matrix)))
|
645 |
-
# # duplicate_to_original_mapping = {}
|
646 |
-
|
647 |
-
# # # Finding nearest neighbors
|
648 |
-
# # progress(0, desc="Finding nearest neighbors...")
|
649 |
-
# # results = reach.nearest_neighbor_threshold(
|
650 |
-
# # embedding_matrix,
|
651 |
-
# # threshold=threshold,
|
652 |
-
# # batch_size=batch_size,
|
653 |
-
# # show_progressbar=False # Disable internal progress bar
|
654 |
-
# # )
|
655 |
-
|
656 |
-
# # # Processing duplicates with a progress bar
|
657 |
-
# # total_items = len(embedding_matrix)
|
658 |
-
# # for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=total_items)):
|
659 |
-
# # if i not in deduplicated_indices:
|
660 |
-
# # continue
|
661 |
-
|
662 |
-
# # similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
663 |
-
|
664 |
-
# # for sim_idx in similar_indices:
|
665 |
-
# # if sim_idx in deduplicated_indices:
|
666 |
-
# # deduplicated_indices.remove(sim_idx)
|
667 |
-
# # duplicate_to_original_mapping[sim_idx] = i
|
668 |
-
|
669 |
-
# # return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
670 |
-
|
671 |
-
# # def deduplicate_across_datasets(embedding_matrix_1: np.ndarray, embedding_matrix_2: np.ndarray, threshold: float, batch_size: int = 1024, progress=None) -> tuple[list[int], dict[int, int]]:
|
672 |
-
# # """
|
673 |
-
# # Deduplicate embeddings across two datasets and return the indices of duplicates between them.
|
674 |
-
# # """
|
675 |
-
# # # Building the index from Dataset 1
|
676 |
-
# # progress(0, desc="Building search index from Dataset 1...")
|
677 |
-
# # reach = Reach(vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))])
|
678 |
-
|
679 |
-
# # duplicate_indices_in_test = []
|
680 |
-
# # duplicate_to_original_mapping = {}
|
681 |
-
|
682 |
-
# # # Finding nearest neighbors between datasets
|
683 |
-
# # progress(0, desc="Finding nearest neighbors between datasets...")
|
684 |
-
# # results = reach.nearest_neighbor_threshold(
|
685 |
-
# # embedding_matrix_2,
|
686 |
-
# # threshold=threshold,
|
687 |
-
# # batch_size=batch_size,
|
688 |
-
# # show_progressbar=False # Disable internal progress bar
|
689 |
-
# # )
|
690 |
-
|
691 |
-
# # total_items = len(embedding_matrix_2)
|
692 |
-
# # # Processing duplicates with a progress bar
|
693 |
-
# # for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)):
|
694 |
-
# # similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
695 |
-
|
696 |
-
# # if similar_indices:
|
697 |
-
# # duplicate_indices_in_test.append(i)
|
698 |
-
# # duplicate_to_original_mapping[i] = similar_indices[0]
|
699 |
-
|
700 |
-
# # return duplicate_indices_in_test, duplicate_to_original_mapping
|
701 |
-
|
702 |
-
# # def display_word_differences(x: str, y: str) -> str:
|
703 |
-
# # diff = ndiff(x.split(), y.split())
|
704 |
-
# # return " ".join([word for word in diff if word.startswith(('+', '-'))])
|
705 |
-
|
706 |
-
# # def perform_deduplication(
|
707 |
-
# # deduplication_type,
|
708 |
-
# # dataset1_name,
|
709 |
-
# # dataset1_split,
|
710 |
-
# # dataset1_text_column,
|
711 |
-
# # dataset2_name="",
|
712 |
-
# # dataset2_split="",
|
713 |
-
# # dataset2_text_column="",
|
714 |
-
# # threshold=default_threshold,
|
715 |
-
# # progress=gr.Progress(track_tqdm=True)
|
716 |
-
# # ):
|
717 |
-
# # try:
|
718 |
-
# # # Convert threshold to float
|
719 |
-
# # threshold = float(threshold)
|
720 |
-
|
721 |
-
# # # Initialize status message
|
722 |
-
# # status = ""
|
723 |
-
|
724 |
-
# # if deduplication_type == "Single dataset":
|
725 |
-
# # # Load Dataset 1
|
726 |
-
# # status = "Loading Dataset 1..."
|
727 |
-
# # yield status, ""
|
728 |
-
# # if dataset1_name == default_dataset1_name and dataset1_split == default_dataset1_split:
|
729 |
-
# # ds = ds_default1
|
730 |
-
# # else:
|
731 |
-
# # ds = load_dataset(dataset1_name, split=dataset1_split)
|
732 |
-
|
733 |
-
# # # Extract texts
|
734 |
-
# # status = "Extracting texts from Dataset 1..."
|
735 |
-
# # yield status, ""
|
736 |
-
# # texts = [example[dataset1_text_column] for example in ds]
|
737 |
-
|
738 |
-
# # # Compute embeddings
|
739 |
-
# # status = "Computing embeddings for Dataset 1..."
|
740 |
-
# # yield status, ""
|
741 |
-
# # embedding_matrix = compute_embeddings(texts, batch_size=64, progress=progress, desc="Computing embeddings for Dataset 1")
|
742 |
-
|
743 |
-
# # # Deduplicate
|
744 |
-
# # status = "Deduplicating embeddings..."
|
745 |
-
# # yield status, ""
|
746 |
-
# # deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
747 |
-
# # embedding_matrix, threshold, progress=progress
|
748 |
-
# # )
|
749 |
-
|
750 |
-
# # # Prepare the results
|
751 |
-
# # num_duplicates = len(duplicate_to_original_mapping)
|
752 |
-
# # num_total = len(texts)
|
753 |
-
# # num_deduplicated = len(deduplicated_indices)
|
754 |
-
|
755 |
-
# # result_text = f"**Total documents:** {num_total}\n"
|
756 |
-
# # result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
757 |
-
# # result_text += f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
758 |
-
|
759 |
-
# # # Show deduplicated examples
|
760 |
-
# # if num_duplicates > 0:
|
761 |
-
# # result_text += "**Examples of duplicates found:**\n\n"
|
762 |
-
# # num_examples = min(5, num_duplicates)
|
763 |
-
# # for duplicate_idx, original_idx in list(duplicate_to_original_mapping.items())[:num_examples]:
|
764 |
-
# # original_text = texts[original_idx]
|
765 |
-
# # duplicate_text = texts[duplicate_idx]
|
766 |
-
# # differences = display_word_differences(original_text, duplicate_text)
|
767 |
-
# # result_text += f"**Original text:**\n{original_text}\n\n"
|
768 |
-
# # result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
|
769 |
-
# # result_text += f"**Differences:**\n{differences}\n"
|
770 |
-
# # result_text += "-" * 50 + "\n\n"
|
771 |
-
# # else:
|
772 |
-
# # result_text += "No duplicates found."
|
773 |
-
|
774 |
-
# # # Final status
|
775 |
-
# # status = "Deduplication completed."
|
776 |
-
# # yield status, result_text
|
777 |
-
|
778 |
-
# # elif deduplication_type == "Cross-dataset":
|
779 |
-
# # # Load Dataset 1
|
780 |
-
# # status = "Loading Dataset 1..."
|
781 |
-
# # yield status, ""
|
782 |
-
# # if dataset1_name == default_dataset1_name and dataset1_split == default_dataset1_split:
|
783 |
-
# # ds1 = ds_default1
|
784 |
-
# # else:
|
785 |
-
# # ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
786 |
-
|
787 |
-
# # # Load Dataset 2
|
788 |
-
# # status = "Loading Dataset 2..."
|
789 |
-
# # yield status, ""
|
790 |
-
# # if dataset2_name == default_dataset2_name and dataset2_split == default_dataset2_split:
|
791 |
-
# # ds2 = ds_default2
|
792 |
-
# # else:
|
793 |
-
# # ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
794 |
-
|
795 |
-
# # # Extract texts from Dataset 1
|
796 |
-
# # status = "Extracting texts from Dataset 1..."
|
797 |
-
# # yield status, ""
|
798 |
-
# # texts1 = [example[dataset1_text_column] for example in ds1]
|
799 |
-
|
800 |
-
# # # Extract texts from Dataset 2
|
801 |
-
# # status = "Extracting texts from Dataset 2..."
|
802 |
-
# # yield status, ""
|
803 |
-
# # texts2 = [example[dataset2_text_column] for example in ds2]
|
804 |
-
|
805 |
-
# # # Compute embeddings for Dataset 1
|
806 |
-
# # status = "Computing embeddings for Dataset 1..."
|
807 |
-
# # yield status, ""
|
808 |
-
# # embedding_matrix1 = compute_embeddings(texts1, batch_size=64, progress=progress, desc="Computing embeddings for Dataset 1")
|
809 |
-
|
810 |
-
# # # Compute embeddings for Dataset 2
|
811 |
-
# # status = "Computing embeddings for Dataset 2..."
|
812 |
-
# # yield status, ""
|
813 |
-
# # embedding_matrix2 = compute_embeddings(texts2, batch_size=64, progress=progress, desc="Computing embeddings for Dataset 2")
