kargaranamir
commited on
Commit
•
a220ee8
1
Parent(s):
62b1636
Create utils.py
Browse files
utils.py
ADDED
@@ -0,0 +1,490 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
utils for Hengam inference
|
3 |
+
"""
|
4 |
+
|
5 |
+
"""### Import Libraries"""
|
6 |
+
|
7 |
+
# import primitive libraries
|
8 |
+
import os
|
9 |
+
import pandas as pd
|
10 |
+
from tqdm import tqdm
|
11 |
+
import numpy as np
|
12 |
+
import json
|
13 |
+
|
14 |
+
# import seqval to report classifier performance metrics
|
15 |
+
from seqeval.metrics import accuracy_score, precision_score, recall_score, f1_score
|
16 |
+
from seqeval.scheme import IOB2
|
17 |
+
|
18 |
+
# import torch related modules
|
19 |
+
import torch
|
20 |
+
from torch.utils.data import DataLoader
|
21 |
+
from torch.utils.data import Dataset
|
22 |
+
from torch.nn.utils.rnn import pad_sequence
|
23 |
+
import torch.nn as nn
|
24 |
+
|
25 |
+
# import pytorch lightning library
|
26 |
+
import pytorch_lightning as pl
|
27 |
+
from torchcrf import CRF as SUPERCRF
|
28 |
+
|
29 |
+
# import NLTK to create better tokenizer
|
30 |
+
import nltk
|
31 |
+
from nltk.tokenize import RegexpTokenizer
|
32 |
+
|
33 |
+
# Transformers : Roberta Model
|
34 |
+
from transformers import XLMRobertaTokenizerFast
|
35 |
+
from transformers import XLMRobertaModel, XLMRobertaConfig
|
36 |
+
|
37 |
+
|
38 |
+
# import Typings
|
39 |
+
from typing import Union, Dict, List, Tuple, Any, Optional
|
40 |
+
|
41 |
+
import glob
|
42 |
+
|
43 |
+
# for sent tokenizer (nltk)
|
44 |
+
nltk.download('punkt')
|
45 |
+
|
46 |
+
|
47 |
+
"""## XLM-Roberta
|
48 |
+
### TokenFromSubtoken
|
49 |
+
- Code adapted from the following [file](https://github.com/deepmipt/DeepPavlov/blob/master/deeppavlov/models/torch_bert/torch_transformers_sequence_tagger.py)
|
50 |
+
- DeepPavlov is an popular open source library for deep learning end-to-end dialog systems and chatbots.
|
51 |
+
- Licensed under the Apache License, Version 2.0 (the "License");
|
52 |
+
"""
|
53 |
+
|
54 |
+
class TokenFromSubtoken(torch.nn.Module):
|
55 |
+
|
56 |
+
def forward(self, units: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
|
57 |
+
""" Assemble token level units from subtoken level units
|
58 |
+
Args:
|
59 |
+
units: torch.Tensor of shape [batch_size, SUBTOKEN_seq_length, n_features]
|
60 |
+
mask: mask of token beginnings. For example: for tokens
|
61 |
+
[[``[CLS]`` ``My``, ``capybara``, ``[SEP]``],
|
62 |
+
[``[CLS]`` ``Your``, ``aar``, ``##dvark``, ``is``, ``awesome``, ``[SEP]``]]
|
63 |
+
the mask will be
|
64 |
+
[[0, 1, 1, 0, 0, 0, 0],
|
65 |
+
[0, 1, 1, 0, 1, 1, 0]]
|
66 |
+
Returns:
|
67 |
+
word_level_units: Units assembled from ones in the mask. For the
|
68 |
+
example above this units will correspond to the following
|
69 |
+
[[``My``, ``capybara``],
|
70 |
+
[``Your`, ``aar``, ``is``, ``awesome``,]]
|
71 |
+
the shape of this tensor will be [batch_size, TOKEN_seq_length, n_features]
|
72 |
+
"""
|
73 |
+
|
74 |
+
device = units.device
|
75 |
+
nf_int = units.size()[-1]
|
76 |
+
batch_size = units.size()[0]
|
77 |
+
|
78 |
+
# number of TOKENS in each sentence
|
79 |
+
token_seq_lengths = torch.sum(mask, 1).to(torch.int64)
|
80 |
+
# number of words
|
81 |
+
