Upload NotebookPCL.ipynb
Browse files- NotebookPCL.ipynb +1163 -0
NotebookPCL.ipynb
ADDED
@@ -0,0 +1,1163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "markdown",
|
5 |
+
"id": "060994f2",
|
6 |
+
"metadata": {},
|
7 |
+
"source": [
|
8 |
+
"# importing the necessary libraries"
|
9 |
+
]
|
10 |
+
},
|
11 |
+
{
|
12 |
+
"cell_type": "code",
|
13 |
+
"execution_count": null,
|
14 |
+
"id": "033ebd27",
|
15 |
+
"metadata": {},
|
16 |
+
"outputs": [],
|
17 |
+
"source": [
|
18 |
+
"# imports - native Python\n",
|
19 |
+
"import collections\n",
|
20 |
+
"import csv\n",
|
21 |
+
"import os\n",
|
22 |
+
"import re\n",
|
23 |
+
"# imports - 3rd party\n",
|
24 |
+
"from sklearn.metrics import precision_recall_fscore_support, accuracy_score\n",
|
25 |
+
"# installs from 🤗\n",
|
26 |
+
"! pip install transformers\n",
|
27 |
+
"! pip install datasets\n",
|
28 |
+
"from transformers import AutoTokenizer, DataCollatorWithPadding, AutoModelForSequenceClassification, TrainingArguments, Trainer\n",
|
29 |
+
"from datasets import Dataset, DatasetDict"
|
30 |
+
]
|
31 |
+
},
|
32 |
+
{
|
33 |
+
"cell_type": "code",
|
34 |
+
"execution_count": null,
|
35 |
+
"id": "0214c70f",
|
36 |
+
"metadata": {},
|
37 |
+
"outputs": [],
|
38 |
+
"source": [
|
39 |
+
"import torch\n",
|
40 |
+
"torch.cuda.empty_cache()"
|
41 |
+
]
|
42 |
+
},
|
43 |
+
{
|
44 |
+
"cell_type": "markdown",
|
45 |
+
"id": "13732b06",
|
46 |
+
"metadata": {},
|
47 |
+
"source": [
|
48 |
+
"# Loading the data"
|
49 |
+
]
|
50 |
+
},
|
51 |
+
{
|
52 |
+
"cell_type": "code",
|
53 |
+
"execution_count": null,
|
54 |
+
"id": "e5a782b3",
|
55 |
+
"metadata": {},
|
56 |
+
"outputs": [],
|
57 |
+
"source": [
|
58 |
+
"# Using csv instead of pandas for sanity and to do filtering while loading\n",
|
59 |
+
"\n",
|
60 |
+
"# make parallel lists of texts and labels\n",
|
61 |
+
"# texts: strings containing messages\n",
|
62 |
+
"dataset_dict = {'text':[], 'label':[]}\n",
|
63 |
+
"for f in os.listdir():\n",
|
64 |
+
" # use all .tsv files that have been loaded\n",
|
65 |
+
" if f.endswith('dontpatronizeme.tsv'):\n",
|
66 |
+
" with open(f) as tsv_file:\n",
|
67 |
+
" reader = csv.DictReader(tsv_file, dialect='excel-tab')\n",
|
68 |
+
" for line in reader:\n",
|
69 |
+
" text = line['text']\n",
|
70 |
+
" # a few of the Message fields are empty, so we should skip those ones\n",
|
71 |
+
" if text!=None and text.strip()!=\"\":\n",
|
72 |
+
" dataset_dict['text'].append(text)\n",
|
73 |
+
" dataset_dict['label'].append(int(line['label']))\n",
|
74 |
+
"# huggingface function to convert from dict to their Dataset object\n",
|
75 |
+
"# which will work nicely with their model trainer\n",
|
76 |
+
"ds = Dataset.from_dict(dataset_dict)"
|
77 |
+
]
|
78 |
+
},
|
79 |
+
{
|
80 |
+
"cell_type": "markdown",
|
81 |
+
"id": "52379811",
|
82 |
+
"metadata": {},
|
83 |
+
"source": [
|
84 |
+
"# Creating train, valid, test splits"
|
85 |
+
]
|
86 |
+
},
|
87 |
+
{
|
88 |
+
"cell_type": "code",
|
89 |
+
"execution_count": null,
|
90 |
+
"id": "a6f69bc1",
|
91 |
+
"metadata": {},
|
92 |
+
"outputs": [],
|
93 |
+
"source": [
|
94 |
+
"# no function to split into train/validation/test so we do 2 separate splits\n",
|
95 |
+
"# first split 80-20 into train and test+validation\n",
|
96 |
+
"train_testvalid = ds.train_test_split(test_size=0.2)\n",
|
97 |
+
"# then split the 20 into 10-10 validation and test\n",
|
98 |
+
"test_valid = train_testvalid['test'].train_test_split(test_size=0.5)\n",
|
99 |
+
"# finally, make the full dataset the 80-10-10 split as a DatasetDict object\n",
|
100 |
+
"train_test_valid_dataset = DatasetDict({\n",
|
101 |
+
" 'train': train_testvalid['train'],\n",
|
102 |
+
" 'test': test_valid['test'],\n",
|
103 |
+
" 'valid': test_valid['train']})\n",
|
104 |
+
"# quick check (if this doesn't pass, will get an error in the tokenization)\n",
|
105 |
+
"# makes sure we filtered the data correcly at the beginning and removed None\n",
|
106 |
+
"for split in train_test_valid_dataset.keys():\n",
|
107 |
+
" assert not any([x==None for x in train_test_valid_dataset[split]['text']])"
|
108 |
+
]
|
109 |
+
},
|
110 |
+
{
|
111 |
+
"cell_type": "markdown",
|
112 |
+
"id": "0dfcc029",
|
113 |
+
"metadata": {},
|
114 |
+
"source": [
|
115 |
+
"# Tokenizer"
|
116 |
+
]
|
117 |
+
},
|
118 |
+
{
|
119 |
+
"cell_type": "markdown",
|
120 |
+
"id": "b2cb0082",
|
121 |
+
"metadata": {},
|
122 |
+
"source": [
|
123 |
+
"This is the tokenizer for the distilbert model"
|
124 |
+
]
|
125 |
+
},
|
126 |
+
{
|
127 |
+
"cell_type": "code",
|
128 |
+
"execution_count": null,
|
129 |
+
"id": "65a26dc2",
|
130 |
+
"metadata": {},
|
131 |
+