|
814 |
-
|
815 |
-
# # # Deduplicate across datasets
|
816 |
-
# # status = "Deduplicating embeddings across datasets..."
|
817 |
-
# # yield status, ""
|
818 |
-
# # duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(
|
819 |
-
# # embedding_matrix1, embedding_matrix2, threshold, progress=progress
|
820 |
-
# # )
|
821 |
-
|
822 |
-
# # num_duplicates = len(duplicate_indices_in_ds2)
|
823 |
-
# # num_total_ds2 = len(texts2)
|
824 |
-
# # num_unique_ds2 = num_total_ds2 - num_duplicates
|
825 |
-
|
826 |
-
# # result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
827 |
-
# # result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
828 |
-
# # result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
829 |
-
|
830 |
-
# # # Show deduplicated examples
|
831 |
-
# # if num_duplicates > 0:
|
832 |
-
# # result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
833 |
-
# # num_examples = min(5, num_duplicates)
|
834 |
-
# # for duplicate_idx in duplicate_indices_in_ds2[:num_examples]:
|
835 |
-
# # original_idx = duplicate_to_original_mapping[duplicate_idx]
|
836 |
-
# # original_text = texts1[original_idx]
|
837 |
-
# # duplicate_text = texts2[duplicate_idx]
|
838 |
-
# # differences = display_word_differences(original_text, duplicate_text)
|
839 |
-
# # result_text += f"**Original text (Dataset 1):**\n{original_text}\n\n"
|
840 |
-
# # result_text += f"**Duplicate text (Dataset 2):**\n{duplicate_text}\n\n"
|
841 |
-
# # result_text += f"**Differences:**\n{differences}\n"
|
842 |
-
# # result_text += "-" * 50 + "\n\n"
|
843 |
-
# # else:
|
844 |
-
# # result_text += "No duplicates found."
|
845 |
-
|
846 |
-
# # # Final status
|
847 |
-
# # status = "Deduplication completed."
|
848 |
-
# # yield status, result_text
|
849 |
-
|
850 |
-
# # except Exception as e:
|
851 |
-
# # yield f"An error occurred: {e}", ""
|
852 |
-
# # raise e
|
853 |
-
|
854 |
-
# # with gr.Blocks() as demo:
|
855 |
-
# # gr.Markdown("# Semantic Deduplication")
|
856 |
-
|
857 |
-
# # deduplication_type = gr.Radio(
|
858 |
-
# # choices=["Single dataset", "Cross-dataset"],
|
859 |
-
# # label="Deduplication Type",
|
860 |
-
# # value="Single dataset"
|
861 |
-
# # )
|
862 |
-
|
863 |
-
# # with gr.Row():
|
864 |
-
# # dataset1_name = gr.Textbox(value=default_dataset1_name, label="Dataset 1 Name")
|
865 |
-
# # dataset1_split = gr.Textbox(value=default_dataset1_split, label="Dataset 1 Split")
|
866 |
-
# # dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
867 |
-
|
868 |
-
# # dataset2_inputs = gr.Column(visible=False)
|
869 |
-
# # with dataset2_inputs:
|
870 |
-
# # gr.Markdown("### Dataset 2")
|
871 |
-
# # with gr.Row():
|
872 |
-
# # dataset2_name = gr.Textbox(value=default_dataset2_name, label="Dataset 2 Name")
|
873 |
-
# # dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
874 |
-
# # dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
875 |
-
|
876 |
-
# # threshold = gr.Slider(
|
877 |
-
# # minimum=0.0,
|
878 |
-
# # maximum=1.0,
|
879 |
-
# # value=default_threshold,
|
880 |
-
# # label="Similarity Threshold"
|
881 |
-
# # )
|
882 |
-
|
883 |
-
# # compute_button = gr.Button("Compute")
|
884 |
-
|
885 |
-
# # status_output = gr.Markdown()
|
886 |
-
# # result_output = gr.Markdown()
|
887 |
-
|
888 |
-
# # # Function to update the visibility of dataset2_inputs
|
889 |
-
# # def update_visibility(deduplication_type_value):
|
890 |
-
# # if deduplication_type_value == "Cross-dataset":
|
891 |
-
# # return gr.update(visible=True)
|
892 |
-
# # else:
|
893 |
-
# # return gr.update(visible=False)
|
894 |
-
|
895 |
-
# # deduplication_type.change(
|
896 |
-
# # update_visibility,
|
897 |
-
# # inputs=deduplication_type,
|
898 |
-
# # outputs=dataset2_inputs
|
899 |
-
# # )
|
900 |
-
|
901 |
-
# # compute_button.click(
|
902 |
-
# # fn=perform_deduplication,
|
903 |
-
# # inputs=[
|
904 |
-
# # deduplication_type,
|
905 |
-
# # dataset1_name,
|
906 |
-
# # dataset1_split,
|
907 |
-
# # dataset1_text_column,
|
908 |
-
# # dataset2_name,
|
909 |
-
# # dataset2_split,
|
910 |
-
# # dataset2_text_column,
|
911 |
-
# # threshold
|
912 |
-
# # ],
|
913 |
-
# # outputs=[status_output, result_output]
|
914 |
-
# # )
|
915 |
-
|
916 |
-
# # demo.launch()
|
917 |
-
|
918 |
-
|
919 |
-
# # import gradio as gr
|
920 |
-
# # from datasets import load_dataset
|
921 |
-
# # import numpy as np
|
922 |
-
# # import model2vec
|
923 |
-
# # from reach import Reach
|
924 |
-
# # from difflib import ndiff
|
925 |
-
|
926 |
-
# # # Load the model at startup
|
927 |
-
# # model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
928 |
-
|
929 |
-
# # # Default dataset parameters
|
930 |
-
# # default_dataset1_name = "sst2"
|
931 |
-
# # default_dataset1_split = "train"
|
932 |
-
# # default_dataset2_name = "sst2"
|
933 |
-
# # default_dataset2_split = "validation"
|
934 |
-
# # default_text_column = "sentence"
|
935 |
-
# # default_threshold = 0.9
|
936 |
-
|
937 |
-
# # # Load the default datasets at startup
|
938 |
-
# # ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
939 |
-
# # ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
940 |
-
|
941 |
-
# # def batch_iterable(iterable, batch_size):
|
942 |
-
# # """Helper function to create batches from an iterable."""
|
943 |
-
# # for i in range(0, len(iterable), batch_size):
|
944 |
-
# # yield iterable[i:i + batch_size]
|
945 |
-
|
946 |
-
# # def compute_embeddings(texts, batch_size, progress, desc="Computing embeddings"):
|
947 |
-
# # embeddings = []
|
948 |
-
# # total_batches = (len(texts) + batch_size - 1) // batch_size
|
949 |
-
# # for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
950 |
-
# # batch_embeddings = model.encode(batch_texts, show_progressbar=False)
|
951 |
-
# # embeddings.append(batch_embeddings)
|
952 |
-
# # progress((i + 1) / total_batches, desc=desc)
|
953 |
-
# # return np.concatenate(embeddings, axis=0)
|
954 |
-
|
955 |
-
# # def deduplicate(
|
956 |
-
# # embedding_matrix: np.ndarray,
|
957 |
-
# # threshold: float,
|
958 |
-
# # batch_size: int = 1024,
|
959 |
-
# # progress=None
|
960 |
-
# # ) -> tuple[np.ndarray, dict[int, int]]:
|
961 |
-
# # reach = Reach(vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))])
|
962 |
-
|
963 |
-
# # deduplicated_indices = set(range(len(embedding_matrix)))
|
964 |
-
# # duplicate_to_original_mapping = {}
|
965 |
-
|
966 |
-
# # results = reach.nearest_neighbor_threshold(
|
967 |
-
# # embedding_matrix,
|
968 |
-
# # threshold=threshold,
|
969 |
-
# # batch_size=batch_size,
|
970 |
-
# # show_progressbar=False,
|
971 |
-
# # )
|
972 |
-
|
973 |
-
# # total_items = len(embedding_matrix)
|
974 |
-
# # for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=total_items)):
|
975 |
-
# # if i not in deduplicated_indices:
|
976 |
-
# # continue
|
977 |
-
|
978 |
-
# # similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
979 |
-
# # for sim_idx in similar_indices:
|
980 |
-
# # if sim_idx in deduplicated_indices:
|
981 |
-
# # deduplicated_indices.remove(sim_idx)
|
982 |
-
# # duplicate_to_original_mapping[sim_idx] = i
|
983 |
-
|
984 |
-
# # return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
985 |
-
|
986 |
-
# # def display_word_differences(x: str, y: str) -> str:
|
987 |
-
# # diff = ndiff(x.split(), y.split())
|
988 |
-
# # return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
989 |
-
|
990 |
-
# # def perform_deduplication(
|
991 |
-
# # deduplication_type,
|
992 |
-
# # dataset1_name,
|
993 |
-
# # dataset1_split,
|
994 |
-
# # dataset1_text_column,
|
995 |
-
# # dataset2_name="",
|
996 |
-
# # dataset2_split="",
|
997 |
-
# # dataset2_text_column="",
|
998 |
-
# # threshold=default_threshold,
|
999 |
-
# # progress=gr.Progress(track_tqdm=True),
|
1000 |
-
# # ):
|
1001 |
-
# # try:
|
1002 |
-
# # threshold = float(threshold)
|
1003 |
-
|
1004 |
-
# # if deduplication_type == "Single dataset":
|
1005 |
-
# # ds = ds_default1 if dataset1_name == default_dataset1_name and dataset1_split == default_dataset1_split else load_dataset(dataset1_name, split=dataset1_split)
