File size: 14,193 Bytes
d75b840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0887481
d75b840
 
 
 
 
 
 
0887481
 
 
d75b840
 
 
 
 
 
 
 
 
 
 
 
 
 
0887481
 
 
d75b840
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0887481
47d6477
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299

# This file is taken from the original DPST implementation by Meisenbacher et al.
# Paper: https://aclanthology.org/2025.emnlp-main.455/
# Source: https://github.com/sjmeis/DPST

import os
import weaviate
from weaviate.classes.query import MetadataQuery, Filter

from transformers import AutoModel, AutoTokenizer, AutoModelForCausalLM, pipeline
from sentence_transformers import util

os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
from openie import StanfordOpenIE
from collections import defaultdict
from tqdm.auto import tqdm
import multiprocessing as mp
import numpy as np
from functools import partial
import nltk
import json
import importlib_resources as impresources

import spacy
spacy.prefer_gpu()

import torch
from torch.nn import CrossEntropyLoss

from datasketch import MinHash, MinHashLSH
from nltk import ngrams

# globals
properties = {
            "openie.affinity_probability_cap": 2 / 3,
            "openie.triple.strict": False,
        }
IEclient = StanfordOpenIE(properties=properties)

class DPST():
    def __init__(self, mode, hf_token=None, model_checkpoint="meta-llama/Llama-3.2-3B-Instruct"):
        print("Initializing...", flush=True)

        mode_map = {
            "50k": "fiftyk",
            "100k": "hundredk",
            "200k": "twohundredk"
        }

        if mode not in mode_map:
            print("Error: [MODE] must be one of [50k, 100k, 200k].")
            return
        else:
            self.mode = mode_map[mode]

        # db connection
        self.client = weaviate.connect_to_local()
        self.collection = self.client.collections.get("Triples")

        if torch.cuda.is_available() == True:
            self.device = "cuda"
        else:
            self.device = "cpu"

        with open(impresources.files("data") / "clusters" / "{}.json".format(mode), 'r') as f:
            self.centroids = torch.tensor(json.load(f)).to(self.device)

        with open(impresources.files("data") / "clusters" / "{}_counts.json".format(mode), 'r') as f:
            self.cluster_counts = json.load(f)

        self.model_checkpoint = model_checkpoint

        #self.pool = mp.Pool(mp.cpu_count(), initargs=(nlp,))
        self.model = AutoModel.from_pretrained("jinaai/jina-embeddings-v3", trust_remote_code=True).to(self.device)

        self.gen_model = AutoModelForCausalLM.from_pretrained(self.model_checkpoint, token=hf_token, device_map=self.device)
        self.tokenizer = AutoTokenizer.from_pretrained(self.model_checkpoint, token=hf_token)
        
        self.pipe = pipeline(
            "text-generation",
            model=self.gen_model,
            tokenizer=self.tokenizer,
            torch_dtype=torch.bfloat16,
            device_map="auto",
        )

        self.ppl_model = AutoModelForCausalLM.from_pretrained("gpt2").to(self.device)
        self.ppl_tokenizer = AutoTokenizer.from_pretrained("gpt2")

        print("Finished.", flush=True)   

    def cleanup(self):
        self.client.close()
        #IEclient.client.stop()

    def exponential(self, candidates, epsilon, sensitivity=1):
        probabilities = [np.exp(epsilon * x[1] / (2 * sensitivity)) for x in candidates]
        probabilities = probabilities / np.linalg.norm(probabilities, ord=1)
        return np.random.choice([x[0] for x in candidates], 1, p=probabilities)[0]

    def query_db(self, vector, cluster):
        # Convert numpy array to list if needed
        vector_list = vector.tolist() if isinstance(vector, np.ndarray) else vector
        response = self.collection.query.near_vector(near_vector=vector_list, limit=self.cluster_counts[cluster], filters=Filter.by_property(self.mode).equal(cluster), return_metadata=MetadataQuery(distance=True))
        candidates = [(x.properties["text"], max(1 - x.metadata.distance, 0)) for x in response.objects]
        return candidates

