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 | @register_model
class PLMASFModel(BaseModel):
def __init__(
self,
model_args: PLMASfArgs,
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.add_question = data_args.task == PredMode.RC
self.fast_tokenizer = AutoTokenizer.from_pretrained(self.backbone)
self.pad_token_id = self.fast_tokenizer.pad_token_id
self.sep_token_id = self.fast_tokenizer.sep_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.classifier_head = nn.Linear(
self.model_args.lstm_hidden_size * 2, self.num_classes
) # *2 for bidirectional
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 feature index mappings for safe access
self.fixation_feature_indices = {
name: idx for idx, name in enumerate(self.model_args.fixation_features)
}
self.eye_feature_indices = {
name: idx for idx, name in enumerate(self.model_args.eye_features)
}
# Calculate number of additional features dynamically
# Based on organise_sp_fixation_features:
# features_by_sp_idx: 8 features (from fixation_features)
# features_by_word_idx: 2 features (from eye_features)
self.num_features_by_sp_idx = 8
self.num_features_by_word_idx = 2
self.num_additional_fixations_features = (
self.num_features_by_sp_idx + self.num_features_by_word_idx
)
# project bert_dim to bert_dim+num_additional_fixations_features
self.projection_layer = nn.Sequential(
nn.Linear(
self.bert_dim, self.bert_dim + self.num_additional_fixations_features
)
)
# create fse_lstm
self.fse_lstm = nn.LSTM(
input_size=self.bert_dim + self.num_additional_fixations_features,
hidden_size=self.model_args.lstm_hidden_size,
num_layers=self.model_args.lstm_num_layers,
batch_first=True,
bidirectional=True,
dropout=self.model_args.lstm_dropout,
)
self.train()
self.save_hyperparameters()
def organise_sp_fixation_features(
self, fixation_features: torch.Tensor, ia_features: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor]:
"""
Organize scanpath and fixation features for the model.
Args:
fixation_features: Tensor of fixation features indexed by scanpath position
ia_features: Tensor of interest area features indexed by word position
Returns:
features_by_sp_idx: Features indexed by scanpath position (8 features)
features_by_word_idx: Features indexed by word position (2 features)
"""
# Create safe feature getters
def get_fix_feat(name: str) -> torch.Tensor:
if name not in self.fixation_feature_indices:
raise ValueError(
f"Feature '{name}' not found in fixation_features. "
f'Available: {list(self.fixation_feature_indices.keys())}'
)
return fixation_features[..., self.fixation_feature_indices[name]]
def get_eye_feat(name: str) -> torch.Tensor:
if name not in self.eye_feature_indices:
raise ValueError(
f"Feature '{name}' not found in eye_features. "
f'Available: {list(self.eye_feature_indices.keys())}'
)
return ia_features[..., self.eye_feature_indices[name]]
features_by_sp_idx = []
features_by_word_idx = []
# Following the PLMAS-F paper:
# 1. Horizontal position of the fixation (CURRENT_FIX_Y) - fixation_features
features_by_sp_idx.append(get_fix_feat('CURRENT_FIX_Y'))
# 2. Total gaze duration (sum of all fixations on the word) (IA_DWELL_TIME) - ia_features
features_by_word_idx.append(get_eye_feat('IA_DWELL_TIME'))
