File size: 51,541 Bytes
2dcbc14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d9e29f
 
2dcbc14
 
 
4d9e29f
2dcbc14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f3bcea6
2dcbc14
 
 
 
 
f3bcea6
2dcbc14
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4d9e29f
 
 
 
 
 
 
 
 
2dcbc14
 
 
 
 
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
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Sentiment Classification with FHE\n",
    "\n",
    "This notebook tackles sentiment classification with Fully Homomorphic Encryption. Let's imagine some client (could be a user or a company) wants to predict whether a specific text (e.g., a tweet) contains positive, neutral or negative feedback using a cloud service provider without actually revealing the text during the process.\n",
    "\n",
    "To do this, we use a machine learning model that can predict over encrypted data thanks to the Concrete-ML library available on [GitHub](https://github.com/zama-ai/concrete-ml).\n",
    "\n",
    "The dataset we use in this notebook can be found on [Kaggle](https://www.kaggle.com/datasets/crowdflower/twitter-airline-sentiment). \n",
    " \n",
    "We present two different ways to encode the text:\n",
    "1. A basic **TF-IDF** approach, which essentially looks at how often a word appears in the text.\n",
    "2. An advanced **transformer** embedding of the text using the Huggingface repository.\n",
    "\n",
    "The main assumption of this notebook is that clients, who want to have their text analyzed in a privacy preserving manner, can encode the text using a predefined representation before encrypting the data. The FHE-friendly model is thus trained in the clear beforehand for the given task, here classification, over theses representations using a relevant training set."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Import the required packages\n",
    "import os\n",
    "import time\n",
    "\n",
    "import numpy\n",
    "import onnx\n",
    "import pandas as pd\n",
    "from sklearn.metrics import average_precision_score\n",
    "from sklearn.model_selection import GridSearchCV, train_test_split\n",
    "\n",
    "from concrete.ml.sklearn import XGBClassifier"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Proportion of positive examples: 16.14%\n",
      "Proportion of negative examples: 62.69%\n",
      "Proportion of neutral examples: 21.17%\n"
     ]
    }
   ],
   "source": [
    "# Download the datasets\n",
    "# The dataset can be downloaded through the `download_data.sh` script, which requires to set up\n",
    "# Kaggle's CLI, or manually at https://www.kaggle.com/datasets/crowdflower/twitter-airline-sentiment\n",
    "if not os.path.isfile(\"local_datasets/twitter-airline-sentiment/Tweets.csv\"):\n",
    "    raise ValueError(\"Please launch the `download_data.sh` script to get datasets\")\n",
    "\n",
    "\n",
    "train = pd.read_csv(\"local_datasets/twitter-airline-sentiment/Tweets.csv\", index_col=0)\n",
    "text_X = train[\"text\"]\n",
    "y = train[\"airline_sentiment\"]\n",
    "y = y.replace([\"negative\", \"neutral\", \"positive\"], [0, 1, 2])\n",
    "\n",
    "pos_ratio = y.value_counts()[2] / y.value_counts().sum()\n",
    "neg_ratio = y.value_counts()[0] / y.value_counts().sum()\n",
    "neutral_ratio = y.value_counts()[1] / y.value_counts().sum()\n",
    "print(f\"Proportion of positive examples: {round(pos_ratio * 100, 2)}%\")\n",
    "print(f\"Proportion of negative examples: {round(neg_ratio * 100, 2)}%\")\n",
    "print(f\"Proportion of neutral examples: {round(neutral_ratio * 100, 2)}%\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Split in train test\n",
    "text_X_train, text_X_test, y_train, y_test = train_test_split(\n",
    "    text_X, y, test_size=0.1, random_state=42\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1. Text representation using TF-IDF\n",
    "\n",
    "[Term Frequency-Inverse Document Frequency](https://en.wikipedia.org/wiki/Tf%E2%80%93idf)(TF-IDF) also known as  is a numerical statistic that is used to compute the importance of a term in a document. The higher the TF-IDF score, the more important the term is to the document.\n",
    "\n",
    "We compute it as follows:\n",
    "\n",
    "$$ \\mathsf{TF\\textrm{-}IDF}(t,d,D) = \\mathsf{TF}(t,d) * \\mathsf{IDF}(t,D) $$\n",
    "\n",
    "where: $\\mathsf{TF}(t,d)$ is the term frequency of term $t$ in document $d$, $\\mathsf{IDF}(t,D)$ is the inverse document frequency of term $t$ in document collection $D$.\n",
    "\n",
    "Here we use the scikit-learn implementation of TF-IDF vectorizer."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Let's first build a representation vector from the text\n",
    "from sklearn.feature_extraction.text import TfidfVectorizer\n",
    "\n",
    "tfidf_vectorizer = TfidfVectorizer(max_features=500, stop_words=\"english\")\n",
    "X_train = tfidf_vectorizer.fit_transform(text_X_train)\n",
    "X_test = tfidf_vectorizer.transform(text_X_test)\n",
    "\n",
    "# Make our train and test dense array\n",
    "X_train = X_train.toarray()\n",
    "X_test = X_test.toarray()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Let's build our model\n",
    "model = XGBClassifier()\n",
    "\n",
    "# A gridsearch to find the best parameters\n",
    "parameters = {\n",
    "    \"n_bits\": [2, 3],\n",
    "    \"max_depth\": [1],\n",
    "    \"n_estimators\": [10, 30, 50],\n",
    "    \"n_jobs\": [-1],\n",