"outputs": [],
|
132 |
+
"source": [
|
133 |
+
"# just use the default tokenizer for the model\n",
|
134 |
+
"tokenizer = AutoTokenizer.from_pretrained('distilbert-base-uncased')\n",
|
135 |
+
"\n",
|
136 |
+
"# simple wrapper\n",
|
137 |
+
"def tokenize(examples, textfield=\"text\"):\n",
|
138 |
+
" return tokenizer(examples[textfield], padding=\"max_length\", truncation=True)\n",
|
139 |
+
"\n",
|
140 |
+
"# batch tokenization\n",
|
141 |
+
"tokenized_datasets = train_test_valid_dataset.map(tokenize, batched=True)"
|
142 |
+
]
|
143 |
+
},
|
144 |
+
{
|
145 |
+
"cell_type": "markdown",
|
146 |
+
"id": "38a15ebb",
|
147 |
+
"metadata": {},
|
148 |
+
"source": [
|
149 |
+
"Below are the examples for also the RoBERTa model and the BERT model"
|
150 |
+
]
|
151 |
+
},
|
152 |
+
{
|
153 |
+
"cell_type": "code",
|
154 |
+
"execution_count": null,
|
155 |
+
"id": "8f45cf1d",
|
156 |
+
"metadata": {},
|
157 |
+
"outputs": [],
|
158 |
+
"source": [
|
159 |
+
"from transformers import AutoTokenizer, AutoModelForMaskedLM\n",
|
160 |
+
"\n",
|
161 |
+
"tokenizer = AutoTokenizer.from_pretrained(\"bert-base-uncased\")\n",
|
162 |
+
"\n",
|
163 |
+
"model = AutoModelForMaskedLM.from_pretrained(\"bert-base-uncased\")"
|
164 |
+
]
|
165 |
+
},
|
166 |
+
{
|
167 |
+
"cell_type": "code",
|
168 |
+
"execution_count": null,
|
169 |
+
"id": "79d33a06",
|
170 |
+
"metadata": {},
|
171 |
+
"outputs": [],
|
172 |
+
"source": [
|
173 |
+
"from transformers import AutoTokenizer, AutoModelForMaskedLM\n",
|
174 |
+
"\n",
|
175 |
+
"tokenizer = AutoTokenizer.from_pretrained(\"roberta-base\")\n",
|
176 |
+
"\n",
|
177 |
+
"model = AutoModelForMaskedLM.from_pretrained(\"roberta-base\")"
|
178 |
+
]
|
179 |
+
},
|
180 |
+
{
|
181 |
+
"cell_type": "markdown",
|
182 |
+
"id": "9b550e83",
|
183 |
+
"metadata": {},
|
184 |
+
"source": [
|
185 |
+
"# Model "
|
186 |
+
]
|
187 |
+
},
|
188 |
+
{
|
189 |
+
"cell_type": "code",
|
190 |
+
"execution_count": null,
|
191 |
+
"id": "12c960c0",
|
192 |
+
"metadata": {},
|
193 |
+
"outputs": [],
|
194 |
+
"source": [
|
195 |
+
"# Setup collation\n",
|
196 |
+
"data_collator = DataCollatorWithPadding(tokenizer=tokenizer)\n",
|
197 |
+
"\n",
|
198 |
+
"# Load model\n",
|
199 |
+
"model = AutoModelForSequenceClassification.from_pretrained(\"distilbert-base-uncased\", num_labels=2)"
|
200 |
+
]
|
201 |
+
},
|
202 |
+
{
|
203 |
+
"cell_type": "markdown",
|
204 |
+
"id": "d4342956",
|
205 |
+
"metadata": {},
|
206 |
+
"source": [
|
207 |
+
"# Computing the metrics and training args"
|
208 |
+
]
|
209 |
+
},
|
210 |
+
{
|
211 |
+
"cell_type": "code",
|
212 |
+
"execution_count": null,
|
213 |
+
"id": "4c974458",
|
214 |
+
"metadata": {},
|
215 |
+
"outputs": [],
|
216 |
+
"source": [
|
217 |
+
"# using sklearn to compute precision, recall, f1, and accuracy\n",
|
218 |
+
"def compute_metrics(pred):\n",
|
219 |
+
" labels = pred.label_ids\n",
|
220 |
+
" preds = pred.predictions.argmax(-1)\n",
|
221 |
+
" precision, recall, f1, _ = precision_recall_fscore_support(labels, preds, average='binary')\n",
|
222 |
+
" acc = accuracy_score(labels, preds)\n",
|
223 |
+
" return {\n",
|
224 |
+
" 'accuracy': acc,\n",
|
225 |
+
" 'f1': f1,\n",
|
226 |
+
" 'precision': precision,\n",
|
227 |
+
" 'recall': recall\n",
|
228 |
+
" }"
|
229 |
+
]
|
230 |
+
},
|
231 |
+
{
|
232 |
+
"cell_type": "code",
|
233 |
+
"execution_count": null,
|
234 |
+
"id": "8c4fb414",
|
235 |
+
"metadata": {},
|
236 |
+
"outputs": [],
|
237 |
+
"source": [
|
238 |
+
"# Set training args (just using defaults from the following tutorial for now:\n",
|
239 |
+
"# https://huggingface.co/docs/transformers/training )\n",
|
240 |
+
"training_args = TrainingArguments(\n",
|
241 |
+
" output_dir=\"./results\",\n",
|
242 |
+
" learning_rate=2e-5,\n",
|
243 |
+
" per_device_train_batch_size=16,\n",
|
244 |
+
" per_device_eval_batch_size=16,\n",
|
245 |
+
" num_train_epochs=5,\n",
|
246 |
+
" weight_decay=0.01,\n",
|
247 |
+
")\n",
|
248 |
+
"\n",
|
249 |
+
"# setup the trainer\n",
|
250 |
+
"trainer = Trainer(\n",
|
251 |
+
" model=model,\n",
|
252 |
+
" args=training_args,\n",
|
253 |
+
" train_dataset=tokenized_datasets[\"train\"],\n",
|
254 |
+
" eval_dataset=tokenized_datasets[\"valid\"],\n",
|
255 |
+
" tokenizer=tokenizer,\n",
|
256 |
+
" data_collator=data_collator,\n",
|
257 |
+
" compute_metrics=compute_metrics,\n",
|
258 |
+
")"
|
259 |
+
]
|
260 |
+
},
|
261 |
+
{
|
262 |
+
"cell_type": "markdown",
|
263 |
+
"id": "cb346507",
|
264 |
+
"metadata": {},
|
265 |
+
"source": [
|
266 |
+
"# Train model and Evaluate"
|
267 |
+
]
|
268 |
+
},
|
269 |
+
{
|
270 |
+
"cell_type": "code",
|
271 |
+
