Skip to content

plm_as_model

PLM-AS model for fixation sequence encoding and classification.

PLMASModel

Bases: BaseModel

Source code in src/models/plm_as_model.py 
 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
@register_model
class PLMASModel(BaseModel):
    def __init__(
        self,
        model_args: PLMASArgs,
        trainer_args: TrainerDL,
        data_args: DataArgs,
    ):
        super().__init__(
            model_args=model_args, trainer_args=trainer_args, data_args=data_args
        )

        self.model_args = model_args
        self.backbone = model_args.backbone
        self.freeze_bert = model_args.freeze

        self.fast_tokenizer = AutoTokenizer.from_pretrained(self.backbone)
        self.pad_token_id = self.fast_tokenizer.pad_token_id
        # Cache tokenizer with add_prefix_space for scanpath processing
        self.fast_tokenizer_prefix = AutoTokenizer.from_pretrained(
            self.backbone, add_prefix_space=True
        )

        # ? self.preorder = False

        self.classifier_head = nn.Linear(model_args.text_dim, self.num_classes)
        self.bert_dim = model_args.text_dim

        encoder_config = AutoConfig.from_pretrained(self.backbone)
        encoder_config.output_hidden_states = True
        # initiate Bert with pre-trained weights
        print('keeping Bert with pre-trained weights')
        self.bert_encoder = AutoModel.from_pretrained(
            self.backbone, config=encoder_config
        )  # type: ignore

        # freeze the parameters in Bert model
        if self.freeze_bert:
            for param in self.bert_encoder.parameters():  # type: ignore
                param.requires_grad = False

        # create fse_lstm
        self.fse_lstm = nn.GRU(
            input_size=self.bert_dim,
            hidden_size=self.bert_dim,
            num_layers=self.model_args.lstm_num_layers,
            batch_first=True,
            bidirectional=False,
            dropout=self.model_args.lstm_dropout,
        )

        self.train()
        self.save_hyperparameters()

    def fixation_sequence_encoder(
        self,
        sn_emd,
        sn_mask,
        word_ids_sn,
        sn_word_len,
        sp_emd,
        sp_pos,
        sp_fix_dur,
        sp_landing_pos,
        word_ids_sp,
        scanpath,
        features_by_sp_idx=None,
        features_by_word_idx=None,
    ):
        """A LSTM based encoder for the fixation sequence (scanpath)
        Args:
            sp_emd (torch.Tensor): A tensor containing the text input_ids ordered according to the scanpath
            sp_pos (torch.Tensor): The word index of each fixation in the scanpath (the word the fixation is on)
            sp_fix_dur (torch.Tensor): The total fixation duration of each word in the scanpath (fixation)
            sp_landing_pos (torch.Tensor): The landing position of each word in the scanpath (fixation)
            sp_mask (torch.Tensor): The mask for the scanpath
            word_ids_sp (torch.Tensor): The word index of each input_id in the scanpath
            scanpath (torch.Tensor): The scanpath tensor

        """

        # used for computing sp_merged_word_mask
        # x = self.bert_encoder.embeddings.word_embeddings(sp_emd)
        # x[sp_emd == self.pad_token_id] = 0
        # # Pool bert subword to word level for English corpus
        # _, sp_merged_word_mask = self.pool_subword_to_word(
        #     x, word_ids_sp, target='sp', pool_method='sum'
        # )

        with torch.no_grad():
            outputs = self.bert_encoder(input_ids=sn_emd, attention_mask=sn_mask)
        #  Make the embedding of the <pad> token to be zeros
        outputs.last_hidden_state[sn_emd == self.pad_token_id] = 0
        # Pool bert subword to word level for english corpus
        merged_word_emb, sn_mask_word = pool_subword_to_word(
            outputs.last_hidden_state,
            word_ids_sn,
            target='sn',
            max_seq_len=self.actual_max_needed_len,
            bert_dim=self.bert_dim,
            pool_method='sum',
        )
        batch_index = torch.arange(scanpath.shape[0]).unsqueeze(1).expand_as(scanpath)
        scanpath_add1 = scanpath.clone()
        scanpath_add1[scanpath != SCANPATH_PADDING_VAL] += SCANPATH_PADDING_VAL
        x = merged_word_emb[
            batch_index, scanpath_add1
        ]  # [batch, max_sp_length, emb_dim], word_emb_sn

        if features_by_sp_idx is not None:
            x = torch.cat([x, features_by_sp_idx], dim=2)
        if features_by_word_idx is not None:
            x = torch.cat(
                [x, features_by_word_idx[batch_index, scanpath]], dim=2
            )  # we don't need scanpath_add1 here because no <s> token in the beginning

