File size: 6,508 Bytes
afdeeca
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import pytorch_lightning as pl
import torch
import numpy as np
import datasets
from transformers import MaxLengthCriteria, StoppingCriteriaList
from transformers.optimization import AdamW
import itertools
from averitec.models.utils import count_stats, f1_metric, pairwise_meteor
from torchmetrics.text.rouge import ROUGEScore
import torch.nn.functional as F
import torchmetrics
from torchmetrics.classification import F1Score

def freeze_params(model):
  for layer in model.parameters():
    layer.requires_grade = False

class JustificationGenerationModule(pl.LightningModule):
    
  def __init__(self, tokenizer, model, learning_rate=1e-3, gen_num_beams=2, gen_max_length=100, should_pad_gen=True):
    super().__init__()
    self.tokenizer = tokenizer
    self.model = model
    self.learning_rate = learning_rate

    self.gen_num_beams = gen_num_beams
    self.gen_max_length = gen_max_length
    self.should_pad_gen = should_pad_gen

    #self.metrics =  datasets.load_metric('meteor')

    freeze_params(self.model.get_encoder())
    self.freeze_embeds()
  
  def freeze_embeds(self):
    ''' freeze the positional embedding parameters of the model; adapted from finetune.py '''
    freeze_params(self.model.model.shared)
    for d in [self.model.model.encoder, self.model.model.decoder]:
      freeze_params(d.embed_positions)
      freeze_params(d.embed_tokens)

  # Do a forward pass through the model
  def forward(self, input_ids, **kwargs):
    return self.model(input_ids, **kwargs)
  
  def configure_optimizers(self):
    optimizer = AdamW(self.parameters(), lr = self.learning_rate)
    return optimizer

  def shift_tokens_right(self, input_ids: torch.Tensor, pad_token_id: int, decoder_start_token_id: int):
    """
    Shift input ids one token to the right.
    https://github.com/huggingface/transformers/blob/main/src/transformers/models/bart/modeling_bart.py.
    """
    shifted_input_ids = input_ids.new_zeros(input_ids.shape)
    shifted_input_ids[:, 1:] = input_ids[:, :-1].clone()
    shifted_input_ids[:, 0] = decoder_start_token_id

    if pad_token_id is None:
        raise ValueError("self.model.config.pad_token_id has to be defined.")
    # replace possible -100 values in labels by `pad_token_id`
    shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)

    return shifted_input_ids

  def run_model(self, batch):
    src_ids, src_mask, tgt_ids = batch[0], batch[1], batch[2]

    decoder_input_ids = self.shift_tokens_right(
                tgt_ids, self.tokenizer.pad_token_id, self.tokenizer.pad_token_id # BART uses the EOS token to start generation as well. Might have to change for other models.
            )

    outputs = self(src_ids, attention_mask=src_mask, decoder_input_ids=decoder_input_ids, use_cache=False)
    return outputs

  def compute_loss(self, batch):
    tgt_ids = batch[2]
    logits = self.run_model(batch)[0]

    cross_entropy = torch.nn.CrossEntropyLoss(ignore_index=self.tokenizer.pad_token_id)
    loss = cross_entropy(logits.view(-1, logits.shape[-1]), tgt_ids.view(-1))

    return loss

  def training_step(self, batch, batch_idx):
    loss = self.compute_loss(batch)

    self.log("train_loss", loss, on_epoch=True)

    return {'loss':loss}

  def validation_step(self, batch, batch_idx):
    preds, loss, tgts = self.generate_and_compute_loss_and_tgts(batch)
    if self.should_pad_gen:
      preds = F.pad(preds, pad=(0, self.gen_max_length - preds.shape[1]), value=self.tokenizer.pad_token_id)

    self.log('val_loss', loss, prog_bar=True, sync_dist=True)

    return {'loss': loss, 'pred': preds, 'target': tgts}

  def test_step(self, batch, batch_idx):
    test_preds, test_loss, test_tgts = self.generate_and_compute_loss_and_tgts(batch)
    if self.should_pad_gen:
      test_preds = F.pad(test_preds, pad=(0, self.gen_max_length - test_preds.shape[1]), value=self.tokenizer.pad_token_id)

    self.log('test_loss', test_loss, prog_bar=True, sync_dist=True)

    return {'loss': test_loss, 'pred': test_preds, 'target': test_tgts}

  def test_epoch_end(self, outputs):
    self.handle_end_of_epoch_scoring(outputs, "test")

  def validation_epoch_end(self, outputs):
    self.handle_end_of_epoch_scoring(outputs, "val")

  def handle_end_of_epoch_scoring(self, outputs, prefix):
      gen = {}
      tgt = {}
      rouge = ROUGEScore()
      rouge_metric = lambda x, y: rouge(x,y)["rougeL_precision"]
      for out in outputs:
        preds = out['pred']
        tgts = out['target']

        preds = self.do_batch_detokenize(preds)
        tgts = self.do_batch_detokenize(tgts)

        for pred, t in zip(preds, tgts):
          rouge_d = rouge_metric(pred, t)
          self.log(prefix+"_rouge", rouge_d)

          meteor_d = pairwise_meteor(pred, t)
          self.log(prefix+"_meteor", meteor_d)

  def generate_and_compute_loss_and_tgts(self, batch):
    src_ids = batch[0]
    loss = self.compute_loss(batch)
    pred_ids, _ = self.generate_for_batch(src_ids)

    tgt_ids = batch[2]

    return pred_ids, loss, tgt_ids

  def do_batch_detokenize(self, batch):
    tokens = self.tokenizer.batch_decode(
      batch, 
      skip_special_tokens=True, 
      clean_up_tokenization_spaces=True
      )

    # Huggingface skipping of special tokens doesn't work for all models, so we do it manually as well for safety:
    tokens = [p.replace("<pad>", "") for p in tokens]
    tokens = [p.replace("<s>", "") for p in tokens]
    tokens = [p.replace("</s>", "") for p in tokens]

    return [t for t in tokens if t != ""]
  
  def generate_for_batch(self, batch):
    generated_ids = self.model.generate(
      batch, 
      decoder_start_token_id = self.tokenizer.pad_token_id,
      num_beams = self.gen_num_beams,
      max_length = self.gen_max_length
      )

    generated_tokens = self.tokenizer.batch_decode(
      generated_ids, 
      skip_special_tokens=True, 
      clean_up_tokenization_spaces=True
      )

    return generated_ids, generated_tokens


  def generate(self, text, max_input_length=512, device=None):
    encoded_dict = self.tokenizer(
            [text],
            max_length=max_input_length,
            padding="longest",
            truncation=True,
            return_tensors="pt",
            add_prefix_space = True
        )

    input_ids = encoded_dict['input_ids']

    if device is not None:
      input_ids = input_ids.to(device)

    with torch.no_grad():
        _, generated_tokens = self.generate_for_batch(input_ids)
    
    return generated_tokens[0]