Datasets:
Genius1237
commited on
Commit
•
d948b6d
1
Parent(s):
d0e6d22
Add scripts used for trainin/evaluating multilingual models
Browse files- politeness_regressor.py +273 -0
- requirements.txt +5 -0
politeness_regressor.py
ADDED
@@ -0,0 +1,273 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from argparse import ArgumentParser
|
2 |
+
from typing import List, Dict
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
import pytorch_lightning as pl
|
6 |
+
import sklearn.metrics
|
7 |
+
import sklearn.model_selection
|
8 |
+
import torch
|
9 |
+
import torch.optim
|
10 |
+
import torch.utils.data
|
11 |
+
import transformers
|
12 |
+
import pandas as pd
|
13 |
+
import random
|
14 |
+
import sklearn.metrics
|
15 |
+
|
16 |
+
try:
|
17 |
+
from polyglot.text import Text
|
18 |
+
except:
|
19 |
+
print("polyglot not installed. Cannot use --strategy_words")
|
20 |
+
|
21 |
+
class MyDataModule(pl.LightningDataModule):
|
22 |
+
def __init__(self, train_file, test_file, binary, tokenizer, max_length, batch_size, strategy_words_replacement_negate=False, strategy_words=None, random_masking_ratio=None):
|
23 |
+
super().__init__()
|
24 |
+
self.train_file = train_file
|
25 |
+
self.test_file = test_file
|
26 |
+
self.binary = binary
|
27 |
+
self.max_length = max_length
|
28 |
+
self.batch_size = batch_size
|
29 |
+
self.tokenizer = tokenizer
|
30 |
+
|
31 |
+
if strategy_words:
|
32 |
+
self.strategy_words = pd.read_csv(strategy_words)
|
33 |
+
self.strategy_words = set(list(self.strategy_words.values[:, 1:].reshape(-1)))
|
34 |
+
else:
|
35 |
+
self.strategy_words = None
|
36 |
+
self.strategy_words_replacement_negate = strategy_words_replacement_negate
|
37 |
+
self.random_masking_ratio = random_masking_ratio
|
38 |
+
|
39 |
+
@staticmethod
|
40 |
+
def read_file(file_name, text_only=False):
|
41 |
+
if file_name.split(".")[-1] == "csv":
|
42 |
+
df = pd.read_csv(file_name)
|
43 |
+
data = [(a, b) for a, b in zip(list(df['sentence']), df['score'])]
|
44 |
+
if text_only:
|
45 |
+
data = [t[0] for t in data]
|
46 |
+
else:
|
47 |
+
data = open(file_name).read().strip().split('\n')
|
48 |
+
return data
|
49 |
+
|
50 |
+
def setup(self, stage=None):
|
51 |
+
if self.train_file:
|
52 |
+
self.train_data = MyDataModule.read_file(self.train_file)
|
53 |
+
self.train_data, self.val_data = sklearn.model_selection.train_test_split(self.train_data, shuffle=False, test_size=0.2)
|
54 |
+
if self.test_file:
|
55 |
+
self.test_data = MyDataModule.read_file(self.test_file)
|
56 |
+
|
57 |
+
def prepare_dataloader(self, mode):
|
58 |
+
if mode == "train":
|
59 |
+
data = self.train_data
|
60 |
+
elif mode == "val":
|
61 |
+
data = self.val_data
|
62 |
+
else:
|
63 |
+
data = self.test_data
|
64 |
+
|
65 |
+
# tokenized = self.tokenizer([t[0] for t in data], padding="max_length", truncation=True, max_length=self.max_length, return_tensors="pt")
|
66 |
+
tokenized = MyDataModule.tokenize([t[0] for t in data], self.tokenizer, self.max_length, self.strategy_words_replacement_negate, self.strategy_words, self.random_masking_ratio)
|
67 |
+
if self.binary:
|
68 |
+
labels = torch.tensor([t[1] > 0 for t in data], dtype=int)
|
69 |
+
else:
|
70 |
+
labels = torch.tensor([t[1] for t in data])
|
71 |
+
|
72 |
+
if mode == "train":
|
73 |
+
weights = torch.zeros_like(labels)
|
74 |
+
weights[labels == 0] = labels.shape[0] - labels.sum()
|
75 |
+
weights[labels == 1] = labels.sum()
|
76 |
+