    "}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<style>#sk-container-id-3 {color: black;background-color: white;}#sk-container-id-3 pre{padding: 0;}#sk-container-id-3 div.sk-toggleable {background-color: white;}#sk-container-id-3 label.sk-toggleable__label {cursor: pointer;display: block;width: 100%;margin-bottom: 0;padding: 0.3em;box-sizing: border-box;text-align: center;}#sk-container-id-3 label.sk-toggleable__label-arrow:before {content: \"β–Έ\";float: left;margin-right: 0.25em;color: #696969;}#sk-container-id-3 label.sk-toggleable__label-arrow:hover:before {color: black;}#sk-container-id-3 div.sk-estimator:hover label.sk-toggleable__label-arrow:before {color: black;}#sk-container-id-3 div.sk-toggleable__content {max-height: 0;max-width: 0;overflow: hidden;text-align: left;background-color: #f0f8ff;}#sk-container-id-3 div.sk-toggleable__content pre {margin: 0.2em;color: black;border-radius: 0.25em;background-color: #f0f8ff;}#sk-container-id-3 input.sk-toggleable__control:checked~div.sk-toggleable__content {max-height: 200px;max-width: 100%;overflow: auto;}#sk-container-id-3 input.sk-toggleable__control:checked~label.sk-toggleable__label-arrow:before {content: \"β–Ύ\";}#sk-container-id-3 div.sk-estimator input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-3 div.sk-label input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-3 input.sk-hidden--visually {border: 0;clip: rect(1px 1px 1px 1px);clip: rect(1px, 1px, 1px, 1px);height: 1px;margin: -1px;overflow: hidden;padding: 0;position: absolute;width: 1px;}#sk-container-id-3 div.sk-estimator {font-family: monospace;background-color: #f0f8ff;border: 1px dotted black;border-radius: 0.25em;box-sizing: border-box;margin-bottom: 0.5em;}#sk-container-id-3 div.sk-estimator:hover {background-color: #d4ebff;}#sk-container-id-3 div.sk-parallel-item::after {content: \"\";width: 100%;border-bottom: 1px solid gray;flex-grow: 1;}#sk-container-id-3 div.sk-label:hover label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-3 div.sk-serial::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: 0;}#sk-container-id-3 div.sk-serial {display: flex;flex-direction: column;align-items: center;background-color: white;padding-right: 0.2em;padding-left: 0.2em;position: relative;}#sk-container-id-3 div.sk-item {position: relative;z-index: 1;}#sk-container-id-3 div.sk-parallel {display: flex;align-items: stretch;justify-content: center;background-color: white;position: relative;}#sk-container-id-3 div.sk-item::before, #sk-container-id-3 div.sk-parallel-item::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: -1;}#sk-container-id-3 div.sk-parallel-item {display: flex;flex-direction: column;z-index: 1;position: relative;background-color: white;}#sk-container-id-3 div.sk-parallel-item:first-child::after {align-self: flex-end;width: 50%;}#sk-container-id-3 div.sk-parallel-item:last-child::after {align-self: flex-start;width: 50%;}#sk-container-id-3 div.sk-parallel-item:only-child::after {width: 0;}#sk-container-id-3 div.sk-dashed-wrapped {border: 1px dashed gray;margin: 0 0.4em 0.5em 0.4em;box-sizing: border-box;padding-bottom: 0.4em;background-color: white;}#sk-container-id-3 div.sk-label label {font-family: monospace;font-weight: bold;display: inline-block;line-height: 1.2em;}#sk-container-id-3 div.sk-label-container {text-align: center;}#sk-container-id-3 div.sk-container {/* jupyter's `normalize.less` sets `[hidden] { display: none; }` but bootstrap.min.css set `[hidden] { display: none !important; }` so we also need the `!important` here to be able to override the default hidden behavior on the sphinx rendered scikit-learn.org. See: https://github.com/scikit-learn/scikit-learn/issues/21755 */display: inline-block !important;position: relative;}#sk-container-id-3 div.sk-text-repr-fallback {display: none;}</style><div id=\"sk-container-id-3\" class=\"sk-top-container\"><div class=\"sk-text-repr-fallback\"><pre>GridSearchCV(cv=3, estimator=XGBClassifier(), n_jobs=1,\n",
       "             param_grid={&#x27;max_depth&#x27;: [1], &#x27;n_bits&#x27;: [2, 3],\n",
       "                         &#x27;n_estimators&#x27;: [10, 30, 50], &#x27;n_jobs&#x27;: [-1]},\n",
       "             scoring=&#x27;accuracy&#x27;)</pre><b>In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook. <br />On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.</b></div><div class=\"sk-container\" hidden><div class=\"sk-item sk-dashed-wrapped\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-7\" type=\"checkbox\" ><label for=\"sk-estimator-id-7\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">GridSearchCV</label><div class=\"sk-toggleable__content\"><pre>GridSearchCV(cv=3, estimator=XGBClassifier(), n_jobs=1,\n",
       "             param_grid={&#x27;max_depth&#x27;: [1], &#x27;n_bits&#x27;: [2, 3],\n",
       "                         &#x27;n_estimators&#x27;: [10, 30, 50], &#x27;n_jobs&#x27;: [-1]},\n",
       "             scoring=&#x27;accuracy&#x27;)</pre></div></div></div><div class=\"sk-parallel\"><div class=\"sk-parallel-item\"><div class=\"sk-item\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-8\" type=\"checkbox\" ><label for=\"sk-estimator-id-8\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">estimator: XGBClassifier</label><div class=\"sk-toggleable__content\"><pre>XGBClassifier()</pre></div></div></div><div class=\"sk-serial\"><div class=\"sk-item\"><div class=\"sk-estimator sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-9\" type=\"checkbox\" ><label for=\"sk-estimator-id-9\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">XGBClassifier</label><div class=\"sk-toggleable__content\"><pre>XGBClassifier()</pre></div></div></div></div></div></div></div></div></div></div>"
      ],
      "text/plain": [
       "GridSearchCV(cv=3, estimator=XGBClassifier(), n_jobs=1,\n",
       "             param_grid={'max_depth': [1], 'n_bits': [2, 3],\n",
       "                         'n_estimators': [10, 30, 50], 'n_jobs': [-1]},\n",