n_words = torch.sum(token_seq_lengths)
|
82 |
+
# max token seq len
|
83 |
+
max_token_seq_len = torch.max(token_seq_lengths)
|
84 |
+
|
85 |
+
idxs = torch.stack(torch.nonzero(mask, as_tuple=True), dim=1)
|
86 |
+
# padding is for computing change from one sample to another in the batch
|
87 |
+
sample_ids_in_batch = torch.nn.functional.pad(input=idxs[:, 0], pad=[1, 0])
|
88 |
+
|
89 |
+
a = (~torch.eq(sample_ids_in_batch[1:], sample_ids_in_batch[:-1])).to(torch.int64)
|
90 |
+
|
91 |
+
# transforming sample start masks to the sample starts themselves
|
92 |
+
q = a * torch.arange(n_words, device=device).to(torch.int64)
|
93 |
+
count_to_substract = torch.nn.functional.pad(torch.masked_select(q, q.to(torch.bool)), [1, 0])
|
94 |
+
|
95 |
+
new_word_indices = torch.arange(n_words, device=device).to(torch.int64) - count_to_substract[torch.cumsum(a, 0)]
|
96 |
+
|
97 |
+
n_total_word_elements = max_token_seq_len*torch.ones_like(token_seq_lengths, device=device).sum()
|
98 |
+
word_indices_flat = (idxs[:, 0] * max_token_seq_len + new_word_indices).to(torch.int64)
|
99 |
+
#x_mask = torch.sum(torch.nn.functional.one_hot(word_indices_flat, n_total_word_elements), 0)
|
100 |
+
#x_mask = x_mask.to(torch.bool)
|
101 |
+
x_mask = torch.zeros(n_total_word_elements, dtype=torch.bool, device=device)
|
102 |
+
x_mask[word_indices_flat] = torch.ones_like(word_indices_flat, device=device, dtype=torch.bool)
|
103 |
+
# to get absolute indices we add max_token_seq_len:
|
104 |
+
# idxs[:, 0] * max_token_seq_len -> [0, 0, 0, 1, 1, 2] * 2 = [0, 0, 0, 3, 3, 6]
|
105 |
+
# word_indices_flat -> [0, 0, 0, 3, 3, 6] + [0, 1, 2, 0, 1, 0] = [0, 1, 2, 3, 4, 6]
|
106 |
+
# total number of words in the batch (including paddings)
|
107 |
+
# batch_size * max_token_seq_len -> 3 * 3 = 9
|
108 |
+
# tf.one_hot(...) ->
|
109 |
+
# [[1. 0. 0. 0. 0. 0. 0. 0. 0.]
|
110 |
+
# [0. 1. 0. 0. 0. 0. 0. 0. 0.]
|
111 |
+
# [0. 0. 1. 0. 0. 0. 0. 0. 0.]
|
112 |
+
# [0. 0. 0. 1. 0. 0. 0. 0. 0.]
|
113 |
+
# [0. 0. 0. 0. 1. 0. 0. 0. 0.]
|
114 |
+
# [0. 0. 0. 0. 0. 0. 1. 0. 0.]]
|
115 |
+
# x_mask -> [1, 1, 1, 1, 1, 0, 1, 0, 0]
|
116 |
+
nonword_indices_flat = (~x_mask).nonzero().squeeze(-1)
|
117 |
+
|
118 |
+
# get a sequence of units corresponding to the start subtokens of the words
|
119 |
+
# size: [n_words, n_features]
|
120 |
+
|
121 |
+
elements = units[mask.bool()]
|
122 |
+
|
123 |
+
# prepare zeros for paddings
|
124 |
+
# size: [batch_size * TOKEN_seq_length - n_words, n_features]
|
125 |
+
paddings = torch.zeros_like(nonword_indices_flat, dtype=elements.dtype).unsqueeze(-1).repeat(1,nf_int).to(device)
|
126 |
+
|
127 |
+
# tensor_flat -> [x, x, x, x, x, 0, x, 0, 0]
|
128 |
+
tensor_flat_unordered = torch.cat([elements, paddings])
|
129 |
+
_, order_idx = torch.sort(torch.cat([word_indices_flat, nonword_indices_flat]))
|
130 |
+
tensor_flat = tensor_flat_unordered[order_idx]
|
131 |
+
|
132 |
+
tensor = torch.reshape(tensor_flat, (-1, max_token_seq_len, nf_int))
|
133 |
+
# tensor -> [[x, x, x],
|
134 |
+
# [x, x, 0],
|
135 |
+
# [x, 0, 0]]
|
136 |
+
|
137 |
+
return tensor
|
138 |
+
|
139 |
+
"""### Conditional Random Field
|
140 |
+
- Code adopted form [torchcrf library](https://pytorch-crf.readthedocs.io/en/stable/)
|
