Updated app with code for deduplication
Browse files- app.py +179 -4
- requirements.txt +6 -0
app.py
CHANGED
@@ -1,7 +1,182 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
|
3 |
-
def
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# import gradio as gr
|
2 |
+
|
3 |
+
# def greet(name):
|
4 |
+
# return "Hello " + name + "!!"
|
5 |
+
|
6 |
+
# demo = gr.Interface(fn=greet, inputs="text", outputs="text")
|
7 |
+
# demo.launch()
|
8 |
+
|
9 |
+
|
10 |
import gradio as gr
|
11 |
+
from datasets import load_dataset
|
12 |
+
import numpy as np
|
13 |
+
from model2vec import StaticModel
|
14 |
+
from reach import Reach
|
15 |
+
from tqdm import tqdm
|
16 |
+
|
17 |
+
def deduplicate(embedding_matrix: np.ndarray, threshold: float, batch_size: int = 1024) -> tuple[np.ndarray, dict[int, int]]:
|
18 |
+
"""
|
19 |
+
Deduplicate embeddings and return the deduplicated indices and a mapping of removed indices to their corresponding original indices.
|
20 |
+
"""
|
21 |
+
reach = Reach(vectors=embedding_matrix, items=[str(i) for i in range(len(embedding_matrix))])
|
22 |
+
|
23 |
+
# Use a set for deduplicated indices and keep track of duplicates
|
24 |
+
deduplicated_indices = set(range(len(embedding_matrix))) # Start with all indices as deduplicated
|
25 |
+
duplicate_to_original_mapping = {}
|
26 |
+
|
27 |
+
results = reach.nearest_neighbor_threshold(
|
28 |
+
embedding_matrix,
|
29 |
+
threshold=threshold,
|
30 |
+
batch_size=batch_size,
|
31 |
+
show_progressbar=True
|
32 |
+
)
|
33 |
+
|
34 |
+
# Process duplicates
|
35 |
+
for i, similar_items in enumerate(tqdm(results)):
|
36 |
+
if i not in deduplicated_indices:
|
37 |
+
continue # Skip already marked duplicates
|
38 |
+
|
39 |
+
# Similar items are returned as (index, score), we are only interested in the index
|
40 |
+
similar_indices = [int(item[0]) for item in similar_items if int(item[0]) != i]
|
41 |
+
|
42 |
+
# Mark similar documents as duplicates and map them to the original
|
43 |
+
for sim_idx in similar_indices:
|
44 |
+
if sim_idx in deduplicated_indices:
|
45 |
+
deduplicated_indices.remove(sim_idx)
|
46 |
+
duplicate_to_original_mapping[sim_idx] = i # Map duplicate to original
|
47 |
+
|
48 |
+
return np.array(list(deduplicated_indices)), duplicate_to_original_mapping
|
49 |
+
|
50 |
+
def deduplicate_across_datasets(embedding_matrix_1: np.ndarray, embedding_matrix_2: np.ndarray, threshold: float, batch_size: int = 1024) -> tuple[list[int], dict[int, int]]:
|
51 |
+
"""
|
52 |
+
Deduplicate embeddings across two datasets and return the indices of duplicates between them.
|
53 |
+
"""
|
54 |
+
reach = Reach(vectors=embedding_matrix_1, items=[str(i) for i in range(len(embedding_matrix_1))])
|
55 |
+
|
56 |
+
# Keep track of duplicates in the second dataset
|
57 |
+
duplicate_indices_in_test = []
|
58 |
+
duplicate_to_original_mapping = {}
|
59 |
+
|
60 |
+
# Find nearest neighbors from the test set in the train set
|
61 |
+
results = reach.nearest_neighbor_threshold(
|
62 |
+
embedding_matrix_2,
|
63 |
+
threshold=threshold,
|
64 |
+
batch_size=batch_size,
|
65 |
+
show_progressbar=True
|
66 |
+
)
|
67 |
+
|
68 |
+
# Process duplicates
|
69 |
+
for i, similar_items in enumerate(tqdm(results)):
|
70 |
+
# Similar items are returned as (index, score), we are only interested in the index
|
71 |
+
similar_indices = [int(item[0]) for item in similar_items if item[1] >= threshold] # Keep those above the threshold
|
72 |
+
|
73 |
+
# If we find a similar item in the train set, mark it as a duplicate
|
74 |
+
if similar_indices:
|
75 |
+
duplicate_indices_in_test.append(i)
|
76 |
+
duplicate_to_original_mapping[i] = similar_indices[0] # Map duplicate in test to original in train
|
77 |
+
|
78 |
+
return duplicate_indices_in_test, duplicate_to_original_mapping
|
79 |
|
80 |
+
def perform_deduplication(
|
81 |
+
deduplication_type,
|
82 |
+
dataset1_name,
|
83 |
+
dataset1_split,
|
84 |
+
dataset2_name,
|
85 |
+
dataset2_split,
|
86 |
+
threshold
|
87 |
+
):
|
88 |
+
# Convert threshold to float
|
89 |
+
threshold = float(threshold)
|
90 |
+
|
91 |
+
if deduplication_type == "Single dataset":
|
92 |
+
# Load the dataset
|
93 |
+
ds = load_dataset(dataset1_name, split=dataset1_split)
|
94 |
+
|
95 |
+
# Extract texts
|
96 |
+
texts = [example['text'] for example in ds]
|
97 |
