Spaces:
Runtime error
Runtime error
ping yang
commited on
Commit
•
9bb46b0
1
Parent(s):
ab756e2
Add application file
Browse files
app.py
ADDED
@@ -0,0 +1,659 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2021 The IDEA Authors. All rights reserved.
|
3 |
+
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
|
16 |
+
from logging import basicConfig
|
17 |
+
import torch
|
18 |
+
from torch import nn
|
19 |
+
import json
|
20 |
+
from tqdm import tqdm
|
21 |
+
import os
|
22 |
+
import numpy as np
|
23 |
+
from transformers import BertTokenizer, AutoTokenizer
|
24 |
+
import pytorch_lightning as pl
|
25 |
+
|
26 |
+
from pytorch_lightning.callbacks import ModelCheckpoint
|
27 |
+
from pytorch_lightning import loggers
|
28 |
+
from torch.utils.data import Dataset, DataLoader
|
29 |
+
from transformers.optimization import get_linear_schedule_with_warmup
|
30 |
+
from transformers import BertForMaskedLM, AlbertTokenizer
|
31 |
+
from transformers import AutoConfig
|
32 |
+
from transformers import MegatronBertForMaskedLM
|
33 |
+
import argparse
|
34 |
+
import copy
|
35 |
+
import streamlit as st
|
36 |
+
# os.environ["CUDA_VISIBLE_DEVICES"] = '6'
|
37 |
+
|
38 |
+
|
39 |
+
class UniMCDataset(Dataset):
|
40 |
+
def __init__(self, data, yes_token, no_token, tokenizer, args, used_mask=True):
|
41 |
+
super().__init__()
|
42 |
+
|
43 |
+
self.tokenizer = tokenizer
|
44 |
+
self.max_length = args.max_length
|
45 |
+
self.num_labels = args.num_labels
|
46 |
+
self.used_mask = used_mask
|
47 |
+
self.data = data
|
48 |
+
self.args = args
|
49 |
+
self.yes_token = yes_token
|
50 |
+
self.no_token = no_token
|
51 |
+
|
52 |
+
def __len__(self):
|
53 |
+
return len(self.data)
|
54 |
+
|
55 |
+
def __getitem__(self, index):
|
56 |
+
return self.encode(self.data[index], self.used_mask)
|
57 |
+
|
58 |
+
def get_token_type(self, sep_idx, max_length):
|
59 |
+
token_type_ids = np.zeros(shape=(max_length,))
|
60 |
+
for i in range(len(sep_idx)-1):
|
61 |
+
if i % 2 == 0:
|
62 |
+
ty = np.ones(shape=(sep_idx[i+1]-sep_idx[i],))
|
63 |
+
else:
|
64 |
+
ty = np.zeros(shape=(sep_idx[i+1]-sep_idx[i],))
|
65 |
+
token_type_ids[sep_idx[i]:sep_idx[i+1]] = ty
|
66 |
+
|
67 |
+
return token_type_ids
|
68 |
+
|
69 |
+
def get_position_ids(self, label_idx, max_length, question_len):
|
70 |
+
question_position_ids = np.arange(question_len)
|
71 |
+
label_position_ids = np.arange(question_len, label_idx[-1])
|
72 |
+
for i in range(len(label_idx)-1):
|
73 |
+
label_position_ids[label_idx[i]-question_len:label_idx[i+1]-question_len] = np.arange(
|
74 |
+
question_len, question_len+label_idx[i+1]-label_idx[i])
|
75 |
+
max_len_label = max(label_position_ids)
|
76 |
+
text_position_ids = np.arange(
|
77 |
+
max_len_label+1, max_length+max_len_label+1-label_idx[-1])
|
78 |
+
position_ids = list(question_position_ids) + \
|
79 |
+
list(label_position_ids)+list(text_position_ids)
|
80 |
+
if max_length <= 512:
|
81 |
+
return position_ids[:max_length]
|
82 |
+
else:
|
83 |
+
for i in range(512, max_length):
|
84 |
+
if position_ids[i] > 511:
|
85 |
+
position_ids[i] = 511
|
86 |
+
return position_ids[:max_length]
|
87 |
+
|
88 |
+
def get_att_mask(self, attention_mask, label_idx, question_len):
|
89 |
+
max_length = len(attention_mask)
|
90 |
+
attention_mask = np.array(attention_mask)
|
91 |
+
attention_mask = np.tile(attention_mask[None, :], (max_length, 1))
|
92 |
+
|
93 |
+
zeros = np.zeros(
|
94 |
+
shape=(label_idx[-1]-question_len, label_idx[-1]-question_len))