        # pass through the LSTM layer
        sorted_lengths, indices = torch.sort(
            (scanpath != SCANPATH_PADDING_VAL).sum(dim=1), descending=True
        )
        x = x[
            indices
        ]  # reorder sequences according to the descending order of the lengths

        # Pass the entire sequence through the LSTM layer
        packed_x = nn.utils.rnn.pack_padded_sequence(
            input=x,
            lengths=sorted_lengths.to('cpu'),
            batch_first=True,
            enforce_sorted=True,
        )

        # set h0 as [CLS] token embedding of each sequence
        h0 = outputs.last_hidden_state[indices, 0]

        # fit h0 and c0 to the shape of the LSTM
        # h0 is [batch, hidden_size] where h0[i] is the initial hidden state for the sequence i
        h0 = h0.unsqueeze(0).repeat(self.fse_lstm.num_layers, 1, 1)
        c0 = torch.zeros_like(h0)
        # ensure contiguous
        h0 = h0.contiguous()
        c0 = c0.contiguous()

        # * c0 isn't used when using GRU

        packed_output, ht = self.fse_lstm(packed_x, h0)
        lstm_last_hidden = ht[-1].squeeze(
            1
        )  # Take the hidden state of the last LSTM layer.

        # reorder the hidden states to the original order
        lstm_last_hidden = lstm_last_hidden[
            torch.argsort(indices)
        ]  # Tested. Reorders correctly

        # Clear unused tensors to free memory
        del outputs, packed_x, packed_output, ht, x, merged_word_emb

        return lstm_last_hidden  # Take the hidden state of the 8th LSTM layer.

    def forward(
        self,
        sn_emd,
        sn_mask,
        word_ids_sn,
        sn_word_len,
        sp_emd,  # (Batch, Maximum length of the scanpath in TOKENS + 1)
        sp_pos,  # (Batch, Scanpath_length + 1) The +1 is for the <s> token in the beginning
        sp_fix_dur,  # (Batch, Scanpath_length + 1) The +1 is for the <s> token in the beginning
        sp_landing_pos,  # (Batch, Scanpath_length + 1) The +1 is for the <s> token in the beginning
        word_ids_sp,  # (Batch, Maximum length of the scanpath in TOKENS + 1)
        scanpath,  # (Batch, Maximum length of the scanpath in WORDS)
        features_by_sp_idx=None,
        features_by_word_idx=None,
    ):
        assert (
            sn_emd[:, 0].sum().item() == 0
        )  # The CLS token is always present first (and 0 in roberta)

        fse_output = self.fixation_sequence_encoder(
            sn_emd=sn_emd,
            sn_mask=sn_mask,
            word_ids_sn=word_ids_sn,
            sn_word_len=sn_word_len,
            sp_emd=sp_emd,
            sp_pos=sp_pos,
            sp_fix_dur=sp_fix_dur,
            sp_landing_pos=sp_landing_pos,
            word_ids_sp=word_ids_sp,
            scanpath=scanpath,
            features_by_sp_idx=features_by_sp_idx,
            features_by_word_idx=features_by_word_idx,
        )  # [batch, step, dec_o_dim]

        pred = self.classifier_head(fse_output)

        return pred, fse_output

    def shared_step(
        self,
        batch: list,
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """

        Args:
            batch (tuple): _description_

        Returns:
            tuple[torch.Tensor, torch.Tensor, torch.Tensor]: _description_

        Notes:
            - in input ids: 0 is for CLS, 2 is for SEP, 1 is for PAD
        """
        batch_data = self.unpack_batch(batch)
        assert batch_data.input_ids is not None, 'input_ids not in batch_dict'
        assert batch_data.input_masks is not None, 'input_masks not in batch_dict'
        assert batch_data.scanpath is not None, 'scanpath not in batch_dict'
        assert batch_data.fixation_features is not None, 'eyes_tensor not in batch_dict'
        assert batch_data.scanpath_pads is not None, 'scanpath_pads not in batch_dict'
        assert batch_data.eyes is not None, 'eye not in batch_dict'

        shortest_scanpath_pad = batch_data.scanpath_pads.min()
        longest_batch_scanpath: int = int(
            self.max_scanpath_length - shortest_scanpath_pad
        )

        scanpath = batch_data.scanpath[..., :longest_batch_scanpath]
        fixation_features = batch_data.fixation_features[
            ..., :longest_batch_scanpath, :
        ]