|
1006 |
-
# # texts = [example[dataset1_text_column] for example in ds]
|
1007 |
-
|
1008 |
-
# # embedding_matrix = compute_embeddings(texts, batch_size=64, progress=progress)
|
1009 |
-
# # deduplicated_indices, duplicate_to_original_mapping = deduplicate(embedding_matrix, threshold, progress=progress)
|
1010 |
-
|
1011 |
-
# # num_duplicates = len(duplicate_to_original_mapping)
|
1012 |
-
# # num_total = len(texts)
|
1013 |
-
# # num_deduplicated = len(deduplicated_indices)
|
1014 |
-
|
1015 |
-
# # result_text = f"**Total documents:** {num_total}\n"
|
1016 |
-
# # result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
1017 |
-
# # result_text += f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
1018 |
-
|
1019 |
-
# # if num_duplicates > 0:
|
1020 |
-
# # result_text += "**Examples of duplicates found:**\n\n"
|
1021 |
-
# # num_examples = min(5, num_duplicates)
|
1022 |
-
# # for duplicate_idx, original_idx in list(duplicate_to_original_mapping.items())[:num_examples]:
|
1023 |
-
# # original_text = texts[original_idx]
|
1024 |
-
# # duplicate_text = texts[duplicate_idx]
|
1025 |
-
# # differences = display_word_differences(original_text, duplicate_text)
|
1026 |
-
# # result_text += f"**Original text:**\n{original_text}\n\n"
|
1027 |
-
# # result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
|
1028 |
-
# # result_text += f"**Differences:**\n{differences}\n"
|
1029 |
-
# # result_text += "-" * 50 + "\n\n"
|
1030 |
-
# # else:
|
1031 |
-
# # result_text += "No duplicates found."
|
1032 |
-
|
1033 |
-
# # yield result_text
|
1034 |
-
|
1035 |
-
# # except Exception as e:
|
1036 |
-
# # yield f"An error occurred: {e}"
|
1037 |
-
|
1038 |
-
# # # Gradio interface setup
|
1039 |
-
# # with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
1040 |
-
# # gr.Markdown("# Semantic Deduplication")
|
1041 |
-
|
1042 |
-
# # deduplication_type = gr.Radio(
|
1043 |
-
# # choices=["Single dataset", "Cross-dataset"],
|
1044 |
-
# # label="Deduplication Type",
|
1045 |
-
# # value="Single dataset",
|
1046 |
-
# # )
|
1047 |
-
|
1048 |
-
# # with gr.Row():
|
1049 |
-
# # dataset1_name = gr.Textbox(value=default_dataset1_name, label="Dataset 1 Name")
|
1050 |
-
# # dataset1_split = gr.Textbox(value=default_dataset1_split, label="Dataset 1 Split")
|
1051 |
-
# # dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
1052 |
-
|
1053 |
-
# # dataset2_inputs = gr.Column(visible=False)
|
1054 |
-
# # with dataset2_inputs:
|
1055 |
-
# # gr.Markdown("### Dataset 2")
|
1056 |
-
# # with gr.Row():
|
1057 |
-
# # dataset2_name = gr.Textbox(value=default_dataset2_name, label="Dataset 2 Name")
|
1058 |
-
# # dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
1059 |
-
# # dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
1060 |
-
|
1061 |
-
# # threshold = gr.Slider(minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold")
|
1062 |
-
|
1063 |
-
# # compute_button = gr.Button("Compute")
|
1064 |
-
|
1065 |
-
# # result_output = gr.Markdown()
|
1066 |
-
|
1067 |
-
# # def update_visibility(deduplication_type_value):
|
1068 |
-
# # return gr.update(visible=True) if deduplication_type_value == "Cross-dataset" else gr.update(visible=False)
|
1069 |
-
|
1070 |
-
# # deduplication_type.change(
|
1071 |
-
# # update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
1072 |
-
# # )
|
1073 |
-
|
1074 |
-
# # compute_button.click(
|
1075 |
-
# # fn=perform_deduplication,
|
1076 |
-
# # inputs=[
|
1077 |
-
# # deduplication_type,
|
1078 |
-
# # dataset1_name,
|
1079 |
-
# # dataset1_split,
|
1080 |
-
# # dataset1_text_column,
|
1081 |
-
# # dataset2_name,
|
1082 |
-
# # dataset2_split,
|
1083 |
-
# # dataset2_text_column,
|
1084 |
-
# # threshold,
|
1085 |
-
# # ],
|
1086 |
-
# # outputs=[result_output],
|
1087 |
-
# # )
|
1088 |
-
|
1089 |
-
# # demo.launch()
|
1090 |
-
|
1091 |
-
|
1092 |
-
# # # import gradio as gr
|
1093 |
-
# # # from datasets import load_dataset
|
1094 |
-
# # # import numpy as np
|
1095 |
-
# # # import model2vec
|
1096 |
-
# # # from reach import Reach
|
1097 |
-
# # # from difflib import ndiff
|
1098 |
-
# # # import time
|
1099 |
-
|
1100 |
-
# # # # Load the model at startup
|
1101 |
-
# # # model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
1102 |
-
|
1103 |
-
# # # # Default dataset parameters
|
1104 |
-
# # # default_dataset1_name = "sst2"
|
1105 |
-
# # # default_dataset1_split = "train"
|
1106 |
-
# # # default_dataset2_name = "sst2"
|
1107 |
-
# # # default_dataset2_split = "validation"
|
1108 |
-
# # # default_text_column = "sentence"
|
1109 |
-
# # # default_threshold = 0.9
|
1110 |
-
|
1111 |
-
# # # # Load the default datasets at startup
|
1112 |
-
# # # ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
1113 |
-
# # # ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
1114 |
-
|
1115 |
-
# # # def batch_iterable(iterable, batch_size):
|
1116 |
-
# # # """Helper function to create batches from an iterable."""
|
1117 |
-
# # # for i in range(0, len(iterable), batch_size):
|
1118 |
-
# # # yield iterable[i:i + batch_size]
|
1119 |
-
|
1120 |
-
# # # def log_time(message, start_time=None, logs=None):
|
1121 |
-
# # # """Helper function to log the start and end times."""
|
1122 |
-
# # # current_time = time.time()
|
1123 |
-
# # # if start_time is not None:
|
1124 |
-
# # # elapsed = current_time - start_time
|
1125 |
-
# # # log_message = f"{message} - Took {elapsed:.2f} seconds"
|
1126 |
-
# # # else:
|
1127 |
-
# # # log_message = f"{message} - Started"
|
1128 |
-
|
1129 |
-
# # # if logs is not None:
|
1130 |
-
# # # logs.append(log_message)
|
1131 |
-
|
1132 |
-
# # # def compute_embeddings(texts, batch_size, progress, logs, desc="Computing embeddings"):
|
1133 |
-
# # # embeddings = []
|
1134 |
-
# # # total_batches = (len(texts) + batch_size - 1) // batch_size
|
1135 |
-
# # # for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
1136 |
-
# # # batch_embeddings = model.encode(batch_texts, show_progressbar=False)
|
1137 |
-
# # # embeddings.append(batch_embeddings)
|
1138 |
-
# # # progress((i + 1) / total_batches, desc=desc)
|
1139 |
-
# # # return np.concatenate(embeddings, axis=0)
|
1140 |
-
|
1141 |
-
# # # def deduplicate(
|
1142 |
-
# # # embedding_matrix: np.ndarray,
|
1143 |
-
# # # threshold: float,
|
1144 |
-
# # # batch_size: int = 1024,
|
1145 |
-
# # # progress=None,
|
1146 |
-
# # # logs=None
|
1147 |
-
# # # ) -> tuple[np.ndarray, dict[int, int]]:
|
1148 |
-
# # # # Building the index
|
1149 |
-
# # # log_time("Building search index", logs=logs)
|
1150 |
-
# # # reach = Reach(
|
1151 |
-
# # # vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
1152 |
-
# # # )
|
1153 |
-
|
1154 |
-
# # # deduplicated_indices = set(range(len(embedding_matrix)))
|
1155 |
-
# # # duplicate_to_original_mapping = {}
|
1156 |
-
|
1157 |
-
# # # # Finding nearest neighbors
|
1158 |
-
# # # log_time("Finding nearest neighbors", logs=logs)
|
1159 |
-
# # # results = reach.nearest_neighbor_threshold(
|
1160 |
-
# # # embedding_matrix,
|
1161 |
-
# # # threshold=threshold,
|
1162 |
-
# # # batch_size=batch_size,
|
1163 |
-
# # # show_progressbar=False, # Disable internal progress bar
|
1164 |
-
# # # )
|
1165 |
-
|
1166 |
-
# # # # Processing duplicates with a progress bar
|
1167 |
-
# # # total_items = len(embedding_matrix)
|
1168 |
-
# # # log_time("Processing duplicates", logs=logs)
|
1169 |
-
# # # for i, similar_items in enumerate(
|
1170 |
-
# # # progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
1171 |
-
# # # ):
|
1172 |
-
# # # if i not in deduplicated_indices:
|
1173 |
-
# # # continue
|
1174 |
-
|
1175 |
-
# # # similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
1176 |
-
|
1177 |
-
# # # for sim_idx in similar_indices:
|
1178 |
-
# # # if sim_idx in deduplicated_indices:
|
1179 |
-
# # # deduplicated_indices.remove(sim_idx)
|
1180 |
-
# # # duplicate_to_original_mapping[sim_idx] = i
|
1181 |
-
|
1182 |
-
# # # return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
1183 |
-
|
1184 |
-
# # # def display_word_differences(x: str, y: str) -> str:
|
1185 |
-
# # # diff = ndiff(x.split(), y.split())
|
1186 |
-
# # # return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
1187 |
-
|
1188 |
-
# # # def encode_texts(texts, progress=None, logs=None):
|
1189 |
-
# # # embedding_matrix = model.encode(texts, show_progressbar=False)
|
1190 |
-
# # # log_time("Encoding texts completed", logs=logs)
|
1191 |
-
# # # return embedding_matrix
|
1192 |
-
|
1193 |
-
# # # def perform_deduplication(
|
1194 |
-
# # # deduplication_type,
|
1195 |
-
# # # dataset1_name,
|
1196 |
-
# # # dataset1_split,
|
1197 |
-
# # # dataset1_text_column,
|
1198 |
-
# # # dataset2_name="",
|
1199 |
-
# # # dataset2_split="",
|
1200 |
-