|
95 |
+
attention_mask[question_len:label_idx[-1],
|
96 |
+
question_len:label_idx[-1]] = zeros
|
97 |
+
|
98 |
+
for i in range(len(label_idx)-1):
|
99 |
+
label_token_length = label_idx[i+1]-label_idx[i]
|
100 |
+
if label_token_length <= 0:
|
101 |
+
print('label_idx', label_idx)
|
102 |
+
print('question_len', question_len)
|
103 |
+
continue
|
104 |
+
ones = np.ones(shape=(label_token_length, label_token_length))
|
105 |
+
attention_mask[label_idx[i]:label_idx[i+1],
|
106 |
+
label_idx[i]:label_idx[i+1]] = ones
|
107 |
+
|
108 |
+
return attention_mask
|
109 |
+
|
110 |
+
def random_masking(self, token_ids, maks_rate, mask_start_idx, max_length, mask_id, tokenizer):
|
111 |
+
rands = np.random.random(len(token_ids))
|
112 |
+
source, target = [], []
|
113 |
+
for i, (r, t) in enumerate(zip(rands, token_ids)):
|
114 |
+
if i < mask_start_idx:
|
115 |
+
source.append(t)
|
116 |
+
target.append(-100)
|
117 |
+
continue
|
118 |
+
if r < maks_rate * 0.8:
|
119 |
+
source.append(mask_id)
|
120 |
+
target.append(t)
|
121 |
+
elif r < maks_rate * 0.9:
|
122 |
+
source.append(t)
|
123 |
+
target.append(t)
|
124 |
+
elif r < maks_rate:
|
125 |
+
source.append(np.random.choice(tokenizer.vocab_size - 1) + 1)
|
126 |
+
target.append(t)
|
127 |
+
else:
|
128 |
+
source.append(t)
|
129 |
+
target.append(-100)
|
130 |
+
while len(source) < max_length:
|
131 |
+
source.append(0)
|
132 |
+
target.append(-100)
|
133 |
+
return source[:max_length], target[:max_length]
|
134 |
+
|
135 |
+
def encode(self, item, used_mask=False):
|
136 |
+
|
137 |
+
while len(self.tokenizer.encode('[MASK]'.join(item['choice']))) > self.max_length-32:
|
138 |
+
item['choice'] = [c[:int(len(c)/2)] for c in item['choice']]
|
139 |
+
|
140 |
+
if 'textb' in item.keys() and item['textb'] != '':
|
141 |
+
if 'question' in item.keys() and item['question'] != '':
|
142 |
+
texta = '[MASK]' + '[MASK]'.join(item['choice']) + '[SEP]' + \
|
143 |
+
item['question'] + '[SEP]' + \
|
144 |
+
item['texta']+'[SEP]'+item['textb']
|
145 |
+
else:
|
146 |
+
texta = '[MASK]' + '[MASK]'.join(item['choice']) + '[SEP]' + \
|
147 |
+
item['texta']+'[SEP]'+item['textb']
|
148 |
+
|
149 |
+
else:
|
150 |
+
if 'question' in item.keys() and item['question'] != '':
|
151 |
+
texta = '[MASK]' + '[MASK]'.join(item['choice']) + '[SEP]' + \
|
152 |
+
item['question'] + '[SEP]' + item['texta']
|
153 |
+
else:
|
154 |
+
texta = '[MASK]' + '[MASK]'.join(item['choice']) + \
|
155 |
+
'[SEP]' + item['texta']
|
156 |
+
|
157 |
+
encode_dict = self.tokenizer.encode_plus(texta,
|
158 |
+
max_length=self.max_length,
|
159 |
+
padding='max_length',
|
160 |
+
truncation='longest_first')
|
161 |
+
|
162 |
+
encode_sent = encode_dict['input_ids']
|
163 |
+
token_type_ids = encode_dict['token_type_ids']
|
164 |
+
attention_mask = encode_dict['attention_mask']
|
165 |
+
sample_max_length = sum(encode_dict['attention_mask'])
|
166 |
+
|
167 |
+
if 'label' not in item.keys():
|
168 |
+
item['label'] = 0
|
169 |
+
item['answer'] = ''
|
170 |
+
|
171 |
+
question_len = 1
|
172 |
+
label_idx = [question_len]
|
173 |
+
for choice in item['choice']:
|
174 |
+
cur_mask_idx = label_idx[-1] + \
|
175 |
+
len(self.tokenizer.encode(choice, add_special_tokens=False))+1
|
176 |
+
label_idx.append(cur_mask_idx)
|
177 |
+
|
178 |
+
token_type_ids = [0]*question_len+[1] * \
|
179 |
+
(label_idx[-1]-label_idx[0]+1)+[0]*self.max_length
|
180 |
+
token_type_ids = token_type_ids[:self.max_length]
|
181 |
+
|
182 |
+