        # scanpath masks
        sp_masks = torch.ones_like(scanpath)
        sp_masks[scanpath == self.pad_token_id] = 0

        decoded_to_txt_input_ids = self.fast_tokenizer.batch_decode(
            batch_data.input_ids, return_tensors='pt'
        )

        word_ids_sn = align_word_ids_with_input_ids(
            tokenizer=self.fast_tokenizer,
            input_ids=batch_data.input_ids,
            decoded_to_txt_input_ids=decoded_to_txt_input_ids,
        )

        # in the decoded texts, space between <pad><pad>, <pad><s>, etc.
        decoded_to_txt_input_ids = list(
            map(
                lambda x: x.replace('<', ' <').split(' ')[1:],
                decoded_to_txt_input_ids,
            )
        )

        sn_word_len = get_sn_word_lens(
            input_ids=batch_data.input_ids,
            decoded_to_txt_input_ids=decoded_to_txt_input_ids,
        )

        word_ids_sp, sp_input_ids = calc_sp_word_input_ids(
            input_ids=batch_data.input_ids,
            decoded_to_txt_input_ids=decoded_to_txt_input_ids,
            roberta_tokenizer_prefix_space=self.fast_tokenizer_prefix,
            scanpath=scanpath,
        )

        sp_pos, sp_fix_dur, sp_landing_pos = eyettention_legacy_code(
            scanpath=scanpath,
            fixation_features=fixation_features,
        )

        sn_embed = batch_data.input_ids
        sn_mask = batch_data.input_masks
        # if the second dimension of the scanpath is more than the maximum context length (of self.bert_encoder), cut it and notify
        bert_encoder_max_len = self.bert_encoder.config.max_position_embeddings
        if sp_input_ids.shape[1] > bert_encoder_max_len - 1:
            # print(
            #     f'Text length is more than the maximum context length of the model ({bert_encoder_max_len}). Cutting from the BEGINNING of the text to max length.'
            # )
            sn_embed = sn_embed[:, : bert_encoder_max_len - 1]
            sn_mask = sn_mask[:, : bert_encoder_max_len - 1]

        logits, _ = self(
            sn_emd=sn_embed,
            sn_mask=sn_mask,
            word_ids_sn=word_ids_sn,
            sn_word_len=sn_word_len,
            sp_emd=sp_input_ids,
            sp_pos=sp_pos,
            sp_fix_dur=sp_fix_dur,
            sp_landing_pos=sp_landing_pos,
            word_ids_sp=word_ids_sp,
            scanpath=scanpath,
        )

        labels = batch_data.labels

        if logits.ndim == 1:
            logits = logits.unsqueeze(0)
        loss = self.loss(logits, labels)

        # Clear intermediate tensors to prevent memory buildup
        del scanpath, fixation_features, sp_input_ids, word_ids_sp
        del sp_pos, sp_fix_dur, sp_landing_pos, sn_embed, sn_mask
        del word_ids_sn, sn_word_len

        return labels, loss, logits.squeeze()

fixation_sequence_encoder(sn_emd, sn_mask, word_ids_sn, sn_word_len, sp_emd, sp_pos, sp_fix_dur, sp_landing_pos, word_ids_sp, scanpath, features_by_sp_idx=None, features_by_word_idx=None)

A LSTM based encoder for the fixation sequence (scanpath) Args: sp_emd (torch.Tensor): A tensor containing the text input_ids ordered according to the scanpath sp_pos (torch.Tensor): The word index of each fixation in the scanpath (the word the fixation is on) sp_fix_dur (torch.Tensor): The total fixation duration of each word in the scanpath (fixation) sp_landing_pos (torch.Tensor): The landing position of each word in the scanpath (fixation) sp_mask (torch.Tensor): The mask for the scanpath word_ids_sp (torch.Tensor): The word index of each input_id in the scanpath scanpath (torch.Tensor): The scanpath tensor