       "             scoring='accuracy')"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Run the gridsearch\n",
    "grid_search = GridSearchCV(model, parameters, cv=3, n_jobs=1, scoring=\"accuracy\")\n",
    "grid_search.fit(X_train, y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Best score: 0.6842744383727991\n",
      "Best parameters: {'max_depth': 1, 'n_bits': 3, 'n_estimators': 50, 'n_jobs': -1}\n"
     ]
    }
   ],
   "source": [
    "# Check the accuracy of the best model\n",
    "print(f\"Best score: {grid_search.best_score_}\")\n",
    "\n",
    "# Check best hyperparameters\n",
    "print(f\"Best parameters: {grid_search.best_params_}\")\n",
    "\n",
    "# Extract best model\n",
    "best_model = grid_search.best_estimator_"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.6810\n",
      "Average precision score for positive class: 0.5615\n",
      "Average precision score for negative class: 0.8349\n",
      "Average precision score for neutral class: 0.3820\n"
     ]
    }
   ],
   "source": [
    "# Compute the average precision for each class\n",
    "y_proba_test_tfidf = best_model.predict_proba(X_test)\n",
    "\n",
    "# Compute accuracy\n",
    "y_pred_test_tfidf = numpy.argmax(y_proba_test_tfidf, axis=1)\n",
    "accuracy_tfidf = numpy.mean(y_pred_test_tfidf == y_test)\n",
    "print(f\"Accuracy: {accuracy_tfidf:.4f}\")\n",
    "\n",
    "y_pred_positive = y_proba_test_tfidf[:, 2]\n",
    "y_pred_negative = y_proba_test_tfidf[:, 0]\n",
    "y_pred_neutral = y_proba_test_tfidf[:, 1]\n",
    "\n",
    "ap_positive_tfidf = average_precision_score((y_test == 2), y_pred_positive)\n",
    "ap_negative_tfidf = average_precision_score((y_test == 0), y_pred_negative)\n",
    "ap_neutral_tfidf = average_precision_score((y_test == 1), y_pred_neutral)\n",
    "\n",
    "print(f\"Average precision score for positive class: \" f\"{ap_positive_tfidf:.4f}\")\n",
    "print(f\"Average precision score for negative class: \" f\"{ap_negative_tfidf:.4f}\")\n",
    "print(f\"Average precision score for neutral class: \" f\"{ap_neutral_tfidf:.4f}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 most positive tweets (class 2):\n",
      "@united sent a DM just now. Thanks I am incredibly happy the fast response I got via Twitter than via customer care. Thank you\n",
      "@JetBlue Great Thank you, lets hope so! Could you please notify me if flight 2302 leaves JFK? Thank you again\n",
      "@AmericanAir Great, thanks. Followed.\n",
      "@SouthwestAir I continue to be amazed by the amazing customer service.  Thank you SWA!\n",
      "@JetBlue Awesome thanks! Thanks for the quick response. You guys ROCK! :)\n",
      "----------------------------------------------------------------------------------------------------\n",
      "5 most negative tweets (class 0):\n",
      "@USAirways been on hold 2 hours for a Cancelled Flighted flight. I understand the delay. I don't understand you auto-reFlight Booking Problems me on TUESDAY. HELP!\n",
      "@SouthwestAir 2 hours on hold for customer service never us SW again\n",
      "@SouthwestAir  placed on hold for total of two hours today after flight was Cancelled Flightled. Online option not available. What to do?\n",
      "@southwestair I've been on hold for 2 hours to reschedule my Cancelled Flightled flight for the morning. What gives? I need help NOW\n",
      "@USAirways Customer service is dead. Last wk, flts delayed/Cancelled Flighted. Bags lost 4 days. Last nt, flt delayed/Cancelled Flighted.  No meal voucher?\n"
     ]
    }
   ],
   "source": [
    "# Let's see what are the top predictions based on the probabilities in y_pred_test\n",
    "print(\"5 most positive tweets (class 2):\")\n",
    "for i in range(5):\n",
    "    print(text_X_test.iloc[y_proba_test_tfidf[:, 2].argsort()[-1 - i]])\n",
    "\n",
    "print(\"-\" * 100)\n",
    "\n",
    "print(\"5 most negative tweets (class 0):\")\n",
    "for i in range(5):\n",
    "    print(text_X_test.iloc[y_proba_test_tfidf[:, 0].argsort()[-1 - i]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Compilation time: 11.5009 seconds\n",
      "FHE inference time: 48.6880 seconds\n"
     ]
    }
   ],
   "source": [
    "# Compile the model to get the FHE inference engine\n",
    "# (this may take a few minutes depending on the selected model)\n",
    "start = time.perf_counter()\n",
    "best_model.compile(X_train)\n",
    "end = time.perf_counter()\n",
    "print(f\"Compilation time: {end - start:.4f} seconds\")\n",
    "\n",
    "# Let's write a custom example and predict in FHE\n",
    "tested_tweet = [\"AirFrance is awesome, almost as much as Zama!\"]\n",
    "X_tested_tweet = tfidf_vectorizer.transform(numpy.array(tested_tweet)).toarray()\n",
    "clear_proba = best_model.predict_proba(X_tested_tweet)\n",
    "\n",
    "# Now let's predict with FHE over a single tweet and print the time it takes\n",
    "start = time.perf_counter()\n",
    "decrypted_proba = best_model.predict_proba(X_tested_tweet, execute_in_fhe=True)\n",
    "end = time.perf_counter()\n",
    "print(f\"FHE inference time: {end - start:.4f} seconds\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 57,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Probabilities from the FHE inference: [[0.50224707 0.25647676 0.24127617]]\n",
      "Probabilities from the clear model: [[0.50224707 0.25647676 0.24127617]]\n"
     ]
    }
   ],
   "source": [
    "print(f\"Probabilities from the FHE inference: {decrypted_proba}\")\n",
    "print(f\"Probabilities from the clear model: {clear_proba}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To sum up, \n",
    "- We trained a XGBoost model over TF-IDF representation of the tweets and their respective sentiment class. \n",
    "- The grid search gives us a model that achieves around ~70% accuracy.\n",
