Updates
Browse files
app.py
CHANGED
@@ -4,7 +4,6 @@ import numpy as np
|
|
4 |
import model2vec
|
5 |
from reach import Reach
|
6 |
from difflib import ndiff
|
7 |
-
import time
|
8 |
|
9 |
# Load the model at startup
|
10 |
model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
@@ -26,19 +25,7 @@ def batch_iterable(iterable, batch_size):
|
|
26 |
for i in range(0, len(iterable), batch_size):
|
27 |
yield iterable[i:i + batch_size]
|
28 |
|
29 |
-
def
|
30 |
-
"""Helper function to log the start and end times."""
|
31 |
-
current_time = time.time()
|
32 |
-
if start_time is not None:
|
33 |
-
elapsed = current_time - start_time
|
34 |
-
log_message = f"{message} - Took {elapsed:.2f} seconds"
|
35 |
-
else:
|
36 |
-
log_message = f"{message} - Started"
|
37 |
-
|
38 |
-
if logs is not None:
|
39 |
-
logs.append(log_message)
|
40 |
-
|
41 |
-
def compute_embeddings(texts, batch_size, progress, logs, desc="Computing embeddings"):
|
42 |
embeddings = []
|
43 |
total_batches = (len(texts) + batch_size - 1) // batch_size
|
44 |
for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
@@ -51,38 +38,26 @@ def deduplicate(
|
|
51 |
embedding_matrix: np.ndarray,
|
52 |
threshold: float,
|
53 |
batch_size: int = 1024,
|
54 |
-
progress=None
|
55 |
-
logs=None
|
56 |
) -> tuple[np.ndarray, dict[int, int]]:
|
57 |
-
|
58 |
-
log_time("Building search index", logs=logs)
|
59 |
-
reach = Reach(
|
60 |
-
vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
61 |
-
)
|
62 |
|
63 |
deduplicated_indices = set(range(len(embedding_matrix)))
|
64 |
duplicate_to_original_mapping = {}
|
65 |
|
66 |
-
# Finding nearest neighbors
|
67 |
-
log_time("Finding nearest neighbors", logs=logs)
|
68 |
results = reach.nearest_neighbor_threshold(
|
69 |
embedding_matrix,
|
70 |
threshold=threshold,
|
71 |
batch_size=batch_size,
|
72 |
-
show_progressbar=False,
|
73 |
)
|
74 |
|
75 |
-
# Processing duplicates with a progress bar
|
76 |
total_items = len(embedding_matrix)
|
77 |
-
|
78 |
-
for i, similar_items in enumerate(
|
79 |
-
progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
80 |
-
):
|
81 |
if i not in deduplicated_indices:
|
82 |
continue
|
83 |
|
84 |
similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
85 |
-
|
86 |
for sim_idx in similar_indices:
|
87 |
if sim_idx in deduplicated_indices:
|
88 |
deduplicated_indices.remove(sim_idx)
|
@@ -94,11 +69,6 @@ def display_word_differences(x: str, y: str) -> str:
|
|
94 |
diff = ndiff(x.split(), y.split())
|
95 |
return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
96 |
|
97 |
-
def encode_texts(texts, progress=None, logs=None):
|
98 |
-
embedding_matrix = model.encode(texts, show_progressbar=False)
|
99 |
-
log_time("Encoding texts completed", logs=logs)
|
100 |
-
return embedding_matrix
|
101 |
-
|
102 |
def perform_deduplication(
|
103 |
deduplication_type,
|
104 |
dataset1_name,
|
@@ -110,59 +80,24 @@ def perform_deduplication(
|
|
110 |
threshold=default_threshold,
|
111 |
progress=gr.Progress(track_tqdm=True),
|
112 |
):
|
113 |
-
logs = [] # To store log messages
|
114 |
try:
|
115 |
-
# Convert threshold to float
|
116 |
threshold = float(threshold)
|
117 |
|
118 |
-
# Initialize status message
|
119 |
-
log_time("Deduplication started", logs=logs)
|
120 |
-
|
121 |
if deduplication_type == "Single dataset":
|
122 |
-
|
123 |
-
start_time = time.time()
|
124 |
-
log_time("Loading Dataset 1", logs=logs)
|
125 |
-
if (
|
126 |
-
dataset1_name == default_dataset1_name
|
127 |
-
and dataset1_split == default_dataset1_split
|
128 |
-
):
|
129 |
-
ds = ds_default1
|
130 |
-
else:
|
131 |
-
ds = load_dataset(dataset1_name, split=dataset1_split)
|
132 |
-
log_time("Loading Dataset 1 completed", start_time=start_time, logs=logs)
|
133 |
-
|
134 |
-
# Extract texts
|
135 |
-
start_time = time.time()
|
136 |
-
log_time("Extracting texts from Dataset 1", logs=logs)
|
137 |
texts = [example[dataset1_text_column] for example in ds]
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
log_time("Computing embeddings for Dataset 1", logs=logs)
|
143 |
-
embedding_matrix = encode_texts(texts, progress=progress, logs=logs)
|
144 |
-
log_time("Computing embeddings for Dataset 1 completed", start_time=start_time, logs=logs)
|
145 |
-
|
146 |
-
# Deduplicate
|
147 |
-
start_time = time.time()
|
148 |
-
log_time("Deduplicating embeddings", logs=logs)
|
149 |
-
deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
150 |
-
embedding_matrix, threshold, progress=progress, logs=logs
|
151 |
-
)
|
152 |
-
log_time("Deduplication completed", start_time=start_time, logs=logs)
|
153 |
-
|
154 |
-
# Prepare the results
|
155 |
num_duplicates = len(duplicate_to_original_mapping)
|
156 |
num_total = len(texts)
|
157 |
num_deduplicated = len(deduplicated_indices)
|
158 |
|
159 |
result_text = f"**Total documents:** {num_total}\n"
|
160 |
result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
161 |
-
result_text +=
|
162 |
-
f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
163 |
-
)
|
164 |
|
165 |
-
# Show deduplicated examples
|
166 |
if num_duplicates > 0:
|
167 |
result_text += "**Examples of duplicates found:**\n\n"
|
168 |
num_examples = min(5, num_duplicates)
|
@@ -177,16 +112,12 @@ def perform_deduplication(
|
|
177 |
else:
|
178 |
result_text += "No duplicates found."
