File size: 10,706 Bytes
d7332b4
 
 
 
6400d8b
 
d7332b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch, random, itertools, tqdm
import numpy as np
from torch import nn
from torch.utils.data import DataLoader
from .util import mean_pooling, read_corpus, CEFRDataset, convert_numeral_to_six_levels
from .model_base import LevelEstimaterBase


class LevelEstimaterClassification(LevelEstimaterBase):
    def __init__(self, pretrained_model, problem_type, with_ib, with_loss_weight,
                 attach_wlv, num_labels,
                 word_num_labels, alpha,
                 ib_beta,
                 batch_size,
                 learning_rate,
                 warmup,
                 lm_layer, corpus_path=None, test_corpus_path=None,):
        super().__init__(corpus_path, test_corpus_path, pretrained_model, with_ib, attach_wlv, num_labels,
                         word_num_labels, alpha,
                         batch_size,
                         learning_rate, warmup, lm_layer)
        self.save_hyperparameters()

        self.problem_type = problem_type
        self.with_loss_weight = with_loss_weight
        self.ib_beta = ib_beta
        self.dropout = nn.Dropout(0.1)

        if self.problem_type == "regression":
            self.slv_classifier = nn.Linear(self.lm.config.hidden_size, 1)
            self.loss_fct = nn.MSELoss()
        else:
            self.slv_classifier = nn.Linear(self.lm.config.hidden_size, self.CEFR_lvs)
            if self.with_loss_weight and corpus_path is not None:
                train_sentlv_weights = self.precompute_loss_weights()
                self.loss_fct = nn.CrossEntropyLoss(weight=train_sentlv_weights)
            else:
                self.loss_fct = nn.CrossEntropyLoss()

    def forward(self, inputs):
        # in lightning, forward defines the prediction/inference actions
        outputs, information_loss = self.encode(inputs)
        outputs = mean_pooling(outputs, attention_mask=inputs['attention_mask'])
        logits = self.slv_classifier(self.dropout(outputs))

        if self.problem_type == "regression":
            predictions = convert_numeral_to_six_levels(logits.detach().clone().cpu().numpy())
        else:
            predictions = torch.argmax(torch.softmax(logits.detach().clone(), dim=1), dim=1, keepdim=True)

        loss = None
        if 'slabels_high' in inputs:
            if self.problem_type == "regression":
                labels = (inputs['slabels_high'] + inputs['slabels_low']) / 2
                cls_loss = self.loss_fct(logits.squeeze(), labels.squeeze())
            else:
                labels = self.get_gold_labels(predictions, inputs['slabels_low'].detach().clone(),
                                              inputs['slabels_high'].detach().clone())
                cls_loss = self.loss_fct(logits.view(-1, self.CEFR_lvs), labels.view(-1))

            loss = cls_loss
            logs = {"loss": cls_loss}

        predictions = predictions.cpu().numpy()

        return (loss, predictions, logs) if loss is not None else predictions

    def step(self, batch):
        loss, predictions, logs = self.forward(batch)
        return loss, logs

    def _shared_eval_step(self, batch):
        loss, predictions, logs = self.forward(batch)

        gold_labels_low = batch['slabels_low'].cpu().detach().clone().numpy()
        gold_labels_high = batch['slabels_high'].cpu().detach().clone().numpy()
        golds_predictions = {'gold_labels_low': gold_labels_low, 'gold_labels_high': gold_labels_high,
                             'pred_labels': predictions}

        return logs, golds_predictions

    def training_step(self, batch, batch_idx):
        loss, logs = self.step(batch)
        self.log_dict({f"train_{k}": v for k, v in logs.items()})
        return loss

    def validation_step(self, batch, batch_idx):
        logs, golds_predictions = self._shared_eval_step(batch)
        self.log_dict({f"val_{k}": v for k, v in logs.items()})
        return golds_predictions

    def validation_epoch_end(self, outputs):
        logs = self.evaluation(outputs)
        self.log_dict({f"val_{k}": v for k, v in logs.items()})

    def test_step(self, batch, batch_idx):
        logs, golds_predictions = self._shared_eval_step(batch)
        self.log_dict({f"test_{k}": v for k, v in logs.items()})
        return golds_predictions

    def test_epoch_end(self, outputs):
        logs = self.evaluation(outputs, test=True)
        self.log_dict({f"test_{k}": v for k, v in logs.items()})


class LevelEstimaterContrastive(LevelEstimaterBase):
    def __init__(self, corpus_path, test_corpus_path, pretrained_model, problem_type, with_ib, with_loss_weight,
                 attach_wlv, num_labels,
                 word_num_labels,
                 num_prototypes,
                 alpha,
                 ib_beta,
                 batch_size,
                 learning_rate,
                 warmup,
                 lm_layer):
        super().__init__(corpus_path, test_corpus_path, pretrained_model, with_ib, attach_wlv, num_labels,
                         word_num_labels, alpha,
                         batch_size,
                         learning_rate, warmup, lm_layer)
        self.save_hyperparameters()

        self.problem_type = problem_type
        self.num_prototypes = num_prototypes
        self.with_loss_weight = with_loss_weight
        self.ib_beta = ib_beta

        self.prototype = nn.Embedding(self.CEFR_lvs * self.num_prototypes, self.lm.config.hidden_size)
        # nn.init.xavier_normal_(self.prototype.weight)  # Xavier initialization
        # nn.init.orthogonal_(self.prototype.weight)  # Make prototype vectors orthogonal

        if self.with_loss_weight:
            loss_weights = self.precompute_loss_weights()
            self.loss_fct = nn.CrossEntropyLoss(weight=loss_weights)
        else:
            self.loss_fct = nn.CrossEntropyLoss()

    def forward(self, batch):
        # in lightning, forward defines the prediction/inference actions
        outputs, information_loss = self.encode(batch)
        outputs = mean_pooling(outputs, attention_mask=batch['attention_mask'])