    "- Given the imbalance in the classes, we rather compute the average precision per class.\n",
    "\n",
    "Now we will see how we can approach the problem by leveraging the transformers power."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. A transformer approach to text representation\n",
    "\n",
    "[**Transformers**](https://en.wikipedia.org/wiki/Transformer_(machine_learning_model\\)) are neural networks that are often trained to predict the next words to appear in a text (this is commonly called self-supervised learning). \n",
    "\n",
    "They are powerful tools for all kind of Natural Language Processing tasks but supporting a transformer model in FHE might not always be ideal as they are quite big models. However, we can still leverage their hidden representation for any text and feed it to a more FHE friendly machine learning model (in this notebook we will use XGBoost) for classification.\n",
    "\n",
    "Here we will use the transformer model from the amazing [**Huggingface**](https://huggingface.co/) repository."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Some weights of the model checkpoint at cardiffnlp/twitter-roberta-base-sentiment-latest were not used when initializing RobertaForSequenceClassification: ['roberta.pooler.dense.bias', 'roberta.pooler.dense.weight']\n",
      "- This IS expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
      "- This IS NOT expected if you are initializing RobertaForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "import tqdm\n",
    "from transformers import AutoModelForSequenceClassification, AutoTokenizer\n",
    "\n",
    "device = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n",
    "\n",
    "# Load the tokenizer (converts text to tokens)\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"cardiffnlp/twitter-roberta-base-sentiment-latest\")\n",
    "\n",
    "# Load the pre-trained model\n",
    "transformer_model = AutoModelForSequenceClassification.from_pretrained(\n",
    "    \"cardiffnlp/twitter-roberta-base-sentiment-latest\"\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 30/30 [00:33<00:00,  1.10s/it]\n"
     ]
    }
   ],
   "source": [
    "# Let's first see what are the model performance by itself\n",
    "list_text_X_test = text_X_test.tolist()\n",
    "\n",
    "tokenized_text_X_test = tokenizer.batch_encode_plus(\n",
    "    list_text_X_test, pad_to_max_length=True, return_tensors=\"pt\"\n",
    ")[\"input_ids\"]\n",
    "\n",
    "# Depending on the hardware used, the number of examples to be processed can be reduced\n",
    "# Here we split the data into 100 examples per batch\n",
    "tokenized_text_X_test_split = torch.split(tokenized_text_X_test, split_size_or_sections=50)\n",
    "transformer_model = transformer_model.to(device)\n",
    "\n",
    "outputs = []\n",
    "for tokenized_x_test in tqdm.tqdm(tokenized_text_X_test_split):\n",
    "    tokenized_x = tokenized_x_test.to(device)\n",
    "    output_batch = transformer_model(tokenized_x)[\"logits\"]\n",
    "    output_batch = output_batch.detach().cpu().numpy()\n",
    "    outputs.append(output_batch)\n",
    "\n",
    "outputs = numpy.concatenate(outputs, axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Predictions for the first 3 tweets:\n",
      " [[-2.3807464  -0.61802083  2.9900746 ]\n",
      " [ 2.0166504   0.4938078  -2.8006463 ]\n",
      " [ 2.3892698   0.1344364  -2.6873822 ]]\n"
     ]
    }
   ],
   "source": [
    "# Let's see what the transformer model predicts\n",
    "print(f\"Predictions for the first 3 tweets:\\n {outputs[:3]}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.8053\n",
      "Average precision score for positive class: 0.8548\n",
      "Average precision score for negative class: 0.9548\n",
      "Average precision score for neutral class: 0.6801\n"
     ]
    }
   ],
   "source": [
    "# Compute the metrics for each class\n",
    "\n",
    "# Compute accuracy\n",
    "accuracy_transformer_only = numpy.mean(numpy.argmax(outputs, axis=1) == y_test)\n",
    "print(f\"Accuracy: {accuracy_transformer_only:.4f}\")\n",
    "\n",
    "y_pred_positive = outputs[:, 2]\n",
    "y_pred_negative = outputs[:, 0]\n",
    "y_pred_neutral = outputs[:, 1]\n",
    "\n",
    "ap_positive_transformer_only = average_precision_score((y_test == 2), y_pred_positive)\n",
    "ap_negative_transformer_only = average_precision_score((y_test == 0), y_pred_negative)\n",
    "ap_neutral_transformer_only = average_precision_score((y_test == 1), y_pred_neutral)\n",
    "\n",
    "print(f\"Average precision score for positive class: \" f\"{ap_positive_transformer_only:.4f}\")\n",
    "print(f\"Average precision score for negative class: \" f\"{ap_negative_transformer_only:.4f}\")\n",
    "print(f\"Average precision score for neutral class: \" f\"{ap_neutral_transformer_only:.4f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "It looks like the transformer outperforms the model built on TF-IDF reprensentation.\n",
    "Unfortunately, running a transformer that big in FHE would be highly inefficient. \n",
    "\n",
    "Let's see if we can leverage transformer representation and train a FHE model for the given classification task. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 13176/13176 [07:20<00:00, 29.91it/s]\n",
      "100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 1464/1464 [00:47<00:00, 30.75it/s]\n"
     ]
    }
   ],
   "source": [
    "# Function that transforms a list of texts to their representation\n",
    "# learned by the transformer.\n",
    "def text_to_tensor(\n",
    "    list_text_X_train: list,\n",
    "    transformer_model: AutoModelForSequenceClassification,\n",
    "    tokenizer: AutoTokenizer,\n",