# # # dataset2_text_column="",
|
1201 |
-
# # # threshold=default_threshold,
|
1202 |
-
# # # progress=gr.Progress(track_tqdm=True),
|
1203 |
-
# # # ):
|
1204 |
-
# # # logs = [] # To store log messages
|
1205 |
-
# # # try:
|
1206 |
-
# # # # Convert threshold to float
|
1207 |
-
# # # threshold = float(threshold)
|
1208 |
-
|
1209 |
-
# # # # Initialize status message
|
1210 |
-
# # # log_time("Deduplication started", logs=logs)
|
1211 |
-
|
1212 |
-
# # # if deduplication_type == "Single dataset":
|
1213 |
-
# # # # Load Dataset 1
|
1214 |
-
# # # start_time = time.time()
|
1215 |
-
# # # log_time("Loading Dataset 1", logs=logs)
|
1216 |
-
# # # if (
|
1217 |
-
# # # dataset1_name == default_dataset1_name
|
1218 |
-
# # # and dataset1_split == default_dataset1_split
|
1219 |
-
# # # ):
|
1220 |
-
# # # ds = ds_default1
|
1221 |
-
# # # else:
|
1222 |
-
# # # ds = load_dataset(dataset1_name, split=dataset1_split)
|
1223 |
-
# # # log_time("Loading Dataset 1 completed", start_time=start_time, logs=logs)
|
1224 |
-
|
1225 |
-
# # # # Extract texts
|
1226 |
-
# # # start_time = time.time()
|
1227 |
-
# # # log_time("Extracting texts from Dataset 1", logs=logs)
|
1228 |
-
# # # texts = [example[dataset1_text_column] for example in ds]
|
1229 |
-
# # # log_time("Extracting texts from Dataset 1 completed", start_time=start_time, logs=logs)
|
1230 |
-
|
1231 |
-
# # # # Compute embeddings
|
1232 |
-
# # # start_time = time.time()
|
1233 |
-
# # # log_time("Computing embeddings for Dataset 1", logs=logs)
|
1234 |
-
# # # embedding_matrix = encode_texts(texts, progress=progress, logs=logs)
|
1235 |
-
# # # log_time("Computing embeddings for Dataset 1 completed", start_time=start_time, logs=logs)
|
1236 |
-
|
1237 |
-
# # # # Deduplicate
|
1238 |
-
# # # start_time = time.time()
|
1239 |
-
# # # log_time("Deduplicating embeddings", logs=logs)
|
1240 |
-
# # # deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
1241 |
-
# # # embedding_matrix, threshold, progress=progress, logs=logs
|
1242 |
-
# # # )
|
1243 |
-
# # # log_time("Deduplication completed", start_time=start_time, logs=logs)
|
1244 |
-
|
1245 |
-
# # # # Prepare the results
|
1246 |
-
# # # num_duplicates = len(duplicate_to_original_mapping)
|
1247 |
-
# # # num_total = len(texts)
|
1248 |
-
# # # num_deduplicated = len(deduplicated_indices)
|
1249 |
-
|
1250 |
-
# # # result_text = f"**Total documents:** {num_total}\n"
|
1251 |
-
# # # result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
1252 |
-
# # # result_text += (
|
1253 |
-
# # # f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
1254 |
-
# # # )
|
1255 |
-
|
1256 |
-
# # # # Show deduplicated examples
|
1257 |
-
# # # if num_duplicates > 0:
|
1258 |
-
# # # result_text += "**Examples of duplicates found:**\n\n"
|
1259 |
-
# # # num_examples = min(5, num_duplicates)
|
1260 |
-
# # # for duplicate_idx, original_idx in list(duplicate_to_original_mapping.items())[:num_examples]:
|
1261 |
-
# # # original_text = texts[original_idx]
|
1262 |
-
# # # duplicate_text = texts[duplicate_idx]
|
1263 |
-
# # # differences = display_word_differences(original_text, duplicate_text)
|
1264 |
-
# # # result_text += f"**Original text:**\n{original_text}\n\n"
|
1265 |
-
# # # result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
|
1266 |
-
# # # result_text += f"**Differences:**\n{differences}\n"
|
1267 |
-
# # # result_text += "-" * 50 + "\n\n"
|
1268 |
-
# # # else:
|
1269 |
-
# # # result_text += "No duplicates found."
|
1270 |
-
|
1271 |
-
# # # log_time("Deduplication process finished", logs=logs)
|
1272 |
-
# # # full_log = "\n".join(logs) # Combine all logs into one output
|
1273 |
-
# # # yield full_log, result_text
|
1274 |
-
|
1275 |
-
# # # except Exception as e:
|
1276 |
-
# # # full_log = "\n".join(logs) # Combine all logs into one output in case of an error
|
1277 |
-
# # # yield f"An error occurred: {e}", ""
|
1278 |
-
# # # raise e
|
1279 |
-
|
1280 |
-
# # # # Adjust the height of the status_output component using custom CSS
|
1281 |
-
# # # with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
1282 |
-
# # # gr.Markdown("# Semantic Deduplication")
|
1283 |
-
|
1284 |
-
# # # deduplication_type = gr.Radio(
|
1285 |
-
# # # choices=["Single dataset", "Cross-dataset"],
|
1286 |
-
# # # label="Deduplication Type",
|
1287 |
-
# # # value="Single dataset",
|
1288 |
-
# # # )
|
1289 |
-
|
1290 |
-
# # # with gr.Row():
|
1291 |
-
# # # dataset1_name = gr.Textbox(value=default_dataset1_name, label="Dataset 1 Name")
|
1292 |
-
# # # dataset1_split = gr.Textbox(value=default_dataset1_split, label="Dataset 1 Split")
|
1293 |
-
# # # dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
1294 |
-
|
1295 |
-
# # # dataset2_inputs = gr.Column(visible=False)
|
1296 |
-
# # # with dataset2_inputs:
|
1297 |
-
# # # gr.Markdown("### Dataset 2")
|
1298 |
-
# # # with gr.Row():
|
1299 |
-
# # # dataset2_name = gr.Textbox(value=default_dataset2_name, label="Dataset 2 Name")
|
1300 |
-
# # # dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
1301 |
-
# # # dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
1302 |
-
|
1303 |
-
# # # threshold = gr.Slider(
|
1304 |
-
# # # minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold"
|
1305 |
-
# # # )
|
1306 |
-
|
1307 |
-
# # # compute_button = gr.Button("Compute")
|
1308 |
-
|
1309 |
-
# # # # Use 'gr.Markdown' with 'elem_id' and custom CSS to adjust height
|
1310 |
-
# # # status_output = gr.Markdown(elem_id="status_output")
|
1311 |
-
# # # result_output = gr.Markdown()
|
1312 |
-
|
1313 |
-
# # # # Function to update the visibility of dataset2_inputs
|
1314 |
-
# # # def update_visibility(deduplication_type_value):
|
1315 |
-
# # # if deduplication_type_value == "Cross-dataset":
|
1316 |
-
# # # return gr.update(visible=True)
|
1317 |
-
# # # else:
|
1318 |
-
# # # return gr.update(visible=False)
|
1319 |
-
|
1320 |
-
# # # deduplication_type.change(
|
1321 |
-
# # # update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
1322 |
-
# # # )
|
1323 |
-
|
1324 |
-
# # # compute_button.click(
|
1325 |
-
# # # fn=perform_deduplication,
|
1326 |
-
# # # inputs=[
|
1327 |
-
# # # deduplication_type,
|
1328 |
-
# # # dataset1_name,
|
1329 |
-
# # # dataset1_split,
|
1330 |
-
# # # dataset1_text_column,
|
1331 |
-
# # # dataset2_name,
|
1332 |
-
# # # dataset2_split,
|
1333 |
-
# # # dataset2_text_column,
|
1334 |
-
# # # threshold,
|
1335 |
-
# # # ],
|
1336 |
-
# # # outputs=[status_output, result_output],
|
1337 |
-
# # # )
|
1338 |
-
|
1339 |
-
# # # demo.launch()
|
1340 |
-
|
1341 |
-
|
1342 |
-
|
1343 |
-
# # # # import gradio as gr
|
1344 |
-
# # # # from datasets import load_dataset
|
1345 |
-
# # # # import numpy as np
|
1346 |
-
# # # # #from model2vec import StaticModel
|
1347 |
-
# # # # import model2vec
|
1348 |
-
# # # # from reach import Reach
|
1349 |
-
# # # # from difflib import ndiff
|
1350 |
-
|
1351 |
-
|
1352 |
-
# # # # # Load the model at startup
|
1353 |
-
# # # # model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
1354 |
-
|
1355 |
-
# # # # # Default dataset parameters
|
1356 |
-
# # # # default_dataset1_name = "sst2"
|
1357 |
-
# # # # default_dataset1_split = "train"
|
1358 |
-
# # # # default_dataset2_name = "sst2"
|
1359 |
-
# # # # default_dataset2_split = "validation"
|
1360 |
-
# # # # default_text_column = "sentence"
|
1361 |
-
# # # # default_threshold = 0.9
|
1362 |
-
|
1363 |
-
# # # # # Load the default datasets at startup
|
1364 |
-
# # # # ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
1365 |
-
# # # # ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
1366 |
-
|
1367 |
-
|
1368 |
-
# # # # def batch_iterable(iterable, batch_size):
|
1369 |
-
# # # # """Helper function to create batches from an iterable."""
|
1370 |
-
# # # # for i in range(0, len(iterable), batch_size):
|
1371 |
-
# # # # yield iterable[i:i + batch_size]
|
1372 |
-
|
1373 |
-
# # # # def compute_embeddings(texts, batch_size, progress, desc="Computing embeddings"):
|
1374 |
-
# # # # embeddings = []
|
1375 |
-
# # # # total_batches = (len(texts) + batch_size - 1) // batch_size
|
1376 |
-
# # # # for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
1377 |
-
# # # # batch_embeddings = model.encode(batch_texts, show_progressbar=False)
|
1378 |
-
# # # # embeddings.append(batch_embeddings)
|
1379 |
-
# # # # progress((i + 1) / total_batches, desc=desc)
|
1380 |
-
# # # # return np.concatenate(embeddings, axis=0)
|
1381 |
-
|
1382 |
-
# # # # def deduplicate(
|
1383 |
-
# # # # embedding_matrix: np.ndarray,
|
1384 |
-
# # # # threshold: float,
|
1385 |
-
# # # # batch_size: int = 1024,
|
1386 |
-
# # # # progress=None
|
1387 |
-
# # # # ) -> tuple[np.ndarray, dict[int, int]]:
|
1388 |
-