141 |
+
- we override veiterbi decoder in order to make it compatible with our code
|
142 |
+
"""
|
143 |
+
|
144 |
+
class CRF(SUPERCRF):
|
145 |
+
|
146 |
+
# override veiterbi decoder in order to make it compatible with our code
|
147 |
+
def _viterbi_decode(self, emissions: torch.FloatTensor,
|
148 |
+
mask: torch.ByteTensor) -> List[List[int]]:
|
149 |
+
# emissions: (seq_length, batch_size, num_tags)
|
150 |
+
# mask: (seq_length, batch_size)
|
151 |
+
assert emissions.dim() == 3 and mask.dim() == 2
|
152 |
+
assert emissions.shape[:2] == mask.shape
|
153 |
+
assert emissions.size(2) == self.num_tags
|
154 |
+
assert mask[0].all()
|
155 |
+
|
156 |
+
seq_length, batch_size = mask.shape
|
157 |
+
|
158 |
+
# Start transition and first emission
|
159 |
+
# shape: (batch_size, num_tags)
|
160 |
+
score = self.start_transitions + emissions[0]
|
161 |
+
history = []
|
162 |
+
|
163 |
+
# score is a tensor of size (batch_size, num_tags) where for every batch,
|
164 |
+
# value at column j stores the score of the best tag sequence so far that ends
|
165 |
+
# with tag j
|
166 |
+
# history saves where the best tags candidate transitioned from; this is used
|
167 |
+
# when we trace back the best tag sequence
|
168 |
+
|
169 |
+
# Viterbi algorithm recursive case: we compute the score of the best tag sequence
|
170 |
+
# for every possible next tag
|
171 |
+
for i in range(1, seq_length):
|
172 |
+
# Broadcast viterbi score for every possible next tag
|
173 |
+
# shape: (batch_size, num_tags, 1)
|
174 |
+
broadcast_score = score.unsqueeze(2)
|
175 |
+
|
176 |
+
# Broadcast emission score for every possible current tag
|
177 |
+
# shape: (batch_size, 1, num_tags)
|
178 |
+
broadcast_emission = emissions[i].unsqueeze(1)
|
179 |
+
|
180 |
+
# Compute the score tensor of size (batch_size, num_tags, num_tags) where
|
181 |
+
# for each sample, entry at row i and column j stores the score of the best
|
182 |
+
# tag sequence so far that ends with transitioning from tag i to tag j and emitting
|
183 |
+
# shape: (batch_size, num_tags, num_tags)
|
184 |
+
next_score = broadcast_score + self.transitions + broadcast_emission
|
185 |
+
|
186 |
+
# Find the maximum score over all possible current tag
|
187 |
+
# shape: (batch_size, num_tags)
|
188 |
+
next_score, indices = next_score.max(dim=1)
|
189 |
+
|
190 |
+
# Set score to the next score if this timestep is valid (mask == 1)
|
191 |
+
# and save the index that produces the next score
|
192 |
+
# shape: (batch_size, num_tags)
|
193 |
+
score = torch.where(mask[i].unsqueeze(1), next_score, score)
|
194 |
+
history.append(indices)
|
195 |
+
|
196 |
+
history = torch.stack(history, dim=0)
|
197 |
+
|
198 |
+
# End transition score
|
199 |
+
# shape: (batch_size, num_tags)
|
200 |
+
score += self.end_transitions
|
201 |
+
|
202 |
+
# Now, compute the best path for each sample
|
203 |
+
|
204 |
+
# shape: (batch_size,)
|
205 |
+
seq_ends = mask.long().sum(dim=0) - 1
|
206 |
+
best_tags_list = []
|
207 |
+
|
208 |
+
for idx in range(batch_size):
|
209 |
+
# Find the tag which maximizes the score at the last timestep; this is our best tag
|
210 |
+
# for the last timestep
|
211 |
+
_, best_last_tag = score[idx].max(dim=0)
|
212 |
+