attention_mask = self.get_att_mask(
|
183 |
+
attention_mask, label_idx, question_len)
|
184 |
+
|
185 |
+
position_ids = self.get_position_ids(
|
186 |
+
label_idx, self.max_length, question_len)
|
187 |
+
|
188 |
+
clslabels_mask = np.zeros(shape=(len(encode_sent),))
|
189 |
+
clslabels_mask[label_idx[:-1]] = 10000
|
190 |
+
clslabels_mask = clslabels_mask-10000
|
191 |
+
|
192 |
+
mlmlabels_mask = np.zeros(shape=(len(encode_sent),))
|
193 |
+
mlmlabels_mask[label_idx[0]] = 1
|
194 |
+
|
195 |
+
used_mask = False
|
196 |
+
if used_mask:
|
197 |
+
mask_rate = 0.1*np.random.choice(4, p=[0.3, 0.3, 0.25, 0.15])
|
198 |
+
source, target = self.random_masking(token_ids=encode_sent, maks_rate=mask_rate,
|
199 |
+
mask_start_idx=label_idx[-1], max_length=self.max_length,
|
200 |
+
mask_id=self.tokenizer.mask_token_id, tokenizer=self.tokenizer)
|
201 |
+
else:
|
202 |
+
source, target = encode_sent[:], encode_sent[:]
|
203 |
+
|
204 |
+
source = np.array(source)
|
205 |
+
target = np.array(target)
|
206 |
+
source[label_idx[:-1]] = self.tokenizer.mask_token_id
|
207 |
+
target[label_idx[:-1]] = self.no_token
|
208 |
+
target[label_idx[item['label']]] = self.yes_token
|
209 |
+
|
210 |
+
input_ids = source[:sample_max_length]
|
211 |
+
token_type_ids = token_type_ids[:sample_max_length]
|
212 |
+
attention_mask = attention_mask[:sample_max_length, :sample_max_length]
|
213 |
+
position_ids = position_ids[:sample_max_length]
|
214 |
+
mlmlabels = target[:sample_max_length]
|
215 |
+
clslabels = label_idx[item['label']]
|
216 |
+
clslabels_mask = clslabels_mask[:sample_max_length]
|
217 |
+
mlmlabels_mask = mlmlabels_mask[:sample_max_length]
|
218 |
+
|
219 |
+
return {
|
220 |
+
"input_ids": torch.tensor(input_ids).long(),
|
221 |
+
"token_type_ids": torch.tensor(token_type_ids).long(),
|
222 |
+
"attention_mask": torch.tensor(attention_mask).float(),
|
223 |
+
"position_ids": torch.tensor(position_ids).long(),
|
224 |
+
"mlmlabels": torch.tensor(mlmlabels).long(),
|
225 |
+
"clslabels": torch.tensor(clslabels).long(),
|
226 |
+
"clslabels_mask": torch.tensor(clslabels_mask).float(),
|
227 |
+
"mlmlabels_mask": torch.tensor(mlmlabels_mask).float(),
|
228 |
+
}
|
229 |
+
|
230 |
+
|
231 |
+
class UniMCDataModel(pl.LightningDataModule):
|
232 |
+
@staticmethod
|
233 |
+
def add_data_specific_args(parent_args):
|
234 |
+
parser = parent_args.add_argument_group('TASK NAME DataModel')
|
235 |
+
parser.add_argument('--num_workers', default=8, type=int)
|
236 |
+
parser.add_argument('--batchsize', default=16, type=int)
|
237 |
+
parser.add_argument('--max_length', default=512, type=int)
|
238 |
+
return parent_args
|
239 |
+
|
240 |
+
def __init__(self, train_data, val_data, yes_token, no_token, tokenizer, args):
|
241 |
+
super().__init__()
|
242 |
+
self.batchsize = args.batchsize
|
243 |
+
|
244 |
+
self.train_data = UniMCDataset(
|
245 |
+
train_data, yes_token, no_token, tokenizer, args, True)
|
246 |
+
self.valid_data = UniMCDataset(
|
247 |
+
val_data, yes_token, no_token, tokenizer, args, False)
|
248 |
+
|
249 |
+
def train_dataloader(self):
|
250 |
+
return DataLoader(self.train_data, shuffle=True, collate_fn=self.collate_fn, batch_size=self.batchsize, pin_memory=False)
|
251 |
+
|
252 |
+
def val_dataloader(self):
|
253 |
+
return DataLoader(self.valid_data, shuffle=False, collate_fn=self.collate_fn, batch_size=self.batchsize, pin_memory=False)
|
254 |
+
|
255 |
+
def collate_fn(self, batch):
|
256 |
+
'''
|
257 |
+
Aggregate a batch data.