"execution_count": null,
|
272 |
+
"id": "de170b1e",
|
273 |
+
"metadata": {},
|
274 |
+
"outputs": [],
|
275 |
+
"source": [
|
276 |
+
"# train the model\n",
|
277 |
+
"trainer.train()"
|
278 |
+
]
|
279 |
+
},
|
280 |
+
{
|
281 |
+
"cell_type": "code",
|
282 |
+
"execution_count": null,
|
283 |
+
"id": "48adbaed",
|
284 |
+
"metadata": {},
|
285 |
+
"outputs": [],
|
286 |
+
"source": [
|
287 |
+
"# evaluate on the test set\n",
|
288 |
+
"# should only do for _best_ model of each type \n",
|
289 |
+
"# after selecting hyperparameters that work best on validation set\n",
|
290 |
+
"trainer.evaluate(tokenized_datasets[\"test\"])"
|
291 |
+
]
|
292 |
+
},
|
293 |
+
{
|
294 |
+
"cell_type": "code",
|
295 |
+
"execution_count": null,
|
296 |
+
"id": "c3dea644",
|
297 |
+
"metadata": {},
|
298 |
+
"outputs": [],
|
299 |
+
"source": [
|
300 |
+
"##!pip install huggingface_hub\n",
|
301 |
+
"#!sudo apt-get install fit-lfs\n",
|
302 |
+
"#!huggingface-cli login\n",
|
303 |
+
"#!git clone https://huggingface.co/achyut/patronizing_detection\n",
|
304 |
+
"#cd /content/patronizing_detection"
|
305 |
+
]
|
306 |
+
},
|
307 |
+
{
|
308 |
+
"cell_type": "markdown",
|
309 |
+
"id": "539c8683",
|
310 |
+
"metadata": {},
|
311 |
+
"source": [
|
312 |
+
"# LIME for Deep Learning Models"
|
313 |
+
]
|
314 |
+
},
|
315 |
+
{
|
316 |
+
"cell_type": "code",
|
317 |
+
"execution_count": null,
|
318 |
+
"id": "9f7c2cab",
|
319 |
+
"metadata": {},
|
320 |
+
"outputs": [],
|
321 |
+
"source": [
|
322 |
+
"# LIME importing all the necessary libraries\n",
|
323 |
+
"import numpy as np\n",
|
324 |
+
"import lime\n",
|
325 |
+
"import torch\n",
|
326 |
+
"import torch.nn.functional as F\n",
|
327 |
+
"from lime.lime_text import LimeTextExplainer\n",
|
328 |
+
"from transformers import AutoTokenizer, AutoModelForSequenceClassification"
|
329 |
+
]
|
330 |
+
},
|
331 |
+
{
|
332 |
+
"cell_type": "code",
|
333 |
+
"execution_count": null,
|
334 |
+
"id": "d53f4b7d",
|
335 |
+
"metadata": {},
|
336 |
+
"outputs": [],
|
337 |
+
"source": [
|
338 |
+
"# Set the class names\n",
|
339 |
+
"class_names = ['non-patronizing','patronizing']"
|
340 |
+
]
|
341 |
+
},
|
342 |
+
{
|
343 |
+
"cell_type": "markdown",
|
344 |
+
"id": "2d91f290",
|
345 |
+
"metadata": {},
|
346 |
+
"source": [
|
347 |
+
"For LIME and other interpretable AI models, we Have to use the tokenizer and the model of the fine-tuned pretrained model. Not the Huggingface un fine tuned model. That is because we want to use the model with the trained weights, tokens and vocab"
|
348 |
+
]
|
349 |
+
},
|
350 |
+
{
|
351 |
+
"cell_type": "code",
|
352 |
+
"execution_count": null,
|
353 |
+
"id": "e2381d7b",
|
354 |
+
"metadata": {},
|
355 |
+
"outputs": [],
|
356 |
+
"source": [
|
357 |
+
"tokenizer = AutoTokenizer.from_pretrained(\"achyut/patronizing_detection\")\n",
|
358 |
+
"\n",
|
359 |
+
"model = AutoModelForSequenceClassification.from_pretrained(\"achyut/patronizing_detection\")"
|
360 |
+
]
|
361 |
+
},
|
362 |
+
{
|
363 |
+
"cell_type": "code",
|
364 |
+
"execution_count": null,
|
365 |
+
"id": "318859d6",
|
366 |
+
"metadata": {},
|
367 |
+
"outputs": [],
|
368 |
+
"source": [
|
369 |
+
"model.cuda()"
|
370 |
+
]
|
371 |
+
},
|
372 |
+
{
|
373 |
+
"cell_type": "code",
|
374 |
+
"execution_count": null,
|
375 |
+
"id": "99a7e69f",
|
376 |
+
"metadata": {},
|
377 |
+
"outputs": [],
|
378 |
+
"source": [
|
379 |
+
"!pip install more_itertools\n"
|
380 |
+
]
|
381 |
+
},
|
382 |
+
{
|
383 |
+
"cell_type": "markdown",
|
384 |
+
"id": "c810588c",
|
385 |
+
"metadata": {},
|
386 |
+
"source": [
|
387 |
+
"# The function that calculates the logits for each sequence. "
|
388 |
+
]
|
389 |
+
},
|
390 |
+
{
|
391 |
+
"cell_type": "code",
|
392 |
+
"execution_count": null,
|
393 |
+
"id": "c3db6441",
|
394 |
+
"metadata": {},
|
395 |
+
"outputs": [],
|
396 |
+
"source": [
|
397 |
+
"import more_itertools\n",
|
398 |
+
"def predictor4(texts, batch_size=64):\n",
|
399 |
+
" probas = []\n",
|
400 |
+
" for chunk in more_itertools.chunked(texts, batch_size):\n",
|
401 |
+
" tokenized = tokenizer(chunk, return_tensors=\"pt\", padding=True)\n",
|
402 |
+
" outputs = model(tokenized['input_ids'].to('cuda'), tokenized['attention_mask'].to('cuda'))\n",
|
403 |
+
" probas.append(F.softmax(outputs.logits).cpu().detach().numpy())\n",
|
404 |
+
" return np.vstack(probas)"
|
405 |
+
]
|
406 |
+
},
|
407 |
+
{
|
408 |
+
"cell_type": "code",
|
409 |
+
"execution_count": null,
|
410 |
+
"id": "1074572d",
|
411 |
+
"metadata": {},
|
412 |
+
"outputs": [],
|
413 |
+
"source": [
|
414 |
+
"predictor4([\"I have two dogs\",\"The keep barking\"])"
|
415 |
+
]
|
416 |