# # # # # Building the index
|
1389 |
-
# # # # progress(0, desc="Building search index...")
|
1390 |
-
# # # # reach = Reach(
|
1391 |
-
# # # # vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
1392 |
-
# # # # )
|
1393 |
-
|
1394 |
-
# # # # deduplicated_indices = set(range(len(embedding_matrix)))
|
1395 |
-
# # # # duplicate_to_original_mapping = {}
|
1396 |
-
|
1397 |
-
# # # # # Finding nearest neighbors
|
1398 |
-
# # # # progress(0, desc="Finding nearest neighbors...")
|
1399 |
-
# # # # results = reach.nearest_neighbor_threshold(
|
1400 |
-
# # # # embedding_matrix,
|
1401 |
-
# # # # threshold=threshold,
|
1402 |
-
# # # # batch_size=batch_size,
|
1403 |
-
# # # # show_progressbar=False, # Disable internal progress bar
|
1404 |
-
# # # # )
|
1405 |
-
|
1406 |
-
# # # # # Processing duplicates with a progress bar
|
1407 |
-
# # # # total_items = len(embedding_matrix)
|
1408 |
-
# # # # for i, similar_items in enumerate(
|
1409 |
-
# # # # progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
1410 |
-
# # # # ):
|
1411 |
-
# # # # if i not in deduplicated_indices:
|
1412 |
-
# # # # continue
|
1413 |
-
|
1414 |
-
# # # # similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
1415 |
-
|
1416 |
-
# # # # for sim_idx in similar_indices:
|
1417 |
-
# # # # if sim_idx in deduplicated_indices:
|
1418 |
-
# # # # deduplicated_indices.remove(sim_idx)
|
1419 |
-
# # # # duplicate_to_original_mapping[sim_idx] = i
|
1420 |
-
|
1421 |
-
# # # # return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
1422 |
-
|
1423 |
-
# # # # def display_word_differences(x: str, y: str) -> str:
|
1424 |
-
# # # # diff = ndiff(x.split(), y.split())
|
1425 |
-
# # # # return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
1426 |
-
|
1427 |
-
|
1428 |
-
# # # # def encode_texts(texts, progress=None):
|
1429 |
-
# # # # embedding_matrix = model.encode(texts, show_progressbar=False)
|
1430 |
-
# # # # return embedding_matrix
|
1431 |
-
|
1432 |
-
# # # # def perform_deduplication(
|
1433 |
-
# # # # deduplication_type,
|
1434 |
-
# # # # dataset1_name,
|
1435 |
-
# # # # dataset1_split,
|
1436 |
-
# # # # dataset1_text_column,
|
1437 |
-
# # # # dataset2_name="",
|
1438 |
-
# # # # dataset2_split="",
|
1439 |
-
# # # # dataset2_text_column="",
|
1440 |
-
# # # # threshold=default_threshold,
|
1441 |
-
# # # # progress=gr.Progress(track_tqdm=True),
|
1442 |
-
# # # # ):
|
1443 |
-
# # # # try:
|
1444 |
-
# # # # # Convert threshold to float
|
1445 |
-
# # # # threshold = float(threshold)
|
1446 |
-
|
1447 |
-
# # # # # Initialize status message
|
1448 |
-
# # # # status = ""
|
1449 |
-
|
1450 |
-
# # # # if deduplication_type == "Single dataset":
|
1451 |
-
# # # # # Load Dataset 1
|
1452 |
-
# # # # status = "Loading Dataset 1..."
|
1453 |
-
# # # # yield status, ""
|
1454 |
-
# # # # if (
|
1455 |
-
# # # # dataset1_name == default_dataset1_name
|
1456 |
-
# # # # and dataset1_split == default_dataset1_split
|
1457 |
-
# # # # ):
|
1458 |
-
# # # # ds = ds_default1
|
1459 |
-
# # # # else:
|
1460 |
-
# # # # ds = load_dataset(dataset1_name, split=dataset1_split)
|
1461 |
-
|
1462 |
-
# # # # # Extract texts
|
1463 |
-
# # # # status = "Extracting texts from Dataset 1..."
|
1464 |
-
# # # # yield status, ""
|
1465 |
-
# # # # texts = [example[dataset1_text_column] for example in ds]
|
1466 |
-
# # # # # Compute embeddings
|
1467 |
-
# # # # status = "Computing embeddings for Dataset 1..."
|
1468 |
-
# # # # yield status, ""
|
1469 |
-
# # # # embedding_matrix = encode_texts(texts, progress=progress)
|
1470 |
-
# # # # #embedding_matrix = model.encode(texts, show_progressbar=True)
|
1471 |
-
# # # # # embedding_matrix = compute_embeddings(
|
1472 |
-
# # # # # texts,
|
1473 |
-
# # # # # batch_size=64,
|
1474 |
-
# # # # # progress=progress,
|
1475 |
-
# # # # # desc="Computing embeddings for Dataset 1",
|
1476 |
-
# # # # # )
|
1477 |
-
|
1478 |
-
# # # # # Deduplicate
|
1479 |
-
# # # # status = "Deduplicating embeddings..."
|
1480 |
-
# # # # yield status, ""
|
1481 |
-
# # # # deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
1482 |
-
# # # # embedding_matrix, threshold, progress=progress
|
1483 |
-
# # # # )
|
1484 |
-
|
1485 |
-
# # # # # Prepare the results
|
1486 |
-
# # # # num_duplicates = len(duplicate_to_original_mapping)
|
1487 |
-
# # # # num_total = len(texts)
|
1488 |
-
# # # # num_deduplicated = len(deduplicated_indices)
|
1489 |
-
|
1490 |
-
# # # # result_text = f"**Total documents:** {num_total}\n"
|
1491 |
-
# # # # result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
1492 |
-
# # # # result_text += (
|
1493 |
-
# # # # f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
1494 |
-
# # # # )
|
1495 |
-
|
1496 |
-
# # # # # Show deduplicated examples
|
1497 |
-
# # # # if num_duplicates > 0:
|
1498 |
-
# # # # result_text += "**Examples of duplicates found:**\n\n"
|
1499 |
-
# # # # num_examples = min(5, num_duplicates)
|
1500 |
-
# # # # for duplicate_idx, original_idx in list(duplicate_to_original_mapping.items())[:num_examples]:
|
1501 |
-
# # # # original_text = texts[original_idx]
|
1502 |
-
# # # # duplicate_text = texts[duplicate_idx]
|
1503 |
-
# # # # differences = display_word_differences(original_text, duplicate_text)
|
1504 |
-
# # # # result_text += f"**Original text:**\n{original_text}\n\n"
|
1505 |
-
# # # # result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
|
1506 |
-
# # # # result_text += f"**Differences:**\n{differences}\n"
|
1507 |
-
# # # # result_text += "-" * 50 + "\n\n"
|
1508 |
-
# # # # else:
|
1509 |
-
# # # # result_text += "No duplicates found."
|
1510 |
-
|
1511 |
-
# # # # # Final status
|
1512 |
-
# # # # status = "Deduplication completed."
|
1513 |
-
# # # # yield status, result_text
|
1514 |
-
|
1515 |
-
# # # # elif deduplication_type == "Cross-dataset":
|
1516 |
-
# # # # # Similar code for cross-dataset deduplication
|
1517 |
-
# # # # # Load Dataset 1
|
1518 |
-
# # # # status = "Loading Dataset 1..."
|
1519 |
-
# # # # yield status, ""
|
1520 |
-
# # # # if (
|
1521 |
-
# # # # dataset1_name == default_dataset1_name
|
1522 |
-
# # # # and dataset1_split == default_dataset1_split
|
1523 |
-
# # # # ):
|
1524 |
-
# # # # ds1 = ds_default1
|
1525 |
-
# # # # else:
|
1526 |
-
# # # # ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
1527 |
-
|
1528 |
-
# # # # # Load Dataset 2
|
1529 |
-
# # # # status = "Loading Dataset 2..."
|
1530 |
-
# # # # yield status, ""
|
1531 |
-
# # # # if (
|
1532 |
-
# # # # dataset2_name == default_dataset2_name
|
1533 |
-
# # # # and dataset2_split == default_dataset2_split
|
1534 |
-
# # # # ):
|
1535 |
-
# # # # ds2 = ds_default2
|
1536 |
-
# # # # else:
|
1537 |
-
# # # # ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
1538 |
-
|
1539 |
-
# # # # # Extract texts from Dataset 1
|
1540 |
-
# # # # status = "Extracting texts from Dataset 1..."
|
1541 |
-
# # # # yield status, ""
|
1542 |
-
# # # # texts1 = [example[dataset1_text_column] for example in ds1]
|
1543 |
-
|
1544 |
-
# # # # # Extract texts from Dataset 2
|
1545 |
-
# # # # status = "Extracting texts from Dataset 2..."
|
1546 |
-
# # # # yield status, ""
|
1547 |
-
# # # # texts2 = [example[dataset2_text_column] for example in ds2]
|
1548 |
-
|
1549 |
-
# # # # # Compute embeddings for Dataset 1
|
1550 |
-
# # # # status = "Computing embeddings for Dataset 1..."
|
1551 |
-
# # # # yield status, ""
|
1552 |
-
# # # # embedding_matrix1 = compute_embeddings(
|
1553 |
-
# # # # texts1,
|
1554 |
-
# # # # batch_size=64,
|
1555 |
-
# # # # progress=progress,
|
1556 |
-
# # # # desc="Computing embeddings for Dataset 1",
|
1557 |
-
# # # # )
|
1558 |
-
|
1559 |
-
# # # # # Compute embeddings for Dataset 2
|
1560 |
-
# # # # status = "Computing embeddings for Dataset 2..."
|
1561 |
-
# # # # yield status, ""
|
1562 |
-
# # # # embedding_matrix2 = compute_embeddings(
|
1563 |
-
# # # # texts2,
|
1564 |
-
# # # # batch_size=64,
|
1565 |
-
# # # # progress=progress,
|
1566 |
-
# # # # desc="Computing embeddings for Dataset 2",
|
1567 |
-
# # # # )
|
1568 |
-
|
1569 |
-
# # # # # Deduplicate across datasets
|
1570 |
-
# # # # status = "Deduplicating embeddings across datasets..."
|
1571 |
-
# # # # yield status, ""
|
1572 |
-
# # # # duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(
|
1573 |
-
# # # # embedding_matrix1, embedding_matrix2, threshold, progress=progress
|
1574 |
-
# # # # )
|
1575 |
-
|
1576 |
-
# # # # num_duplicates = len(duplicate_indices_in_ds2)
|
1577 |
-
# # # # num_total_ds2 = len(texts2)
|
1578 |
-
# # # # num_unique_ds2 = num_total_ds2 - num_duplicates
|
1579 |
-
|
1580 |
-
# # # # result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
1581 |
-
# # # # result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
1582 |
