File size: 19,765 Bytes
9d42f47
 
0681bc6
40bc3ae
1e660fe
b97751f
b6bed34
10396f5
a51b600
99b2dac
691657e
30eb466
e284b45
64226c2
 
187bb52
 
ac05d72
d818c99
171817a
7437ef5
 
 
171817a
 
74e821f
 
 
ea2c6ba
 
 
336121a
 
da6f7d4
 
 
 
 
8e6d41d
 
08eb467
 
168d045
 
df20fa3
 
 
b62e2d9
 
fae8953
 
15505de
 
 
5686d0a
 
 
 
d50b787
 
 
87281da
 
 
4acf537
 
 
 
027f3d0
 
 
 
a2f8e94
 
 
 
2ad8ac8
9e7288d
6a7186e
 
4fd5b35
 
55f1dff
 
53082d2
22ddb9e
 
 
 
 
9fc882b
3315012
 
 
 
 
 
 
 
 
da1164c
 
 
 
 
 
 
f030196
 
 
 
 
 
 
 
05f01c9
 
59662d1
05f01c9
 
 
 
0b6db0c
 
 
 
 
 
453434f
 
 
 
 
 
30f6bdb
f9e196a
a1bbdbc
 
 
 
f9e196a
68248ca
 
2c57b14
68248ca
3b5d713
 
 
 
ae4f8b6
 
 
b3810dd
ae4f8b6
b3810dd
cd7fa2f
 
 
9ed09a4
 
 
 
0015a3c
9ed09a4
43e0847
 
 
 
 
 
 
 
 
9ed09a4
6e46269
 
 
 
 
 
 
 
 
 
7002b11
1b559c6
9978534
6e46269
 
0015a3c
6e46269
 
 
056c147
 
 
 
 
 
 
0015a3c
056c147
 
 
 
 
 
 
 
 
fafca07
 
 
 
 
 
0015a3c
fafca07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0015a3c
fafca07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0015a3c
fafca07
 
 
 
2679662
 
 
 
 
e8d776b
fafca07
 
2679662
 
 
 
 
fafca07
 
 
2679662
 
 
fafca07
 
 
 
 
2679662
fafca07
 
2679662
fafca07
 
8cdbfed
fafca07
 
 
 
07c0bde
fafca07
cbf3592
dd1727d
cf4737f
26d8504
5ecdcfb
26d09e7
d8aa17a
83a1f94
c30637a
0f1daf7
e4c6ec6
b9ee162
c8a440e
4fb639f
a88fd83
fafca07
 
0bc3261
 
 
 
 
 
 
 
 
 
 
 
 
 
fafca07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7f81350
adf7210
fafca07
 
 
 
 
 
 
 
 
df90151
fafca07
 
 
5e3ef77
 
 
 
 
 
fafca07
 
bcf1243
 
 
 
fafca07
63f745e
fafca07
7010421
ea40e7f
fafca07
 
 
 
 
 
 
 
 
 
 
 
 
 
b48d762
fafca07
 
25a6cb6
fafca07
 
 
 
 
 
 
 
 
5866066
fafca07
 
 
 
 
 
 
 
c22b8b7
b4f8dd3
bf102c2
8291502
272671c
ae6c335
c4b22ed
c59f6db
5288717
3744f32
744530a
c28c196
7491802
981a6e0
7f493d7
fafca07
 
 
 
 
96f677f
 
 
fafca07
 
 
 
 
16c3afe
fafca07
 
 
 
 
 
 
 
 
 
7c59938
 
 
 
 
 
 
 
 
 
 
 
 
fafca07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2757c53
fb1c8a9
ffdf601
116ab57
b974160
3e38b94
f7095dd
502f04f
5ffa1ac
19db832
 
daf0155
 
 
ab10bc6
 
 
 
d7bb16b
2981c19
 
 
1cf9cc4
 
 
cf02d2a
 
 
ac42dfa
 
 
1875c38
 
 
 
cd07047
 
 
 
d513beb
 
 
b4a6e7e
 
 
 
 
 
 
 
 
 
 
 
be8895a
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
import torch
import pandas as pd
from transformers import (
    AdamW,
    T5ForConditionalGeneration,
    T5TokenizerFast as T5Tokenizer,
    MT5Tokenizer,
    MT5ForConditionalGeneration,
    ByT5Tokenizer,
)
from torch.utils.data import Dataset, DataLoader
import pytorch_lightning as pl
from pytorch_lightning.loggers import MLFlowLogger, WandbLogger
from pytorch_lightning import Trainer
from pytorch_lightning.callbacks.early_stopping import EarlyStopping
from pytorch_lightning import LightningDataModule
from pytorch_lightning import LightningModule
from datasets import load_metric
from tqdm.auto import tqdm