+
},
|
417 |
+
{
|
418 |
+
"cell_type": "code",
|
419 |
+
"execution_count": null,
|
420 |
+
"id": "661d8281",
|
421 |
+
"metadata": {},
|
422 |
+
"outputs": [],
|
423 |
+
"source": [
|
424 |
+
"explainer = LimeTextExplainer(class_names=class_names)"
|
425 |
+
]
|
426 |
+
},
|
427 |
+
{
|
428 |
+
"cell_type": "code",
|
429 |
+
"execution_count": null,
|
430 |
+
"id": "abb9b201",
|
431 |
+
"metadata": {},
|
432 |
+
"outputs": [],
|
433 |
+
"source": [
|
434 |
+
"str_to_predict = ds[6]['text']\n",
|
435 |
+
"exp = explainer.explain_instance(str_to_predict, predictor4, num_features= 25, num_samples = 2000)\n",
|
436 |
+
"exp.show_in_notebook(text=str_to_predict)"
|
437 |
+
]
|
438 |
+
},
|
439 |
+
{
|
440 |
+
"cell_type": "code",
|
441 |
+
"execution_count": null,
|
442 |
+
"id": "1885619b",
|
443 |
+
"metadata": {},
|
444 |
+
"outputs": [],
|
445 |
+
"source": [
|
446 |
+
"exp.as_list()"
|
447 |
+
]
|
448 |
+
},
|
449 |
+
{
|
450 |
+
"cell_type": "code",
|
451 |
+
"execution_count": null,
|
452 |
+
"id": "5f004287",
|
453 |
+
"metadata": {},
|
454 |
+
"outputs": [],
|
455 |
+
"source": []
|
456 |
+
},
|
457 |
+
{
|
458 |
+
"cell_type": "markdown",
|
459 |
+
"id": "42dfbb84",
|
460 |
+
"metadata": {},
|
461 |
+
"source": [
|
462 |
+
"# classical Machine Learning"
|
463 |
+
]
|
464 |
+
},
|
465 |
+
{
|
466 |
+
"cell_type": "code",
|
467 |
+
"execution_count": null,
|
468 |
+
"id": "94835013",
|
469 |
+
"metadata": {},
|
470 |
+
"outputs": [],
|
471 |
+
"source": [
|
472 |
+
"import collections\n",
|
473 |
+
"import csv\n",
|
474 |
+
"import os\n",
|
475 |
+
"import re\n",
|
476 |
+
"import pandas as pd\n",
|
477 |
+
"import numpy as np\n",
|
478 |
+
"from nltk.tokenize import word_tokenize\n",
|
479 |
+
"from sklearn.preprocessing import LabelEncoder\n",
|
480 |
+
"from collections import defaultdict\n",
|
481 |
+
"from nltk.corpus import wordnet as wn\n",
|
482 |
+
"from sklearn.feature_extraction.text import TfidfVectorizer\n",
|
483 |
+
"from sklearn import model_selection, naive_bayes, svm\n",
|
484 |
+
"from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score\n",
|
485 |
+
"from nltk import pos_tag\n",
|
486 |
+
"from nltk.corpus import stopwords\n",
|
487 |
+
"from nltk.stem import WordNetLemmatizer"
|
488 |
+
]
|
489 |
+
},
|
490 |
+
{
|
491 |
+
"cell_type": "code",
|
492 |
+
"execution_count": null,
|
493 |
+
"id": "8605ed57",
|
494 |
+
"metadata": {},
|
495 |
+
"outputs": [],
|
496 |
+
"source": [
|
497 |
+
"# We can use a seed if we want reproducibility\n",
|
498 |
+
"#np.random.seed(500)"
|
499 |
+
]
|
500 |
+
},
|
501 |
+
{
|
502 |
+
"cell_type": "code",
|
503 |
+
"execution_count": null,
|
504 |
+
"id": "5475808d",
|
505 |
+
"metadata": {},
|
506 |
+
"outputs": [],
|
507 |
+
"source": [
|
508 |
+
"import nltk\n",
|
509 |
+
"nltk.download('wordnet')"
|
510 |
+
]
|
511 |
+
},
|
512 |
+
{
|
513 |
+
"cell_type": "code",
|
514 |
+
"execution_count": null,
|
515 |
+
"id": "c3745eee",
|
516 |
+
"metadata": {},
|
517 |
+
"outputs": [],
|
518 |
+
"source": [
|
519 |
+
"import nltk\n",
|
520 |
+
"nltk.download('averaged_perceptron_tagger')"
|
521 |
+
]
|
522 |
+
},
|
523 |
+
{
|
524 |
+
"cell_type": "code",
|
525 |
+
"execution_count": null,
|
526 |
+
"id": "180f42bf",
|
527 |
+
"metadata": {},
|
528 |
+
"outputs": [],
|
529 |
+
"source": [
|
530 |
+
"Corpus = pd.read_csv(\"patro_downsampled.csv\", names = ['text','label'])\n",
|
531 |
+
"# change it to str, lower case and drop the na values\n",
|
532 |
+
"Corpus.text = Corpus.text.astype(str)\n",
|
533 |
+
"Corpus['text'] = Corpus['text'].str.lower()\n",
|
534 |
+
"Corpus = Corpus.dropna()\n",
|
535 |
+
"Corpus.head()"
|
536 |
+
]
|
537 |
+
},
|
538 |
+
{
|
539 |
+
"cell_type": "code",
|
540 |
+
"execution_count": null,
|
541 |
+
"id": "5f9d00c8",
|
542 |
+
"metadata": {},
|
543 |
+
"outputs": [],
|
544 |
+
"source": [
|
545 |
+
"Corpus.info()"
|
546 |
+
]
|
547 |
+
},
|
548 |
+
{
|
549 |
+
"cell_type": "code",
|
550 |
+
"execution_count": null,
|
551 |
+
"id": "659d463e",
|
552 |
+
"metadata": {},
|
553 |
+
"outputs": [],
|
554 |
+
"source": [
|
555 |
+
"#tokenizing our para text column here\n",
|
556 |
+
"Corpus['text'] = Corpus['text'].apply(nltk.word_tokenize)\n",
|
557 |
+
"\n",
|
558 |
+
"# Tagging to understand if the word is a noun, verb, adverb etc\n",
|
559 |
+
"\n",
|
560 |
+
"tag_map = defaultdict(lambda : wn.NOUN)\n",
|
561 |
+
"tag_map['J'] = wn.ADJ\n",
|
562 |
+
"tag_map['V'] = wn.VERB\n",
|
563 |
+
"tag_map['R'] = wn.ADV"
|
564 |
+
]
|
565 |
+
},
|
566 |
+
{
|
567 |
+
"cell_type": "code",
|
568 |
+
"execution_count": null,
|
569 |
+
"id": "5af9ea94",
|
570 |
+
"metadata": {},
|
571 |