best_tags = [best_last_tag]
|
213 |
+
|
214 |
+
# We trace back where the best last tag comes from, append that to our best tag
|
215 |
+
# sequence, and trace it back again, and so on
|
216 |
+
for i, hist in enumerate(torch.flip(history[:seq_ends[idx]], dims=(0,))):
|
217 |
+
best_last_tag = hist[idx][best_tags[-1]]
|
218 |
+
best_tags.append(best_last_tag)
|
219 |
+
|
220 |
+
best_tags = torch.stack(best_tags, dim=0)
|
221 |
+
|
222 |
+
# Reverse the order because we start from the last timestep
|
223 |
+
best_tags_list.append(torch.flip(best_tags, dims=(0,)))
|
224 |
+
|
225 |
+
best_tags_list = nn.utils.rnn.pad_sequence(best_tags_list, batch_first=True, padding_value=0)
|
226 |
+
|
227 |
+
return best_tags_list
|
228 |
+
|
229 |
+
"""### CRFLayer
|
230 |
+
- Forward: decide output logits basaed on backbone network
|
231 |
+
- Decode: decode based on CRF weights
|
232 |
+
"""
|
233 |
+
|
234 |
+
class CRFLayer(nn.Module):
|
235 |
+
def __init__(self, embedding_size, n_labels):
|
236 |
+
|
237 |
+
super(CRFLayer, self).__init__()
|
238 |
+
self.dropout = nn.Dropout(0.1)
|
239 |
+
self.output_dense = nn.Linear(embedding_size,n_labels)
|
240 |
+
self.crf = CRF(n_labels, batch_first=True)
|
241 |
+
self.token_from_subtoken = TokenFromSubtoken()
|
242 |
+
|
243 |
+
# Forward: decide output logits basaed on backbone network
|
244 |
+
def forward(self, embedding, mask):
|
245 |
+
logits = self.output_dense(self.dropout(embedding))
|
246 |
+
logits = self.token_from_subtoken(logits, mask)
|
247 |
+
pad_mask = self.token_from_subtoken(mask.unsqueeze(-1), mask).squeeze(-1).bool()
|
248 |
+
return logits, pad_mask
|
249 |
+
|
250 |
+
# Decode: decode based on CRF weights
|
251 |
+
def decode(self, logits, pad_mask):
|
252 |
+
return self.crf.decode(logits, pad_mask)
|
253 |
+
|
254 |
+
# Evaluation Loss: calculate mean log likelihood of CRF layer
|
255 |
+
def eval_loss(self, logits, targets, pad_mask):
|
256 |
+
mean_log_likelihood = self.crf(logits, targets, pad_mask, reduction='sum').mean()
|
257 |
+
return -mean_log_likelihood
|
258 |
+
|
259 |
+
"""### NERModel
|
260 |
+
- Roberta Model with CRF Layer
|
261 |
+
"""
|
262 |
+
|
263 |
+
class NERModel(nn.Module):
|
264 |
+
|
265 |
+
def __init__(self, n_labels:int, roberta_path:str):
|
266 |
+
super(NERModel,self).__init__()
|
267 |
+
self.roberta = XLMRobertaModel.from_pretrained(roberta_path)
|
268 |
+
self.crf = CRFLayer(self.roberta.config.hidden_size, n_labels)
|
269 |
+
|
270 |
+
# Forward: pass embedings to CRF layer in order to evaluate logits from suboword sequence
|
271 |
+
def forward(self,
|
272 |
+
input_ids:torch.Tensor,
|
273 |
+
attention_mask:torch.Tensor,
|
274 |
+
token_type_ids:torch.Tensor,
|
275 |
+
mask:torch.Tensor) -> torch.Tensor:
|
276 |
+
|
277 |
+
embedding = self.roberta(input_ids=input_ids,
|
278 |
+
attention_mask=attention_mask,
|
279 |
+
token_type_ids=token_type_ids)[0]
|
280 |
+
logits, pad_mask = self.crf(embedding, mask)
|
281 |
+
return logits, pad_mask
|
282 |
+
|
283 |
+
# Disable Gradient and Predict with model
|
284 |
+
@torch.no_grad()
|
285 |
+
def predict(self, inputs:Tuple[torch.Tensor]) -> torch.Tensor:
|
286 |
+
input_ids, attention_mask, token_type_ids, mask = inputs
|
287 |
+