# from dagshub.pytorch_lightning import DAGsHubLogger


torch.cuda.empty_cache()
pl.seed_everything(42)


class DataModule(Dataset):
    """
    Data Module for pytorch
    """

    def __init__(
        self,
        data: pd.DataFrame,
        tokenizer: T5Tokenizer,
        source_max_token_len: int = 512,
        target_max_token_len: int = 512,
    ):
        """
        :param data:
        :param tokenizer:
        :param source_max_token_len:
        :param target_max_token_len:
        """
        self.data = data
        self.target_max_token_len = target_max_token_len
        self.source_max_token_len = source_max_token_len
        self.tokenizer = tokenizer

    def __len__(self):
        return len(self.data)

    def __getitem__(self, index: int):
        data_row = self.data.iloc[index]

        input_encoding = self.tokenizer(
            data_row["input_text"],
            max_length=self.source_max_token_len,
            padding="max_length",
            truncation=True,
            return_attention_mask=True,
            add_special_tokens=True,
            return_tensors="pt",
        )

        output_encoding = self.tokenizer(
            data_row["output_text"],
            max_length=self.target_max_token_len,
            padding="max_length",
            truncation=True,
            return_attention_mask=True,
            add_special_tokens=True,
            return_tensors="pt",
        )

        labels = output_encoding["input_ids"]
        labels[labels == 0] = -100

        return dict(
            keywords=data_row["input_text"],
            text=data_row["output_text"],
            keywords_input_ids=input_encoding["input_ids"].flatten(),
            keywords_attention_mask=input_encoding["attention_mask"].flatten(),
            labels=labels.flatten(),
            labels_attention_mask=output_encoding["attention_mask"].flatten(),
        )


class PLDataModule(LightningDataModule):
    def __init__(
        self,
        train_df: pd.DataFrame,
        test_df: pd.DataFrame,
        tokenizer: T5Tokenizer,
        source_max_token_len: int = 512,
        target_max_token_len: int = 512,
        batch_size: int = 4,
        split: float = 0.1,
        num_workers: int = 2,
    ):
        """
        :param data_df:
        :param tokenizer:
        :param source_max_token_len:
        :param target_max_token_len:
        :param batch_size:
        :param split:
        """
        super().__init__()
        self.train_df = train_df
        self.test_df = test_df
        self.split = split
        self.batch_size = batch_size
        self.target_max_token_len = target_max_token_len
        self.source_max_token_len = source_max_token_len
        self.tokenizer = tokenizer
        self.num_workers = num_workers

    def setup(self, stage=None):
        self.train_dataset = DataModule(
            self.train_df,
            self.tokenizer,
            self.source_max_token_len,
            self.target_max_token_len,
        )
        self.test_dataset = DataModule(
            self.test_df,
            self.tokenizer,
            self.source_max_token_len,
            self.target_max_token_len,
        )

    def train_dataloader(self):
        """training dataloader"""
        return DataLoader(
            self.train_dataset,
            batch_size=self.batch_size,
            shuffle=True,
            num_workers=self.num_workers,
        )

    def test_dataloader(self):
        """test dataloader"""
        return DataLoader(
            self.test_dataset,
            batch_size=self.batch_size,
            shuffle=False,
            num_workers=self.num_workers,
        )

    def val_dataloader(self):
        """validation dataloader"""
        return DataLoader(
            self.test_dataset,
            batch_size=self.batch_size,
            shuffle=False,
            num_workers=self.num_workers,
        )


class LightningModel(LightningModule):
    """PyTorch Lightning Model class"""

    def __init__(
        self,
        tokenizer,
        model,
        learning_rate,
        adam_epsilon,
        weight_decay,
        output: str = "outputs",
    ):
        """
        initiates a PyTorch Lightning Model
        Args:
            tokenizer : T5 tokenizer
            model : T5 model
            output (str, optional): output directory to save model checkpoints. Defaults to "outputs".
        """
        super().__init__()
        self.model = model
        self.tokenizer = tokenizer
        self.output = output
        self.learning_rate = learning_rate
        self.adam_epsilon = adam_epsilon
        self.weight_decay = weight_decay