-
# # # # result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
1583 |
-
|
1584 |
-
# # # # # Show deduplicated examples
|
1585 |
-
# # # # if num_duplicates > 0:
|
1586 |
-
# # # # result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
1587 |
-
# # # # num_examples = min(5, num_duplicates)
|
1588 |
-
# # # # for duplicate_idx in duplicate_indices_in_ds2[:num_examples]:
|
1589 |
-
# # # # original_idx = duplicate_to_original_mapping[duplicate_idx]
|
1590 |
-
# # # # original_text = texts1[original_idx]
|
1591 |
-
# # # # duplicate_text = texts2[duplicate_idx]
|
1592 |
-
# # # # differences = display_word_differences(original_text, duplicate_text)
|
1593 |
-
# # # # result_text += f"**Original text (Dataset 1):**\n{original_text}\n\n"
|
1594 |
-
# # # # result_text += f"**Duplicate text (Dataset 2):**\n{duplicate_text}\n\n"
|
1595 |
-
# # # # result_text += f"**Differences:**\n{differences}\n"
|
1596 |
-
# # # # result_text += "-" * 50 + "\n\n"
|
1597 |
-
# # # # else:
|
1598 |
-
# # # # result_text += "No duplicates found."
|
1599 |
-
|
1600 |
-
# # # # # Final status
|
1601 |
-
# # # # status = "Deduplication completed."
|
1602 |
-
# # # # yield status, result_text
|
1603 |
-
|
1604 |
-
# # # # except Exception as e:
|
1605 |
-
# # # # yield f"An error occurred: {e}", ""
|
1606 |
-
# # # # raise e
|
1607 |
-
|
1608 |
-
# # # # def deduplicate_across_datasets(
|
1609 |
-
# # # # embedding_matrix_1: np.ndarray,
|
1610 |
-
# # # # embedding_matrix_2: np.ndarray,
|
1611 |
-
# # # # threshold: float,
|
1612 |
-
# # # # batch_size: int = 1024,
|
1613 |
-
# # # # progress=None
|
1614 |
-
# # # # ) -> tuple[list[int], dict[int, int]]:
|
1615 |
-
# # # # # Building the index from Dataset 1
|
1616 |
-
# # # # progress(0, desc="Building search index from Dataset 1...")
|
1617 |
-
# # # # reach = Reach(
|
1618 |
-
# # # # vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))]
|
1619 |
-
# # # # )
|
1620 |
-
|
1621 |
-
# # # # duplicate_indices_in_test = []
|
1622 |
-
# # # # duplicate_to_original_mapping = {}
|
1623 |
-
|
1624 |
-
# # # # # Finding nearest neighbors between datasets
|
1625 |
-
# # # # progress(0, desc="Finding nearest neighbors between datasets...")
|
1626 |
-
# # # # results = reach.nearest_neighbor_threshold(
|
1627 |
-
# # # # embedding_matrix_2,
|
1628 |
-
# # # # threshold=threshold,
|
1629 |
-
# # # # batch_size=batch_size,
|
1630 |
-
# # # # show_progressbar=False, # Disable internal progress bar
|
1631 |
-
# # # # )
|
1632 |
-
|
1633 |
-
# # # # total_items = len(embedding_matrix_2)
|
1634 |
-
# # # # # Processing duplicates with a progress bar
|
1635 |
-
# # # # for i, similar_items in enumerate(
|
1636 |
-
# # # # progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)
|
1637 |
-
# # # # ):
|
1638 |
-
# # # # similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
1639 |
-
|
1640 |
-
# # # # if similar_indices:
|
1641 |
-
# # # # duplicate_indices_in_test.append(i)
|
1642 |
-
# # # # duplicate_to_original_mapping[i] = similar_indices[0]
|
1643 |
-
|
1644 |
-
# # # # return duplicate_indices_in_test, duplicate_to_original_mapping
|
1645 |
-
|
1646 |
-
# # # # # Adjust the height of the status_output component using custom CSS
|
1647 |
-
# # # # with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
1648 |
-
# # # # gr.Markdown("# Semantic Deduplication")
|
1649 |
-
|
1650 |
-
# # # # deduplication_type = gr.Radio(
|
1651 |
-
# # # # choices=["Single dataset", "Cross-dataset"],
|
1652 |
-
# # # # label="Deduplication Type",
|
1653 |
-
# # # # value="Single dataset",
|
1654 |
-
# # # # )
|
1655 |
-
|
1656 |
-
# # # # with gr.Row():
|
1657 |
-
# # # # dataset1_name = gr.Textbox(value=default_dataset1_name, label="Dataset 1 Name")
|
1658 |
-
# # # # dataset1_split = gr.Textbox(value=default_dataset1_split, label="Dataset 1 Split")
|
1659 |
-
# # # # dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
1660 |
-
|
1661 |
-
# # # # dataset2_inputs = gr.Column(visible=False)
|
1662 |
-
# # # # with dataset2_inputs:
|
1663 |
-
# # # # gr.Markdown("### Dataset 2")
|
1664 |
-
# # # # with gr.Row():
|
1665 |
-
# # # # dataset2_name = gr.Textbox(value=default_dataset2_name, label="Dataset 2 Name")
|
1666 |
-
# # # # dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
1667 |
-
# # # # dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
1668 |
-
|
1669 |
-
# # # # threshold = gr.Slider(
|
1670 |
-
# # # # minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold"
|
1671 |
-
# # # # )
|
1672 |
-
|
1673 |
-
# # # # compute_button = gr.Button("Compute")
|
1674 |
-
|
1675 |
-
# # # # # Use 'gr.Markdown' with 'elem_id' and custom CSS to adjust height
|
1676 |
-
# # # # status_output = gr.Markdown(elem_id="status_output")
|
1677 |
-
# # # # result_output = gr.Markdown()
|
1678 |
-
|
1679 |
-
# # # # # Function to update the visibility of dataset2_inputs
|
1680 |
-
# # # # def update_visibility(deduplication_type_value):
|
1681 |
-
# # # # if deduplication_type_value == "Cross-dataset":
|
1682 |
-
# # # # return gr.update(visible=True)
|
1683 |
-
# # # # else:
|
1684 |
-
# # # # return gr.update(visible=False)
|
1685 |
-
|
1686 |
-
# # # # deduplication_type.change(
|
1687 |
-
# # # # update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
1688 |
-
# # # # )
|
1689 |
-
|
1690 |
-
# # # # compute_button.click(
|
1691 |
-
# # # # fn=perform_deduplication,
|
1692 |
-
# # # # inputs=[
|
1693 |
-
# # # # deduplication_type,
|
1694 |
-
# # # # dataset1_name,
|
1695 |
-
# # # # dataset1_split,
|
1696 |
-
# # # # dataset1_text_column,
|
1697 |
-
# # # # dataset2_name,
|
1698 |
-
# # # # dataset2_split,
|
1699 |
-
# # # # dataset2_text_column,
|
1700 |
-
# # # # threshold,
|
1701 |
-
# # # # ],
|
1702 |
-
# # # # outputs=[status_output, result_output],
|
1703 |
-
# # # # )
|
1704 |
-
|
1705 |
-
# # # # demo.launch()
|
|
|
5 |
from reach import Reach
|
6 |
from difflib import ndiff
|
7 |
|
8 |
+
# Load the model
|
9 |
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
10 |
|
11 |
+
# Default parameters
|
12 |
+
default_dataset_name = "sst2"
|
13 |
+
default_dataset_split = "train"
|
|
|
|
|
14 |
default_text_column = "sentence"
|
15 |
default_threshold = 0.9
|
16 |
|
|
|
|
|
|
|
|
|
17 |
def batch_iterable(iterable, batch_size):
|
18 |
+
"""Yield successive batches from an iterable."""
|
19 |
for i in range(0, len(iterable), batch_size):
|
20 |
yield iterable[i:i + batch_size]
|
21 |
|
22 |
+
def compute_embeddings(texts, batch_size, progress, desc):
|
23 |
+
"""Compute embeddings for a list of texts with progress tracking."""
|
24 |
embeddings = []
|
25 |
total_batches = (len(texts) + batch_size - 1) // batch_size
|
26 |
for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
27 |
+
embeddings.append(model.encode(batch_texts, show_progressbar=False))
|
|
|
28 |
progress((i + 1) / total_batches, desc=desc)
|
29 |
return np.concatenate(embeddings, axis=0)
|
30 |
|
31 |
+
def deduplicate_embeddings(
|
32 |
+
embeddings_a: np.ndarray,
|
33 |
+
embeddings_b: np.ndarray = None,
|
34 |
+
threshold: float = 0.9,
|
35 |
batch_size: int = 1024,
|
36 |
progress=None
|
37 |
+
):
|
38 |
+
"""Deduplicate within one dataset or across two datasets."""
|
39 |
+
if embeddings_b is None:
|
40 |
+
reach = Reach(vectors=embeddings_a, items=[str(i) for i in range(len(embeddings_a))])
|
41 |
+
duplicate_to_original = {}
|
42 |
+
results = reach.nearest_neighbor_threshold(
|
43 |
+
embeddings_a, threshold=threshold, batch_size=batch_size, show_progressbar=False
|
44 |
+
)
|
45 |
+
for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=len(embeddings_a))):
|
46 |
+
for sim_idx, _ in similar_items:
|
47 |
+
sim_idx = int(sim_idx)
|
48 |
+
if sim_idx != i and sim_idx not in duplicate_to_original:
|
49 |
+
duplicate_to_original[sim_idx] = i
|
50 |
+
deduplicated_indices = set(range(len(embeddings_a))) - set(duplicate_to_original.keys())
|
51 |
+
return deduplicated_indices, duplicate_to_original
|
52 |
+
else:
|
53 |
+
reach = Reach(vectors=embeddings_a, items=[str(i) for i in range(len(embeddings_a))])
|
54 |
+
duplicate_indices_in_b = []
|
55 |
+
duplicate_to_original = {}
|
56 |
+
results = reach.nearest_neighbor_threshold(
|
57 |
+
embeddings_b, threshold=threshold, batch_size=batch_size, show_progressbar=False
|
58 |
+
)
|
59 |
+
for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=len(embeddings_b))):
|
60 |
+
if similar_items:
|
61 |
+
duplicate_indices_in_b.append(i)
|
62 |
+
duplicate_to_original[i] = int(similar_items[0][0])
|
63 |
+
return duplicate_indices_in_b, duplicate_to_original
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
|
65 |
def display_word_differences(x: str, y: str) -> str:
|
66 |
+
"""Display differences between two texts."""