return torch.utils.data.DataLoader(torch.utils.data.TensorDataset(tokenized['input_ids'], tokenized['attention_mask'], labels), batch_size=self.batch_size, sampler=torch.utils.data.WeightedRandomSampler(1 / weights, len(weights), replacement=True))
|
77 |
+
else:
|
78 |
+
return torch.utils.data.DataLoader(torch.utils.data.TensorDataset(tokenized['input_ids'], tokenized['attention_mask'], labels), batch_size=self.batch_size)
|
79 |
+
|
80 |
+
@staticmethod
|
81 |
+
def tokenize(data: List[str], tokenizer, max_length, strategy_words_replacement_negate, strategy_words, random_masking_ratio):
|
82 |
+
if strategy_words is not None or random_masking_ratio is not None:
|
83 |
+
tokenized_data = []
|
84 |
+
for sentence in data:
|
85 |
+
words = Text(sentence).words
|
86 |
+
words = [t.lower() for t in words]
|
87 |
+
if strategy_words:
|
88 |
+
words = [t if ((t in strategy_words) != strategy_words_replacement_negate) else tokenizer.mask_token for t in words]
|
89 |
+
elif random_masking_ratio:
|
90 |
+
words = [t if random.random() <= random_masking_ratio else tokenizer.mask_token for t in words]
|
91 |
+
tokenized_data.append(' '.join(words))
|
92 |
+
out = tokenizer(tokenized_data, padding="max_length", truncation=True, max_length=max_length, return_tensors="pt")
|
93 |
+
# out['attention_mask'] = torch.tensor(out['input_ids'] != tokenizer.pad_token_id, dtype=int)
|
94 |
+
return out
|
95 |
+
else:
|
96 |
+
return tokenizer(data, padding="max_length", truncation=True, max_length=max_length, return_tensors="pt")
|
97 |
+
|
98 |
+
def train_dataloader(self):
|
99 |
+
return self.prepare_dataloader("train")
|
100 |
+
# return torch.utils.data.DataLoader(MyDataModule.CustomDataset1(self.tokenizer, self.train_data, self.max_length), batch_size=self.batch_size)
|
101 |
+
|
102 |
+
def test_dataloader(self):
|
103 |
+
return self.prepare_dataloader("test")
|
104 |
+
# return torch.utils.data.DataLoader(MyDataModule.CustomDataset1(self.tokenizer, self.test_data, self.max_length), batch_size=self.batch_size)
|
105 |
+
|
106 |
+
def val_dataloader(self):
|
107 |
+
return self.prepare_dataloader("val")
|
108 |
+
# return torch.utils.data.DataLoader(MyDataModule.CustomDataset1(self.tokenizer, self.val_data, self.max_length), batch_size=self.batch_size)
|
109 |
+
|
110 |
+
|
111 |
+
class RegressionModel(pl.LightningModule):
|
112 |
+
def __init__(self, pretrained_model, binary, learning_rate, num_warmup_steps, tokenizer):
|
113 |
+
super(RegressionModel, self).__init__()
|
114 |
+
self.save_hyperparameters()
|
115 |
+
self.pretrained_model = pretrained_model
|
116 |
+
self.binary = binary
|
117 |
+
self.learning_rate = learning_rate
|
118 |
+
self.num_warmup_steps = num_warmup_steps
|
119 |
+
self.tokenizer = tokenizer
|
120 |
+
self.model = transformers.AutoModelForSequenceClassification.from_pretrained(self.pretrained_model, num_labels=2 if self.binary else 1)
|
121 |
+
|
122 |
+
def forward(self, **kwargs):
|
123 |
+
return self.model(**kwargs)
|
124 |
+
|
125 |
+
def training_step(self, batch, batch_idx):
|
126 |
+
outputs = self.forward(input_ids=batch[0], attention_mask=batch[1], labels=batch[2])
|
127 |
+
loss = outputs['loss']
|
128 |
+
ret = {"loss": loss}
|
129 |
+
if self.binary:
|
130 |
+
acc = torch.tensor(batch[2] == torch.argmax(outputs['logits']), dtype=float).mean().item()
|
131 |
+
ret["acc"] = acc
|
132 |
+
else:
|
133 |
+
rmse = (torch.mean((batch[2] - outputs['logits'])**2)**0.5).item()
|
134 |
+
ret["rmse"] = rmse
|
135 |
+
|
136 |
+
return {"loss": loss, "log": ret}
|
137 |
+
|