logits, pad_mask = self(input_ids, attention_mask, token_type_ids, mask)
|
288 |
+
decoded = self.crf.decode(logits, pad_mask)
|
289 |
+
return decoded, pad_mask
|
290 |
+
|
291 |
+
# Decode: pass to crf decoder and decode based on CRF weights
|
292 |
+
def decode(self, logits, pad_mask):
|
293 |
+
"""Decode logits using CRF weights
|
294 |
+
"""
|
295 |
+
return self.crf.decode(logits, pad_mask)
|
296 |
+
|
297 |
+
# Evaluation Loss: pass to crf eval_loss and calculate mean log likelihood of CRF layer
|
298 |
+
def eval_loss(self, logits, targets, pad_mask):
|
299 |
+
return self.crf.eval_loss(logits, targets, pad_mask)
|
300 |
+
|
301 |
+
# Determine number of layers to be fine-tuned (!freeze)
|
302 |
+
def freeze_roberta(self, n_freeze:int=6):
|
303 |
+
for param in self.roberta.parameters():
|
304 |
+
param.requires_grad = False
|
305 |
+
|
306 |
+
for param in self.roberta.encoder.layer[n_freeze:].parameters():
|
307 |
+
param.requires_grad = True
|
308 |
+
|
309 |
+
"""### NERTokenizer
|
310 |
+
- NLTK tokenizer along with XLMRobertaTokenizerFast tokenizer
|
311 |
+
- Code adapted from the following [file](https://github.com/ugurcanozalp/multilingual-ner/blob/main/multiner/utils/custom_tokenizer.py)
|
312 |
+
"""
|
313 |
+
|
314 |
+
class NERTokenizer(object):
|
315 |
+
|
316 |
+
MAX_LEN=512
|
317 |
+
BATCH_LENGTH_LIMT = 380 # Max number of roberta tokens in one sentence.
|
318 |
+
|
319 |
+
# Modified version of http://stackoverflow.com/questions/36353125/nltk-regular-expression-tokenizer
|
320 |
+
PATTERN = r'''(?x) # set flag to allow verbose regexps
|
321 |
+
(?:[A-Z]\.)+ # abbreviations, e.g. U.S.A. or U.S.A #
|
322 |
+
| (?:\d+\.) # numbers
|
323 |
+
| \w+(?:[-.]\w+)* # words with optional internal hyphens
|
324 |
+
| \$?\d+(?:.\d+)?%? # currency and percentages, e.g. $12.40, 82%
|
325 |
+
| \.\.\. # ellipsis, and special chars below, includes ], [
|
326 |
+
| [-\]\[.؟،؛;"'?,():_`“”/°º‘’″…#$%()*+<>=@\\^_{}|~❑&§\!]
|
327 |
+
| \u200c
|
328 |
+
'''
|
329 |
+
|
330 |
+
def __init__(self, base_model:str, to_device:str='cpu'):
|
331 |
+
super(NERTokenizer,self).__init__()
|
332 |
+
self.roberta_tokenizer = XLMRobertaTokenizerFast.from_pretrained(base_model, do_lower_case=False, padding=True, truncation=True)
|
333 |
+
self.to_device = to_device
|
334 |
+
|
335 |
+
self.word_tokenizer = RegexpTokenizer(self.PATTERN)
|
336 |
+
self.sent_tokenizer = nltk.data.load('tokenizers/punkt/english.pickle')
|
337 |
+
|
338 |
+
# tokenize batch of tokens
|
339 |
+
def tokenize_batch(self, inputs, pad_to = None) -> torch.Tensor:
|
340 |
+
batch = [inputs] if isinstance(inputs[0], str) else inputs
|
341 |
+
|
342 |
+
input_ids, attention_mask, token_type_ids, mask = [], [], [], []
|
343 |
+
for tokens in batch:
|
344 |
+
input_ids_tmp, attention_mask_tmp, token_type_ids_tmp, mask_tmp = self._tokenize_words(tokens)
|
345 |
+
input_ids.append(input_ids_tmp)
|
346 |
+
attention_mask.append(attention_mask_tmp)
|
347 |
+
token_type_ids.append(token_type_ids_tmp)
|
348 |
+
mask.append(mask_tmp)
|
349 |
+
|
350 |
+
input_ids = pad_sequence(input_ids, batch_first=True, padding_value=self.roberta_tokenizer.pad_token_id)
|
351 |
+
attention_mask = pad_sequence(attention_mask, batch_first=True, padding_value=0)