Source code in src/models/plm_as_model.py 
 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
def fixation_sequence_encoder(
    self,
    sn_emd,
    sn_mask,
    word_ids_sn,
    sn_word_len,
    sp_emd,
    sp_pos,
    sp_fix_dur,
    sp_landing_pos,
    word_ids_sp,
    scanpath,
    features_by_sp_idx=None,
    features_by_word_idx=None,
):
    """A LSTM based encoder for the fixation sequence (scanpath)
    Args:
        sp_emd (torch.Tensor): A tensor containing the text input_ids ordered according to the scanpath
        sp_pos (torch.Tensor): The word index of each fixation in the scanpath (the word the fixation is on)
        sp_fix_dur (torch.Tensor): The total fixation duration of each word in the scanpath (fixation)
        sp_landing_pos (torch.Tensor): The landing position of each word in the scanpath (fixation)
        sp_mask (torch.Tensor): The mask for the scanpath
        word_ids_sp (torch.Tensor): The word index of each input_id in the scanpath
        scanpath (torch.Tensor): The scanpath tensor

    """

    # used for computing sp_merged_word_mask
    # x = self.bert_encoder.embeddings.word_embeddings(sp_emd)
    # x[sp_emd == self.pad_token_id] = 0
    # # Pool bert subword to word level for English corpus
    # _, sp_merged_word_mask = self.pool_subword_to_word(
    #     x, word_ids_sp, target='sp', pool_method='sum'
    # )

    with torch.no_grad():
        outputs = self.bert_encoder(input_ids=sn_emd, attention_mask=sn_mask)
    #  Make the embedding of the <pad> token to be zeros
    outputs.last_hidden_state[sn_emd == self.pad_token_id] = 0
    # Pool bert subword to word level for english corpus
    merged_word_emb, sn_mask_word = pool_subword_to_word(
        outputs.last_hidden_state,
        word_ids_sn,
        target='sn',
        max_seq_len=self.actual_max_needed_len,
        bert_dim=self.bert_dim,
        pool_method='sum',
    )
    batch_index = torch.arange(scanpath.shape[0]).unsqueeze(1).expand_as(scanpath)
    scanpath_add1 = scanpath.clone()
    scanpath_add1[scanpath != SCANPATH_PADDING_VAL] += SCANPATH_PADDING_VAL
    x = merged_word_emb[
        batch_index, scanpath_add1
    ]  # [batch, max_sp_length, emb_dim], word_emb_sn

    if features_by_sp_idx is not None:
        x = torch.cat([x, features_by_sp_idx], dim=2)
    if features_by_word_idx is not None:
        x = torch.cat(
            [x, features_by_word_idx[batch_index, scanpath]], dim=2
        )  # we don't need scanpath_add1 here because no <s> token in the beginning

    # pass through the LSTM layer
    sorted_lengths, indices = torch.sort(
        (scanpath != SCANPATH_PADDING_VAL).sum(dim=1), descending=True
    )
    x = x[
        indices
    ]  # reorder sequences according to the descending order of the lengths

    # Pass the entire sequence through the LSTM layer
    packed_x = nn.utils.rnn.pack_padded_sequence(
        input=x,
        lengths=sorted_lengths.to('cpu'),
        batch_first=True,
        enforce_sorted=True,
    )

    # set h0 as [CLS] token embedding of each sequence
    h0 = outputs.last_hidden_state[indices, 0]

    # fit h0 and c0 to the shape of the LSTM
    # h0 is [batch, hidden_size] where h0[i] is the initial hidden state for the sequence i
    h0 = h0.unsqueeze(0).repeat(self.fse_lstm.num_layers, 1, 1)
    c0 = torch.zeros_like(h0)
    # ensure contiguous
    h0 = h0.contiguous()
    c0 = c0.contiguous()

    # * c0 isn't used when using GRU

    packed_output, ht = self.fse_lstm(packed_x, h0)
    lstm_last_hidden = ht[-1].squeeze(
        1
    )  # Take the hidden state of the last LSTM layer.