|
179 |
|
180 |
-
|
181 |
-
full_log = "\n".join(logs) # Combine all logs into one output
|
182 |
-
yield full_log, result_text
|
183 |
|
184 |
except Exception as e:
|
185 |
-
|
186 |
-
yield f"An error occurred: {e}", ""
|
187 |
-
raise e
|
188 |
|
189 |
-
#
|
190 |
with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
191 |
gr.Markdown("# Semantic Deduplication")
|
192 |
|
@@ -209,22 +140,14 @@ with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
|
209 |
dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
210 |
dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
211 |
|
212 |
-
threshold = gr.Slider(
|
213 |
-
minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold"
|
214 |
-
)
|
215 |
|
216 |
compute_button = gr.Button("Compute")
|
217 |
|
218 |
-
# Use 'gr.Markdown' with 'elem_id' and custom CSS to adjust height
|
219 |
-
status_output = gr.Markdown(elem_id="status_output")
|
220 |
result_output = gr.Markdown()
|
221 |
|
222 |
-
# Function to update the visibility of dataset2_inputs
|
223 |
def update_visibility(deduplication_type_value):
|
224 |
-
if deduplication_type_value == "Cross-dataset"
|
225 |
-
return gr.update(visible=True)
|
226 |
-
else:
|
227 |
-
return gr.update(visible=False)
|
228 |
|
229 |
deduplication_type.change(
|
230 |
update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
@@ -242,21 +165,19 @@ with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
|
242 |
dataset2_text_column,
|
243 |
threshold,
|
244 |
],
|
245 |
-
outputs=[
|
246 |
)
|
247 |
|
248 |
demo.launch()
|
249 |
|
250 |
|
251 |
-
|
252 |
# import gradio as gr
|
253 |
# from datasets import load_dataset
|
254 |
# import numpy as np
|
255 |
-
# #from model2vec import StaticModel
|
256 |
# import model2vec
|
257 |
# from reach import Reach
|
258 |
# from difflib import ndiff
|
259 |
-
|
260 |
|
261 |
# # Load the model at startup
|
262 |
# model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
@@ -273,13 +194,24 @@ demo.launch()
|
|
273 |
# ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
274 |
# ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
275 |
|
276 |
-
|
277 |
# def batch_iterable(iterable, batch_size):
|
278 |
# """Helper function to create batches from an iterable."""
|
279 |
# for i in range(0, len(iterable), batch_size):
|
280 |
# yield iterable[i:i + batch_size]
|
281 |
|
282 |
-
# def
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
283 |
# embeddings = []
|
284 |
# total_batches = (len(texts) + batch_size - 1) // batch_size
|
285 |
# for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
@@ -292,10 +224,11 @@ demo.launch()
|
|
292 |
# embedding_matrix: np.ndarray,
|
293 |
# threshold: float,
|
294 |
# batch_size: int = 1024,
|
295 |
-
# progress=None
|
|
|
296 |
# ) -> tuple[np.ndarray, dict[int, int]]:
|
297 |
# # Building the index
|
298 |
-
#
|
299 |
# reach = Reach(
|
300 |
# vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
301 |
# )
|
@@ -304,7 +237,7 @@ demo.launch()
|
|
304 |
# duplicate_to_original_mapping = {}
|
305 |
|
306 |
# # Finding nearest neighbors
|
307 |
-
#
|
308 |
# results = reach.nearest_neighbor_threshold(
|
309 |
# embedding_matrix,
|
310 |
# threshold=threshold,
|
@@ -314,6 +247,7 @@ demo.launch()
|
|
314 |
|
315 |
# # Processing duplicates with a progress bar
|
316 |
# total_items = len(embedding_matrix)
|
|
|
317 |
# for i, similar_items in enumerate(
|
318 |
# progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
319 |
# ):
|
@@ -333,9 +267,9 @@ demo.launch()
|
|
333 |
# diff = ndiff(x.split(), y.split())
|
334 |
# return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
335 |
|
336 |
-
|
337 |
-
# def encode_texts(texts, progress=None):
|
338 |
# embedding_matrix = model.encode(texts, show_progressbar=False)
|
|
|
339 |
# return embedding_matrix
|
340 |
|
341 |
# def perform_deduplication(
|
@@ -349,17 +283,18 @@ demo.launch()
|
|
349 |
# threshold=default_threshold,
|
350 |
# progress=gr.Progress(track_tqdm=True),
|
351 |
# ):
|
|
|
352 |
# try:
|
353 |
# # Convert threshold to float
|
354 |
# threshold = float(threshold)
|
355 |
|
356 |
# # Initialize status message
|
357 |
-
#
|
358 |
|
359 |
# if deduplication_type == "Single dataset":
|
360 |
# # Load Dataset 1
|
361 |
-
#
|
362 |
-
#
|
363 |
# if (
|
364 |
# dataset1_name == default_dataset1_name
|
365 |
# and dataset1_split == default_dataset1_split
|
@@ -367,29 +302,27 @@ demo.launch()
|
|
367 |
# ds = ds_default1
|
368 |
# else:
|
369 |
# ds = load_dataset(dataset1_name, split=dataset1_split)
|
|
|
370 |
|
371 |
# # Extract texts
|
372 |
-
#
|
373 |
-
#
|
374 |
# texts = [example[dataset1_text_column] for example in ds]
|
|
|
|
|
375 |
# # Compute embeddings
|
376 |
-
#
|
377 |
-
#
|
378 |
-
# embedding_matrix = encode_texts(texts, progress=progress)
|
379 |
-
#
|
380 |
-
# # embedding_matrix = compute_embeddings(
|
381 |
-
# # texts,
|
382 |
-
# # batch_size=64,
|
383 |
-
# # progress=progress,
|
384 |
-
# # desc="Computing embeddings for Dataset 1",
|
385 |
-
# # )
|
386 |
|
387 |
# # Deduplicate
|
388 |
-
#
|
389 |
-
#
|
390 |
# deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
391 |
-
# embedding_matrix, threshold, progress=progress
|
392 |
# )
|
|
|
393 |
|
394 |
# # Prepare the results
|
395 |
# num_duplicates = len(duplicate_to_original_mapping)
|
@@ -417,141 +350,15 @@ demo.launch()