+
"outputs": [],
|
572 |
+
"source": [
|
573 |
+
"for index,entry in enumerate(Corpus['text']):\n",
|
574 |
+
" # empty list which I will append to the df in the end.\n",
|
575 |
+
" Final_words = []\n",
|
576 |
+
" \n",
|
577 |
+
" word_Lemmatized = WordNetLemmatizer()\n",
|
578 |
+
" for word, tag in pos_tag(entry):\n",
|
579 |
+
" # check for Stop words and consider only alphabets\n",
|
580 |
+
" if word not in stopwords.words('english') and word.isalpha():\n",
|
581 |
+
" word_Final = word_Lemmatized.lemmatize(word,tag_map[tag[0]])\n",
|
582 |
+
" Final_words.append(word_Final)\n",
|
583 |
+
" # The final processed set of words for each iteration will be stored in 'text_final'\n",
|
584 |
+
" Corpus.loc[index,'text_final'] = str(Final_words)"
|
585 |
+
]
|
586 |
+
},
|
587 |
+
{
|
588 |
+
"cell_type": "code",
|
589 |
+
"execution_count": null,
|
590 |
+
"id": "8c6d9bc6",
|
591 |
+
"metadata": {},
|
592 |
+
"outputs": [],
|
593 |
+
"source": [
|
594 |
+
"Corpus.head()"
|
595 |
+
]
|
596 |
+
},
|
597 |
+
{
|
598 |
+
"cell_type": "code",
|
599 |
+
"execution_count": null,
|
600 |
+
"id": "f654c4ab",
|
601 |
+
"metadata": {},
|
602 |
+
"outputs": [],
|
603 |
+
"source": [
|
604 |
+
"#Train, test split\n",
|
605 |
+
"Train_X, Test_X, Train_Y, Test_Y = model_selection.train_test_split(Corpus['text_final'],\n",
|
606 |
+
" Corpus['label'],\n",
|
607 |
+
" test_size=0.2)"
|
608 |
+
]
|
609 |
+
},
|
610 |
+
{
|
611 |
+
"cell_type": "code",
|
612 |
+
"execution_count": null,
|
613 |
+
"id": "00747dbd",
|
614 |
+
"metadata": {},
|
615 |
+
"outputs": [],
|
616 |
+
"source": [
|
617 |
+
"#Encoding our labels\n",
|
618 |
+
"Encoder = LabelEncoder()\n",
|
619 |
+
"Train_Y = Encoder.fit_transform(Train_Y)\n",
|
620 |
+
"Test_Y = Encoder.fit_transform(Test_Y)\n",
|
621 |
+
"\n",
|
622 |
+
"# Vectorizer\n",
|
623 |
+
"Tfidf_vect = TfidfVectorizer()\n",
|
624 |
+
"\n",
|
625 |
+
"Tfidf_vect.fit(Corpus['text_final'])"
|
626 |
+
]
|
627 |
+
},
|
628 |
+
{
|
629 |
+
"cell_type": "code",
|
630 |
+
"execution_count": null,
|
631 |
+
"id": "95b89126",
|
632 |
+
"metadata": {},
|
633 |
+
"outputs": [],
|
634 |
+
"source": [
|
635 |
+
"# Transforming the train and test inputs into vectors\n",
|
636 |
+
"Train_X_Tfidf = Tfidf_vect.transform(Train_X)\n",
|
637 |
+
"Test_X_Tfidf = Tfidf_vect.transform(Test_X)\n",
|
638 |
+
"print(len(Tfidf_vect.vocabulary_))"
|
639 |
+
]
|
640 |
+
},
|
641 |
+
{
|
642 |
+
"cell_type": "markdown",
|
643 |
+
"id": "1da1f215",
|
644 |
+
"metadata": {},
|
645 |
+
"source": [
|
646 |
+
"# Fitting Models"
|
647 |
+
]
|
648 |
+
},
|
649 |
+
{
|
650 |
+
"cell_type": "markdown",
|
651 |
+
"id": "b8d618cd",
|
652 |
+
"metadata": {},
|
653 |
+
"source": [
|
654 |
+
"## NaiveBayes"
|
655 |
+
]
|
656 |
+
},
|
657 |
+
{
|
658 |
+
"cell_type": "code",
|
659 |
+
"execution_count": null,
|
660 |
+
"id": "7613821b",
|
661 |
+
"metadata": {},
|
662 |
+
"outputs": [],
|
663 |
+
"source": [
|
664 |
+
"# fit the NB classifier\n",
|
665 |
+
"Naive = naive_bayes.MultinomialNB()\n",
|
666 |
+
"naive_model = Naive.fit(Train_X_Tfidf,Train_Y)\n",
|
667 |
+
"predictions_NB = Naive.predict(Test_X_Tfidf)\n",
|
668 |
+
"print(\"Naive Bayes Accuracy Score -> \",accuracy_score(predictions_NB, Test_Y)*100)"
|
669 |
+
]
|
670 |
+
},
|
671 |
+
{
|
672 |
+
"cell_type": "code",
|
673 |
+
"execution_count": null,
|
674 |
+
"id": "d04b0813",
|
675 |
+
"metadata": {},
|
676 |
+
"outputs": [],
|
677 |
+
"source": [
|
678 |
+
"print(f1_score(predictions_NB, Test_Y),precision_score(predictions_NB, Test_Y),recall_score(predictions_NB, Test_Y))"
|
679 |
+
]
|
680 |
+
},
|
681 |
+
{
|
682 |
+
"cell_type": "markdown",
|
683 |
+
"id": "539cb258",
|
684 |
+
"metadata": {},
|
685 |
+
"source": [
|
686 |
+
"## SVM"
|
687 |
+
]
|
688 |
+
},
|
689 |
+
{
|
690 |
+
"cell_type": "code",
|
691 |
+
"execution_count": null,
|
692 |
+
"id": "cf9ebed3",
|
693 |
+
"metadata": {},
|
694 |
+
"outputs": [],
|
695 |
+
"source": [
|
696 |
+
"#SVM classifier\n",
|
697 |
+
"SVM = svm.SVC(C=2.0, kernel='poly',degree=2, gamma='scale')\n",
|
698 |
+
"svm_model = SVM.fit(Train_X_Tfidf,Train_Y)\n",
|
699 |
+
"predictions_SVM = SVM.predict(Test_X_Tfidf)\n",
|
700 |
+
"print(\"SVM Accuracy Score -> \",accuracy_score(predictions_SVM, Test_Y)*100)"
|
701 |
+
]
|
702 |
+
},
|
703 |
+
{
|
704 |
+
"cell_type": "code",
|
705 |
+
"execution_count": null,
|
706 |
+
"id": "1fbf3e41",
|
707 |
+
"metadata": {},
|
708 |
+
"outputs": [],
|
709 |
+
"source": [
|
710 |
+
"print(f1_score(predictions_SVM, Test_Y),precision_score(predictions_SVM, Test_Y),recall_score(predictions_SVM, Test_Y))"
|
711 |
+
]
|
712 |
+
},
|
713 |
+
{
|
714 |
+