    def forward(self, input_ids, attention_mask, decoder_attention_mask, labels=None):
        """forward step"""
        output = self.model(
            input_ids,
            attention_mask=attention_mask,
            labels=labels,
            decoder_attention_mask=decoder_attention_mask,
        )

        return output.loss, output.logits

    def training_step(self, batch, batch_size):
        """training step"""
        input_ids = batch["keywords_input_ids"]
        attention_mask = batch["keywords_attention_mask"]
        labels = batch["labels"]
        labels_attention_mask = batch["labels_attention_mask"]

        loss, outputs = self(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_attention_mask=labels_attention_mask,
            labels=labels,
        )
        self.log("train_loss", loss, prog_bar=True, logger=True)
        return loss

    def validation_step(self, batch, batch_size):
        """validation step"""
        input_ids = batch["keywords_input_ids"]
        attention_mask = batch["keywords_attention_mask"]
        labels = batch["labels"]
        labels_attention_mask = batch["labels_attention_mask"]

        loss, outputs = self(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_attention_mask=labels_attention_mask,
            labels=labels,
        )
        self.log("val_loss", loss, prog_bar=True, logger=True)
        return loss

    def test_step(self, batch, batch_size):
        """test step"""
        input_ids = batch["keywords_input_ids"]
        attention_mask = batch["keywords_attention_mask"]
        labels = batch["labels"]
        labels_attention_mask = batch["labels_attention_mask"]

        loss, outputs = self(
            input_ids=input_ids,
            attention_mask=attention_mask,
            decoder_attention_mask=labels_attention_mask,
            labels=labels,
        )

        self.log("test_loss", loss, prog_bar=True, logger=True)
        return loss

    def configure_optimizers(self):
        """configure optimizers"""
        model = self.model
        no_decay = ["bias", "LayerNorm.weight"]
        optimizer_grouped_parameters = [
            {
                "params": [
                    p
                    for n, p in model.named_parameters()
                    if not any(nd in n for nd in no_decay)
                ],
                "weight_decay": self.weight_decay,
            },
            {
                "params": [
                    p
                    for n, p in model.named_parameters()
                    if any(nd in n for nd in no_decay)
                ],
                "weight_decay": 0.0,
            },
        ]
        optimizer = AdamW(
            optimizer_grouped_parameters, lr=self.learning_rate, eps=self.adam_epsilon
        )
        self.opt = optimizer
        return [optimizer]


class Summarization:
    """Custom Summarization class"""

    def __init__(self) -> None:
        """initiates Summarization class"""
        pass

    def from_pretrained(self, model_type="t5", model_name="t5-base") -> None:
        """
        loads T5/MT5 Model model for training/finetuning
        Args:
            model_name (str, optional): exact model architecture name, "t5-base" or "t5-large". Defaults to "t5-base".
            :param model_type:
        """
        if model_type == "t5":
            self.tokenizer = T5Tokenizer.from_pretrained(f"{model_name}")
            self.model = T5ForConditionalGeneration.from_pretrained(
                f"{model_name}", return_dict=True
            )
        elif model_type == "mt5":
            self.tokenizer = MT5Tokenizer.from_pretrained(f"{model_name}")
            self.model = MT5ForConditionalGeneration.from_pretrained(
                f"{model_name}", return_dict=True
            )
        elif model_type == "byt5":
            self.tokenizer = ByT5Tokenizer.from_pretrained(f"{model_name}")
            self.model = T5ForConditionalGeneration.from_pretrained(
                f"{model_name}", return_dict=True
            )