|
|
417 |
# else:
|
418 |
# result_text += "No duplicates found."
|
419 |
|
420 |
-
#
|
421 |
-
#
|
422 |
-
# yield
|
423 |
-
|
424 |
-
# elif deduplication_type == "Cross-dataset":
|
425 |
-
# # Similar code for cross-dataset deduplication
|
426 |
-
# # Load Dataset 1
|
427 |
-
# status = "Loading Dataset 1..."
|
428 |
-
# yield status, ""
|
429 |
-
# if (
|
430 |
-
# dataset1_name == default_dataset1_name
|
431 |
-
# and dataset1_split == default_dataset1_split
|
432 |
-
# ):
|
433 |
-
# ds1 = ds_default1
|
434 |
-
# else:
|
435 |
-
# ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
436 |
-
|
437 |
-
# # Load Dataset 2
|
438 |
-
# status = "Loading Dataset 2..."
|
439 |
-
# yield status, ""
|
440 |
-
# if (
|
441 |
-
# dataset2_name == default_dataset2_name
|
442 |
-
# and dataset2_split == default_dataset2_split
|
443 |
-
# ):
|
444 |
-
# ds2 = ds_default2
|
445 |
-
# else:
|
446 |
-
# ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
447 |
-
|
448 |
-
# # Extract texts from Dataset 1
|
449 |
-
# status = "Extracting texts from Dataset 1..."
|
450 |
-
# yield status, ""
|
451 |
-
# texts1 = [example[dataset1_text_column] for example in ds1]
|
452 |
-
|
453 |
-
# # Extract texts from Dataset 2
|
454 |
-
# status = "Extracting texts from Dataset 2..."
|
455 |
-
# yield status, ""
|
456 |
-
# texts2 = [example[dataset2_text_column] for example in ds2]
|
457 |
-
|
458 |
-
# # Compute embeddings for Dataset 1
|
459 |
-
# status = "Computing embeddings for Dataset 1..."
|
460 |
-
# yield status, ""
|
461 |
-
# embedding_matrix1 = compute_embeddings(
|
462 |
-
# texts1,
|
463 |
-
# batch_size=64,
|
464 |
-
# progress=progress,
|
465 |
-
# desc="Computing embeddings for Dataset 1",
|
466 |
-
# )
|
467 |
-
|
468 |
-
# # Compute embeddings for Dataset 2
|
469 |
-
# status = "Computing embeddings for Dataset 2..."
|
470 |
-
# yield status, ""
|
471 |
-
# embedding_matrix2 = compute_embeddings(
|
472 |
-
# texts2,
|
473 |
-
# batch_size=64,
|
474 |
-
# progress=progress,
|
475 |
-
# desc="Computing embeddings for Dataset 2",
|
476 |
-
# )
|
477 |
-
|
478 |
-
# # Deduplicate across datasets
|
479 |
-
# status = "Deduplicating embeddings across datasets..."
|
480 |
-
# yield status, ""
|
481 |
-
# duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(
|
482 |
-
# embedding_matrix1, embedding_matrix2, threshold, progress=progress
|
483 |
-
# )
|
484 |
-
|
485 |
-
# num_duplicates = len(duplicate_indices_in_ds2)
|
486 |
-
# num_total_ds2 = len(texts2)
|
487 |
-
# num_unique_ds2 = num_total_ds2 - num_duplicates
|
488 |
-
|
489 |
-
# result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
490 |
-
# result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
491 |
-
# result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
492 |
-
|
493 |
-
# # Show deduplicated examples
|
494 |
-
# if num_duplicates > 0:
|
495 |
-
# result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
496 |
-
# num_examples = min(5, num_duplicates)
|
497 |
-
# for duplicate_idx in duplicate_indices_in_ds2[:num_examples]:
|
498 |
-
# original_idx = duplicate_to_original_mapping[duplicate_idx]
|
499 |
-
# original_text = texts1[original_idx]
|
500 |
-
# duplicate_text = texts2[duplicate_idx]
|
501 |
-
# differences = display_word_differences(original_text, duplicate_text)
|
502 |
-
# result_text += f"**Original text (Dataset 1):**\n{original_text}\n\n"
|
503 |
-
# result_text += f"**Duplicate text (Dataset 2):**\n{duplicate_text}\n\n"
|
504 |
-
# result_text += f"**Differences:**\n{differences}\n"
|
505 |
-
# result_text += "-" * 50 + "\n\n"
|
506 |
-
# else:
|
507 |
-
# result_text += "No duplicates found."
|
508 |
-
|
509 |
-
# # Final status
|
510 |
-
# status = "Deduplication completed."
|
511 |
-
# yield status, result_text
|
512 |
|
513 |
# except Exception as e:
|
|
|
514 |
# yield f"An error occurred: {e}", ""
|
515 |
# raise e
|
516 |
|
517 |
-
# def deduplicate_across_datasets(
|
518 |
-
# embedding_matrix_1: np.ndarray,
|
519 |
-
# embedding_matrix_2: np.ndarray,
|
520 |
-
# threshold: float,
|
521 |
-
# batch_size: int = 1024,
|
522 |
-
# progress=None
|
523 |
-
# ) -> tuple[list[int], dict[int, int]]:
|
524 |
-
# # Building the index from Dataset 1
|
525 |
-
# progress(0, desc="Building search index from Dataset 1...")
|
526 |
-
# reach = Reach(
|
527 |
-
# vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))]
|
528 |
-
# )
|
529 |
-
|
530 |
-
# duplicate_indices_in_test = []
|
531 |
-
# duplicate_to_original_mapping = {}
|
532 |
-
|
533 |
-
# # Finding nearest neighbors between datasets
|
534 |
-
# progress(0, desc="Finding nearest neighbors between datasets...")
|
535 |
-
# results = reach.nearest_neighbor_threshold(
|
536 |
-
# embedding_matrix_2,
|
537 |
-
# threshold=threshold,
|
538 |
-
# batch_size=batch_size,
|
539 |
-
# show_progressbar=False, # Disable internal progress bar
|
540 |
-
# )
|
541 |
-
|
542 |
-
# total_items = len(embedding_matrix_2)
|
543 |
-