|
258 |
+
batch = [ins1_dict, ins2_dict, ..., insN_dict]
|
259 |
+
batch_data = {'sentence':[ins1_sentence, ins2_sentence...], 'input_ids':[ins1_input_ids, ins2_input_ids...], ...}
|
260 |
+
'''
|
261 |
+
batch_data = {}
|
262 |
+
for key in batch[0]:
|
263 |
+
batch_data[key] = [example[key] for example in batch]
|
264 |
+
|
265 |
+
batch_data['input_ids'] = nn.utils.rnn.pad_sequence(batch_data['input_ids'],
|
266 |
+
batch_first=True,
|
267 |
+
padding_value=0)
|
268 |
+
batch_data['clslabels_mask'] = nn.utils.rnn.pad_sequence(batch_data['clslabels_mask'],
|
269 |
+
batch_first=True,
|
270 |
+
padding_value=-10000)
|
271 |
+
|
272 |
+
batch_size, batch_max_length = batch_data['input_ids'].shape
|
273 |
+
for k, v in batch_data.items():
|
274 |
+
if k == 'input_ids' or k == 'clslabels_mask':
|
275 |
+
continue
|
276 |
+
if k == 'clslabels':
|
277 |
+
batch_data[k] = torch.tensor(v).long()
|
278 |
+
continue
|
279 |
+
if k != 'attention_mask':
|
280 |
+
batch_data[k] = nn.utils.rnn.pad_sequence(v,
|
281 |
+
batch_first=True,
|
282 |
+
padding_value=0)
|
283 |
+
else:
|
284 |
+
attention_mask = torch.zeros(
|
285 |
+
(batch_size, batch_max_length, batch_max_length))
|
286 |
+
for i, att in enumerate(v):
|
287 |
+
sample_length, _ = att.shape
|
288 |
+
attention_mask[i, :sample_length, :sample_length] = att
|
289 |
+
batch_data[k] = attention_mask
|
290 |
+
return batch_data
|
291 |
+
|
292 |
+
|
293 |
+
class UniMCModel(nn.Module):
|
294 |
+
def __init__(self, pre_train_dir, yes_token):
|
295 |
+
super().__init__()
|
296 |
+
self.config = AutoConfig.from_pretrained(pre_train_dir)
|
297 |
+
if self.config.model_type == 'megatron-bert':
|
298 |
+
self.bert = MegatronBertForMaskedLM.from_pretrained(pre_train_dir)
|
299 |
+
else:
|
300 |
+
self.bert = BertForMaskedLM.from_pretrained(pre_train_dir)
|
301 |
+
|
302 |
+
self.loss_func = torch.nn.CrossEntropyLoss()
|
303 |
+
self.yes_token = yes_token
|
304 |
+
|
305 |
+
def forward(self, input_ids, attention_mask, token_type_ids, position_ids=None, mlmlabels=None, clslabels=None, clslabels_mask=None, mlmlabels_mask=None):
|
306 |
+
|
307 |
+
batch_size, seq_len = input_ids.shape
|
308 |
+
outputs = self.bert(input_ids=input_ids,
|
309 |
+
attention_mask=attention_mask,
|
310 |
+
position_ids=position_ids,
|
311 |
+
token_type_ids=token_type_ids,
|
312 |
+
labels=mlmlabels) # (bsz, seq, dim)
|
313 |
+
mask_loss = outputs.loss
|
314 |
+
mlm_logits = outputs.logits
|
315 |
+
cls_logits = mlm_logits[:, :,
|
316 |
+
self.yes_token].view(-1, seq_len)+clslabels_mask
|
317 |
+
|
318 |
+
if mlmlabels == None:
|
319 |
+
return 0, mlm_logits, cls_logits
|
320 |
+
else:
|
321 |
+
cls_loss = self.loss_func(cls_logits, clslabels)
|
322 |
+
all_loss = mask_loss+cls_loss
|
323 |
+
return all_loss, mlm_logits, cls_logits
|
324 |
+
|
325 |
+
|
326 |
+
class UniMCLitModel(pl.LightningModule):
|
327 |
+
|
328 |
+
@staticmethod
|
329 |
+
def add_model_specific_args(parent_args):
|
330 |
+
parser = parent_args.add_argument_group('BaseModel')
|
331 |
+
|
332 |
+
parser.add_argument('--learning_rate', default=1e-5, type=float)
|
333 |
+
parser.add_argument('--weight_decay', default=0.1, type=float)