|
67 |
diff = ndiff(x.split(), y.split())
|
68 |
+
return " ".join(word for word in diff if word.startswith(("+", "-")))
|
69 |
+
|
70 |
+
def load_dataset_texts(dataset_name, dataset_split, text_column):
|
71 |
+
"""Load texts from a specified dataset."""
|
72 |
+
ds = load_dataset(dataset_name, split=dataset_split)
|
73 |
+
return [example[text_column] for example in ds]
|
74 |
|
75 |
def perform_deduplication(
|
76 |
deduplication_type,
|
|
|
84 |
progress=gr.Progress(track_tqdm=True),
|
85 |
):
|
86 |
try:
|
|
|
87 |
threshold = float(threshold)
|
88 |
|
89 |
+
# Load and process Dataset 1
|
90 |
+
yield "Loading Dataset 1...", ""
|
91 |
+
texts1 = load_dataset_texts(dataset1_name, dataset1_split, dataset1_text_column)
|
92 |
+
yield "Computing embeddings for Dataset 1...", ""
|
93 |
+
embeddings1 = compute_embeddings(texts1, batch_size=64, progress=progress, desc="Dataset 1 embeddings")
|
94 |
|
95 |
if deduplication_type == "Single dataset":
|
96 |
+
# Deduplicate within Dataset 1
|
97 |
+
yield "Deduplicating within Dataset 1...", ""
|
98 |
+
deduplicated_indices, duplicate_mapping = deduplicate_embeddings(
|
99 |
+
embeddings1, threshold=threshold, progress=progress
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
)
|
101 |
|
102 |
+
num_duplicates = len(duplicate_mapping)
|
103 |
+
result_text = (
|
104 |
+
f"**Total documents:** {len(texts1)}\n"
|
105 |
+
f"**Duplicates found:** {num_duplicates}\n"
|
106 |
+
f"**Unique documents after deduplication:** {len(deduplicated_indices)}\n\n"
|
|
|
|
|
|
|
|
|
107 |
)
|
108 |
|
|
|
109 |
if num_duplicates > 0:
|
110 |
+
result_text += "**Sample duplicates:**\n\n"
|
111 |
+
for dup_idx, orig_idx in list(duplicate_mapping.items())[:5]:
|
112 |
+
orig_text = texts1[orig_idx]
|
113 |
+
dup_text = texts1[dup_idx]
|
114 |
+
differences = display_word_differences(orig_text, dup_text)
|
115 |
+
result_text += (
|
116 |
+
f"**Original:**\n{orig_text}\n\n"
|
117 |
+
f"**Duplicate:**\n{dup_text}\n\n"
|
118 |
+
f"**Differences:**\n{differences}\n"
|
119 |
+
+ "-" * 50 + "\n\n"
|
120 |
+
)
|
121 |
else:
|
122 |
result_text += "No duplicates found."
|
123 |
|
124 |
+
yield "Deduplication completed.", result_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
125 |
|
126 |
+
else:
|
127 |
+
# Load and process Dataset 2
|
128 |
+
yield "Loading Dataset 2...", ""
|
129 |
+
texts2 = load_dataset_texts(dataset2_name, dataset2_split, dataset2_text_column)
|
130 |
+
yield "Computing embeddings for Dataset 2...", ""
|
131 |
+
embeddings2 = compute_embeddings(texts2, batch_size=64, progress=progress, desc="Dataset 2 embeddings")
|
132 |
+
|
133 |
+
# Deduplicate Dataset 2 against Dataset 1
|
134 |
+
yield "Deduplicating Dataset 2 against Dataset 1...", ""
|
135 |
+
duplicate_indices, duplicate_mapping = deduplicate_embeddings(
|
136 |
+
embeddings1, embeddings_b=embeddings2, threshold=threshold, progress=progress
|
137 |
)
|
138 |
|
139 |
+
num_duplicates = len(duplicate_indices)
|
140 |
+
result_text = (
|
141 |
+
f"**Total documents in {dataset2_name}/{dataset2_split}:** {len(texts2)}\n"
|
142 |
+
f"**Duplicates found in Dataset 2:** {num_duplicates}\n"
|
143 |
+
f"**Unique documents after deduplication:** {len(texts2) - num_duplicates}\n\n"
|
144 |
)
|
145 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
146 |
if num_duplicates > 0:
|
147 |
+
result_text += "**Sample duplicates from Dataset 2:**\n\n"
|
148 |
+
for idx in duplicate_indices[:5]:
|
149 |
+
orig_text = texts1[duplicate_mapping[idx]]
|
150 |
+
dup_text = texts2[idx]
|
151 |
+
differences = display_word_differences(orig_text, dup_text)
|
152 |
+
result_text += (
|
153 |
+
f"**Original (Dataset 1):**\n{orig_text}\n\n"
|
154 |
+
f"**Duplicate (Dataset 2):**\n{dup_text}\n\n"
|
155 |
+
f"**Differences:**\n{differences}\n"
|
156 |
+
+ "-" * 50 + "\n\n"
|
157 |
+
)
|
158 |
else:
|
159 |
result_text += "No duplicates found."
|
160 |
|
161 |
+
yield "Deduplication completed.", result_text
|
|
|
|
|
162 |
|
163 |
except Exception as e:
|
164 |
yield f"An error occurred: {e}", ""
|
165 |
raise e
|
166 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
167 |
with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
168 |
gr.Markdown("# Semantic Deduplication")
|
169 |
|
|
|
174 |
)
|
175 |
|
176 |
with gr.Row():
|
177 |
+
dataset1_name = gr.Textbox(value=default_dataset_name, label="Dataset 1 Name")
|
178 |
+
dataset1_split = gr.Textbox(value=default_dataset_split, label="Dataset 1 Split")
|
179 |
dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
180 |
|
181 |
dataset2_inputs = gr.Column(visible=False)
|
182 |
with dataset2_inputs:
|
183 |
gr.Markdown("### Dataset 2")
|
184 |
with gr.Row():
|
185 |
+
dataset2_name = gr.Textbox(value=default_dataset_name, label="Dataset 2 Name")
|
186 |
+
dataset2_split = gr.Textbox(value=default_dataset_split, label="Dataset 2 Split")
|
187 |
dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
188 |
|
189 |
+
threshold = gr.Slider(0.0, 1.0, value=default_threshold, label="Similarity Threshold")
|
|
|
|
|
|
|
190 |
compute_button = gr.Button("Compute")
|
|
|
|
|
191 |
status_output = gr.Markdown(elem_id="status_output")
|
192 |
result_output = gr.Markdown()
|
193 |
|
194 |
+
def update_visibility(choice):
|
195 |
+
return gr.update(visible=choice == "Cross-dataset")
|
|
|
|
|
|
|
|
|
196 |
|
197 |
+
deduplication_type.change(update_visibility, inputs=deduplication_type, outputs=dataset2_inputs)
|
|
|
|
|
198 |
|
199 |
compute_button.click(
|
200 |
fn=perform_deduplication,
|
|
|
213 |
|
214 |
demo.launch()