    def train(
        self,
        train_df: pd.DataFrame,
        eval_df: pd.DataFrame,
        source_max_token_len: int = 512,
        target_max_token_len: int = 512,
        batch_size: int = 8,
        max_epochs: int = 5,
        use_gpu: bool = True,
        outputdir: str = "models",
        early_stopping_patience_epochs: int = 0,  # 0 to disable early stopping feature
        learning_rate: float = 0.0001,
        adam_epsilon: float = 0.01,
        num_workers: int = 2,
        weight_decay: float = 0.0001,
    ):
        """
        trains T5/MT5 model on custom dataset
        Args:
            train_df (pd.DataFrame): training datarame. Dataframe must have 2 column --> "input_text" and "output_text"
            eval_df ([type], optional): validation datarame. Dataframe must have 2 column --> "input_text" and
            "output_text"
            source_max_token_len (int, optional): max token length of source text. Defaults to 512.
            target_max_token_len (int, optional): max token length of target text. Defaults to 512.
            batch_size (int, optional): batch size. Defaults to 8.
            max_epochs (int, optional): max number of epochs. Defaults to 5.
            use_gpu (bool, optional): if True, model uses gpu for training. Defaults to True.
            outputdir (str, optional): output directory to save model checkpoints. Defaults to "outputs".
            early_stopping_patience_epochs (int, optional): monitors val_loss on epoch end and stops training,
            if val_loss does not improve after the specied number of epochs. set 0 to disable early stopping.
            Defaults to 0 (disabled)
            :param learning_rate:
            :param adam_epsilon:
        """
        self.target_max_token_len = target_max_token_len
        self.data_module = PLDataModule(
            train_df,
            eval_df,
            self.tokenizer,
            batch_size=batch_size,
            source_max_token_len=source_max_token_len,
            target_max_token_len=target_max_token_len,
            num_workers=num_workers,
        )

        self.T5Model = LightningModel(
            tokenizer=self.tokenizer,
            model=self.model,
            output=outputdir,
            learning_rate=learning_rate,
            adam_epsilon=adam_epsilon,
            weight_decay=weight_decay,
        )

        MLlogger = MLFlowLogger(
            experiment_name="Summarization",
            tracking_uri="https://dagshub.com/gagan3012/summarization.mlflow",
        )

        WandLogger = WandbLogger(project="summarization-dagshub")

        # logger = DAGsHubLogger(metrics_path='reports/training_metrics.txt')

        early_stop_callback = (
            [
                EarlyStopping(
                    monitor="val_loss",
                    min_delta=0.00,
                    patience=early_stopping_patience_epochs,
                    verbose=True,
                    mode="min",
                )
            ]
            if early_stopping_patience_epochs > 0
            else None
        )

        gpus = -1 if use_gpu and torch.cuda.is_available() else 0

        trainer = Trainer(
            logger=[WandLogger, MLlogger],
            callbacks=early_stop_callback,
            max_epochs=max_epochs,
            gpus=gpus,
            progress_bar_refresh_rate=5,
        )

        trainer.fit(self.T5Model, self.data_module)

    def load_model(
        self, model_type: str = "t5", model_dir: str = "models", use_gpu: bool = False
    ):
        """
        loads a checkpoint for inferencing/prediction
        Args:
            model_type (str, optional): "t5" or "mt5". Defaults to "t5".
            model_dir (str, optional): path to model directory. Defaults to "outputs".
            use_gpu (bool, optional): if True, model uses gpu for inferencing/prediction. Defaults to True.
        """
        if model_type == "t5":
            self.tokenizer = T5Tokenizer.from_pretrained(f"{model_dir}")
            self.model = T5ForConditionalGeneration.from_pretrained(
                f"{model_dir}", return_dict=True
            )
        elif model_type == "mt5":
            self.tokenizer = MT5Tokenizer.from_pretrained(f"{model_dir}")
            self.model = MT5ForConditionalGeneration.from_pretrained(
                f"{model_dir}", return_dict=True
            )
        elif model_type == "byt5":
            self.tokenizer = ByT5Tokenizer.from_pretrained(f"{model_dir}")
            self.model = T5ForConditionalGeneration.from_pretrained(
                f"{model_dir}", return_dict=True
            )

        if use_gpu:
            if torch.cuda.is_available():
                self.device = torch.device("cuda")
            else:
                raise Exception(
                    "exception ---> no gpu found. set use_gpu=False, to use CPU"
                )
        else:
            self.device = torch.device("cpu")

        self.model = self.model.to(self.device)

    def save_model(self, model_dir="models"):
        """
        Save model to dir
        :param model_dir:
        :return: model is saved
        """
        path = f"{model_dir}"
        self.tokenizer.save_pretrained(path)
        self.model.save_pretrained(path)