|
334 |
+
parser.add_argument('--warmup', default=0.01, type=float)
|
335 |
+
parser.add_argument('--num_labels', default=2, type=int)
|
336 |
+
|
337 |
+
return parent_args
|
338 |
+
|
339 |
+
def __init__(self, args, yes_token, num_data=100):
|
340 |
+
super().__init__()
|
341 |
+
self.args = args
|
342 |
+
self.num_data = num_data
|
343 |
+
self.model = UniMCModel(self.args.pretrained_model_path, yes_token)
|
344 |
+
|
345 |
+
def setup(self, stage) -> None:
|
346 |
+
if stage == 'fit':
|
347 |
+
num_gpus = self.trainer.gpus if self.trainer.gpus is not None else 0
|
348 |
+
self.total_step = int(self.trainer.max_epochs * self.num_data /
|
349 |
+
(max(1, num_gpus) * self.trainer.accumulate_grad_batches))
|
350 |
+
print('Total training step:', self.total_step)
|
351 |
+
|
352 |
+
def training_step(self, batch, batch_idx):
|
353 |
+
loss, logits, cls_logits = self.model(**batch)
|
354 |
+
cls_acc = self.comput_metrix(
|
355 |
+
cls_logits, batch['clslabels'], batch['mlmlabels_mask'])
|
356 |
+
self.log('train_loss', loss)
|
357 |
+
self.log('train_acc', cls_acc)
|
358 |
+
return loss
|
359 |
+
|
360 |
+
def validation_step(self, batch, batch_idx):
|
361 |
+
loss, logits, cls_logits = self.model(**batch)
|
362 |
+
cls_acc = self.comput_metrix(
|
363 |
+
cls_logits, batch['clslabels'], batch['mlmlabels_mask'])
|
364 |
+
self.log('val_loss', loss)
|
365 |
+
self.log('val_acc', cls_acc)
|
366 |
+
|
367 |
+
def configure_optimizers(self):
|
368 |
+
|
369 |
+
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
|
370 |
+
paras = list(
|
371 |
+
filter(lambda p: p[1].requires_grad, self.named_parameters()))
|
372 |
+
paras = [{
|
373 |
+
'params':
|
374 |
+
[p for n, p in paras if not any(nd in n for nd in no_decay)],
|
375 |
+
'weight_decay': self.args.weight_decay
|
376 |
+
}, {
|
377 |
+
'params': [p for n, p in paras if any(nd in n for nd in no_decay)],
|
378 |
+
'weight_decay': 0.0
|
379 |
+
}]
|
380 |
+
optimizer = torch.optim.AdamW(paras, lr=self.args.learning_rate)
|
381 |
+
scheduler = get_linear_schedule_with_warmup(
|
382 |
+
optimizer, int(self.total_step * self.args.warmup),
|
383 |
+
self.total_step)
|
384 |
+
|
385 |
+
return [{
|
386 |
+
'optimizer': optimizer,
|
387 |
+
'lr_scheduler': {
|
388 |
+
'scheduler': scheduler,
|
389 |
+
'interval': 'step',
|
390 |
+
'frequency': 1
|
391 |
+
}
|
392 |
+
}]
|
393 |
+
|
394 |
+
def comput_metrix(self, logits, labels, mlmlabels_mask):
|
395 |
+
logits = torch.nn.functional.softmax(logits, dim=-1)
|
396 |
+
logits = torch.argmax(logits, dim=-1)
|
397 |
+
y_pred = logits.view(size=(-1,))
|
398 |
+
y_true = labels.view(size=(-1,))
|
399 |
+
corr = torch.eq(y_pred, y_true).float()
|
400 |
+
return torch.sum(corr.float())/labels.size(0)
|
401 |
+
|
402 |
+
|
403 |
+
class TaskModelCheckpoint:
|
404 |
+
@staticmethod
|
405 |
+
def add_argparse_args(parent_args):
|
406 |
+
parser = parent_args.add_argument_group('BaseModel')
|
407 |
+
|
408 |
+
parser.add_argument('--monitor', default='val_acc', type=str)
|
409 |
+
parser.add_argument('--mode', default='max', type=str)
|
410 |
+
parser.add_argument('--dirpath', default='./log/', type=str)
|