    "    device: str,\n",
    ") -> numpy.ndarray:\n",
    "    # Tokenize each text in the list one by one\n",
    "    tokenized_text_X_train_split = []\n",
    "    for text_x_train in list_text_X_train:\n",
    "        tokenized_text_X_train_split.append(tokenizer.encode(text_x_train, return_tensors=\"pt\"))\n",
    "\n",
    "    # Send the model to the device\n",
    "    transformer_model = transformer_model.to(device)\n",
    "    output_hidden_states_list = []\n",
    "\n",
    "    for tokenized_x in tqdm.tqdm(tokenized_text_X_train_split):\n",
    "        # Pass the tokens through the transformer model and get the hidden states\n",
    "        # Only keep the last hidden layer state for now\n",
    "        output_hidden_states = transformer_model(tokenized_x.to(device), output_hidden_states=True)[\n",
    "            1\n",
    "        ][-1]\n",
    "        # Average over the tokens axis to get a representation at the text level.\n",
    "        output_hidden_states = output_hidden_states.mean(dim=1)\n",
    "        output_hidden_states = output_hidden_states.detach().cpu().numpy()\n",
    "        output_hidden_states_list.append(output_hidden_states)\n",
    "\n",
    "    return numpy.concatenate(output_hidden_states_list, axis=0)\n",
    "\n",
    "\n",
    "# Let's vectorize the text using the transformer\n",
    "list_text_X_train = text_X_train.tolist()\n",
    "list_text_X_test = text_X_test.tolist()\n",
    "\n",
    "X_train_transformer = text_to_tensor(list_text_X_train, transformer_model, tokenizer, device)\n",
    "X_test_transformer = text_to_tensor(list_text_X_test, transformer_model, tokenizer, device)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<style>#sk-container-id-2 {color: black;background-color: white;}#sk-container-id-2 pre{padding: 0;}#sk-container-id-2 div.sk-toggleable {background-color: white;}#sk-container-id-2 label.sk-toggleable__label {cursor: pointer;display: block;width: 100%;margin-bottom: 0;padding: 0.3em;box-sizing: border-box;text-align: center;}#sk-container-id-2 label.sk-toggleable__label-arrow:before {content: \"β–Έ\";float: left;margin-right: 0.25em;color: #696969;}#sk-container-id-2 label.sk-toggleable__label-arrow:hover:before {color: black;}#sk-container-id-2 div.sk-estimator:hover label.sk-toggleable__label-arrow:before {color: black;}#sk-container-id-2 div.sk-toggleable__content {max-height: 0;max-width: 0;overflow: hidden;text-align: left;background-color: #f0f8ff;}#sk-container-id-2 div.sk-toggleable__content pre {margin: 0.2em;color: black;border-radius: 0.25em;background-color: #f0f8ff;}#sk-container-id-2 input.sk-toggleable__control:checked~div.sk-toggleable__content {max-height: 200px;max-width: 100%;overflow: auto;}#sk-container-id-2 input.sk-toggleable__control:checked~label.sk-toggleable__label-arrow:before {content: \"β–Ύ\";}#sk-container-id-2 div.sk-estimator input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-2 div.sk-label input.sk-toggleable__control:checked~label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-2 input.sk-hidden--visually {border: 0;clip: rect(1px 1px 1px 1px);clip: rect(1px, 1px, 1px, 1px);height: 1px;margin: -1px;overflow: hidden;padding: 0;position: absolute;width: 1px;}#sk-container-id-2 div.sk-estimator {font-family: monospace;background-color: #f0f8ff;border: 1px dotted black;border-radius: 0.25em;box-sizing: border-box;margin-bottom: 0.5em;}#sk-container-id-2 div.sk-estimator:hover {background-color: #d4ebff;}#sk-container-id-2 div.sk-parallel-item::after {content: \"\";width: 100%;border-bottom: 1px solid gray;flex-grow: 1;}#sk-container-id-2 div.sk-label:hover label.sk-toggleable__label {background-color: #d4ebff;}#sk-container-id-2 div.sk-serial::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: 0;}#sk-container-id-2 div.sk-serial {display: flex;flex-direction: column;align-items: center;background-color: white;padding-right: 0.2em;padding-left: 0.2em;position: relative;}#sk-container-id-2 div.sk-item {position: relative;z-index: 1;}#sk-container-id-2 div.sk-parallel {display: flex;align-items: stretch;justify-content: center;background-color: white;position: relative;}#sk-container-id-2 div.sk-item::before, #sk-container-id-2 div.sk-parallel-item::before {content: \"\";position: absolute;border-left: 1px solid gray;box-sizing: border-box;top: 0;bottom: 0;left: 50%;z-index: -1;}#sk-container-id-2 div.sk-parallel-item {display: flex;flex-direction: column;z-index: 1;position: relative;background-color: white;}#sk-container-id-2 div.sk-parallel-item:first-child::after {align-self: flex-end;width: 50%;}#sk-container-id-2 div.sk-parallel-item:last-child::after {align-self: flex-start;width: 50%;}#sk-container-id-2 div.sk-parallel-item:only-child::after {width: 0;}#sk-container-id-2 div.sk-dashed-wrapped {border: 1px dashed gray;margin: 0 0.4em 0.5em 0.4em;box-sizing: border-box;padding-bottom: 0.4em;background-color: white;}#sk-container-id-2 div.sk-label label {font-family: monospace;font-weight: bold;display: inline-block;line-height: 1.2em;}#sk-container-id-2 div.sk-label-container {text-align: center;}#sk-container-id-2 div.sk-container {/* jupyter's `normalize.less` sets `[hidden] { display: none; }` but bootstrap.min.css set `[hidden] { display: none !important; }` so we also need the `!important` here to be able to override the default hidden behavior on the sphinx rendered scikit-learn.org. See: https://github.com/scikit-learn/scikit-learn/issues/21755 */display: inline-block !important;position: relative;}#sk-container-id-2 div.sk-text-repr-fallback {display: none;}</style><div id=\"sk-container-id-2\" class=\"sk-top-container\"><div class=\"sk-text-repr-fallback\"><pre>GridSearchCV(cv=3, estimator=XGBClassifier(), n_jobs=1,\n",