# # Processing duplicates with a progress bar
|
544 |
-
# for i, similar_items in enumerate(
|
545 |
-
# progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)
|
546 |
-
# ):
|
547 |
-
# similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
548 |
-
|
549 |
-
# if similar_indices:
|
550 |
-
# duplicate_indices_in_test.append(i)
|
551 |
-
# duplicate_to_original_mapping[i] = similar_indices[0]
|
552 |
-
|
553 |
-
# return duplicate_indices_in_test, duplicate_to_original_mapping
|
554 |
-
|
555 |
# # Adjust the height of the status_output component using custom CSS
|
556 |
# with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
557 |
# gr.Markdown("# Semantic Deduplication")
|
@@ -612,3 +419,369 @@ demo.launch()
|
|
612 |
# )
|
613 |
|
614 |
# demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
import model2vec
|
5 |
from reach import Reach
|
6 |
from difflib import ndiff
|
|
|
7 |
|
8 |
# Load the model at startup
|
9 |
model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
|
|
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="Computing embeddings"):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
embeddings = []
|
30 |
total_batches = (len(texts) + batch_size - 1) // batch_size
|
31 |
for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
|
|
38 |
embedding_matrix: np.ndarray,
|
39 |
threshold: float,
|
40 |
batch_size: int = 1024,
|
41 |
+
progress=None
|
|
|
42 |
) -> tuple[np.ndarray, dict[int, int]]:
|
43 |
+
reach = Reach(vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))])
|
|
|
|
|
|
|
|
|
44 |
|
45 |
deduplicated_indices = set(range(len(embedding_matrix)))
|
46 |
duplicate_to_original_mapping = {}
|
47 |
|
|
|
|
|
48 |
results = reach.nearest_neighbor_threshold(
|
49 |
embedding_matrix,
|
50 |
threshold=threshold,
|
51 |
batch_size=batch_size,
|
52 |
+
show_progressbar=False,
|
53 |
)
|
54 |
|
|
|
55 |
total_items = len(embedding_matrix)
|
56 |
+
for i, similar_items in enumerate(progress.tqdm(results, desc="Processing duplicates", total=total_items)):
|
|
|
|
|
|
|
57 |
if i not in deduplicated_indices:
|
58 |
continue
|
59 |
|
60 |
similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
|
|
61 |
for sim_idx in similar_indices:
|
62 |
if sim_idx in deduplicated_indices:
|
63 |
deduplicated_indices.remove(sim_idx)
|
|
|
69 |
diff = ndiff(x.split(), y.split())
|
70 |
return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
71 |
|
|
|
|
|
|
|
|
|
|
|
72 |
def perform_deduplication(
|
73 |
deduplication_type,
|
74 |
dataset1_name,
|
|
|
80 |
threshold=default_threshold,
|
81 |
progress=gr.Progress(track_tqdm=True),
|
82 |
):
|
|
|
83 |
try:
|
|
|
84 |
threshold = float(threshold)
|
85 |
|
|
|
|
|
|
|
86 |
if deduplication_type == "Single dataset":
|
87 |
+
ds = ds_default1 if dataset1_name == default_dataset1_name and dataset1_split == default_dataset1_split else load_dataset(dataset1_name, split=dataset1_split)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
88 |
texts = [example[dataset1_text_column] for example in ds]
|
89 |
+
|
90 |
+
embedding_matrix = compute_embeddings(texts, batch_size=64, progress=progress)
|
91 |
+
deduplicated_indices, duplicate_to_original_mapping = deduplicate(embedding_matrix, threshold, progress=progress)
|
92 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
num_duplicates = len(duplicate_to_original_mapping)
|
94 |
num_total = len(texts)
|
95 |
num_deduplicated = len(deduplicated_indices)
|
96 |
|
97 |
result_text = f"**Total documents:** {num_total}\n"
|
98 |
result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
99 |
+
result_text += f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
|
|
|
|
100 |
|
|
|
101 |
if num_duplicates > 0:
|
102 |
result_text += "**Examples of duplicates found:**\n\n"
|
103 |
num_examples = min(5, num_duplicates)
|
|
|
112 |
else:
|
113 |
result_text += "No duplicates found."
|
114 |
|
115 |
+
yield result_text
|
|
|
|
|
116 |
|
117 |
except Exception as e:
|
118 |
+
yield f"An error occurred: {e}"
|
|
|
|
|
119 |
|
120 |
+
# Gradio interface setup
|
121 |
with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
122 |
gr.Markdown("# Semantic Deduplication")
|
123 |
|
|
|
140 |
dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
141 |
dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
142 |
|
143 |
+
threshold = gr.Slider(minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold")
|
|
|
|
|
144 |
|
145 |
compute_button = gr.Button("Compute")
|
146 |
|
|
|
|
|
147 |
result_output = gr.Markdown()
|
148 |
|
|
|
149 |
def update_visibility(deduplication_type_value):
|
150 |
+
return gr.update(visible=True) if deduplication_type_value == "Cross-dataset" else gr.update(visible=False)
|
|
|
|
|
|
|
151 |
|
152 |
deduplication_type.change(
|
153 |
update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
|
|
165 |
dataset2_text_column,
|
166 |
threshold,
|
167 |
],
|
168 |
+
outputs=[result_output],
|
169 |
)
|
170 |
|
171 |
demo.launch()