411 |
+
parser.add_argument(
|
412 |
+
'--filename', default='model-{epoch:02d}-{val_acc:.4f}', type=str)
|
413 |
+
parser.add_argument('--save_top_k', default=3, type=float)
|
414 |
+
parser.add_argument('--every_n_epochs', default=1, type=float)
|
415 |
+
parser.add_argument('--every_n_train_steps', default=100, type=float)
|
416 |
+
parser.add_argument('--save_weights_only', default=True, type=bool)
|
417 |
+
return parent_args
|
418 |
+
|
419 |
+
def __init__(self, args):
|
420 |
+
self.callbacks = ModelCheckpoint(monitor=args.monitor,
|
421 |
+
save_top_k=args.save_top_k,
|
422 |
+
mode=args.mode,
|
423 |
+
save_last=True,
|
424 |
+
every_n_train_steps=args.every_n_train_steps,
|
425 |
+
save_weights_only=args.save_weights_only,
|
426 |
+
dirpath=args.dirpath,
|
427 |
+
filename=args.filename)
|
428 |
+
|
429 |
+
|
430 |
+
class UniMCPredict:
|
431 |
+
def __init__(self, yes_token, no_token, model, tokenizer, args):
|
432 |
+
self.tokenizer = tokenizer
|
433 |
+
self.args = args
|
434 |
+
self.data_model = UniMCDataModel(
|
435 |
+
[], [], yes_token, no_token, tokenizer, args)
|
436 |
+
self.model = model
|
437 |
+
|
438 |
+
def predict(self, batch_data):
|
439 |
+
batch = [self.data_model.train_data.encode(
|
440 |
+
sample) for sample in batch_data]
|
441 |
+
batch = self.data_model.collate_fn(batch)
|
442 |
+
batch = {k: v.cuda() for k, v in batch.items()}
|
443 |
+
_, _, logits = self.model.model(**batch)
|
444 |
+
soft_logits = torch.nn.functional.softmax(logits, dim=-1)
|
445 |
+
logits = torch.argmax(soft_logits, dim=-1).detach().cpu().numpy()
|
446 |
+
|
447 |
+
soft_logits = soft_logits.detach().cpu().numpy()
|
448 |
+
clslabels_mask = batch['clslabels_mask'].detach(
|
449 |
+
).cpu().numpy().tolist()
|
450 |
+
clslabels = batch['clslabels'].detach().cpu().numpy().tolist()
|
451 |
+
for i, v in enumerate(batch_data):
|
452 |
+
label_idx = [idx for idx, v in enumerate(
|
453 |
+
clslabels_mask[i]) if v == 0.]
|
454 |
+
label = label_idx.index(logits[i])
|
455 |
+
answer = batch_data[i]['choice'][label]
|
456 |
+
score = {}
|
457 |
+
for c in range(len(batch_data[i]['choice'])):
|
458 |
+
score[batch_data[i]['choice'][c]] = float(
|
459 |
+
soft_logits[i][label_idx[c]])
|
460 |
+
|
461 |
+
batch_data[i]['label_ori'] = copy.deepcopy(batch_data[i]['label'])
|
462 |
+
batch_data[i]['label'] = label
|
463 |
+
batch_data[i]['answer'] = answer
|
464 |
+
batch_data[i]['score'] = score
|
465 |
+
|
466 |
+
return batch_data
|
467 |
+
|
468 |
+
|
469 |
+
class UniMCPipelines:
|
470 |
+
@staticmethod
|
471 |
+
def pipelines_args(parent_args):
|
472 |
+
total_parser = parent_args.add_argument_group("pipelines args")
|
473 |
+
total_parser.add_argument(
|
474 |
+
'--pretrained_model_path', default='', type=str)
|
475 |
+
total_parser.add_argument('--load_checkpoints_path',
|
476 |
+
default='', type=str)
|
477 |
+
total_parser.add_argument('--train', action='store_true')
|
478 |
+
total_parser.add_argument('--language',
|
479 |
+
default='chinese', type=str)
|
480 |
+
|
481 |
+
total_parser = UniMCDataModel.add_data_specific_args(total_parser)
|
482 |
+
total_parser = TaskModelCheckpoint.add_argparse_args(total_parser)