"cell_type": "code",
|
715 |
+
"execution_count": null,
|
716 |
+
"id": "81cf1425",
|
717 |
+
"metadata": {},
|
718 |
+
"outputs": [],
|
719 |
+
"source": [
|
720 |
+
"scores = cross_val_score(SVM,Train_X_Tfidf,Train_Y, cv = 5 , scoring = 'f1_macro')\n",
|
721 |
+
"scores"
|
722 |
+
]
|
723 |
+
},
|
724 |
+
{
|
725 |
+
"cell_type": "code",
|
726 |
+
"execution_count": null,
|
727 |
+
"id": "c5a07117",
|
728 |
+
"metadata": {},
|
729 |
+
"outputs": [],
|
730 |
+
"source": [
|
731 |
+
"scores = cross_val_score(SVM,Train_X_Tfidf,Train_Y, cv = 10 , scoring = 'f1_macro')\n",
|
732 |
+
"scores"
|
733 |
+
]
|
734 |
+
},
|
735 |
+
{
|
736 |
+
"cell_type": "markdown",
|
737 |
+
"id": "a4dea60f",
|
738 |
+
"metadata": {},
|
739 |
+
"source": [
|
740 |
+
"## Logistic Regression"
|
741 |
+
]
|
742 |
+
},
|
743 |
+
{
|
744 |
+
"cell_type": "code",
|
745 |
+
"execution_count": null,
|
746 |
+
"id": "7c96b88d",
|
747 |
+
"metadata": {},
|
748 |
+
"outputs": [],
|
749 |
+
"source": [
|
750 |
+
"from sklearn.linear_model import LogisticRegression\n",
|
751 |
+
"logisticReg = LogisticRegression()\n",
|
752 |
+
"logisticReg.fit(Train_X_Tfidf,Train_Y)\n",
|
753 |
+
"predictions_LR = logisticReg.predict(Test_X_Tfidf)\n",
|
754 |
+
"print(\"LR Accuracy Score -> \",accuracy_score(predictions_LR, Test_Y)*100)"
|
755 |
+
]
|
756 |
+
},
|
757 |
+
{
|
758 |
+
"cell_type": "code",
|
759 |
+
"execution_count": null,
|
760 |
+
"id": "47750ca0",
|
761 |
+
"metadata": {},
|
762 |
+
"outputs": [],
|
763 |
+
"source": [
|
764 |
+
"print(f1_score(predictions_LR, Test_Y), precision_score(predictions_LR, Test_Y),recall_score(predictions_LR, Test_Y))"
|
765 |
+
]
|
766 |
+
},
|
767 |
+
{
|
768 |
+
"cell_type": "markdown",
|
769 |
+
"id": "75efc6b3",
|
770 |
+
"metadata": {},
|
771 |
+
"source": [
|
772 |
+
"## RandomForest"
|
773 |
+
]
|
774 |
+
},
|
775 |
+
{
|
776 |
+
"cell_type": "code",
|
777 |
+
"execution_count": null,
|
778 |
+
"id": "144104e6",
|
779 |
+
"metadata": {},
|
780 |
+
"outputs": [],
|
781 |
+
"source": [
|
782 |
+
"# Apply random forest on the data\n",
|
783 |
+
"from sklearn.ensemble import RandomForestClassifier\n",
|
784 |
+
"randomForest = RandomForestClassifier(n_estimators = 50) \n",
|
785 |
+
"randomForest.fit(Train_X_Tfidf,Train_Y)\n",
|
786 |
+
"predictions_RF = logisticReg.predict(Test_X_Tfidf)\n",
|
787 |
+
"print(\"LR Accuracy Score -> \",accuracy_score(predictions_RF, Test_Y)*100)"
|
788 |
+
]
|
789 |
+
},
|
790 |
+
{
|
791 |
+
"cell_type": "code",
|
792 |
+
"execution_count": null,
|
793 |
+
"id": "1f083f5e",
|
794 |
+
"metadata": {},
|
795 |
+
"outputs": [],
|
796 |
+
"source": [
|
797 |
+
"print(f1_score(predictions_RF, Test_Y),precision_score(predictions_RF, Test_Y),recall_score(predictions_RF, Test_Y))"
|
798 |
+
]
|
799 |
+
},
|
800 |
+
{
|
801 |
+
"cell_type": "markdown",
|
802 |
+
"id": "03fb7cc8",
|
803 |
+
"metadata": {},
|
804 |
+
"source": [
|
805 |
+
"# LIME for classical ML"
|
806 |
+
]
|
807 |
+
},
|
808 |
+
{
|
809 |
+
"cell_type": "code",
|
810 |
+
"execution_count": null,
|
811 |
+
"id": "41fa18be",
|
812 |
+
"metadata": {},
|
813 |
+
"outputs": [],
|
814 |
+
"source": [
|
815 |
+
"import lime\n",
|
816 |
+
"import sklearn.ensemble\n",
|
817 |
+
"from __future__ import print_function\n",
|
818 |
+
"from lime import lime_text\n",
|
819 |
+
"from sklearn.pipeline import make_pipeline\n",
|
820 |
+
"from lime.lime_text import LimeTextExplainer"
|
821 |
+
]
|
822 |
+
},
|
823 |
+
{
|
824 |
+
"cell_type": "markdown",
|
825 |
+
"id": "d952eb5d",
|
826 |
+
"metadata": {},
|
827 |
+
"source": [
|
828 |
+
"## Make the pipeline"
|
829 |
+
]
|
830 |
+
},
|
831 |
+
{
|
832 |
+
"cell_type": "code",
|
833 |
+
"execution_count": null,
|
834 |
+
"id": "f96a244e",
|
835 |
+
"metadata": {},
|
836 |
+
"outputs": [],
|
837 |
+
"source": [
|
838 |
+
"c = make_pipeline(Tfidf_vect, logisticred_model)\n",
|
839 |
+
"ls_X_test= list(Test_X)\n",
|
840 |
+
"class_names = {0: 'patro', 1:'non-patro'}\n",
|
841 |
+
"LIME_explainer = LimeTextExplainer(class_names=class_names)\n"
|
842 |
+
]
|
843 |
+
},
|
844 |
+
{
|
845 |
+
"cell_type": "code",
|
846 |
+
"execution_count": null,
|
847 |
+
"id": "c0a727a1",
|
848 |
+
"metadata": {},
|
849 |
+
"outputs": [],
|
850 |
+
"source": [
|
851 |
+
"idx = 15\n",
|
852 |
+
"LIME_exp = LIME_explainer.explain_instance(ls_X_test[idx], c.predict_proba)"
|
853 |
+
]
|
854 |
+
},
|
855 |
+
{
|
856 |
+
"cell_type": "code",
|
857 |
+
"execution_count": null,
|
858 |
+
"id": "b1755fc8",
|
859 |
+
"metadata": {},
|
860 |
+
"outputs": [],
|
861 |
+
"source": [
|
862 |
+
"print('Document id: %d' % idx)\n",
|
863 |
+
"print('Text: ', ls_X_test[idx])\n",