+
|
98 |
+
# Compute embeddings
|
99 |
+
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
100 |
+
embedding_matrix = model.encode(texts, show_progressbar=True)
|
101 |
+
|
102 |
+
# Deduplicate
|
103 |
+
deduplicated_indices, duplicate_to_original_mapping = deduplicate(embedding_matrix, threshold)
|
104 |
+
|
105 |
+
# Prepare the results
|
106 |
+
num_duplicates = len(duplicate_to_original_mapping)
|
107 |
+
num_total = len(texts)
|
108 |
+
num_deduplicated = len(deduplicated_indices)
|
109 |
+
|
110 |
+
result_text = f"**Total documents:** {num_total}\n"
|
111 |
+
result_text += f"**Number of duplicates found:** {num_duplicates}\n"
|
112 |
+
result_text += f"**Number of unique documents after deduplication:** {num_deduplicated}\n\n"
|
113 |
+
result_text += f"**Deduplicated indices:** {deduplicated_indices.tolist()}\n\n"
|
114 |
+
result_text += f"**Duplicate to original mapping:** {duplicate_to_original_mapping}\n"
|
115 |
+
|
116 |
+
return result_text
|
117 |
+
|
118 |
+
elif deduplication_type == "Cross-dataset":
|
119 |
+
# Load datasets
|
120 |
+
ds1 = load_dataset(dataset1_name, split=dataset1_split)
|
121 |
+
ds2 = load_dataset(dataset2_name, split=dataset2_split)
|
122 |
+
|
123 |
+
# Extract texts
|
124 |
+
texts1 = [example['text'] for example in ds1]
|
125 |
+
texts2 = [example['text'] for example in ds2]
|
126 |
+
|
127 |
+
# Compute embeddings
|
128 |
+
model = StaticModel.from_pretrained("minishlab/M2V_base_output")
|
129 |
+
embedding_matrix1 = model.encode(texts1, show_progressbar=True)
|
130 |
+
embedding_matrix2 = model.encode(texts2, show_progressbar=True)
|
131 |
+
|
132 |
+
# Deduplicate across datasets
|
133 |
+
duplicate_indices_in_ds2, duplicate_to_original_mapping = deduplicate_across_datasets(embedding_matrix1, embedding_matrix2, threshold)
|
134 |
+
|
135 |
+
num_duplicates = len(duplicate_indices_in_ds2)
|
136 |
+
num_total_ds2 = len(texts2)
|
137 |
+
num_unique_ds2 = num_total_ds2 - num_duplicates
|
138 |
+
|
139 |
+
result_text = f"**Total documents in {dataset2_name}/{dataset2_split}:** {num_total_ds2}\n"
|
140 |
+
result_text += f"**Number of duplicates found in {dataset2_name}/{dataset2_split}:** {num_duplicates}\n"
|
141 |
+
result_text += f"**Number of unique documents in {dataset2_name}/{dataset2_split} after deduplication:** {num_unique_ds2}\n\n"
|
142 |
+
result_text += f"**Duplicate indices in {dataset2_name}/{dataset2_split}:** {duplicate_indices_in_ds2}\n\n"
|
143 |
+
result_text += f"**Duplicate to original mapping:** {duplicate_to_original_mapping}\n"
|
144 |
+
|
145 |
+
return result_text
|
146 |
|
147 |
+
with gr.Blocks() as demo:
|
148 |
+
gr.Markdown("# Semantic Deduplication")
|
149 |
+
|
150 |
+
deduplication_type = gr.Radio(choices=["Single dataset", "Cross-dataset"], label="Deduplication Type", value="Single dataset")
|
151 |
+
|
152 |
+
with gr.Row():
|
153 |
+
dataset1_name = gr.Textbox(value="ag_news", label="Dataset 1 Name")
|
154 |
+
dataset1_split = gr.Textbox(value="train", label="Dataset 1 Split")
|
155 |
+
|
156 |
+
dataset2_row = gr.Row(visible=False)
|
157 |
+
with dataset2_row:
|
158 |
+
dataset2_name = gr.Textbox(value="ag_news", label="Dataset 2 Name")
|
159 |
+
dataset2_split = gr.Textbox(value="test", label="Dataset 2 Split")
|
160 |
+
|
161 |
+
threshold = gr.Slider(minimum=0.0, maximum=1.0, value=0.8, label="Similarity Threshold")
|
162 |
+
|
163 |
+
compute_button = gr.Button("Compute")
|
164 |
+
|
165 |
+
output = gr.Markdown()
|
166 |
+
|
167 |
+
# Function to update the visibility of dataset2_row
|
168 |
+
def update_visibility(deduplication_type):
|
169 |
+
if deduplication_type == "Cross-dataset":
|
170 |
+
return {dataset2_row: gr.update(visible=True)}
|
171 |
+
else:
|
172 |
+
return {dataset2_row: gr.update(visible=False)}
|
173 |
+
|
174 |
+
deduplication_type.change(update_visibility, inputs=deduplication_type, outputs=[dataset2_row])
|
175 |
+
|
176 |
+
compute_button.click(
|
177 |
+
fn=perform_deduplication,
|
178 |
+
inputs=[deduplication_type, dataset1_name, dataset1_split, dataset2_name, dataset2_split, threshold],
|
179 |
+
outputs=output
|
180 |
+
)
|
181 |
+
|
182 |
+
demo.launch()
|
requirements.txt
ADDED
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
reach < 5
|
2 |
+
model2vec
|
3 |
+
numpy
|
4 |
+
datasets
|
5 |
+
tqdm
|
6 |
+
|