|
172 |
|
173 |
|
|
|
174 |
# import gradio as gr
|
175 |
# from datasets import load_dataset
|
176 |
# import numpy as np
|
|
|
177 |
# import model2vec
|
178 |
# from reach import Reach
|
179 |
# from difflib import ndiff
|
180 |
+
# import time
|
181 |
|
182 |
# # Load the model at startup
|
183 |
# model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
|
|
194 |
# ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
195 |
# ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
196 |
|
|
|
197 |
# def batch_iterable(iterable, batch_size):
|
198 |
# """Helper function to create batches from an iterable."""
|
199 |
# for i in range(0, len(iterable), batch_size):
|
200 |
# yield iterable[i:i + batch_size]
|
201 |
|
202 |
+
# def log_time(message, start_time=None, logs=None):
|
203 |
+
# """Helper function to log the start and end times."""
|
204 |
+
# current_time = time.time()
|
205 |
+
# if start_time is not None:
|
206 |
+
# elapsed = current_time - start_time
|
207 |
+
# log_message = f"{message} - Took {elapsed:.2f} seconds"
|
208 |
+
# else:
|
209 |
+
# log_message = f"{message} - Started"
|
210 |
+
|
211 |
+
# if logs is not None:
|
212 |
+
# logs.append(log_message)
|
213 |
+
|
214 |
+
# def compute_embeddings(texts, batch_size, progress, logs, desc="Computing embeddings"):
|
215 |
# embeddings = []
|
216 |
# total_batches = (len(texts) + batch_size - 1) // batch_size
|
217 |
# for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
|
|
224 |
# embedding_matrix: np.ndarray,
|
225 |
# threshold: float,
|
226 |
# batch_size: int = 1024,
|
227 |
+
# progress=None,
|
228 |
+
# logs=None
|
229 |
# ) -> tuple[np.ndarray, dict[int, int]]:
|
230 |
# # Building the index
|
231 |
+
# log_time("Building search index", logs=logs)
|
232 |
# reach = Reach(
|
233 |
# vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
234 |
# )
|
|
|
237 |
# duplicate_to_original_mapping = {}
|
238 |
|
239 |
# # Finding nearest neighbors
|
240 |
+
# log_time("Finding nearest neighbors", logs=logs)
|
241 |
# results = reach.nearest_neighbor_threshold(
|
242 |
# embedding_matrix,
|
243 |
# threshold=threshold,
|
|
|
247 |
|
248 |
# # Processing duplicates with a progress bar
|
249 |
# total_items = len(embedding_matrix)
|
250 |
+
# log_time("Processing duplicates", logs=logs)
|
251 |
# for i, similar_items in enumerate(
|
252 |
# progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
253 |
# ):
|
|
|
267 |
# diff = ndiff(x.split(), y.split())
|
268 |
# return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
269 |
|
270 |
+
# def encode_texts(texts, progress=None, logs=None):
|
|
|
271 |
# embedding_matrix = model.encode(texts, show_progressbar=False)
|
272 |
+
# log_time("Encoding texts completed", logs=logs)
|
273 |
# return embedding_matrix
|
274 |
|
275 |
# def perform_deduplication(
|
|
|
283 |
# threshold=default_threshold,
|
284 |
# progress=gr.Progress(track_tqdm=True),
|
285 |
# ):
|
286 |
+
# logs = [] # To store log messages
|
287 |
# try:
|
288 |
# # Convert threshold to float
|
289 |
# threshold = float(threshold)
|
290 |
|
291 |
# # Initialize status message
|
292 |
+
# log_time("Deduplication started", logs=logs)
|
293 |
|
294 |
# if deduplication_type == "Single dataset":
|
295 |
# # Load Dataset 1
|
296 |
+
# start_time = time.time()
|
297 |
+
# log_time("Loading Dataset 1", logs=logs)
|
298 |
# if (
|
299 |
# dataset1_name == default_dataset1_name
|
300 |
# and dataset1_split == default_dataset1_split
|
|
|
302 |
# ds = ds_default1
|
303 |
# else:
|
304 |
# ds = load_dataset(dataset1_name, split=dataset1_split)
|
305 |
+
# log_time("Loading Dataset 1 completed", start_time=start_time, logs=logs)
|
306 |
|
307 |
# # Extract texts
|
308 |
+
# start_time = time.time()
|
309 |
+
# log_time("Extracting texts from Dataset 1", logs=logs)
|
310 |
# texts = [example[dataset1_text_column] for example in ds]
|
311 |
+
# log_time("Extracting texts from Dataset 1 completed", start_time=start_time, logs=logs)
|
312 |
+
|
313 |
# # Compute embeddings
|
314 |
+
# start_time = time.time()
|
315 |
+
# log_time("Computing embeddings for Dataset 1", logs=logs)
|
316 |
+
# embedding_matrix = encode_texts(texts, progress=progress, logs=logs)
|
317 |
+
# log_time("Computing embeddings for Dataset 1 completed", start_time=start_time, logs=logs)
|
|
|
|
|
|
|
|
|
|
|
|
|
318 |
|
319 |
# # Deduplicate
|
320 |
+
# start_time = time.time()
|
321 |
+
# log_time("Deduplicating embeddings", logs=logs)
|
322 |
# deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
323 |
+
# embedding_matrix, threshold, progress=progress, logs=logs
|
324 |
# )
|
325 |
+
# log_time("Deduplication completed", start_time=start_time, logs=logs)
|
326 |
|
327 |
# # Prepare the results
|
328 |
# num_duplicates = len(duplicate_to_original_mapping)
|
|
|
350 |
# else:
|
351 |
# result_text += "No duplicates found."
|
352 |
|
353 |
+
# log_time("Deduplication process finished", logs=logs)
|
354 |
+
# full_log = "\n".join(logs) # Combine all logs into one output
|
355 |
+
# yield full_log, result_text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
356 |
|
357 |
# except Exception as e:
|
358 |
+
# full_log = "\n".join(logs) # Combine all logs into one output in case of an error
|
359 |
# yield f"An error occurred: {e}", ""
|