|
864 |
+
"print('Probability =', c.predict_proba([ls_X_test[idx]]).round(3)[0,1])\n",
|
865 |
+
"print('True class: %s' % class_names.get(list(Test_Y)[idx]))"
|
866 |
+
]
|
867 |
+
},
|
868 |
+
{
|
869 |
+
"cell_type": "code",
|
870 |
+
"execution_count": null,
|
871 |
+
"id": "78b0d22e",
|
872 |
+
"metadata": {},
|
873 |
+
"outputs": [],
|
874 |
+
"source": [
|
875 |
+
"print(\"1 = non-Patro class, 0 = Patro class\")\n",
|
876 |
+
"# show the explainability results with highlighted text\n",
|
877 |
+
"LIME_exp.show_in_notebook(text=True)"
|
878 |
+
]
|
879 |
+
},
|
880 |
+
{
|
881 |
+
"cell_type": "code",
|
882 |
+
"execution_count": null,
|
883 |
+
"id": "e3e16b80",
|
884 |
+
"metadata": {},
|
885 |
+
"outputs": [],
|
886 |
+
"source": [
|
887 |
+
"idx = 45\n",
|
888 |
+
"LIME_exp = LIME_explainer.explain_instance(ls_X_test[idx], c.predict_proba)\n",
|
889 |
+
"print('Document id: %d' % idx)\n",
|
890 |
+
"print('Text: ', ls_X_test[idx])\n",
|
891 |
+
"print('Probability =', c.predict_proba([ls_X_test[idx]]).round(3)[0,1])\n",
|
892 |
+
"print('True class: %s' % class_names.get(list(Test_Y)[idx]))"
|
893 |
+
]
|
894 |
+
},
|
895 |
+
{
|
896 |
+
"cell_type": "code",
|
897 |
+
"execution_count": null,
|
898 |
+
"id": "bd8e838a",
|
899 |
+
"metadata": {},
|
900 |
+
"outputs": [],
|
901 |
+
"source": [
|
902 |
+
"print(\"1 = non-Patro class, 0 = Patro class\")\n",
|
903 |
+
"# show the explainability results with highlighted text\n",
|
904 |
+
"LIME_exp.show_in_notebook(text=True)"
|
905 |
+
]
|
906 |
+
},
|
907 |
+
{
|
908 |
+
"cell_type": "markdown",
|
909 |
+
"id": "f8f07e74",
|
910 |
+
"metadata": {},
|
911 |
+
"source": [
|
912 |
+
"# Topic Modeling"
|
913 |
+
]
|
914 |
+
},
|
915 |
+
{
|
916 |
+
"cell_type": "code",
|
917 |
+
"execution_count": null,
|
918 |
+
"id": "2825b328",
|
919 |
+
"metadata": {},
|
920 |
+
"outputs": [],
|
921 |
+
"source": [
|
922 |
+
"import pandas as pd\n",
|
923 |
+
"import numpy as np \n",
|
924 |
+
"import re\n",
|
925 |
+
"from wordcloud import WordCloud\n",
|
926 |
+
"import gensim\n",
|
927 |
+
"from gensim.utils import simple_preprocess\n",
|
928 |
+
"from nltk.corpus import stopwords\n",
|
929 |
+
"import gensim.corpora as corpora\n",
|
930 |
+
"from pprint import pprint\n",
|
931 |
+
"import pyLDAvis.gensim_models\n",
|
932 |
+
"import pickle\n",
|
933 |
+
"import pyLDAvis"
|
934 |
+
]
|
935 |
+
},
|
936 |
+
{
|
937 |
+
"cell_type": "code",
|
938 |
+
"execution_count": null,
|
939 |
+
"id": "71ab6908",
|
940 |
+
"metadata": {},
|
941 |
+
"outputs": [],
|
942 |
+
"source": [
|
943 |
+
"df = pd.read_csv(\"dontpatronizeme.csv\", names = ['Message','label'])"
|
944 |
+
]
|
945 |
+
},
|
946 |
+
{
|
947 |
+
"cell_type": "code",
|
948 |
+
"execution_count": null,
|
949 |
+
"id": "0c4a0602",
|
950 |
+
"metadata": {},
|
951 |
+
"outputs": [],
|
952 |
+
"source": [
|
953 |
+
"df[\"Message_processed\"] = df[\"Message\"].map(lambda x: re.sub('[,\\.!?]', '', str(x)))\n",
|
954 |
+
"df['Message_processed'] = df['Message_processed'].map(lambda x: x.lower())\n",
|
955 |
+
"df['Message_processed'].head()"
|
956 |
+
]
|
957 |
+
},
|
958 |
+
{
|
959 |
+
"cell_type": "code",
|
960 |
+
"execution_count": null,
|
961 |
+
"id": "0e507f49",
|
962 |
+
"metadata": {},
|
963 |
+
"outputs": [],
|
964 |
+
"source": [
|
965 |
+
"long_string = ','.join(list(df['Message_processed'].values))# Create a WordCloud object\n",
|
966 |
+
"wordcloud = WordCloud(background_color=\"white\", max_words=5000, contour_width=3, contour_color='steelblue')# Generate a word cloud\n",
|
967 |
+
"wordcloud.generate(long_string)# Visualize the word cloud\n",
|
968 |
+
"wordcloud.to_image()"
|
969 |
+
]
|
970 |
+
},
|
971 |
+
{
|
972 |
+
"cell_type": "code",
|
973 |
+
"execution_count": null,
|
974 |
+
"id": "76a3f280",
|
975 |
+
"metadata": {},
|
976 |
+
"outputs": [],
|
977 |
+
"source": [
|
978 |
+
"stop_words = stopwords.words('english')\n",
|
979 |
+
"stop_words.extend(['from', 'subject', 're', 'edu', 'use'])\n",
|
980 |
+
"def sent_to_words(sentences):\n",
|
981 |
+
" for sentence in sentences:\n",
|
982 |
+
" # deacc=True removes punctuations\n",
|
983 |
+
" yield(gensim.utils.simple_preprocess(str(sentence), deacc=True))\n",
|
984 |
+
" \n",
|
985 |
+
"def remove_stopwords(texts):\n",
|
986 |
+
" return [[word for word in simple_preprocess(str(doc)) \n",
|
987 |
+
" if word not in stop_words] for doc in texts]\n",
|
988 |
+
"data = df.Message_processed.values.tolist()\n",
|
989 |
+
"data_words = list(sent_to_words(data))# remove stop words\n",
|
990 |
+
"data_words = remove_stopwords(data_words)"
|
991 |
+
]
|
992 |
+
},
|
993 |
+
{
|
994 |
+
"cell_type": "code",