|
352 |
+
token_type_ids = pad_sequence(token_type_ids, batch_first=True, padding_value=0)
|
353 |
+
mask = pad_sequence(mask, batch_first=True, padding_value=0)
|
354 |
+
# truncate MAX_LEN
|
355 |
+
if input_ids.shape[-1]>self.MAX_LEN:
|
356 |
+
input_ids = input_ids[:,:,:self.MAX_LEN]
|
357 |
+
attention_mask = attention_mask[:,:,:self.MAX_LEN]
|
358 |
+
token_type_ids = token_type_ids[:,:,:self.MAX_LEN]
|
359 |
+
mask = mask[:,:,:self.MAX_LEN]
|
360 |
+
|
361 |
+
# extend pad
|
362 |
+
elif pad_to is not None and pad_to>input_ids.shape[1]:
|
363 |
+
bs = input_ids.shape[0]
|
364 |
+
padlen = pad_to-input_ids.shape[1]
|
365 |
+
|
366 |
+
input_ids_append = torch.tensor([self.roberta_tokenizer.pad_token_id], dtype=torch.long).repeat([bs, padlen]).to(self.to_device)
|
367 |
+
input_ids = torch.cat([input_ids, input_ids_append], dim=-1)
|
368 |
+
|
369 |
+
attention_mask_append = torch.tensor([0], dtype=torch.long).repeat([bs, padlen]).to(self.to_device)
|
370 |
+
attention_mask = torch.cat([attention_mask, attention_mask_append], dim=-1)
|
371 |
+
|
372 |
+
token_type_ids_append = torch.tensor([0], dtype=torch.long).repeat([bs, padlen]).to(self.to_device)
|
373 |
+
token_type_ids = torch.cat([token_type_ids, token_type_ids_append], dim=-1)
|
374 |
+
|
375 |
+
mask_append = torch.tensor([0], dtype=torch.long).repeat([bs, padlen]).to(self.to_device)
|
376 |
+
mask = torch.cat([mask, mask_append], dim=-1)
|
377 |
+
|
378 |
+
# truncate pad
|
379 |
+
elif pad_to is not None and pad_to<input_ids.shape[1]:
|
380 |
+
input_ids = input_ids[:,:,:pad_to]
|
381 |
+
attention_mask = attention_mask[:,:,:pad_to]
|
382 |
+
token_type_ids = token_type_ids[:,:,:pad_to]
|
383 |
+
mask = mask[:,:,:pad_to]
|
384 |
+
|
385 |
+
if isinstance(inputs[0], str):
|
386 |
+
return input_ids[0], attention_mask[0], token_type_ids[0], mask[0]
|
387 |
+
else:
|
388 |
+
return input_ids, attention_mask, token_type_ids, mask
|
389 |
+
|
390 |
+
# tokenize list of words with roberta tokenizer
|
391 |
+
def _tokenize_words(self, words):
|
392 |
+
roberta_tokens = []
|
393 |
+
mask = []
|
394 |
+
for word in words:
|
395 |
+
subtokens = self.roberta_tokenizer.tokenize(word)
|
396 |
+
roberta_tokens+=subtokens
|
397 |
+
n_subtoken = len(subtokens)
|
398 |
+
if n_subtoken>=1:
|
399 |
+
mask = mask + [1] + [0]*(n_subtoken-1)
|
400 |
+
|
401 |
+
# add special tokens [CLS] and [SeP]
|
402 |
+
roberta_tokens = [self.roberta_tokenizer.cls_token] + roberta_tokens + [self.roberta_tokenizer.sep_token]
|
403 |
+
mask = [0] + mask + [0]
|
404 |
+
input_ids = torch.tensor(self.roberta_tokenizer.convert_tokens_to_ids(roberta_tokens), dtype=torch.long).to(self.to_device)
|
405 |
+
attention_mask = torch.ones(len(mask), dtype=torch.long).to(self.to_device)
|
406 |
+
token_type_ids = torch.zeros(len(mask), dtype=torch.long).to(self.to_device)
|
407 |
+
mask = torch.tensor(mask, dtype=torch.long).to(self.to_device)
|
408 |
+
return input_ids, attention_mask, token_type_ids, mask
|
409 |
+
|
410 |
+
# sent_to_token: yield each sentence token with positional span using nltk
|
411 |
+
def sent_to_token(self, raw_text):
|
412 |
+
for offset, ending in self.sent_tokenizer.span_tokenize(raw_text):
|
413 |
+
sub_text = raw_text[offset:ending]
