Upload instruction_template_retriever.py
Browse files
instruction_template_retriever.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import itertools
|
2 |
+
import json
|
3 |
+
|
4 |
+
from datasets import load_dataset
|
5 |
+
import faiss
|
6 |
+
import pandas as pd
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
|
10 |
+
from huggingface_hub import hf_hub_download
|
11 |
+
from sentence_transformers import SentenceTransformer
|
12 |
+
from pooling_coverage import use_gaussian_coverage_pooling
|
13 |
+
|
14 |
+
|
15 |
+
class InstructionTemplateRetriever:
|
16 |
+
FINETEMPLATES_REVISION = "831ab22c90f9da011bd972585afdf609f40fa54b"
|
17 |
+
RETRIEVAL_EMBEDDING_NAME = "fineinstructions/matching_embedding"
|
18 |
+
RETRIEVAL_EMBEDDING_REVISION = "db4efbde126216250ffa5a356663fc7da3bf7856"
|
19 |
+
|
20 |
+
def __init__(
|
21 |
+
self,
|
22 |
+
coverage_chunks=10,
|
23 |
+
sigma=0.05,
|
24 |
+
alpha=1.0,
|
25 |
+
nprobe=150,
|
26 |
+
):
|
27 |
+
"""
|
28 |
+
Computes embeddings that cover a document to find relevant
|
29 |
+
instruction templates using Gaussian-weighted embeddings that cover
|
30 |
+
different parts of the document.
|
31 |
+
|
32 |
+
Args:
|
33 |
+
coverage_chunks (int): The number of equally sized chunks/sections
|
34 |
+
to get coverage over the entire document.
|
35 |
+
sigma (float): Standard deviation for Gaussian weighting, this
|
36 |
+
will essentially control how "wide" / "focused" each chunk is.
|
37 |
+
alpha (float): A weighting factor to control how much to balance
|
38 |
+
the representation of a single chunk, versus the representation of
|
39 |
+
the entire document.
|
40 |
+
nprobe (int): The number of probes to use when searching the FAISS
|
41 |
+
index (larger is more accurate, but slower).
|
42 |
+
"""
|
43 |
+
self.d = load_dataset(
|
44 |
+
"fineinstructions/finetemplates",
|
45 |
+
revision=InstructionTemplateRetriever.FINETEMPLATES_REVISION,
|
46 |
+
split="full",
|
47 |
+
)
|
48 |
+
self.m = SentenceTransformer(
|
49 |
+
InstructionTemplateRetriever.RETRIEVAL_EMBEDDING_NAME,
|
50 |
+
revision=InstructionTemplateRetriever.RETRIEVAL_EMBEDDING_REVISION,
|
51 |
+
device="cpu",
|
52 |
+
)
|
53 |
+
self.m = use_gaussian_coverage_pooling(
|
54 |
+
self.m, coverage_chunks=coverage_chunks, sigma=sigma, alpha=alpha
|
55 |
+
)
|
56 |
+
self.index = faiss.read_index(
|
57 |
+
hf_hub_download(
|
58 |
+
"fineinstructions/finetemplates",
|
59 |
+
"faiss_index/finetemplates.index",
|
60 |
+
revision=InstructionTemplateRetriever.FINETEMPLATES_REVISION,
|
61 |
+
repo_type="dataset",
|
62 |
+
),
|
63 |
+
faiss.IO_FLAG_MMAP | faiss.IO_FLAG_READ_ONLY,
|
64 |
+
)
|
65 |
+
self.index.nprobe = nprobe
|
66 |
+
if torch.cuda.is_available():
|
67 |
+
self.m = self.m.to("cuda")
|
68 |
+
elif torch.backends.mps.is_available():
|
69 |
+
self.m = self.m.to("mps")
|
70 |
+
|
71 |
+
def _filter_rows(self, rows, filter_string):
|
72 |
+
if not rows:
|
73 |
+
return []
|
74 |
+
df = pd.DataFrame(rows)
|
75 |
+
try:
|
76 |
+
filtered_df = df.query(filter_string)
|
77 |
+
return filtered_df.to_dict(orient="records")
|
78 |
+
except Exception as e:
|
79 |
+
return rows
|
80 |
+
|
81 |
+
def search(
|
82 |
+
self, document, filters="", search_k=20000, max_results=250, deduplicate=True
|
83 |
+
):
|
84 |
+
"""
|
85 |
+
Given a document
|
86 |
+
|
87 |
+
Args:
|
88 |
+
document (str): The document to retrieve relevant instruction templates for.