|
483 |
+
total_parser = UniMCLitModel.add_model_specific_args(total_parser)
|
484 |
+
total_parser = pl.Trainer.add_argparse_args(parent_args)
|
485 |
+
return parent_args
|
486 |
+
|
487 |
+
def __init__(self, args):
|
488 |
+
self.args = args
|
489 |
+
self.checkpoint_callback = TaskModelCheckpoint(args).callbacks
|
490 |
+
self.logger = loggers.TensorBoardLogger(save_dir=args.default_root_dir)
|
491 |
+
self.trainer = pl.Trainer.from_argparse_args(args,
|
492 |
+
logger=self.logger,
|
493 |
+
callbacks=[self.checkpoint_callback])
|
494 |
+
self.config = AutoConfig.from_pretrained(args.pretrained_model_path)
|
495 |
+
if self.config.model_type == 'albert':
|
496 |
+
self.tokenizer = AlbertTokenizer.from_pretrained(
|
497 |
+
args.pretrained_model_path)
|
498 |
+
else:
|
499 |
+
if args.language == 'chinese':
|
500 |
+
self.tokenizer = BertTokenizer.from_pretrained(
|
501 |
+
args.pretrained_model_path)
|
502 |
+
else:
|
503 |
+
self.tokenizer = AutoTokenizer.from_pretrained(
|
504 |
+
args.pretrained_model_path, is_split_into_words=True, add_prefix_space=True)
|
505 |
+
|
506 |
+
if args.language == 'chinese':
|
507 |
+
self.yes_token = self.tokenizer.encode('是')[1]
|
508 |
+
self.no_token = self.tokenizer.encode('非')[1]
|
509 |
+
else:
|
510 |
+
self.yes_token = self.tokenizer.encode('yes')[1]
|
511 |
+
self.no_token = self.tokenizer.encode('no')[1]
|
512 |
+
|
513 |
+
if args.load_checkpoints_path != '':
|
514 |
+
self.model = UniMCLitModel.load_from_checkpoint(
|
515 |
+
args.load_checkpoints_path, args=args, yes_token=self.yes_token)
|
516 |
+
print('load model from: ', args.load_checkpoints_path)
|
517 |
+
else:
|
518 |
+
self.model = UniMCLitModel(args, yes_token=self.yes_token)
|
519 |
+
|
520 |
+
def fit(self, train_data, dev_data, process=True):
|
521 |
+
if process:
|
522 |
+
train_data = self.preprocess(train_data)
|
523 |
+
dev_data = self.preprocess(dev_data)
|
524 |
+
data_model = UniMCDataModel(
|
525 |
+
train_data, dev_data, self.yes_token, self.no_token, self.tokenizer, self.args)
|
526 |
+
self.model.num_data = len(train_data)
|
527 |
+
self.trainer.fit(self.model, data_model)
|
528 |
+
|
529 |
+
def predict(self, test_data, cuda=True, process=True):
|
530 |
+
if process:
|
531 |
+
test_data = self.preprocess(test_data)
|
532 |
+
|
533 |
+
result = []
|
534 |
+
start = 0
|
535 |
+
if cuda:
|
536 |
+
self.model = self.model.cuda()
|
537 |
+
self.model.model.eval()
|
538 |
+
predict_model = UniMCPredict(
|
539 |
+
self.yes_token, self.no_token, self.model, self.tokenizer, self.args)
|
540 |
+
while start < len(test_data):
|
541 |
+
batch_data = test_data[start:start+self.args.batchsize]
|
542 |
+
start += self.args.batchsize
|
543 |
+
batch_result = predict_model.predict(batch_data)
|
544 |
+
result.extend(batch_result)
|
545 |
+
if process:
|
546 |
+
result = self.postprocess(result)
|
547 |
+
return result
|
548 |
+
|
549 |
+
def preprocess(self, data):
|
550 |
+
|
551 |
+
for i, line in enumerate(data):
|
552 |
+
if 'task_type' in line.keys() and line['task_type'] == '语义匹配':
|
553 |
+
data[i]['choice'] = ['不能理解为:'+data[i]
|
554 |
+
['textb'], '可以理解为:'+data[i]['textb']]