360 |
# raise e
|
361 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
362 |
# # Adjust the height of the status_output component using custom CSS
|
363 |
# with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
364 |
# gr.Markdown("# Semantic Deduplication")
|
|
|
419 |
# )
|
420 |
|
421 |
# demo.launch()
|
422 |
+
|
423 |
+
|
424 |
+
|
425 |
+
# # import gradio as gr
|
426 |
+
# # from datasets import load_dataset
|
427 |
+
# # import numpy as np
|
428 |
+
# # #from model2vec import StaticModel
|
429 |
+
# # import model2vec
|
430 |
+
# # from reach import Reach
|
431 |
+
# # from difflib import ndiff
|
432 |
+
|
433 |
+
|
434 |
+
# # # Load the model at startup
|
435 |
+
# # model = model2vec.StaticModel.from_pretrained("minishlab/M2V_base_output")
|
436 |
+
|
437 |
+
# # # Default dataset parameters
|
438 |
+
# # default_dataset1_name = "sst2"
|
439 |
+
# # default_dataset1_split = "train"
|
440 |
+
# # default_dataset2_name = "sst2"
|
441 |
+
# # default_dataset2_split = "validation"
|
442 |
+
# # default_text_column = "sentence"
|
443 |
+
# # default_threshold = 0.9
|
444 |
+
|
445 |
+
# # # Load the default datasets at startup
|
446 |
+
# # ds_default1 = load_dataset(default_dataset1_name, split=default_dataset1_split)
|
447 |
+
# # ds_default2 = load_dataset(default_dataset2_name, split=default_dataset2_split)
|
448 |
+
|
449 |
+
|
450 |
+
# # def batch_iterable(iterable, batch_size):
|
451 |
+
# # """Helper function to create batches from an iterable."""
|
452 |
+
# # for i in range(0, len(iterable), batch_size):
|
453 |
+
# # yield iterable[i:i + batch_size]
|
454 |
+
|
455 |
+
# # def compute_embeddings(texts, batch_size, progress, desc="Computing embeddings"):
|
456 |
+
# # embeddings = []
|
457 |
+
# # total_batches = (len(texts) + batch_size - 1) // batch_size
|
458 |
+
# # for i, batch_texts in enumerate(batch_iterable(texts, batch_size)):
|
459 |
+
# # batch_embeddings = model.encode(batch_texts, show_progressbar=False)
|
460 |
+
# # embeddings.append(batch_embeddings)
|
461 |
+
# # progress((i + 1) / total_batches, desc=desc)
|
462 |
+
# # return np.concatenate(embeddings, axis=0)
|
463 |
+
|
464 |
+
# # def deduplicate(
|
465 |
+
# # embedding_matrix: np.ndarray,
|
466 |
+
# # threshold: float,
|
467 |
+
# # batch_size: int = 1024,
|
468 |
+
# # progress=None
|
469 |
+
# # ) -> tuple[np.ndarray, dict[int, int]]:
|
470 |
+
# # # Building the index
|
471 |
+
# # progress(0, desc="Building search index...")
|
472 |
+
# # reach = Reach(
|
473 |
+
# # vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))]
|
474 |
+
# # )
|
475 |
+
|
476 |
+
# # deduplicated_indices = set(range(len(embedding_matrix)))
|
477 |
+
# # duplicate_to_original_mapping = {}
|
478 |
+
|
479 |
+
# # # Finding nearest neighbors
|
480 |
+
# # progress(0, desc="Finding nearest neighbors...")
|
481 |
+
# # results = reach.nearest_neighbor_threshold(
|
482 |
+
# # embedding_matrix,
|
483 |
+
# # threshold=threshold,
|
484 |
+
# # batch_size=batch_size,
|
485 |
+
# # show_progressbar=False, # Disable internal progress bar
|
486 |
+
# # )
|
487 |
+
|
488 |
+
# # # Processing duplicates with a progress bar
|
489 |
+
# # total_items = len(embedding_matrix)
|
490 |
+
# # for i, similar_items in enumerate(
|
491 |
+
# # progress.tqdm(results, desc="Processing duplicates", total=total_items)
|
492 |
+
# # ):
|
493 |
+
# # if i not in deduplicated_indices:
|
494 |
+
# # continue
|
495 |
+
|
496 |
+
# # similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
497 |
+
|
498 |
+
# # for sim_idx in similar_indices:
|
499 |
+
# # if sim_idx in deduplicated_indices:
|
500 |
+
# # deduplicated_indices.remove(sim_idx)
|
501 |
+
# # duplicate_to_original_mapping[sim_idx] = i
|
502 |
+
|
503 |
+
# # return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
504 |
+
|
505 |
+
# # def display_word_differences(x: str, y: str) -> str:
|
506 |
+
# # diff = ndiff(x.split(), y.split())
|
507 |
+
# # return " ".join([word for word in diff if word.startswith(("+", "-"))])
|
508 |
+
|
509 |
+
|
510 |
+
# # def encode_texts(texts, progress=None):
|
511 |
+
# # embedding_matrix = model.encode(texts, show_progressbar=False)
|
512 |
+
# # return embedding_matrix
|
513 |
+
|
514 |
+
# # def perform_deduplication(
|
515 |
+
# # deduplication_type,
|
516 |
+
# # dataset1_name,
|
517 |
+
# # dataset1_split,
|
518 |
+
# # dataset1_text_column,
|
519 |
+
# # dataset2_name="",
|
520 |
+
# # dataset2_split="",
|
521 |
+
# # dataset2_text_column="",
|
522 |
+
# # threshold=default_threshold,
|
523 |
+
# # progress=gr.Progress(track_tqdm=True),
|
524 |
+
# # ):
|
525 |
+
# # try:
|
526 |
+
# # # Convert threshold to float
|
527 |
+
# # threshold = float(threshold)
|
528 |
+
|
529 |
+
# # # Initialize status message
|
530 |
+
# # status = ""
|
531 |
+
|
532 |
+
# # if deduplication_type == "Single dataset":
|
533 |
+
# # # Load Dataset 1
|
534 |
+
# # status = "Loading Dataset 1..."
|
535 |
+
# # yield status, ""
|
536 |
+
# # if (
|
537 |
+
# # dataset1_name == default_dataset1_name
|
538 |
+
# # and dataset1_split == default_dataset1_split
|
539 |
+
# # ):
|
540 |
+
# # ds = ds_default1
|
541 |
+
# # else:
|
542 |
+
# # ds = load_dataset(dataset1_name, split=dataset1_split)
|
543 |
+
|
544 |
+
# # # Extract texts
|
545 |
+