|
89 |
+
filters (str): A query string in the format of pandas.DataFrame.query()
|
90 |
+
search_k (int): The number of search results to pull when retrieving from FAISS.
|
91 |
+
max_results (int): The max number of results to return.
|
92 |
+
deduplicate (bool): Deduplicate results between coverage sections.
|
93 |
+
"""
|
94 |
+
|
95 |
+
# Search FAISS index
|
96 |
+
vecs = self.m.encode([document], normalize_embeddings=False).reshape(
|
97 |
+
-1, self.m[0].auto_model.config.hidden_size
|
98 |
+
)
|
99 |
+
scores_batch, indices_batch = self.index.search(np.vstack(vecs), k=search_k)
|
100 |
+
|
101 |
+
# Pull in FineTemplates rows into memory
|
102 |
+
to_select = [i.item() for i in itertools.chain.from_iterable(indices_batch)]
|
103 |
+
d_in_mem = {
|
104 |
+
i: row for i, row in zip(to_select, self.d.select(to_select).to_list())
|
105 |
+
}
|
106 |
+
|
107 |
+
# Group by coverage chunk
|
108 |
+
true_coverage_chunks = self.m[1].coverage_chunks + 1
|
109 |
+
scores_per_input, indices_per_input = (
|
110 |
+
[
|
111 |
+
scores_batch[i : i + true_coverage_chunks]
|
112 |
+
for i in range(0, len(scores_batch), true_coverage_chunks)
|
113 |
+
],
|
114 |
+
[
|
115 |
+
indices_batch[i : i + true_coverage_chunks]
|
116 |
+
for i in range(0, len(indices_batch), true_coverage_chunks)
|
117 |
+
],
|
118 |
+
)
|
119 |
+
|
120 |
+
# Get the results for the first result in the batch (assuming bz=1)
|
121 |
+
scores_per_input, indices_per_input = scores_per_input[0], indices_per_input[0]
|
122 |
+
|
123 |
+
# Create result rows
|
124 |
+
rows = [
|
125 |
+
[
|
126 |
+
{
|
127 |
+
"coverage_section": f"{chunk_idx}/{self.m[1].coverage_chunks}"
|
128 |
+
if chunk_idx > 0
|
129 |
+
else "Entire Document",
|
130 |
+
"score": s.item(),
|
131 |
+
**d_in_mem[i.item()],
|
132 |
+
}
|
133 |
+
for i, s in zip(indices, scores)
|
134 |
+
]
|
135 |
+
for chunk_idx, (indices, scores) in enumerate(
|
136 |
+
zip(indices_per_input, scores_per_input)
|
137 |
+
)
|
138 |
+
]
|
139 |
+
|
140 |
+
# Deduplicate
|
141 |
+
if deduplicate:
|
142 |
+
seen = set()
|
143 |
+
rows = [
|
144 |
+
r
|
145 |
+
for r in itertools.chain.from_iterable(zip(*rows))
|
146 |
+
if (len(seen) != len(seen.add(r["template_id"]) or seen))
|
147 |
+
]
|
148 |
+
else:
|
149 |
+
rows = list(itertools.chain.from_iterable(zip(*rows)))
|
150 |
+
|
151 |
+
# Filter
|
152 |
+
rows = self._filter_rows(rows, filters)[:max_results]
|
153 |
+
|
154 |
+
# Return rows
|
155 |
+
return rows
|