    def predict(
        self,
        source_text: str,
        max_length: int = 512,
        num_return_sequences: int = 1,
        num_beams: int = 2,
        top_k: int = 50,
        top_p: float = 0.95,
        do_sample: bool = True,
        repetition_penalty: float = 2.5,
        length_penalty: float = 1.0,
        early_stopping: bool = True,
        skip_special_tokens: bool = True,
        clean_up_tokenization_spaces: bool = True,
    ):
        """
        generates prediction for T5/MT5 model
        Args:
            source_text (str): any text for generating predictions
            max_length (int, optional): max token length of prediction. Defaults to 512.
            num_return_sequences (int, optional): number of predictions to be returned. Defaults to 1.
            num_beams (int, optional): number of beams. Defaults to 2.
            top_k (int, optional): Defaults to 50.
            top_p (float, optional): Defaults to 0.95.
            do_sample (bool, optional): Defaults to True.
            repetition_penalty (float, optional): Defaults to 2.5.
            length_penalty (float, optional): Defaults to 1.0.
            early_stopping (bool, optional): Defaults to True.
            skip_special_tokens (bool, optional): Defaults to True.
            clean_up_tokenization_spaces (bool, optional): Defaults to True.
        Returns:
            list[str]: returns predictions
        """
        input_ids = self.tokenizer.encode(
            source_text, return_tensors="pt", add_special_tokens=True
        )

        input_ids = input_ids.to(self.device)
        generated_ids = self.model.generate(
            input_ids=input_ids,
            num_beams=num_beams,
            max_length=max_length,
            repetition_penalty=repetition_penalty,
            length_penalty=length_penalty,
            early_stopping=early_stopping,
            top_p=top_p,
            top_k=top_k,
            num_return_sequences=num_return_sequences,
        )
        preds = self.tokenizer.decode(
            generated_ids[0],
            skip_special_tokens=skip_special_tokens,
            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
        )
        return preds

    def evaluate(self, test_df: pd.DataFrame, metrics: str = "rouge"):
        metric = load_metric(metrics)
        input_text = test_df["input_text"]
        references = test_df["output_text"]
        references = references.to_list()

        predictions = [self.predict(x) for x in tqdm(input_text)]

        results = metric.compute(predictions=predictions, references=references)

        output = {
            "Rouge_1 Low Precision": results["rouge1"].low.precision,
            "Rouge_1 Low recall": results["rouge1"].low.recall,
            "Rouge_1 Low F1": results["rouge1"].low.fmeasure,
            "Rouge_1 Mid Precision": results["rouge1"].mid.precision,
            "Rouge_1 Mid recall": results["rouge1"].mid.recall,
            "Rouge_1 Mid F1": results["rouge1"].mid.fmeasure,
            "Rouge_1 High Precision": results["rouge1"].high.precision,
            "Rouge_1 High recall": results["rouge1"].high.recall,
            "Rouge_1 High F1": results["rouge1"].high.fmeasure,
            "Rouge_2 Low Precision": results["rouge2"].low.precision,
            "Rouge_2 Low recall": results["rouge2"].low.recall,
            "Rouge_2 Low F1": results["rouge2"].low.fmeasure,
            "Rouge_2 Mid Precision": results["rouge2"].mid.precision,
            "Rouge_2 Mid recall": results["rouge2"].mid.recall,
            "Rouge_2 Mid F1": results["rouge2"].mid.fmeasure,
            "Rouge_2 High Precision": results["rouge2"].high.precision,
            "Rouge_2 High recall": results["rouge2"].high.recall,
            "Rouge_2 High F1": results["rouge2"].high.fmeasure,
            "Rouge_L Low Precision": results["rougeL"].low.precision,
            "Rouge_L Low recall": results["rougeL"].low.recall,
            "Rouge_L Low F1": results["rougeL"].low.fmeasure,
            "Rouge_L Mid Precision": results["rougeL"].mid.precision,
            "Rouge_L Mid recall": results["rougeL"].mid.recall,
            "Rouge_L Mid F1": results["rougeL"].mid.fmeasure,
            "Rouge_L High Precision": results["rougeL"].high.precision,
            "Rouge_L High recall": results["rougeL"].high.recall,
            "Rouge_L High F1": results["rougeL"].high.fmeasure,
            "rougeLsum Low Precision": results["rougeLsum"].low.precision,
            "rougeLsum Low recall": results["rougeLsum"].low.recall,
            "rougeLsum Low F1": results["rougeLsum"].low.fmeasure,
            "rougeLsum Mid Precision": results["rougeLsum"].mid.precision,
            "rougeLsum Mid recall": results["rougeLsum"].mid.recall,
            "rougeLsum Mid F1": results["rougeLsum"].mid.fmeasure,
            "rougeLsum High Precision": results["rougeLsum"].high.precision,
            "rougeLsum High recall": results["rougeLsum"].high.recall,
            "rougeLsum High F1": results["rougeLsum"].high.fmeasure,
        }
        return output