138 |
+
def configure_optimizers(self):
|
139 |
+
optimizer = torch.optim.AdamW(self.parameters(), lr=self.learning_rate)
|
140 |
+
scheduler = transformers.get_linear_schedule_with_warmup(optimizer, self.num_warmup_steps, len(self.trainer.datamodule.train_dataloader()) // self.trainer.accumulate_grad_batches)
|
141 |
+
return [optimizer], [scheduler]
|
142 |
+
|
143 |
+
def test_step(self, batch, batch_idx):
|
144 |
+
return self.validation_step(batch, batch_idx, mode="test")
|
145 |
+
|
146 |
+
def validation_step(self, batch, batch_idx, mode="val"):
|
147 |
+
outputs = self.forward(input_ids=batch[0], attention_mask=batch[1], labels=batch[2])
|
148 |
+
loss = outputs['loss']
|
149 |
+
self.log("{}_loss".format(mode), loss, prog_bar=True)
|
150 |
+
|
151 |
+
ret = {"loss": loss}
|
152 |
+
if self.binary:
|
153 |
+
preds = torch.argmax(outputs['logits'], axis=1).tolist()
|
154 |
+
gold = batch[2].tolist()
|
155 |
+
ret["preds"] = preds
|
156 |
+
ret["gold"] = gold
|
157 |
+
# f1 = sklearn.metrics.f1_score(gold, preds)
|
158 |
+
# acc = sklearn.metrics.accuracy_score(gold, preds)
|
159 |
+
# ret["acc"] = acc
|
160 |
+
# ret["f1"] = f1
|
161 |
+
# self.log("{}_acc".format(mode), acc, prog_bar=True)
|
162 |
+
# self.log("{}_f1".format(mode), f1, prog_bar=True)
|
163 |
+
else:
|
164 |
+
preds = outputs['logits'].tolist()
|
165 |
+
gold = batch[2].tolist()
|
166 |
+
ret['preds'] = preds
|
167 |
+
ret['gold'] = gold
|
168 |
+
# rmse = (torch.mean((batch[2] - outputs['logits'])**2)**0.5).item()
|
169 |
+
# self.log("{}_rmse".format(mode), rmse, prog_bar=True)
|
170 |
+
# ret["rmse"] = rmse
|
171 |
+
|
172 |
+
return {"loss": loss, "log": ret}
|
173 |
+
|
174 |
+
def validation_epoch_end(self, outputs, mode="val"):
|
175 |
+
gold = []
|
176 |
+
preds = []
|
177 |
+
for batch in outputs:
|
178 |
+
gold.extend(batch['log']['gold'])
|
179 |
+
preds.extend(batch['log']['preds'])
|
180 |
+
if self.binary:
|
181 |
+
f1 = sklearn.metrics.f1_score(gold, preds)
|
182 |
+
acc = sklearn.metrics.accuracy_score(gold, preds)
|
183 |
+
self.log("{}_acc".format(mode), acc, prog_bar=True)
|
184 |
+
self.log("{}_f1".format(mode), f1, prog_bar=True)
|
185 |
+
else:
|
186 |
+
rmse = (torch.mean((torch.tensor(gold) - torch.tensor(preds))**2)**0.5).item()
|
187 |
+
self.log("{}_rmse".format(mode), rmse, prog_bar=True)
|
188 |
+
|
189 |
+
def test_epoch_end(self, outputs):
|
190 |
+
return self.validation_epoch_end(outputs, mode="test")
|
191 |
+
|
192 |
+
def predict_step(self, batch, batch_idx):
|
193 |
+
preds = self.forward(input_ids=batch[0], attention_mask=batch[1])
|
194 |
+
if self.binary:
|
195 |
+
ret = preds['logits'].tolist()
|
196 |
+
else:
|
197 |
+
ret = preds['logits'].view(-1).tolist()
|
198 |
+
return ret
|
199 |
+
|
200 |
+
@staticmethod
|
201 |
+
def add_model_specific_args(parent_parser):
|
202 |
+
parser = parent_parser.add_argument_group("RegressionModel")
|
203 |
+
parser.add_argument('--pretrained_model', type=str)
|
204 |
+
parser.add_argument('--learning_rate', type=float, default="5e-6")
|
205 |
+
parser.add_argument('--num_warmup_steps', type=float, default="0")
|
206 |
+
return parent_parser
|
207 |
+
|
208 |
+
|
209 |
+
if __name__ == "__main__":
|
210 |
+
parser = ArgumentParser()
|
211 |
+
parser.add_argument("--train", action="store_true")
|
212 |
+
parser.add_argument("--test", action="store_true")
|
213 |
+
parser.add_argument("--load_model", type=str)
|
214 |
+
parser.add_argument("--train_file", type=str)
|
215 |
+