|
414 |
+
words, spans = [], []
|
415 |
+
flush = False
|
416 |
+
total_subtoken = 0
|
417 |
+
for start, end in self.word_tokenizer.span_tokenize(sub_text):
|
418 |
+
flush = True
|
419 |
+
start += offset
|
420 |
+
end += offset
|
421 |
+
words.append(raw_text[start:end])
|
422 |
+
spans.append((start,end))
|
423 |
+
total_subtoken += len(self.roberta_tokenizer.tokenize(words[-1]))
|
424 |
+
if (total_subtoken > self.BATCH_LENGTH_LIMT):
|
425 |
+
# Print
|
426 |
+
yield words[:-1],spans[:-1]
|
427 |
+
spans = spans[len(spans)-1:]
|
428 |
+
words = words[len(words)-1:]
|
429 |
+
total_subtoken = sum([len(self.roberta_tokenizer.tokenize(word)) for word in words])
|
430 |
+
flush = False
|
431 |
+
|
432 |
+
if flush and len(spans) > 0:
|
433 |
+
yield words,spans
|
434 |
+
|
435 |
+
# Extract (batch words span() from a raw sentence
|
436 |
+
def prepare_row_text(self, raw_text, batch_size=16):
|
437 |
+
words_list, spans_list = [], []
|
438 |
+
end_batch = False
|
439 |
+
for words, spans in self.sent_to_token(raw_text):
|
440 |
+
end_batch = True
|
441 |
+
words_list.append(words)
|
442 |
+
spans_list.append(spans)
|
443 |
+
if len(spans_list) >= batch_size:
|
444 |
+
input_ids, attention_mask, token_type_ids, mask = self.tokenize_batch(words_list)
|
445 |
+
yield (input_ids, attention_mask, token_type_ids, mask), words_list, spans_list
|
446 |
+
words_list, spans_list = [], []
|
447 |
+
if end_batch and len(words_list) > 0:
|
448 |
+
input_ids, attention_mask, token_type_ids, mask = self.tokenize_batch(words_list)
|
449 |
+
yield (input_ids, attention_mask, token_type_ids, mask), words_list, spans_list
|
450 |
+
|
451 |
+
"""### NER
|
452 |
+
NER Interface : We Use this interface to infer sentence Time-Date tags.
|
453 |
+
"""
|
454 |
+
|
455 |
+
class NER(object):
|
456 |
+
|
457 |
+
def __init__(self, model_path, tags):
|
458 |
+
|
459 |
+
self.tags = tags
|
460 |
+
self.device = "cuda" if torch.cuda.is_available() else "cpu"
|
461 |
+
# Load Pre-Trained model
|
462 |
+
roberta_path = "xlm-roberta-base"
|
463 |
+
self.model = NERModel(n_labels=len(self.tags), roberta_path=roberta_path).to(self.device)
|
464 |
+
# Load Fine-Tuned model
|
465 |
+
state_dict = torch.load(model_path)
|
466 |
+
self.model.load_state_dict(state_dict, strict=False)
|
467 |
+
# Enable Evaluation mode
|
468 |
+
self.model.eval()
|
469 |
+
self.tokenizer = NERTokenizer(base_model=roberta_path, to_device=self.device)
|
470 |
+
|
471 |
+
# Predict and Pre/Post-Process the input/output
|
472 |
+
@torch.no_grad()
|
473 |
+
def __call__(self, raw_text):
|
474 |
+
|
475 |
+
outputs_flat, spans_flat, entities = [], [], []
|
476 |
+
for batch, words, spans in self.tokenizer.prepare_row_text(raw_text):
|
477 |
+
output, pad_mask = self.model.predict(batch)
|
478 |
+
outputs_flat.extend(output[pad_mask.bool()].reshape(-1).tolist())
|
479 |
+
spans_flat += sum(spans, [])
|
480 |
+
|
481 |
+
for tag_idx,(start,end) in zip(outputs_flat,spans_flat):
|
482 |
+
tag = self.tags[tag_idx]
|
483 |
+
# filter out O tags
|
484 |
+
if tag != 'O':
|
485 |
+
entities.append({'Text': raw_text[start:end],
|
486 |
+
'Tag': tag,
|
487 |
+
'Start':start,
|
488 |
+
'End': end})
|
489 |
+
|
490 |
+
return entities
|