# 3. Landing position of first fixation within the word (IA_FIRST_RUN_LANDING_POSITION) - ia_features
# features_by_word_idx.append(ia_features[..., 1]) #! removed because of nans
# 4. Landing position of last fixation within the word (IA_LAST_RUN_LANDING_POSITION) - ia_features
# features_by_word_idx.append(ia_features[..., 2]) #! removed because of nans
# 5. Duration of first fixation (IA_FIRST_FIXATION_DURATION) - ia_features
features_by_word_idx.append(get_eye_feat('IA_FIRST_FIXATION_DURATION'))
# 6. Duration of outgoing saccade (NEXT_SAC_DURATION) - fixation_features
features_by_sp_idx.append(get_fix_feat('NEXT_SAC_DURATION'))
# 7. Horizontal distance of outgoing saccade (NEXT_SAC_END_X - NEXT_SAC_START_X) - fixation_features
features_by_sp_idx.append(
get_fix_feat('NEXT_SAC_END_X') - get_fix_feat('NEXT_SAC_START_X')
)
# 8. Vertical distance of outgoing saccade (NEXT_SAC_END_Y - NEXT_SAC_START_Y) - fixation_features
features_by_sp_idx.append(
get_fix_feat('NEXT_SAC_END_Y') - get_fix_feat('NEXT_SAC_START_Y')
)
# 9. Total distance of outgoing saccade (pitagoras of 7 and 8) - fixation_features
features_by_sp_idx.append(
torch.sqrt(
torch.pow(
get_fix_feat('NEXT_SAC_END_X') - get_fix_feat('NEXT_SAC_START_X'), 2
)
+ torch.pow(
get_fix_feat('NEXT_SAC_END_Y') - get_fix_feat('NEXT_SAC_START_Y'), 2
)
)
)
# 10. Duration of incoming saccade (NEXT_SAC_DURATION of previous fixation) - fixation_features
## v[i] = v[i-1], v[0] = 0
tmp = torch.zeros_like(get_fix_feat('NEXT_SAC_DURATION')).to(
fixation_features.device
)
tmp[..., 1:] = get_fix_feat('NEXT_SAC_DURATION')[..., :-1]
features_by_sp_idx.append(tmp)
# 11. Horizontal distance of incoming saccade (NEXT_SAC_END_X - NEXT_SAC_START_X of previous) - fixation_features
prev_sac_end_x = torch.zeros_like(get_fix_feat('NEXT_SAC_END_X')).to(
fixation_features.device
)
prev_sac_end_x[..., 1:] = get_fix_feat('NEXT_SAC_END_X')[..., :-1]
prev_sac_start_x = torch.zeros_like(get_fix_feat('NEXT_SAC_START_X')).to(
fixation_features.device
)
prev_sac_start_x[..., 1:] = get_fix_feat('NEXT_SAC_START_X')[..., :-1]
features_by_sp_idx.append(prev_sac_end_x - prev_sac_start_x)
# 12. Vertical distance of incoming saccade (NEXT_SAC_END_Y - NEXT_SAC_START_Y of previous) - fixation_features
prev_sac_end_y = torch.zeros_like(get_fix_feat('NEXT_SAC_END_Y')).to(
fixation_features.device
)
prev_sac_end_y[..., 1:] = get_fix_feat('NEXT_SAC_END_Y')[..., :-1]
prev_sac_start_y = torch.zeros_like(get_fix_feat('NEXT_SAC_START_Y')).to(
fixation_features.device
)
prev_sac_start_y[..., 1:] = get_fix_feat('NEXT_SAC_START_Y')[..., :-1]
features_by_sp_idx.append(prev_sac_end_y - prev_sac_start_y)
features_by_sp_idx_tensor = torch.stack(features_by_sp_idx, dim=2)
features_by_word_idx_tensor = torch.stack(features_by_word_idx, dim=2)
return features_by_sp_idx_tensor, features_by_word_idx_tensor
def split_context_embeds(
self,
encoded_word_seq: torch.Tensor,
input_ids: torch.Tensor,
) -> tuple[
Union[torch.Tensor, None], Union[torch.Tensor, None], Union[torch.Tensor, None]
]:
# Find the positions of the separator tokens
sep_positions = (
(input_ids == self.sep_token_id).nonzero(as_tuple=True)[0].cpu().numpy()
)
# Calculate the sizes of the splits
split_sizes = np.diff(a=sep_positions, prepend=0, append=input_ids.size(dim=0))