|
555 |
+
# data[i]['question']='怎么理解这段话?'
|
556 |
+
data[i]['textb'] = ''
|
557 |
+
|
558 |
+
if 'task_type' in line.keys() and line['task_type'] == '自然语言推理':
|
559 |
+
data[i]['choice'] = ['不能推断出:'+data[i]['textb'],
|
560 |
+
'很难推断出:'+data[i]['textb'], '可以推断出:'+data[i]['textb']]
|
561 |
+
# data[i]['question']='根据这段话'
|
562 |
+
data[i]['textb'] = ''
|
563 |
+
|
564 |
+
return data
|
565 |
+
|
566 |
+
def postprocess(self, data):
|
567 |
+
for i, line in enumerate(data):
|
568 |
+
if 'task_type' in line.keys() and line['task_type'] == '语义匹配':
|
569 |
+
data[i]['textb'] = data[i]['choice'][0].replace('不能理解为:', '')
|
570 |
+
data[i]['choice'] = ['不相似', '相似']
|
571 |
+
ns = {}
|
572 |
+
for k, v in data[i]['score'].items():
|
573 |
+
if '不能' in k:
|
574 |
+
k = '不相似'
|
575 |
+
if '可以' in k:
|
576 |
+
k = '相似'
|
577 |
+
ns[k] = v
|
578 |
+
data[i]['score'] = ns
|
579 |
+
data[i]['answer'] = data[i]['choice'][data[i]['label']]
|
580 |
+
|
581 |
+
if 'task_type' in line.keys() and line['task_type'] == '自然语言推理':
|
582 |
+
data[i]['textb'] = data[i]['choice'][0].replace('不能推断出:', '')
|
583 |
+
data[i]['choice'] = ['矛盾', '自然', '蕴含']
|
584 |
+
ns = {}
|
585 |
+
for k, v in data[i]['score'].items():
|
586 |
+
if '不能' in k:
|
587 |
+
k = '矛盾'
|
588 |
+
if '很难' in k:
|
589 |
+
k = '自然'
|
590 |
+
if '可以' in k:
|
591 |
+
k = '蕴含'
|
592 |
+
ns[k] = v
|
593 |
+
data[i]['score'] = ns
|
594 |
+
data[i]['answer'] = data[i]['choice'][data[i]['label']]
|
595 |
+
|
596 |
+
return data
|
597 |
+
|
598 |
+
|
599 |
+
def load_data(data_path):
|
600 |
+
with open(data_path, 'r', encoding='utf8') as f:
|
601 |
+
lines = f.readlines()
|
602 |
+
samples = [json.loads(line) for line in tqdm(lines)]
|
603 |
+
return samples
|
604 |
+
|
605 |
+
|
606 |
+
def comp_acc(pred_data, test_data):
|
607 |
+
corr = 0
|
608 |
+
for i in range(len(pred_data)):
|
609 |
+
if pred_data[i]['label'] == test_data[i]['label']:
|
610 |
+
corr += 1
|
611 |
+
return corr/len(pred_data)
|
612 |
+
|
613 |
+
|
614 |
+
@st.experimental_memo()
|
615 |
+
def load_model():
|
616 |
+
total_parser = argparse.ArgumentParser("TASK NAME")
|
617 |
+
total_parser = UniMCPipelines.pipelines_args(total_parser)
|
618 |
+
args = total_parser.parse_args()
|
619 |
+
|
620 |
+
args.pretrained_model_path = 'IDEA-CCNL/Erlangshen-UniMC-RoBERTa-110M-Chinese'
|
621 |
+
args.max_length = 512
|
622 |
+
args.batchsize = 8
|
623 |
+
args.default_root_dir = './'
|
624 |
+
|
625 |
+
model = UniMCPipelines(args)
|
626 |
+
return model
|
627 |
+
|
628 |
+
|
629 |
+
def main():
|
630 |
+
|
631 |
+
model = load_model()
|
632 |
+
|
633 |
+
st.subheader("UniMC Zero-shot 体验")
|
634 |
+
st.info("请输入以下信息...")
|
635 |
+
|
636 |
+
sentences = st.text_area("请输入句子:", """彭于晏不着急,胡歌也不着急,他俩都不着急,那我也不着急""")
|
637 |
+
question = st.text_input("请输入问题(不输入问题也可以):", "请问下面的新闻属于哪个类别?")
|
638 |
+
choice = st.text_input("输入标签(以中文;分割):", "娱乐;军事;体育;财经")
|
639 |
+
choice = choice.split(';')
|
640 |
+
|
641 |
+
data = [{"texta": sentences,
|
642 |
+
"textb": "",
|
643 |
+
"question": question,
|
644 |
+
"choice": choice,
|
645 |
+
"answer": "", "label": 0,
|
646 |
+
"id": 0}]
|
647 |
+
|
648 |
+
if st.button("点击一下,开始预测!"):
|
649 |
+
result = model.predict(data, cuda=False)
|
650 |
+
st.success("预测成功!")
|
651 |
+
st.json(result[0])
|
652 |
+
else:
|
653 |
+
st.info(
|
654 |
+
"**Enter a text** above and **press the button** to predict the category."
|
655 |
+
)
|
656 |
+
|
657 |
+
|
658 |
+
if __name__ == "__main__":
|
659 |
+
main()
|