# # status = "Extracting texts from Dataset 1..."
|
546 |
+
# # yield status, ""
|
547 |
+
# # texts = [example[dataset1_text_column] for example in ds]
|
548 |
+
# # # Compute embeddings
|
549 |
+
# # status = "Computing embeddings for Dataset 1..."
|
550 |
+
# # yield status, ""
|
551 |
+
# # embedding_matrix = encode_texts(texts, progress=progress)
|
552 |
+
# # #embedding_matrix = model.encode(texts, show_progressbar=True)
|
553 |
+
# # # embedding_matrix = compute_embeddings(
|
554 |
+
# # # texts,
|
555 |
+
# # # batch_size=64,
|
556 |
+
# # # progress=progress,
|
557 |
+
# # # desc="Computing embeddings for Dataset 1",
|
558 |
+
# # # )
|
559 |
+
|
560 |
+
# # # Deduplicate
|
561 |
+
# # status = "Deduplicating embeddings..."
|
562 |
+
# # yield status, ""
|
563 |
+
# # deduplicated_indices, duplicate_to_original_mapping = deduplicate(
|
564 |
+
# # embedding_matrix, threshold, progress=progress
|
565 |
+
# # )
|
566 |
+
|
567 |
+
# # # Prepare the results
|
568 |
+
# # num_duplicates = len(duplicate_to_original_mapping)
|
569 |
+
# # num_total = len(texts)
|
570 |
+
# # num_deduplicated = len(deduplicated_indices)
|
571 |
+
|
572 |
+
# # result_text = f"**Total documents:** {num_total}\n"
|
573 |
+
# # result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
574 |
+
# # result_text += (
|
575 |
+
# # f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
576 |
+
# # )
|
577 |
+
|
578 |
+
# # # Show deduplicated examples
|
579 |
+
# # if num_duplicates > 0:
|
580 |
+
# # result_text += "**Examples of duplicates found:**\n\n"
|
581 |
+
# # num_examples = min(5, num_duplicates)
|
582 |
+
# # for duplicate_idx, original_idx in list(duplicate_to_original_mapping.items())[:num_examples]:
|
583 |
+
# # original_text = texts[original_idx]
|
584 |
+
# # duplicate_text = texts[duplicate_idx]
|
585 |
+
# # differences = display_word_differences(original_text, duplicate_text)
|
586 |
+
# # result_text += f"**Original text:**\n{original_text}\n\n"
|
587 |
+
# # result_text += f"**Duplicate text:**\n{duplicate_text}\n\n"
|
588 |
+
# # result_text += f"**Differences:**\n{differences}\n"
|
589 |
+
# # result_text += "-" * 50 + "\n\n"
|
590 |
+
# # else:
|
591 |
+
# # result_text += "No duplicates found."
|
592 |
+
|
593 |
+
# # # Final status
|
594 |
+
# # status = "Deduplication completed."
|
595 |
+
# # yield status, result_text
|
596 |
+
|
597 |
+
# # elif deduplication_type == "Cross-dataset":
|
598 |
+
# # # Similar code for cross-dataset deduplication
|
599 |
+
# # # Load Dataset 1
|
600 |
+
# # status = "Loading Dataset 1..."
|
601 |
+
# # yield status, ""
|
602 |
+
# # if (
|
603 |
+
# # dataset1_name == default_dataset1_name
|
604 |
+
# # and dataset1_split == default_dataset1_split
|
605 |
+
# # ):
|
606 |
+
# # ds1 = ds_default1
|
607 |
+
# # else:
|
608 |
+
# # ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
609 |
+
|
610 |
+
# # # Load Dataset 2
|
611 |
+
# # status = "Loading Dataset 2..."
|
612 |
+
# # yield status, ""
|
613 |
+
# # if (
|
614 |
+
# # dataset2_name == default_dataset2_name
|
615 |
+
# # and dataset2_split == default_dataset2_split
|
616 |
+
# # ):
|
617 |
+
# # ds2 = ds_default2
|
618 |
+
# # else:
|
619 |
+
# # ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
620 |
+
|
621 |
+
# # # Extract texts from Dataset 1
|
622 |
+
# # status = "Extracting texts from Dataset 1..."
|
623 |
+
# # yield status, ""
|
624 |
+
# # texts1 = [example[dataset1_text_column] for example in ds1]
|
625 |
+
|
626 |
+
# # # Extract texts from Dataset 2
|
627 |
+
# # status = "Extracting texts from Dataset 2..."
|
628 |
+
# # yield status, ""
|
629 |
+
# # texts2 = [example[dataset2_text_column] for example in ds2]
|
630 |
+
|
631 |
+
# # # Compute embeddings for Dataset 1
|
632 |
+
# # status = "Computing embeddings for Dataset 1..."
|
633 |
+
# # yield status, ""
|
634 |
+
# # embedding_matrix1 = compute_embeddings(
|
635 |
+
# # texts1,
|
636 |
+
# # batch_size=64,
|
637 |
+
# # progress=progress,
|
638 |
+
# # desc="Computing embeddings for Dataset 1",
|
639 |
+
# # )
|
640 |
+
|
641 |
+
# # # Compute embeddings for Dataset 2
|
642 |
+
# # status = "Computing embeddings for Dataset 2..."
|
643 |
+
# # yield status, ""
|
644 |
+
# # embedding_matrix2 = compute_embeddings(
|
645 |
+
# # texts2,
|
646 |
+
# # batch_size=64,
|
647 |
+
# # progress=progress,
|
648 |
+
# # desc="Computing embeddings for Dataset 2",
|
649 |
+
# # )
|
650 |
+
|
651 |
+
# # # Deduplicate across datasets
|
652 |
+
# # status = "Deduplicating embeddings across datasets..."
|
653 |
+
# # yield status, ""
|
654 |
+
# # duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(
|
655 |
+
# # embedding_matrix1, embedding_matrix2, threshold, progress=progress
|
656 |
+
# # )
|
657 |
+
|
658 |
+
# # num_duplicates = len(duplicate_indices_in_ds2)
|
659 |
+
# # num_total_ds2 = len(texts2)
|
660 |
+
# # num_unique_ds2 = num_total_ds2 - num_duplicates
|
661 |
+
|
662 |
+
# # result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
663 |
+
# # result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
664 |
+
# # result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
665 |
+
|
666 |
+
# # # Show deduplicated examples
|
667 |
+
# # if num_duplicates > 0:
|
668 |
+
# # result_text += "**Examples of duplicates found in Dataset 2:**\n\n"
|