    def get_prompt(self, triples, messages=True):
        PROMPT = [
            {"role": "system", "content": "Generate a concise text for the given set of triples. Ensure that the generated output only includes the provided information from the triples, but feel free to fill in the gaps where sensible. If necessary, ignore triples that do not fit into the larger context. It is very important that the output is grammatically correct, natural, and logical. Provide a text that captures the semantic meaning of the triples, without being too verbose or lengthy. Do not provide any further explanation, only provide the output text."},
            {"role": "user", "content": "Input triples: [{’object’: ’Mike_Mularkey’,’property’: ’coach’,’subject’: ’Tennessee_Titans’}]"},
            {"role": "assistant", "content": "Output text: Mike Mularkey is the coach of the Tennessee Titans."},
            {"role": "user", "content": "Input triples: [{’object’: ’Albert_E._Austin’, ’property’: ’successor’, ’subject’: ’Alfred_N._Phillips’}, {’object’: ’Connecticut’, ’property’: ’birthPlace’, ’subject’: ’Alfred_N._Phillips’}, {’object’: ’United_States_House_of_Representatives’, ’property’: ’office’, ’subject’: ’Alfred_N._Phillips’}]"},
            {"role": "assistant", "content": "Output text: Albert E. Austin succeeded Alfred N. Phillips who was born in Connecticut and worked at the United States House of Representatives."},
            {"role": "user", "content": "Input triples: [{’object’: ’College_of_William_&_Mary’, ’property’: ’owner’, ’subject’: ’Alan_B._Miller_Hall’}, {’object’: ’2009-06-01’, ’property’: ’completionDate’, ’subject’: ’Alan_B._Miller_Hall’}, {’object’: ’101 Ukrop Way’, ’property’: ’address’, ’subject’: ’Alan_B._Miller_Hall’}, {’object’: ’Williamsburg,_Virginia’, ’property’: ’location’, ’subject’: ’Alan_B._Miller_Hall’}, {’object’: ’Robert_A._M._Stern’, ’property’: ’architect’, ’subject’: ’Alan_B._Miller_Hall’}]"},
            {"role": "assistant", "content": "Output text: The Alan B Miller Hall’s location is 101 Ukrop Way, Williamsburg, Virginia. It was designed by Robert A.M. Stern and was completed on 1 June 2009. Its owner is the College of William and Mary."},
            {"role": "user", "content": "Input Triples: {}".format(str(triples))}
        ]
        return PROMPT
    
    def compute_ppl(self, predictions, batch_size: int = 16, add_start_token: bool = True, max_length=32):
        # if batch_size > 1 (which generally leads to padding being required), and
        # if there is not an already assigned pad_token, assign an existing
        # special token to also be the padding token
        if self.ppl_tokenizer.pad_token is None and batch_size > 1:
            existing_special_tokens = list(self.ppl_tokenizer.special_tokens_map_extended.values())
            # check that the model already has at least one special token defined
            assert (
                len(existing_special_tokens) > 0
            ), "If batch_size > 1, model must have at least one special token to use for padding. Please use a different model or set batch_size=1."
            # assign one of the special tokens to also be the pad token
            self.ppl_tokenizer.add_special_tokens({"pad_token": existing_special_tokens[0]})

        if add_start_token and max_length:
            # leave room for <BOS> token to be added:
            assert (
                self.ppl_tokenizer.bos_token is not None
            ), "Input model must already have a BOS token if using add_start_token=True. Please use a different model, or set add_start_token=False"
            max_tokenized_len = max_length - 1
        else:
            max_tokenized_len = max_length

        encodings = self.ppl_tokenizer(
            predictions,
            add_special_tokens=False,
            padding=True,
            truncation=True if max_tokenized_len else False,
            max_length=max_tokenized_len,
            return_tensors="pt",
            return_attention_mask=True,
        ).to(self.device)

        encoded_texts = encodings["input_ids"]
        attn_masks = encodings["attention_mask"]

        # check that each input is long enough:
        if add_start_token:
            assert torch.all(torch.ge(attn_masks.sum(1), 1)), "Each input text must be at least one token long."
        else:
            assert torch.all(
                torch.ge(attn_masks.sum(1), 2)
            ), "When add_start_token=False, each input text must be at least two tokens long. Run with add_start_token=True if inputting strings of only one token, and remove all empty input strings."