       "             param_grid={&#x27;max_depth&#x27;: [1], &#x27;n_bits&#x27;: [2, 3],\n",
       "                         &#x27;n_estimators&#x27;: [10, 30, 50], &#x27;n_jobs&#x27;: [-1]},\n",
       "             scoring=&#x27;accuracy&#x27;)</pre><b>In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook. <br />On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.</b></div><div class=\"sk-container\" hidden><div class=\"sk-item sk-dashed-wrapped\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-4\" type=\"checkbox\" ><label for=\"sk-estimator-id-4\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">GridSearchCV</label><div class=\"sk-toggleable__content\"><pre>GridSearchCV(cv=3, estimator=XGBClassifier(), n_jobs=1,\n",
       "             param_grid={&#x27;max_depth&#x27;: [1], &#x27;n_bits&#x27;: [2, 3],\n",
       "                         &#x27;n_estimators&#x27;: [10, 30, 50], &#x27;n_jobs&#x27;: [-1]},\n",
       "             scoring=&#x27;accuracy&#x27;)</pre></div></div></div><div class=\"sk-parallel\"><div class=\"sk-parallel-item\"><div class=\"sk-item\"><div class=\"sk-label-container\"><div class=\"sk-label sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-5\" type=\"checkbox\" ><label for=\"sk-estimator-id-5\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">estimator: XGBClassifier</label><div class=\"sk-toggleable__content\"><pre>XGBClassifier()</pre></div></div></div><div class=\"sk-serial\"><div class=\"sk-item\"><div class=\"sk-estimator sk-toggleable\"><input class=\"sk-toggleable__control sk-hidden--visually\" id=\"sk-estimator-id-6\" type=\"checkbox\" ><label for=\"sk-estimator-id-6\" class=\"sk-toggleable__label sk-toggleable__label-arrow\">XGBClassifier</label><div class=\"sk-toggleable__content\"><pre>XGBClassifier()</pre></div></div></div></div></div></div></div></div></div></div>"
      ],
      "text/plain": [
       "GridSearchCV(cv=3, estimator=XGBClassifier(), n_jobs=1,\n",
       "             param_grid={'max_depth': [1], 'n_bits': [2, 3],\n",
       "                         'n_estimators': [10, 30, 50], 'n_jobs': [-1]},\n",
       "             scoring='accuracy')"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# Now we have a representation for each tweet, we can train a model on these.\n",
    "grid_search = GridSearchCV(model, parameters, cv=3, n_jobs=1, scoring=\"accuracy\")\n",
    "grid_search.fit(X_train_transformer, y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Best score: 0.8378111718275654\n",
      "Best parameters: {'max_depth': 1, 'n_bits': 3, 'n_estimators': 50, 'n_jobs': -1}\n"
     ]
    }
   ],
   "source": [
    "# Check the accuracy of the best model\n",
    "print(f\"Best score: {grid_search.best_score_}\")\n",
    "\n",
    "# Check best hyperparameters\n",
    "print(f\"Best parameters: {grid_search.best_params_}\")\n",
    "\n",
    "# Extract best model\n",
    "best_model = grid_search.best_estimator_"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy: 0.8504\n",
      "Average precision score for positive class: 0.8917\n",
      "Average precision score for negative class: 0.9597\n",
      "Average precision score for neutral class: 0.7341\n"
     ]
    }
   ],
   "source": [
    "# Compute the metrics for each class\n",
    "\n",
    "y_proba = best_model.predict_proba(X_test_transformer)\n",
    "\n",
    "# Compute the accuracy\n",
    "y_pred = numpy.argmax(y_proba, axis=1)\n",
    "accuracy_transformer_xgboost = numpy.mean(y_pred == y_test)\n",
    "print(f\"Accuracy: {accuracy_transformer_xgboost:.4f}\")\n",
    "\n",
    "y_pred_positive = y_proba[:, 2]\n",
    "y_pred_negative = y_proba[:, 0]\n",
    "y_pred_neutral = y_proba[:, 1]\n",
    "\n",
    "ap_positive_transformer_xgboost = average_precision_score((y_test == 2), y_pred_positive)\n",
    "ap_negative_transformer_xgboost = average_precision_score((y_test == 0), y_pred_negative)\n",
    "ap_neutral_transformer_xgboost = average_precision_score((y_test == 1), y_pred_neutral)\n",
    "\n",
    "print(f\"Average precision score for positive class: \" f\"{ap_positive_transformer_xgboost:.4f}\")\n",
    "print(f\"Average precision score for negative class: \" f\"{ap_negative_transformer_xgboost:.4f}\")\n",
    "print(f\"Average precision score for neutral class: \" f\"{ap_neutral_transformer_xgboost:.4f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Our FHE-friendly XGBoost model does 38% better than the XGBoost model built over TF-IDF representation of the text. Note that here we are still not using FHE and only evaluating the model.\n",
    "Interestingly, using XGBoost over the transformer representation of the text matches the performance of the transformer model alone."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 most positive tweets (class 2):\n",
      "@SouthwestAir love them! Always get the best deals!\n",
      "@AmericanAir THANK YOU FOR ALL THE HELP!  :P You guys are the best.  #americanairlines #americanair\n",
      "@SouthwestAir - Great flight from Phoenix to Dallas tonight!Great service and ON TIME! Makes @timieyancey very happy! http://t.co/TkVCMhbPim\n",
      "@AmericanAir AA2416 on time and awesome flight. Great job American!\n",
      "@SouthwestAir AMAZING c/s today by SW thank you SO very much. This is the reason we fly you #southwest\n",
      "----------------------------------------------------------------------------------------------------\n",
      "5 most negative tweets (class 0):\n",
      "@AmericanAir This entire process took sooooo long that no decent seats are left.  #customerservice\n",
      "@USAirways Not only did u lose the flight plan! Now ur flight crew is FAA timed out! Thx for havin us sit on the tarmac for an hr! #Pathetic\n",