|
995 |
+
"execution_count": null,
|
996 |
+
"id": "1e257cc3",
|
997 |
+
"metadata": {},
|
998 |
+
"outputs": [],
|
999 |
+
"source": [
|
1000 |
+
"print(data_words[:1][0][:30])"
|
1001 |
+
]
|
1002 |
+
},
|
1003 |
+
{
|
1004 |
+
"cell_type": "code",
|
1005 |
+
"execution_count": null,
|
1006 |
+
"id": "98c5203f",
|
1007 |
+
"metadata": {},
|
1008 |
+
"outputs": [],
|
1009 |
+
"source": [
|
1010 |
+
"id2word = corpora.Dictionary(data_words)\n",
|
1011 |
+
"texts = data_words# Term Document Frequency\n",
|
1012 |
+
"corpus = [id2word.doc2bow(text) for text in texts]# View\n",
|
1013 |
+
"print(corpus[:1][0][:30])"
|
1014 |
+
]
|
1015 |
+
},
|
1016 |
+
{
|
1017 |
+
"cell_type": "code",
|
1018 |
+
"execution_count": null,
|
1019 |
+
"id": "b4a35025",
|
1020 |
+
"metadata": {},
|
1021 |
+
"outputs": [],
|
1022 |
+
"source": [
|
1023 |
+
"num_topics = 10# Build LDA model\n",
|
1024 |
+
"lda_model = gensim.models.LdaMulticore(corpus=corpus,\n",
|
1025 |
+
" id2word=id2word,\n",
|
1026 |
+
" num_topics=num_topics)\n",
|
1027 |
+
"# Print the Keyword in the 10 topics\n",
|
1028 |
+
"pprint(lda_model.print_topics())\n",
|
1029 |
+
"doc_lda = lda_model[corpus]"
|
1030 |
+
]
|
1031 |
+
},
|
1032 |
+
{
|
1033 |
+
"cell_type": "code",
|
1034 |
+
"execution_count": null,
|
1035 |
+
"id": "00346a62",
|
1036 |
+
"metadata": {},
|
1037 |
+
"outputs": [],
|
1038 |
+
"source": [
|
1039 |
+
"pyLDAvis.enable_notebook()\n"
|
1040 |
+
]
|
1041 |
+
},
|
1042 |
+
{
|
1043 |
+
"cell_type": "code",
|
1044 |
+
"execution_count": null,
|
1045 |
+
"id": "f6f7889b",
|
1046 |
+
"metadata": {},
|
1047 |
+
"outputs": [],
|
1048 |
+
"source": [
|
1049 |
+
"vis = pyLDAvis.gensim_models.prepare(lda_model, corpus, id2word, mds=\"mmds\", R=30)\n",
|
1050 |
+
"vis\n"
|
1051 |
+
]
|
1052 |
+
},
|
1053 |
+
{
|
1054 |
+
"cell_type": "code",
|
1055 |
+
"execution_count": null,
|
1056 |
+
"id": "e4b7ca16",
|
1057 |
+
"metadata": {},
|
1058 |
+
"outputs": [],
|
1059 |
+
"source": []
|
1060 |
+
},
|
1061 |
+
{
|
1062 |
+
"cell_type": "code",
|
1063 |
+
"execution_count": null,
|
1064 |
+
"id": "1b214796",
|
1065 |
+
"metadata": {},
|
1066 |
+
"outputs": [],
|
1067 |
+
"source": []
|
1068 |
+
},
|
1069 |
+
{
|
1070 |
+
"cell_type": "code",
|
1071 |
+
"execution_count": null,
|
1072 |
+
"id": "e7f8e54c",
|
1073 |
+
"metadata": {},
|
1074 |
+
"outputs": [],
|
1075 |
+
"source": []
|
1076 |
+
},
|
1077 |
+
{
|
1078 |
+
"cell_type": "code",
|
1079 |
+
"execution_count": null,
|
1080 |
+
"id": "021b015f",
|
1081 |
+
"metadata": {},
|
1082 |
+
"outputs": [],
|
1083 |
+
"source": []
|
1084 |
+
},
|
1085 |
+
{
|
1086 |
+
"cell_type": "code",
|
1087 |
+
"execution_count": null,
|
1088 |
+
"id": "ab1a9490",
|
1089 |
+
"metadata": {},
|
1090 |
+
"outputs": [],
|
1091 |
+
"source": []
|
1092 |
+
},
|
1093 |
+
{
|
1094 |
+
"cell_type": "code",
|
1095 |
+
"execution_count": null,
|
1096 |
+
"id": "0da95a15",
|
1097 |
+
"metadata": {},
|
1098 |
+
"outputs": [],
|
1099 |
+
"source": []
|
1100 |
+
},
|
1101 |
+
{
|
1102 |
+
"cell_type": "code",
|
1103 |
+
"execution_count": null,
|
1104 |
+
"id": "22c069c0",
|
1105 |
+
"metadata": {},
|
1106 |
+
"outputs": [],
|
1107 |
+
"source": []
|
1108 |
+
},
|
1109 |
+
{
|
1110 |
+
"cell_type": "code",
|
1111 |
+
"execution_count": null,
|
1112 |
+
"id": "c02c30f3",
|
1113 |
+
"metadata": {},
|
1114 |
+
"outputs": [],
|
1115 |
+
"source": []
|
1116 |
+
},
|
1117 |
+
{
|
1118 |
+
"cell_type": "code",
|
1119 |
+
"execution_count": null,
|
1120 |
+
"id": "9cdde3ad",
|
1121 |
+
"metadata": {},
|
1122 |
+
"outputs": [],
|
1123 |
+
"source": []
|
1124 |
+
},
|
1125 |
+
{
|
1126 |
+
"cell_type": "code",
|
1127 |
+
"execution_count": null,
|
1128 |
+
"id": "717270ef",
|
1129 |
+
"metadata": {},
|
1130 |
+
"outputs": [],
|
1131 |
+
"source": []
|
1132 |
+
},
|
1133 |
+
{
|
1134 |
+
"cell_type": "code",
|
1135 |
+
"execution_count": null,
|
1136 |
+
"id": "25a8f105",
|
1137 |
+
"metadata": {},
|
1138 |
+
"outputs": [],
|
1139 |
+
"source": []
|
1140 |
+
}
|
1141 |
+
],
|
1142 |
+
"metadata": {
|
1143 |
+
"kernelspec": {
|
1144 |
+
"display_name": "Python 3 (ipykernel)",
|
1145 |
+
"language": "python",
|
1146 |
+
"name": "python3"
|
1147 |
+
},
|
1148 |
+
"language_info": {
|
1149 |
+
"codemirror_mode": {
|
1150 |
+
"name": "ipython",
|
1151 |
+
"version": 3
|
1152 |
+
},
|
1153 |
+
"file_extension": ".py",
|
1154 |
+
"mimetype": "text/x-python",
|
1155 |
+
"name": "python",
|
1156 |
+
"nbconvert_exporter": "python",
|
1157 |
+
"pygments_lexer": "ipython3",
|
1158 |
+
"version": "3.9.7"
|
1159 |
+
}
|
1160 |
+
},
|
1161 |
+
"nbformat": 4,
|
1162 |
+
"nbformat_minor": 5
|
1163 |
+
}
|