        # positive: compute cosine similarity
        outputs = torch.nn.functional.normalize(outputs)
        positive_prototypes = torch.nn.functional.normalize(self.prototype.weight)
        logits = torch.mm(outputs, positive_prototypes.T)
        logits = logits.reshape((-1, self.num_prototypes, self.CEFR_lvs))
        logits = logits.mean(dim=1)

        # prediction
        predictions = torch.argmax(torch.softmax(logits.detach().clone(), dim=1), dim=1, keepdim=True)

        loss = None
        if 'slabels_high' in batch:
            labels = self.get_gold_labels(predictions, batch['slabels_low'].detach().clone(),
                                          batch['slabels_high'].detach().clone())
            # cross-entropy loss
            cls_loss = self.loss_fct(logits.view(-1, self.CEFR_lvs), labels.view(-1))

            loss = cls_loss
            logs = {"loss": loss}

        predictions = predictions.cpu().numpy()

        return (loss, predictions, logs) if loss is not None else predictions

    def _shared_eval_step(self, batch):
        loss, predictions, logs = self.forward(batch)

        gold_labels_low = batch['slabels_low'].cpu().detach().clone().numpy()
        gold_labels_high = batch['slabels_high'].cpu().detach().clone().numpy()
        golds_predictions = {'gold_labels_low': gold_labels_low, 'gold_labels_high': gold_labels_high,
                             'pred_labels': predictions}

        return logs, golds_predictions

    def on_train_start(self) -> None:
        # Init with BERT embeddings
        epcilon = 1.0e-6
        higher_labels, lower_labels = [], []
        prototype_initials = torch.full((self.CEFR_lvs, self.lm.config.hidden_size), fill_value=epcilon).to(self.device)

        self.lm.eval()
        for batch in tqdm.tqdm(self.train_dataloader(), leave=False, desc='init prototypes'):
            higher_labels += batch['slabels_high'].squeeze().detach().clone().numpy().tolist()
            lower_labels += batch['slabels_low'].squeeze().detach().clone().numpy().tolist()
            batch = {k: v.cuda() for k, v in batch.items()}
            with torch.no_grad():
                outputs = self.lm(batch['input_ids'], attention_mask=batch['attention_mask'], output_hidden_states=True)
                outputs_mean = mean_pooling(outputs.hidden_states[self.lm_layer],
                                            attention_mask=batch['attention_mask'])
            for lv in range(self.CEFR_lvs):
                prototype_initials[lv] += outputs_mean[
                    (batch['slabels_low'].squeeze() == lv) | (batch['slabels_high'].squeeze() == lv)].sum(0)
        if not self.with_ib:
            self.lm.train()

        higher_labels = torch.tensor(higher_labels)
        lower_labels = torch.tensor(lower_labels)
        for lv in range(self.CEFR_lvs):
            denom = torch.count_nonzero((higher_labels == lv) | (lower_labels == lv)) + epcilon
            prototype_initials[lv] = prototype_initials[lv] / denom

        var = torch.var(prototype_initials).item() * 0.05 # Add Gaussian noize with 5% variance of the original tensor
        # prototype_initials = torch.repeat_interleave(prototype_initials, self.num_prototypes, dim=0)
        prototype_initials = prototype_initials.repeat(self.num_prototypes, 1)
        noise = (var ** 0.5) * torch.randn(prototype_initials.size()).to(self.device)
        prototype_initials = prototype_initials + noise  # Add Gaussian noize
        self.prototype.weight = nn.Parameter(prototype_initials)
        nn.init.orthogonal_(self.prototype.weight)  # Make prototype vectors orthogonal

        # # Init with Xavier
        # nn.init.xavier_normal_(self.prototype.weight)  # Xavier initialization

    def training_step(self, batch, batch_idx):
        loss, predictions, logs = self.forward(batch)
        self.log_dict({f"train_{k}": v for k, v in logs.items()})
        return loss

    def validation_step(self, batch, batch_idx):
        logs, golds_predictions = self._shared_eval_step(batch)
        self.log_dict({f"val_{k}": v for k, v in logs.items()})
        return golds_predictions

    def validation_epoch_end(self, outputs):
        logs = self.evaluation(outputs)
        self.log_dict({f"val_{k}": v for k, v in logs.items()})

    def test_step(self, batch, batch_idx):
        logs, golds_predictions = self._shared_eval_step(batch)
        self.log_dict({f"test_{k}": v for k, v in logs.items()})
        return golds_predictions

    def test_epoch_end(self, outputs):
        logs = self.evaluation(outputs, test=True)
        self.log_dict({f"test_{k}": v for k, v in logs.items()})