File size: 5,441 Bytes
d758c99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
from transformers.modeling_bart import BartForConditionalGeneration
import logging
import os
from relogic.logickit.modules.span_extractors.average_span_extractor import AverageSpanExtractor
import torch.nn.functional as F

logger = logging.getLogger(__name__)
WEIGHTS_NAME = "pytorch_model.bin"

class TaBARTModel(nn.Module):
  """
  output: tuple: (loss, ) in training
  """
  def __init__(self):
    super().__init__()
    self.bert = BartForConditionalGeneration.from_pretrained("facebook/bart-large")
    self.average_span_extractor = AverageSpanExtractor()
    self.column_mlp = nn.Linear(self.bert.config.d_model, self.bert.config.d_model)
    self.column_to_prob = nn.Linear(self.bert.config.d_model, 1)

  def column_prediction(self, input_ids, attention_mask, column_spans):
    column_mask = (column_spans[:, :, 0] > 0).long()
    features = self.bert.model.encoder(input_ids=input_ids,
                                 attention_mask=attention_mask)[0].contiguous()
    column_features = self.average_span_extractor(
      sequence_tensor=features,
      span_indices=column_spans,
      span_indices_mask=column_mask)
    column_selection_logits = self.column_to_prob(torch.relu(self.column_mlp(column_features)))
    column_selection_prob = torch.sigmoid(column_selection_logits)
    return column_selection_prob


  def forward(self, *input, **kwargs):
    input_ids = kwargs.pop("input_ids")

    pad_token_id = kwargs.pop("pad_token_id")
    attention_mask = (input_ids != pad_token_id).long()

    if self.training:
      task = kwargs.pop("task")
      if task == "mlm":
        output_ids = kwargs.pop('labels')
        y_ids = output_ids[:, :-1].contiguous()
        lm_labels = output_ids[:, 1:].clone()
        lm_labels[output_ids[:, 1:] == pad_token_id] = -100

        outputs = self.bert(input_ids,
                            attention_mask=attention_mask, decoder_input_ids=y_ids, lm_labels=lm_labels, )
        return (outputs[0],)
      elif task == "col_pred":
        label_ids = kwargs.pop("labels")
        column_spans = kwargs.pop("column_spans")
        column_selection_prob = self.column_prediction(input_ids, attention_mask, column_spans)
        label_mask = column_spans.view(-1, 2)[:,0] > 0

        column_selection_loss = F.binary_cross_entropy(column_selection_prob.view(-1)[label_mask], label_ids.view(-1)[label_mask].float(),
                                                       reduction="sum") / label_ids.size(0)
        return (column_selection_loss, )
      else:
        raise NotImplementedError("Unknown task {}".format(task))

    else:
      task = kwargs.pop("task")

      if task == "mlm":
        label_eos_id = kwargs.pop("label_eos_id")
        label_bos_id = kwargs.pop("label_bos_id")
        label_padding_id = kwargs.pop("label_padding_id")
        generated_ids = self.bert.generate(
          input_ids=input_ids,
          attention_mask=attention_mask,
          num_beams=3,
          max_length=input_ids.size(1) + 5,
          length_penalty=2.0,
          early_stopping=True,
          use_cache=True,
          decoder_start_token_id=label_bos_id,
          eos_token_id=label_eos_id,
          pad_token_id=label_padding_id
        )

        output_ids = kwargs.pop('labels')
        y_ids = output_ids[:, :-1].contiguous()
        lm_labels = output_ids[:, 1:].clone()
        lm_labels[output_ids[:, 1:] == pad_token_id] = -100

        outputs = self.bert(input_ids,
                            attention_mask=attention_mask, decoder_input_ids=y_ids, lm_labels=lm_labels, )

        return (outputs[0].detach(), generated_ids)
      elif task == "col_pred":
        label_ids = kwargs.pop("labels")
        column_spans = kwargs.pop("column_spans")
        column_selection_prob = self.column_prediction(input_ids, attention_mask, column_spans)

        generated_ids = (column_selection_prob.squeeze(-1) > 0.5).long()
        generated_ids[column_spans[:,:,0]==0] = -100

        label_mask = column_spans.view(-1, 2)[:, 0] > 0

        column_selection_loss = F.binary_cross_entropy(column_selection_prob.view(-1)[label_mask],
                                                       label_ids.view(-1)[label_mask].float(),
                                                       reduction="sum") / label_ids.size(0)
        return (column_selection_loss.detach(), generated_ids)


      else:
        raise NotImplementedError()
  def save_pretrained(self, save_directory):
    """ Save a model and its configuration file to a directory, so that it
        can be re-loaded using the `:func:`~transformers.PreTrainedModel.from_pretrained`` class method.

        Arguments:
            save_directory: directory to which to save.
    """
    assert os.path.isdir(
      save_directory
    ), "Saving path should be a directory where the model and configuration can be saved"

    # Only save the model itself if we are using distributed training
    model_to_save = self.module if hasattr(self, "module") else self

    # Attach architecture to the config
    # model_to_save.config.architectures = [model_to_save.__class__.__name__]

    # If we save using the predefined names, we can load using `from_pretrained`
    output_model_file = os.path.join(save_directory, WEIGHTS_NAME)

    torch.save(model_to_save.state_dict(), output_model_file)

    logger.info("Model weights saved in {}".format(output_model_file))