    # reorder the hidden states to the original order
    lstm_last_hidden = lstm_last_hidden[
        torch.argsort(indices)
    ]  # Tested. Reorders correctly

    # Clear unused tensors to free memory
    del outputs, packed_x, packed_output, ht, x, merged_word_emb

    return lstm_last_hidden  # Take the hidden state of the 8th LSTM layer.

shared_step(batch)

Parameters:

Name Type Description Default
batch tuple

description

 required

Returns:

Type Description
tuple[Tensor, Tensor, Tensor]

tuple[torch.Tensor, torch.Tensor, torch.Tensor]: description
Notes
  • in input ids: 0 is for CLS, 2 is for SEP, 1 is for PAD
Source code in src/models/plm_as_model.py 
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
def shared_step(
    self,
    batch: list,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """

    Args:
        batch (tuple): _description_

    Returns:
        tuple[torch.Tensor, torch.Tensor, torch.Tensor]: _description_

    Notes:
        - in input ids: 0 is for CLS, 2 is for SEP, 1 is for PAD
    """
    batch_data = self.unpack_batch(batch)
    assert batch_data.input_ids is not None, 'input_ids not in batch_dict'
    assert batch_data.input_masks is not None, 'input_masks not in batch_dict'
    assert batch_data.scanpath is not None, 'scanpath not in batch_dict'
    assert batch_data.fixation_features is not None, 'eyes_tensor not in batch_dict'
    assert batch_data.scanpath_pads is not None, 'scanpath_pads not in batch_dict'
    assert batch_data.eyes is not None, 'eye not in batch_dict'

    shortest_scanpath_pad = batch_data.scanpath_pads.min()
    longest_batch_scanpath: int = int(
        self.max_scanpath_length - shortest_scanpath_pad
    )

    scanpath = batch_data.scanpath[..., :longest_batch_scanpath]
    fixation_features = batch_data.fixation_features[
        ..., :longest_batch_scanpath, :
    ]

    # scanpath masks
    sp_masks = torch.ones_like(scanpath)
    sp_masks[scanpath == self.pad_token_id] = 0

    decoded_to_txt_input_ids = self.fast_tokenizer.batch_decode(
        batch_data.input_ids, return_tensors='pt'
    )

    word_ids_sn = align_word_ids_with_input_ids(
        tokenizer=self.fast_tokenizer,
        input_ids=batch_data.input_ids,
        decoded_to_txt_input_ids=decoded_to_txt_input_ids,
    )

    # in the decoded texts, space between <pad><pad>, <pad><s>, etc.
    decoded_to_txt_input_ids = list(
        map(
            lambda x: x.replace('<', ' <').split(' ')[1:],
            decoded_to_txt_input_ids,
        )
    )

    sn_word_len = get_sn_word_lens(
        input_ids=batch_data.input_ids,
        decoded_to_txt_input_ids=decoded_to_txt_input_ids,
    )

    word_ids_sp, sp_input_ids = calc_sp_word_input_ids(
        input_ids=batch_data.input_ids,
        decoded_to_txt_input_ids=decoded_to_txt_input_ids,
        roberta_tokenizer_prefix_space=self.fast_tokenizer_prefix,
        scanpath=scanpath,
    )

    sp_pos, sp_fix_dur, sp_landing_pos = eyettention_legacy_code(
        scanpath=scanpath,
        fixation_features=fixation_features,
    )

    sn_embed = batch_data.input_ids
    sn_mask = batch_data.input_masks
    # if the second dimension of the scanpath is more than the maximum context length (of self.bert_encoder), cut it and notify
    bert_encoder_max_len = self.bert_encoder.config.max_position_embeddings
    if sp_input_ids.shape[1] > bert_encoder_max_len - 1:
        # print(
        #     f'Text length is more than the maximum context length of the model ({bert_encoder_max_len}). Cutting from the BEGINNING of the text to max length.'
        # )
        sn_embed = sn_embed[:, : bert_encoder_max_len - 1]
        sn_mask = sn_mask[:, : bert_encoder_max_len - 1]

    logits, _ = self(
        sn_emd=sn_embed,
        sn_mask=sn_mask,
        word_ids_sn=word_ids_sn,
        sn_word_len=sn_word_len,
        sp_emd=sp_input_ids,
        sp_pos=sp_pos,
        sp_fix_dur=sp_fix_dur,
        sp_landing_pos=sp_landing_pos,
        word_ids_sp=word_ids_sp,
        scanpath=scanpath,
    )

    labels = batch_data.labels

    if logits.ndim == 1:
        logits = logits.unsqueeze(0)
    loss = self.loss(logits, labels)