assert split_sizes.sum() == input_ids.size(dim=0), (
f'split_sizes.sum(): {split_sizes.sum()}'
)
if self.add_question:
assert split_sizes.size == 4, (
f'split_sizes.size: {split_sizes.size} but expected 4'
)
# Split the encoded_word_seq tensor at the separator positions
p, _, q, _ = torch.split(
tensor=encoded_word_seq,
split_size_or_sections=split_sizes.tolist(),
dim=0,
)
a = torch.zeros_like(q)
else:
# only paragraph
assert split_sizes.size == 4, (
f'split_sizes.size: {split_sizes.size} but expected 4'
)
# Split the encoded_word_seq tensor at the separator positions
p, _, _, _ = torch.split(
tensor=encoded_word_seq,
split_size_or_sections=split_sizes.tolist(),
dim=0,
)
# q = torch.zeros_like(p)
# q = q[1:, :].mean(dim=0)
# a = torch.zeros_like(p)
q = None
a = None
return p, q, a
def split_context_embds_batched(self, encoded_word_seq, input_ids):
p_embds, q_embds, a_embeds, p_masks, q_masks, a_masks = (
None,
None,
None,
None,
None,
None,
) # initialize so that the return variables are defined
p_embds_batches, q_embds_batches, a_embeds_batches = [], [], []
# Process each batch separately
for ewsb, iib in zip(encoded_word_seq, input_ids):
p_embds, q_embds, a_embeds = self.split_context_embeds(
encoded_word_seq=ewsb, input_ids=iib
)
p_embds_batches.append(p_embds)
q_embds_batches.append(q_embds)
a_embeds_batches.append(a_embeds)
# pad the embeddings to the maximum sequence length of each list
p_max_len = self.actual_max_needed_len
p_masks = torch.stack(
[
torch.cat([torch.ones(p.shape[0]), torch.zeros(p_max_len - p.shape[0])])
for p in p_embds_batches
],
dim=0,
)
p_embds_batches = [
F.pad(p, (0, 0, 0, p_max_len - p.shape[0])) for p in p_embds_batches
]
p_embds = torch.stack(p_embds_batches, dim=0)
if self.add_question:
q_max_len = max([q.shape[0] for q in q_embds_batches])
q_masks = torch.stack(
[
torch.cat(
[torch.ones(q.shape[0]), torch.zeros(q_max_len - q.shape[0])]
)
for q in q_embds_batches
],
dim=0,
)
q_embds_batches = [
F.pad(q, (0, 0, 0, q_max_len - q.shape[0])) for q in q_embds_batches
]
q_embds = torch.stack(q_embds_batches, dim=0)
return p_embds, q_embds, q_embds, p_masks, q_masks, a_masks
def forward(
self,
input_ids,
input_masks,
p_input_ids,
scanpath,
fixation_features,
scanpath_pads,
eyes,
):
assert (
input_ids[:, 0].sum().item() == 0
) # The CLS token is always present first (and 0 in roberta)
scanpath, fixation_features = trim_scanpath_and_fixation_features(
scanpath=scanpath,
fixation_features=fixation_features,
scanpath_pads=scanpath_pads,
max_scanpath_length=self.max_scanpath_length,
)
features_by_sp_idx, features_by_word_idx = self.organise_sp_fixation_features(
fixation_features=fixation_features,
ia_features=eyes,
)
# scanpath masks --------------------------
scanpath_masks = torch.ones_like(scanpath)
scanpath_masks[scanpath == self.pad_token_id] = 0
# -----------------------------------------
# get the decoded texts from the input_ids
p_input_ids_decoded_txts = self.fast_tokenizer.batch_decode(
p_input_ids, return_tensors='pt'
) # previously decoded_to_txt_input_ids
# -----------------------------------------
paragraph_token_to_word_idxs_in_p_text = (
plm_as_model.align_word_ids_with_input_ids(
tokenizer=self.fast_tokenizer,
input_ids=p_input_ids,
decoded_to_txt_input_ids=p_input_ids_decoded_txts,
)