      "@United site errored out at last step of changing award. Now can't even pull up reservation. 60 minute wait time.  Thanks @United!\n",
      "@united OKC ticket agent Roger McLarren(sp?) LESS than helpful with our Intl group travel problems Can't find a supervisor for help.\n",
      "@AmericanAir the dinner and called me \"hon\". Not the service I would expect from 1st class.  #disappointed\n"
     ]
    }
   ],
   "source": [
    "# Get probabilities predictions in clear\n",
    "y_pred_test = best_model.predict_proba(X_test_transformer)\n",
    "\n",
    "# Let's see what are the top predictions based on the probabilities in y_pred_test\n",
    "print(\"5 most positive tweets (class 2):\")\n",
    "for i in range(5):\n",
    "    print(text_X_test.iloc[y_pred_test[:, 2].argsort()[-1 - i]])\n",
    "\n",
    "print(\"-\" * 100)\n",
    "\n",
    "print(\"5 most negative tweets (class 0):\")\n",
    "for i in range(5):\n",
    "    print(text_X_test.iloc[y_pred_test[:, 0].argsort()[-1 - i]])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 most positive (predicted) tweets that are actually negative (ground truth class 0):\n",
      "@USAirways as far as being delayed goes… Looks like tailwinds are going to make up for it. Good news!\n",
      "@united thanks for the link, now finally arrived in Brussels, 9 h after schedule...\n",
      "@USAirways your saving grace was our flight attendant Dallas who was amazing. wish he would transfer to Delta where I would see him again\n",
      "@AmericanAir that luggage you forgot...#mia.....he just won an oscarπŸ˜„πŸ’πŸ’πŸ’\n",
      "@united thanks for having changed me. Managed to arrive with only 8 hours of delay and exhausted\n",
      "----------------------------------------------------------------------------------------------------\n",
      "5 most negative (predicted) tweets that are actually positive (ground truth class 2):\n",
      "@united thanks for updating me about the 1+ hour delay the exact second I got to ATL. πŸ™…πŸ™…πŸ™…\n",
      "@JetBlue you don't remember our date Monday night back to NYC? #heartbroken\n",
      "@SouthwestAir save mile to visit family in 2015 and this will impact how many times I can see my mother.  I planned and you change the rules\n",
      "@SouthwestAir hot stewardess flipped me off\n",
      "@SouthwestAir - We left iPad in a seat pocket.  Filed lost item report. Received it exactly 1 week Late Flightr.  Is that a record?  #unbelievable\n"
     ]
    }
   ],
   "source": [
    "# Now let's see where the model is wrong\n",
    "y_pred_test_0 = y_pred_test[y_test == 0]\n",
    "text_X_test_0 = text_X_test[y_test == 0]\n",
    "\n",
    "print(\"5 most positive (predicted) tweets that are actually negative (ground truth class 0):\")\n",
    "for i in range(5):\n",
    "    print(text_X_test_0.iloc[y_pred_test_0[:, 2].argsort()[-1 - i]])\n",
    "\n",
    "print(\"-\" * 100)\n",
    "\n",
    "y_pred_test_2 = y_pred_test[y_test == 2]\n",
    "text_X_test_2 = text_X_test[y_test == 2]\n",
    "print(\"5 most negative (predicted) tweets that are actually positive (ground truth class 2):\")\n",
    "for i in range(5):\n",
    "    print(text_X_test_2.iloc[y_pred_test_2[:, 0].argsort()[-1 - i]])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Interestingly, these misclassifications are not obvious and some actually look rather like mislabeled. Also, it seems that the model is having a hard time to find ironic tweets.\n",
    "\n",
    "Now we have our model trained which has some great accuracy. Let's have it predict over the encrypted representation."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Sentiment Analysis of the Tweet with Fully Homomorphic Encryption\n",
    "\n",
    "Now that we have our model ready for FHE inference and our data ready for encryption let's use the model in a privacy preserving manner with FHE."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Compilation time: 12.6855 seconds\n"
     ]
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|β–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆβ–ˆ| 1/1 [00:00<00:00, 36.43it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "FHE inference time: 53.0192 seconds\n"
     ]
    }
   ],
   "source": [
    "# Compile the model to get the FHE inference engine\n",
    "# (this may take a few minutes depending on the selected model)\n",
    "start = time.perf_counter()\n",
    "best_model.compile(X_train_transformer)\n",
    "end = time.perf_counter()\n",
    "print(f\"Compilation time: {end - start:.4f} seconds\")\n",
    "\n",
    "\n",
    "# Let's write a custom example and predict in FHE\n",
    "tested_tweet = [\"AirFrance is awesome, almost as much as Zama!\"]\n",
    "X_tested_tweet = text_to_tensor(tested_tweet, transformer_model, tokenizer, device)\n",
    "clear_proba = best_model.predict_proba(X_tested_tweet)\n",
    "\n",
    "# Now let's predict with FHE over a single tweet and print the time it takes\n",
    "start = time.perf_counter()\n",
    "decrypted_proba = best_model.predict_proba(X_tested_tweet, execute_in_fhe=True)\n",
    "end = time.perf_counter()\n",
    "fhe_exec_time = end - start\n",
    "print(f\"FHE inference time: {fhe_exec_time:.4f} seconds\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Probabilities from the FHE inference: [[0.08434131 0.05571389 0.8599448 ]]\n",
      "Probabilities from the clear model: [[0.08434131 0.05571389 0.8599448 ]]\n"
     ]
    }
   ],
   "source": [
    "print(f\"Probabilities from the FHE inference: {decrypted_proba}\")\n",
    "print(f\"Probabilities from the clear model: {clear_proba}\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Let's export the final model such that we can reuse it in a client/server environment\n",