|
215 |
|
|
|
216 |
# import gradio as gr
|
217 |
# from datasets import load_dataset
|
218 |
# import numpy as np
|
219 |
# from model2vec import StaticModel
|
220 |
# from reach import Reach
|
221 |
# from difflib import ndiff
|
|
|
222 |
|
223 |
# # Load the model at startup
|
224 |
# model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
225 |
|
226 |
+
# # Default dataset parameters
|
227 |
# default_dataset1_name = "sst2"
|
228 |
# default_dataset1_split = "train"
|
229 |
# default_dataset2_name = "sst2"
|
|
|
242 |
|
243 |
# def compute_embeddings(texts, batch_size, progress, desc="Computing embeddings"):
|
244 |
# embeddings = []
|
245 |
+
# total_batches = (len(texts) + batch_size - 1) // batch_size
|
246 |
+
# for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
247 |
+
# batch_embeddings = model.encode(batch_texts, show_progressbar=False)
|
248 |
# embeddings.append(batch_embeddings)
|
249 |
+
# progress((i + 1) / total_batches, desc=desc)
|
250 |
# return np.concatenate(embeddings, axis=0)
|
251 |
|
252 |
+
# def deduplicate(
|
253 |
+
# embedding_matrix: np.ndarray,
|
254 |
+
# threshold: float,
|
255 |
+
# batch_size: int = 1024,
|
256 |
+
# progress=None
|
257 |
+
# ) -> tuple[np.ndarray, dict[int, int]]:
|
258 |
+
# # Building the index
|
259 |
+
# progress(0, desc="Building search index...")
|
260 |
+
# reach = Reach(
|
261 |
+
# vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
262 |
+
# )
|
263 |
|
264 |
# deduplicated_indices = set(range(len(embedding_matrix)))
|
265 |
# duplicate_to_original_mapping = {}
|
266 |
|
267 |
+
# # Finding nearest neighbors
|
268 |
+
# progress(0, desc="Finding nearest neighbors...")
|
269 |
# results = reach.nearest_neighbor_threshold(
|
270 |
# embedding_matrix,
|
271 |
# threshold=threshold,
|
272 |
# batch_size=batch_size,
|
273 |
+
# show_progressbar=False, # Disable internal progress bar
|
274 |
# )
|
275 |
|
276 |
+
# # Processing duplicates with a progress bar
|
277 |
# total_items = len(embedding_matrix)
|
278 |
+
# for i, similar_items in enumerate(
|
279 |
+
# progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
280 |
+
# ):
|
281 |
# if i not in deduplicated_indices:
|
282 |
# continue
|
283 |
|
|
|
290 |
|
291 |
# return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
292 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
293 |
# def display_word_differences(x: str, y: str) -> str:
|
294 |
# diff = ndiff(x.split(), y.split())
|
295 |
+
# return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
296 |
|
297 |
# def perform_deduplication(
|
298 |
# deduplication_type,
|
|
|
303 |
# dataset2_split="",
|
304 |
# dataset2_text_column="",
|
305 |
# threshold=default_threshold,
|
306 |
+
# progress=gr.Progress(track_tqdm=True),
|
307 |
# ):
|
308 |
# try:
|
309 |
+
# # Convert threshold to float
|
310 |
# threshold = float(threshold)
|
311 |
|
312 |
+
# # Initialize status message
|
313 |
+
# status = ""
|
314 |
+
|
315 |
# if deduplication_type == "Single dataset":
|
316 |
+
# # Load Dataset 1
|
317 |
+
# status = "Loading Dataset 1..."
|
318 |
+
# yield status, ""
|
319 |
+
# if (
|
320 |
+
# dataset1_name == default_dataset1_name
|
321 |
+
# and dataset1_split == default_dataset1_split
|
322 |
+
# ):
|
323 |
+
# ds = ds_default1
|
324 |
+
# else:
|
325 |
+
# ds = load_dataset(dataset1_name, split=dataset1_split)
|
326 |
|
327 |
+
# # Extract texts
|
328 |
+
# status = "Extracting texts from Dataset 1..."
|
329 |
+
# yield status, ""
|
330 |
+
# texts = [example[dataset1_text_column] for example in ds]
|
331 |
|
332 |
+
# # Compute embeddings
|
333 |
+
# status = "Computing embeddings for Dataset 1..."
|
334 |
+
# yield status, ""
|
335 |
+
# embedding_matrix = compute_embeddings(
|
336 |
+
# texts,
|
337 |
+
# batch_size=64,
|
338 |
+
# progress=progress,
|
339 |
+
# desc="Computing embeddings for Dataset 1",
|
340 |
+
# )
|
341 |
+
|
342 |
+
# # Deduplicate
|
343 |
+
# status = "Deduplicating embeddings..."
|
344 |
+
# yield status, ""
|
345 |
+
# deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
346 |
+
# embedding_matrix, threshold, progress=progress
|
347 |
+
# )
|
348 |
+
|
349 |
+
# # Prepare the results
|
350 |
# num_duplicates = len(duplicate_to_original_mapping)
|
351 |
# num_total = len(texts)
|
352 |
# num_deduplicated = len(deduplicated_indices)
|
353 |
|
354 |
# result_text = f"**Total documents:** {num_total}\n"
|
355 |
# result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
356 |
+
# result_text += (
|
357 |
+
# f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
358 |
+
# )
|
359 |
|
360 |
+
# # Show deduplicated examples
|
361 |
# if num_duplicates > 0:
|
362 |
# result_text += "**Examples of duplicates found:**\n\n"
|
363 |
# num_examples = min(5, num_duplicates)
|
|
|
372 |
# else:
|
373 |
# result_text += "No duplicates found."
|
374 |
|
375 |
+
# # Final status
|
376 |
+
# status = "Deduplication completed."
|
377 |
+
# yield status, result_text
|
378 |
|
379 |
# elif deduplication_type == "Cross-dataset":
|
380 |
+
# # Similar code for cross-dataset deduplication
|
381 |
+
# # Load Dataset 1
|
382 |
+
# status = "Loading Dataset 1..."
|
383 |
+
# yield status, ""
|
384 |
+
# if (
|
385 |
+
# dataset1_name == default_dataset1_name
|
386 |
+
# and dataset1_split == default_dataset1_split
|
387 |
+
# ):
|
388 |
+
# ds1 = ds_default1
|
389 |
+
# else:
|
390 |
+
# ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
391 |
+
|
392 |
+
# # Load Dataset 2
|
393 |
+
# status = "Loading Dataset 2..."
|
394 |
+
# yield status, ""
|
395 |
+
# if (
|
396 |
+
# dataset2_name == default_dataset2_name
|
397 |
+
# and dataset2_split == default_dataset2_split
|
398 |
+
# ):
|
399 |
+
# ds2 = ds_default2
|
400 |
+
# else:
|
401 |
+
# ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
402 |
|
403 |
+
# # Extract texts from Dataset 1
|
404 |
+
# status = "Extracting texts from Dataset 1..."
|
405 |
+
# yield status, ""
|
406 |
# texts1 = [example[dataset1_text_column] for example in ds1]
|
|
|
407 |
|
408 |
+
# # Extract texts from Dataset 2
|
409 |
+
# status = "Extracting texts from Dataset 2..."
|
410 |
+
# yield status, ""
|
411 |
+
# texts2 = [example[dataset2_text_column] for example in ds2]
|
412 |
|
413 |
+
# # Compute embeddings for Dataset 1
|
414 |
+
# status = "Computing embeddings for Dataset 1..."
|
415 |
+
# yield status, ""
|
416 |
+
# embedding_matrix1 = compute_embeddings(
|
417 |
+
# texts1,
|
418 |
+
# batch_size=64,
|
419 |
+
# progress=progress,
|
420 |
+
# desc="Computing embeddings for Dataset 1",
|
421 |
+
# )
|
422 |
+
|
423 |
+
# # Compute embeddings for Dataset 2
|
424 |
+
# status = "Computing embeddings for Dataset 2..."
|
425 |
+
# yield status, ""
|
426 |
+
# embedding_matrix2 = compute_embeddings(
|
427 |
+
# texts2,
|
428 |
+
# batch_size=64,
|
429 |
+
# progress=progress,
|
430 |
+
# desc="Computing embeddings for Dataset 2",
|
431 |
+
# )
|
432 |
+
|
433 |
+
# # Deduplicate across datasets
|
434 |
+
# status = "Deduplicating embeddings across datasets..."
|
435 |
+
# yield status, ""
|
436 |
+
# duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(
|
437 |
+
# embedding_matrix1, embedding_matrix2, threshold, progress=progress
|
438 |
+
# )
|
439 |
|
440 |
# num_duplicates = len(duplicate_indices_in_ds2)
|
441 |
# num_total_ds2 = len(texts2)
|
|
|
445 |
# result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
446 |
# result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
447 |
|
448 |
+
# # Show deduplicated examples
|
449 |
# if num_duplicates > 0:
|
450 |
# result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
451 |
# num_examples = min(5, num_duplicates)
|
|
|
461 |
# else:
|
462 |
# result_text += "No duplicates found."
|
463 |
|
464 |
+
# # Final status
|
465 |
+
# status = "Deduplication completed."
|
466 |
+
# yield status, result_text
|
467 |
|
468 |
# except Exception as e:
|
469 |
# yield f"An error occurred: {e}", ""
|
470 |
+
# raise e
|
471 |
+
|
472 |
+
# def deduplicate_across_datasets(
|
473 |
+
# embedding_matrix_1: np.ndarray,
|
474 |
+
# embedding_matrix_2: np.ndarray,
|
475 |
+
# threshold: float,
|
476 |
+
# batch_size: int = 1024,
|
477 |
+
# progress=None
|
478 |
+
# ) -> tuple[list[int], dict[int, int]]:
|
479 |
+
# # Building the index from Dataset 1
|
480 |
+
# progress(0, desc="Building search index from Dataset 1...")
|
481 |
+
# reach = Reach(
|
482 |
+
# vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))]
|
483 |
+
# )
|
484 |
+
|
485 |
+
# duplicate_indices_in_test = []
|
486 |
+
# duplicate_to_original_mapping = {}
|
487 |
+
|
488 |
+
# # Finding nearest neighbors between datasets
|
489 |
+
# progress(0, desc="Finding nearest neighbors between datasets...")
|
490 |
+
# results = reach.nearest_neighbor_threshold(
|
491 |
+
# embedding_matrix_2,
|
492 |
+
# threshold=threshold,
|
493 |
+
# batch_size=batch_size,
|
494 |
+
# show_progressbar=False, # Disable internal progress bar
|
495 |
+
# )
|
496 |
|
497 |
+
# total_items = len(embedding_matrix_2)
|
498 |
+
# # Processing duplicates with a progress bar
|
499 |
+
# for i, similar_items in enumerate(
|
500 |
+
# progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)
|
501 |
+
# ):
|
502 |
+
# similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
503 |
+
|
504 |
+
# if similar_indices:
|
505 |
+
# duplicate_indices_in_test.append(i)
|
506 |
+
# duplicate_to_original_mapping[i] = similar_indices[0]
|
507 |
+
|
508 |
+
# return duplicate_indices_in_test, duplicate_to_original_mapping
|
509 |
+
|
510 |
+
# # Adjust the height of the status_output component using custom CSS
|
511 |
+
# with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
512 |
# gr.Markdown("# Semantic Deduplication")
|
513 |
|
514 |
# deduplication_type = gr.Radio(
|
515 |
# choices=["Single dataset", "Cross-dataset"],
|
516 |
# label="Deduplication Type",
|
517 |
+
# value="Single dataset",
|
518 |
# )
|
519 |
|
520 |
# with gr.Row():
|
|
|
531 |
# dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
532 |
|
533 |
# threshold = gr.Slider(
|
534 |
+
# minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold"
|
|
|
|
|
|
|
535 |
# )
|
536 |
|
537 |
# compute_button = gr.Button("Compute")
|
538 |
|
539 |
+
# # Use 'gr.Markdown' with 'elem_id' and custom CSS to adjust height
|
540 |
# status_output = gr.Markdown(elem_id="status_output")
|
541 |
+
# result_output = gr.Markdown()
|
542 |
|
543 |
+
# # Function to update the visibility of dataset2_inputs
|
544 |
# def update_visibility(deduplication_type_value):
|
545 |
# if deduplication_type_value == "Cross-dataset":
|
546 |
# return gr.update(visible=True)
|
|
|
548 |
# return gr.update(visible=False)
|
549 |
|
550 |
# deduplication_type.change(
|
551 |
+
# update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
|
|
|
|
552 |
# )
|
553 |
|
554 |
# compute_button.click(
|
|
|
561 |
# dataset2_name,
|
562 |
# dataset2_split,
|
563 |
# dataset2_text_column,
|
564 |
+
# threshold,
|
565 |
# ],
|
566 |
+
# outputs=[status_output, result_output],
|
567 |
# )
|
568 |
|
569 |
# demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|