    # Clear intermediate tensors to prevent memory buildup
    del scanpath, fixation_features, sp_input_ids, word_ids_sp
    del sp_pos, sp_fix_dur, sp_landing_pos, sn_embed, sn_mask
    del word_ids_sn, sn_word_len

    return labels, loss, logits.squeeze()

align_word_ids_with_input_ids(tokenizer, input_ids, decoded_to_txt_input_ids)

Returns a tensor of the same shape as input_ids, containing the word index of each input_id (token)

Source code in src/models/plm_as_model.py 
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
def align_word_ids_with_input_ids(
    tokenizer,
    input_ids: torch.Tensor,
    decoded_to_txt_input_ids: list,
):
    """
    Returns a tensor of the same shape as input_ids, containing the word index of each input_id (token)
    """
    word_ids_sn_lst = []
    retokenized_sn = tokenizer(
        decoded_to_txt_input_ids,
        return_tensors='pt',
    )
    for i in range(input_ids.shape[0]):
        word_ids_sn_lst.append(retokenized_sn.word_ids(i)[1:-1])

    word_ids_sn = torch.tensor(word_ids_sn_lst).to(input_ids.device)

    return word_ids_sn

calc_sp_word_input_ids(input_ids, decoded_to_txt_input_ids, roberta_tokenizer_prefix_space, scanpath)

This function calculates the word input ids for the scanpath

Returns:

Name Type Description
word_ids_sp Tensor

The word index of each input_id (token) in the scanpath text
sp_input_ids Tensor

The input ids of the scanpath text

Parameters:

Name Type Description Default
input_ids Tensor

The word sequence input ids. Tensor of (batch_size, max_text_length_in_tokens)

 required
decoded_to_txt_input_ids list

The decoded input ids. (list of lists of strings)

 required
roberta_tokenizer_prefix_space object

The tokenizer with add_prefix_space=True

 required
scanpath Tensor

A scanpath tensor containing the word indices in the scanpath order Tensor of (batch_size, max_scanpath_length_in_words)

 required
Source code in src/models/plm_as_model.py 
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
def calc_sp_word_input_ids(
    input_ids: torch.Tensor,
    decoded_to_txt_input_ids: List[List[str]],
    roberta_tokenizer_prefix_space: object,
    scanpath: torch.Tensor,
):
    """This function calculates the word input ids for the scanpath

    Returns:
        word_ids_sp (torch.Tensor): The word index of each input_id (token) in the scanpath text
        sp_input_ids (torch.Tensor): The input ids of the scanpath text

    Args:
        input_ids (torch.Tensor): The word sequence input ids.
                Tensor of (batch_size, max_text_length_in_tokens)
        decoded_to_txt_input_ids (list): The decoded input ids.
                (list of lists of strings)
        roberta_tokenizer_prefix_space: The tokenizer with add_prefix_space=True
        scanpath (torch.Tensor): A scanpath tensor containing the word indices in the scanpath order
                Tensor of (batch_size, max_scanpath_length_in_words)
    """
    SP_word_ids, SP_input_ids = [], []

    sp_tokens_strs = convert_positions_to_words_sp(
        scanpath=scanpath,
        decoded_to_txt_input_ids=decoded_to_txt_input_ids,
        roberta_tokenizer_prefix_space=roberta_tokenizer_prefix_space,
    )

    tokenized_SPs = roberta_tokenizer_prefix_space.batch_encode_plus(
        sp_tokens_strs,
        add_special_tokens=False,
        truncation=False,
        padding='longest',
        return_attention_mask=True,
        is_split_into_words=True,
    )
    for i in range(scanpath.shape[0]):
        encoded_sp = tokenized_SPs['input_ids'][i]
        word_ids_sp = tokenized_SPs.word_ids(i)  # -> Take the <sep> into account
        word_ids_sp = [
            val if val is not None else SCANPATH_PADDING_VAL for val in word_ids_sp
        ]

        SP_word_ids.append(word_ids_sp)
        SP_input_ids.append(encoded_sp)

    word_ids_sp = torch.tensor(SP_word_ids).to(input_ids.device)
    sp_input_ids = torch.tensor(SP_input_ids).to(input_ids.device)

    return word_ids_sp, sp_input_ids