    "\n",
    "# Export the model to ONNX\n",
    "onnx.save(best_model._onnx_model_, \"server_model.onnx\")  # pylint: disable=protected-access\n",
    "\n",
    "# Export some data to be used for compilation\n",
    "X_train_numpy = X_train_transformer[:100]\n",
    "\n",
    "# Merge the two arrays in a pandas dataframe\n",
    "X_test_numpy_df = pd.DataFrame(X_train_numpy)\n",
    "\n",
    "# to csv\n",
    "X_test_numpy_df.to_csv(\"samples_for_compilation.csv\")\n",
    "\n",
    "# Let's save the model to be pushed to a server later\n",
    "from concrete.ml.deployment import FHEModelDev\n",
    "\n",
    "fhe_api = FHEModelDev(\"sentiment_fhe_model\", best_model)\n",
    "fhe_api.save()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Accuracy</th>\n",
       "      <th>Average Precision (positive)</th>\n",
       "      <th>Average Precision (negative)</th>\n",
       "      <th>Average Precision (neutral)</th>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>Model</th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "      <th></th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>TF-IDF + XGBoost</th>\n",
       "      <td>0.681011</td>\n",
       "      <td>0.561521</td>\n",
       "      <td>0.834914</td>\n",
       "      <td>0.382002</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>Transformer Only</th>\n",
       "      <td>0.805328</td>\n",
       "      <td>0.854827</td>\n",
       "      <td>0.954804</td>\n",
       "      <td>0.680110</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>Transformer + XGBoost</th>\n",
       "      <td>0.850410</td>\n",
       "      <td>0.891691</td>\n",
       "      <td>0.959747</td>\n",
       "      <td>0.734144</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                       Accuracy  Average Precision (positive)  \\\n",
       "Model                                                           \n",
       "TF-IDF + XGBoost       0.681011                      0.561521   \n",
       "Transformer Only       0.805328                      0.854827   \n",
       "Transformer + XGBoost  0.850410                      0.891691   \n",
       "\n",
       "                       Average Precision (negative)  \\\n",
       "Model                                                 \n",
       "TF-IDF + XGBoost                           0.834914   \n",
       "Transformer Only                           0.954804   \n",
       "Transformer + XGBoost                      0.959747   \n",
       "\n",
       "                       Average Precision (neutral)  \n",
       "Model                                               \n",
       "TF-IDF + XGBoost                          0.382002  \n",
       "Transformer Only                          0.680110  \n",
       "Transformer + XGBoost                     0.734144  "
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "# Let's print the results obtained in this notebook\n",
    "df_results = pd.DataFrame(\n",
    "    {\n",
    "        \"Model\": [\"TF-IDF + XGBoost\", \"Transformer Only\", \"Transformer + XGBoost\"],\n",
    "        \"Accuracy\": [accuracy_tfidf, accuracy_transformer_only, accuracy_transformer_xgboost],\n",
    "        \"Average Precision (positive)\": [\n",
    "            ap_positive_tfidf,\n",
    "            ap_positive_transformer_only,\n",
    "            ap_positive_transformer_xgboost,\n",
    "        ],\n",
    "        \"Average Precision (negative)\": [\n",
    "            ap_negative_tfidf,\n",
    "            ap_negative_transformer_only,\n",
    "            ap_negative_transformer_xgboost,\n",
    "        ],\n",
    "        \"Average Precision (neutral)\": [\n",
    "            ap_neutral_tfidf,\n",
    "            ap_neutral_transformer_only,\n",
    "            ap_neutral_transformer_xgboost,\n",
    "        ],\n",
    "    }\n",
    ")\n",
    "df_results.set_index(\"Model\", inplace=True)\n",
    "df_results  # pylint: disable=pointless-statement"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Conclusion\n",
    "\n",
    "In this notebook we presented two different ways to represent a text.\n",
    "1. Using TF-IDF vectorization\n",
    "2. Using the hidden layers from a transformer\n",
    "\n",
    "Both representation are then used to train a machine learning model will run in FHE (here XGBoost)\n",
    "\n",
    "Once the model is trained, clients can send encrypted text representation to the server to get a sentiment analysis done and they receive the probability for each class (negative, neutral and positive) in an encrypted format which can then be decrypted by the client. For now, all the FHE magic (encrypt, predict and decrypt) is done within the `predict_proba` function with the argument `execute_in_fhe=True`. In the next release, an API will be provided to split the server/client parts.\n",
    "\n",
    "Regarding the FHE execution times, the final XGboost model can predict over an encrypted data point in ~40 seconds. This will change depending on the number of threads available. In the future, more hardware acceleration will be available to speed up the execution time.\n",
    "\n",
    "It seems that the combination of a transformer (thanks Huggingface!) with a \"simpler\" model such as XGBoost works pretty well. Thanks to Concrete-ML library, we can easily use this text representation on the client machine and then encrypt it to send it to a remote server without having to deal with a transformer runtime in FHE."
   ]
  }
 ],
 "metadata": {
  "execution": {
   "timeout": 10800
  },
  "kernelspec": {
   "display_name": ".venv",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "name": "python",
   "version": "3.10.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}