669 |
+
# # num_examples = min(5, num_duplicates)
|
670 |
+
# # for duplicate_idx in duplicate_indices_in_ds2[:num_examples]:
|
671 |
+
# # original_idx = duplicate_to_original_mapping[duplicate_idx]
|
672 |
+
# # original_text = texts1[original_idx]
|
673 |
+
# # duplicate_text = texts2[duplicate_idx]
|
674 |
+
# # differences = display_word_differences(original_text, duplicate_text)
|
675 |
+
# # result_text += f"**Original text (Dataset 1):**\n{original_text}\n\n"
|
676 |
+
# # result_text += f"**Duplicate text (Dataset 2):**\n{duplicate_text}\n\n"
|
677 |
+
# # result_text += f"**Differences:**\n{differences}\n"
|
678 |
+
# # result_text += "-" * 50 + "\n\n"
|
679 |
+
# # else:
|
680 |
+
# # result_text += "No duplicates found."
|
681 |
+
|
682 |
+
# # # Final status
|
683 |
+
# # status = "Deduplication completed."
|
684 |
+
# # yield status, result_text
|
685 |
+
|
686 |
+
# # except Exception as e:
|
687 |
+
# # yield f"An error occurred: {e}", ""
|
688 |
+
# # raise e
|
689 |
+
|
690 |
+
# # def deduplicate_across_datasets(
|
691 |
+
# # embedding_matrix_1: np.ndarray,
|
692 |
+
# # embedding_matrix_2: np.ndarray,
|
693 |
+
# # threshold: float,
|
694 |
+
# # batch_size: int = 1024,
|
695 |
+
# # progress=None
|
696 |
+
# # ) -> tuple[list[int], dict[int, int]]:
|
697 |
+
# # # Building the index from Dataset 1
|
698 |
+
# # progress(0, desc="Building search index from Dataset 1...")
|
699 |
+
# # reach = Reach(
|
700 |
+
# # vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))]
|
701 |
+
# # )
|
702 |
+
|
703 |
+
# # duplicate_indices_in_test = []
|
704 |
+
# # duplicate_to_original_mapping = {}
|
705 |
+
|
706 |
+
# # # Finding nearest neighbors between datasets
|
707 |
+
# # progress(0, desc="Finding nearest neighbors between datasets...")
|
708 |
+
# # results = reach.nearest_neighbor_threshold(
|
709 |
+
# # embedding_matrix_2,
|
710 |
+
# # threshold=threshold,
|
711 |
+
# # batch_size=batch_size,
|
712 |
+
# # show_progressbar=False, # Disable internal progress bar
|
713 |
+
# # )
|
714 |
+
|
715 |
+
# # total_items = len(embedding_matrix_2)
|
716 |
+
# # # Processing duplicates with a progress bar
|
717 |
+
# # for i, similar_items in enumerate(
|
718 |
+
# # progress.tqdm(results, desc="Processing duplicates across datasets", total=total_items)
|
719 |
+
# # ):
|
720 |
+
# # similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold]
|
721 |
+
|
722 |
+
# # if similar_indices:
|
723 |
+
# # duplicate_indices_in_test.append(i)
|
724 |
+
# # duplicate_to_original_mapping[i] = similar_indices[0]
|
725 |
+
|
726 |
+
# # return duplicate_indices_in_test, duplicate_to_original_mapping
|
727 |
+
|
728 |
+
# # # Adjust the height of the status_output component using custom CSS
|
729 |
+
# # with gr.Blocks(css="#status_output { height: 150px; overflow: auto; }") as demo:
|
730 |
+
# # gr.Markdown("# Semantic Deduplication")
|
731 |
+
|
732 |
+
# # deduplication_type = gr.Radio(
|
733 |
+
# # choices=["Single dataset", "Cross-dataset"],
|
734 |
+
# # label="Deduplication Type",
|
735 |
+
# # value="Single dataset",
|
736 |
+
# # )
|
737 |
+
|
738 |
+
# # with gr.Row():
|
739 |
+
# # dataset1_name = gr.Textbox(value=default_dataset1_name, label="Dataset 1 Name")
|
740 |
+
# # dataset1_split = gr.Textbox(value=default_dataset1_split, label="Dataset 1 Split")
|
741 |
+
# # dataset1_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
742 |
+
|
743 |
+
# # dataset2_inputs = gr.Column(visible=False)
|
744 |
+
# # with dataset2_inputs:
|
745 |
+
# # gr.Markdown("### Dataset 2")
|
746 |
+
# # with gr.Row():
|
747 |
+
# # dataset2_name = gr.Textbox(value=default_dataset2_name, label="Dataset 2 Name")
|
748 |
+
# # dataset2_split = gr.Textbox(value=default_dataset2_split, label="Dataset 2 Split")
|
749 |
+
# # dataset2_text_column = gr.Textbox(value=default_text_column, label="Text Column Name")
|
750 |
+
|
751 |
+
# # threshold = gr.Slider(
|
752 |
+
# # minimum=0.0, maximum=1.0, value=default_threshold, label="Similarity Threshold"
|
753 |
+
# # )
|
754 |
+
|
755 |
+
# # compute_button = gr.Button("Compute")
|
756 |
+
|
757 |
+
# # # Use 'gr.Markdown' with 'elem_id' and custom CSS to adjust height
|
758 |
+
# # status_output = gr.Markdown(elem_id="status_output")
|
759 |
+
# # result_output = gr.Markdown()
|
760 |
+
|
761 |
+
# # # Function to update the visibility of dataset2_inputs
|
762 |
+
# # def update_visibility(deduplication_type_value):
|
763 |
+
# # if deduplication_type_value == "Cross-dataset":
|
764 |
+
# # return gr.update(visible=True)
|
765 |
+
# # else:
|
766 |
+
# # return gr.update(visible=False)
|
767 |
+
|
768 |
+
# # deduplication_type.change(
|
769 |
+
# # update_visibility, inputs=deduplication_type, outputs=dataset2_inputs
|
770 |
+
# # )
|
771 |
+
|
772 |
+
# # compute_button.click(
|
773 |
+
# # fn=perform_deduplication,
|
774 |
+
# # inputs=[
|
775 |
+
# # deduplication_type,
|
776 |
+
# # dataset1_name,
|
777 |
+
# # dataset1_split,
|
778 |
+
# # dataset1_text_column,
|
779 |
+
# # dataset2_name,
|
780 |
+
# # dataset2_split,
|
781 |
+
# # dataset2_text_column,
|
782 |
+
# # threshold,
|
783 |
+
# # ],
|
784 |
+
# # outputs=[status_output, result_output],
|
785 |
+
# # )
|
786 |
+
|
787 |
+
# # demo.launch()
|