        ppls = []
        loss_fct = CrossEntropyLoss(reduction="none")

        for start_index in range(0, len(encoded_texts), batch_size):
            end_index = min(start_index + batch_size, len(encoded_texts))
            encoded_batch = encoded_texts[start_index:end_index]
            attn_mask = attn_masks[start_index:end_index]

            if add_start_token:
                bos_tokens_tensor = torch.tensor([[self.ppl_tokenizer.bos_token_id]] * encoded_batch.size(dim=0)).to(self.device)
                encoded_batch = torch.cat([bos_tokens_tensor, encoded_batch], dim=1)
                attn_mask = torch.cat(
                    [torch.ones(bos_tokens_tensor.size(), dtype=torch.int64).to(self.device), attn_mask], dim=1
                )

            labels = encoded_batch

            with torch.no_grad():
                out_logits = self.ppl_model(encoded_batch, attention_mask=attn_mask).logits

            shift_logits = out_logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            shift_attention_mask_batch = attn_mask[..., 1:].contiguous()

            perplexity_batch = torch.exp(
                (loss_fct(shift_logits.transpose(1, 2), shift_labels) * shift_attention_mask_batch).sum(1)
                / shift_attention_mask_batch.sum(1)
            )

            ppls += perplexity_batch.tolist()

        return {"perplexities": ppls, "mean_perplexity": np.mean(ppls)}

    def get_triples_ie(self, text):
        res = [x for x in IEclient.annotate(text)]
        temp = [tuple(x.values()) for x in res]

        current = defaultdict(list)
        for t in temp:
            current[(t[0], t[1])].append(t)

        final = []
        for t in temp:
            s = "{} | {} | {}".format(t[0], t[1], t[2])
            if s not in final:
                final.append(s.replace("_", " "))
        
        lsh = MinHashLSH(threshold=0.4, num_perm=128)
        minhashes = {}
        for i, f in enumerate(final):
            minhash = MinHash(num_perm=128)
            for d in ngrams(f, 3):
                minhash.update("".join(d).encode('utf-8'))
            lsh.insert(i, minhash)
            minhashes[i] = minhash

        matches = {}
        for x, y in zip(final, minhashes):
            matches[x] = [final[z] for z in lsh.query(minhashes[y]) if z != y] 

        clusters = []
        covered = []
        for m in sorted(matches, key=lambda x: len(matches[x]), reverse=True):
            if m not in covered and len(matches[m]) > 0:
                clusters.append(matches[m])
                covered.extend(matches[m])

        clean = [x.replace(" | ", " ") for x in covered]
        if len(clean) == 0:
            return []
        ppls = dict(zip(covered, self.compute_ppl(predictions=clean, batch_size=64)["perplexities"]))

        best = []
        for c in clusters:
            scores = [ppls[x] for x in c]
            imin = np.argmin(scores)
            best.append(c[imin])

        ordered = []
        for f in final:
            if f in best:
                ordered.append(f)
        return ordered

    def privatize(self, texts, epsilon=10, DP=True, verbose=False):
        results = []
        for i, t in tqdm(enumerate(texts), total=len(texts)):
            triples = self.get_triples_ie(t)
            if len(triples) == 0:
                results.append(t)
                continue

            if verbose:
                print(f"\n  Extracted Triples : {triples}")

            if DP == True:
                eps = epsilon / len(triples)
                query_vectors = self.model.encode(triples, task="text-matching", truncate_dim=32, max_length=64)

                res = util.semantic_search(query_embeddings=torch.tensor(query_vectors).to(self.device), corpus_embeddings=self.centroids, top_k=1)
                clusters = [r[0]["corpus_id"] for r in res]

                candidates = []
                for q, c in zip(query_vectors, clusters):
                    near = self.query_db(q, c)
                    if len(near) > 0:
                        candidates.append(near)
                private_triples = [self.exponential(c, eps) for c in candidates]

                if verbose:
                    print(f"  Private Triples   : {private_triples}")

                final = []
                for p in private_triples:
                    m = p.split(" | ")
                    final.append({"object": m[0], "property": m[1], "subject": m[2]})
                prompt = self.get_prompt(final)
            else:
                final = []
                for x in triples:
                    m = x.split(" | ")
                    final.append({"object": m[0], "property": m[1], "subject": m[2]})
                prompt = self.get_prompt(final)

            outputs = self.pipe(
                prompt,
                pad_token_id=self.tokenizer.eos_token_id,
                max_new_tokens=int(len(self.tokenizer.encode(texts[i], return_tensors="pt")[0]))
            )
            generated = outputs[0]["generated_text"][-1]["content"]

            generated = generated.split("Output text: ")[-1].strip().replace("\n", "")
            generated = generated.split("USER:")[0].strip().replace("\n", "")
            generated = generated.split("\t")[0].split("ASSISTANT")[0].split("USER")[0].split("###")[0].split("Note:")[0].split("Explanation:")[0].split("```")[0].split("EXPECTED_OUTPUT")[0]
            results.append(generated.strip())
        return results