parser.add_argument("--test_file", type=str)
|
216 |
+
parser.add_argument("--binary", action="store_true")
|
217 |
+
parser.add_argument("--seed", type=int, default=42)
|
218 |
+
parser.add_argument("--batch_size", type=int, default=64)
|
219 |
+
parser.add_argument("--max_length", type=int, default=128)
|
220 |
+
parser.add_argument("--model_save_location", type=str)
|
221 |
+
parser.add_argument("--preds_save_location", type=str)
|
222 |
+
parser.add_argument("--preds_save_logits", action="store_true")
|
223 |
+
parser.add_argument("--strategy_words", type=str)
|
224 |
+
parser.add_argument("--strategy_words_replacement_negate", action="store_true")
|
225 |
+
parser.add_argument("--random_masking_ratio", type=float)
|
226 |
+
parser = RegressionModel.add_model_specific_args(parser)
|
227 |
+
parser = pl.Trainer.add_argparse_args(parser)
|
228 |
+
args = parser.parse_args()
|
229 |
+
print(args)
|
230 |
+
|
231 |
+
pl.utilities.seed.seed_everything(seed=args.seed)
|
232 |
+
if args.load_model:
|
233 |
+
model = RegressionModel.load_from_checkpoint(args.load_model)
|
234 |
+
tokenizer = model.tokenizer
|
235 |
+
else:
|
236 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(args.pretrained_model)
|
237 |
+
model = RegressionModel(pretrained_model=args.pretrained_model, binary=args.binary, learning_rate=args.learning_rate, num_warmup_steps=args.num_warmup_steps, tokenizer=tokenizer)
|
238 |
+
trainer = pl.Trainer.from_argparse_args(args)
|
239 |
+
|
240 |
+
dataset = MyDataModule(train_file=args.train_file, test_file=args.test_file, binary=model.binary, max_length=args.max_length, batch_size=args.batch_size, tokenizer=tokenizer, strategy_words_replacement_negate=args.strategy_words_replacement_negate, strategy_words=args.strategy_words, random_masking_ratio=args.random_masking_ratio)
|
241 |
+
dataset.setup()
|
242 |
+
|
243 |
+
if args.train:
|
244 |
+
trainer.fit(model, dataset)
|
245 |
+
|
246 |
+
if args.test:
|
247 |
+
trainer.test(model, dataset.test_dataloader())
|
248 |
+
|
249 |
+
if args.preds_save_location:
|
250 |
+
data = MyDataModule.read_file(args.test_file, True)
|
251 |
+
strategy_words = None
|
252 |
+
if args.strategy_words:
|
253 |
+
strategy_words = pd.read_csv(args.strategy_words)
|
254 |
+
strategy_words = set(list(args.strategy_words.values[:, 1:].reshape(-1)))
|
255 |
+
tokenized = MyDataModule.tokenize(data, tokenizer, args.max_length, args.strategy_words_replacement_negate, strategy_words, args.random_masking_ratio)
|
256 |
+
input_data = torch.utils.data.DataLoader(torch.utils.data.TensorDataset(tokenized['input_ids'], tokenized['attention_mask']), batch_size=args.batch_size)
|
257 |
+
preds = trainer.predict(model, input_data, return_predictions=True)
|
258 |
+
preds = [t for y in preds for t in y]
|
259 |
+
preds = torch.tensor(preds)
|
260 |
+
if model.binary:
|
261 |
+
if args.preds_save_logits:
|
262 |
+
preds = torch.softmax(preds, axis=1)[:, 1].tolist()
|
263 |
+
else:
|
264 |
+
preds = preds.argmax(axis=1).tolist()
|
265 |
+
else:
|
266 |
+
preds = preds.view(-1).tolist()
|
267 |
+
preds = [str(t) for t in preds]
|
268 |
+
|
269 |
+
with open(args.preds_save_location, 'w') as f:
|
270 |
+
f.write('\n'.join(preds) + '\n')
|
271 |
+
|
272 |
+
if args.model_save_location:
|
273 |
+
trainer.save_checkpoint(args.model_save_location, weights_only=True)
|
requirements.txt
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
numpy
|
2 |
+
pandas
|
3 |
+
torch
|
4 |
+
pytorch-lightning
|
5 |
+
sklearn
|