) # previously word_ids_sn
# in the decoded texts, space between <pad><pad>, <pad><s>, etc.
p_input_ids_decoded_txts = list(
map(
lambda x: x.replace('<', ' <').split(' ')[
1:
], # <s> is the first token so, when doing the replace, we get " <s>", so we split by space and take the second element
p_input_ids_decoded_txts,
)
)
scanpath_token_to_word_idxs_in_sp_text, scanpath_input_ids = (
plm_as_model.calc_sp_word_input_ids(
input_ids=p_input_ids,
decoded_to_txt_input_ids=p_input_ids_decoded_txts,
roberta_tokenizer_prefix_space=self.fast_tokenizer_prefix,
scanpath=scanpath,
)
) # previously word_ids_sp, sp_input_ids
# 1. encode the whole sequence (input_ids)
# 2. then split the embeddings into paragraph, question, and answer embeddings
# 3. reorder the paragraph embeddings according to the scanpath
# 4. combine the paragraph embeddings with features_by_sp_idx, features_by_word_idx
# 5. combine parts into a sequence of embeddings to feed the lstm
# (1) encode the whole sequence
with torch.no_grad():
outputs = self.bert_encoder(input_ids=input_ids, attention_mask=input_masks)
# Make the embedding of the <pad> token to be zeros
outputs.last_hidden_state[input_ids == self.pad_token_id] = 0
encoded_input_ids = outputs.last_hidden_state
# (2) split the embeddings into paragraph, question, and answer embeddings
p_embds, q_embds, a_embeds, p_masks, q_masks, a_masks = (
self.split_context_embds_batched(encoded_input_ids, input_ids)
) # embds shape: [batch, seq_len, emb_dim]
# (3) reorder the paragraph embeddings according to the scanpath
# Pool bert subword to word level for english corpus
merged_p_emb, p_mask = plm_as_model.pool_subword_to_word(
p_embds,
paragraph_token_to_word_idxs_in_p_text,
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
p_embds_sp_order = merged_p_emb[
batch_index, scanpath_add1
] # [batch, max_sp_length, emb_dim]
# note that p_embds_sp_order should never contain the embeddings of the <CLS> token
# (4) combine the paragraph embeddings with features_by_sp_idx, features_by_word_idx
if features_by_sp_idx is not None:
p_embds_sp_order = torch.cat([p_embds_sp_order, features_by_sp_idx], dim=2)
if features_by_word_idx is not None:
p_embds_sp_order = torch.cat(
[p_embds_sp_order, features_by_word_idx[batch_index, scanpath]], dim=2
) # we don't need scanpath_add1 here because no <s> token in the beginning
# (5) combine parts into a sequence of embeddings to feed the lstm
if (
self.prediction_mode
in BINARY_PARAGRAPH_ONLY_TASKS
+ BINARY_P_AND_Q_TASKS
+ REGRESSION_PARAGRAPH_ONLY_TASKS
):
final_seq_embds = p_embds_sp_order
final_seq_lens = (
(scanpath != SCANPATH_PADDING_VAL).sum(dim=1).clone().detach()
)
else:
raise NotImplementedError(
f'Not implemented for prediction_mode: {self.prediction_mode}'
)
# pass through the LSTM layer
sorted_lengths, indices = torch.sort(final_seq_lens, descending=True)
final_seq_embds = final_seq_embds[
indices
] # reorder sequences according to the descending order of the lengths
# Pass the entire sequence through the LSTM layer
packed_final_seq_embds = nn.utils.rnn.pack_padded_sequence(
input=final_seq_embds,
lengths=sorted_lengths.to('cpu'),
batch_first=True,
enforce_sorted=True,
)
packed_output, (hn, cn) = self.fse_lstm(packed_final_seq_embds)
unpacked_output = nn.utils.rnn.unpack_sequence(packed_output)
# unpacked output is a list of batch_size tensors,
# each tensor is shaped [seq_len, no.directions*hidden_size]
# create a tensor of shape [batch_size, no.directions*hidden_size]
# by taking the mean of each element in the list
lstm_mean_hidden = torch.stack([torch.mean(t, dim=0) for t in unpacked_output])
# reorder the hidden states to the original order
lstm_mean_hidden = lstm_mean_hidden[
torch.argsort(indices)
] # Tested. Reorders correctly
logits = self.classifier_head(lstm_mean_hidden)
return logits, lstm_mean_hidden
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'
assert batch_data.p_input_ids is not None, 'p_input_ids not in batch_dict'
# -----------------------------------------
logits, _ = self.forward(
input_ids=batch_data.input_ids,
input_masks=batch_data.input_masks,
p_input_ids=batch_data.p_input_ids,
scanpath=batch_data.scanpath,
fixation_features=batch_data.fixation_features,
scanpath_pads=batch_data.scanpath_pads,
eyes=batch_data.eyes,
)
labels = batch_data.labels
if logits.ndim == 1:
logits = logits.unsqueeze(0)
loss = self.loss(logits, labels)
return labels, loss, logits.squeeze()
|