Skip to content

base_roberta

base_roberta.py - Base class for MAG and RoBERTeye models. See 1. On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines: https://www.semanticscholar.org/reader/8b9d77d5e52a70af37451d3db3d32781b83ea054 for parameters

BaseMultiModalRoberta

Bases: BaseModel

Model for Multiple Choice Question Answering and question prediction tasks.

Parameters:

Name Type Description Default
model_args Roberteye | MAG | PostFusion

Model arguments.

 required
trainer_args TrainerDL

Trainer arguments.

 required
Source code in src/models/base_roberta.py 
 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
class BaseMultiModalRoberta(BaseModel):
    """
    Model for Multiple Choice Question Answering and question prediction tasks.

    Args:
        model_args (Roberteye | MAG | PostFusion): Model arguments.
        trainer_args (TrainerDL): Trainer arguments.
    """

    def __init__(
        self,
        model_args: RoberteyeArgs | MAG | PostFusion,
        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.preorder = model_args.preorder
        self.warmup_proportion = model_args.warmup_proportion

        self.train()
        self.save_hyperparameters()

    def forward(
        self,
        input_ids,
        attention_mask,
        labels,
        gaze_features,
        gaze_positions,
        eye_token_type_ids=None,
        **kwargs,
    ) -> torch.Tensor:
        """
        Forward pass of the model.

        Args:
            input_ids (torch.Tensor): Input IDs.
            attention_mask (torch.Tensor): Attention mask.
            labels (torch.Tensor): Labels.
            gaze_features (torch.Tensor): Gaze features.
            gaze_positions (torch.Tensor): Gaze positions.
            eye_token_type_ids (torch.Tensor, optional): Eye token type IDs. Defaults to None.

        Returns:
            torch.Tensor: Model output.
        """
        return self.model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            labels=labels,
            gaze_features=gaze_features,
            gaze_positions=gaze_positions,
            eye_token_type_ids=eye_token_type_ids,
            **kwargs,
        )

    def finalize_eye_data(
        self, batch
    ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor]:
        """
        Set the final gaze_features, gaze_positions, and attention_mask
        based on whether we use fixation reports or not.

        Args:
            batch (BatchData): Batch data.
            actual_needed_eye_padding_len (int | None): Actual needed eye padding length.

        Returns:
            tuple: Gaze features, gaze positions, and attention mask.
        """
        if self.model_args.use_fixation_report:
            assert batch.fixation_features is not None, (
                'fixation_features must be present if using fixation_report'
            )

            gaze_positions = batch.grouped_inversions
            gaze_features = batch.fixation_features

            attention_mask = batch.input_masks

        else:
            # No fixation report => we must have eyes instead
            assert batch.eyes is not None
            assert batch.input_ids is not None

            gaze_features = batch.eyes
            gaze_positions = batch.grouped_inversions
            attention_mask = batch.input_masks

        # If we do NOT want to prepend_eye_features_to_text for ROBERTEYE_MODEL, set them to None
        if (
            not self.model_args.prepend_eye_features_to_text
            and self.model_args.base_model_name == DLModelNames.ROBERTEYE_MODEL
        ):
            gaze_features = None
            gaze_positions = None

        return gaze_features, gaze_positions, attention_mask

    def shared_step(
        self, batch: list
    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """
        Shared step for training, validation, and testing.

        Args:
            batch (list): Batch data.

        Returns:
            tuple: Ordered labels, loss, ordered logits, labels, and logits.
        """
        # 1. Unpack the batch
        batch_data = self.unpack_batch(batch)

        # 2. Process different modes that affect input_masks/grouped_inversions
        eye_token_type_ids = (
            None  # if it will stay None, it will be created automatically later
        )

        # Safety checks
        assert batch_data.input_masks is not None, 'input_masks cannot be None'
        assert batch_data.grouped_inversions is not None, (
            'grouped_inversions cannot be None'
        )

        # 3. Finalize eye/gaze data
        gaze_features, gaze_positions, attention_mask = self.finalize_eye_data(
            batch_data
        )

        labels = batch_data.labels

        if self.task == TaskTypes.REGRESSION:
            labels = labels.squeeze().float()

        output = self(
            input_ids=batch_data.input_ids,
            attention_mask=attention_mask,
            labels=labels,
            gaze_features=gaze_features,
            gaze_positions=gaze_positions,
            output_hidden_states=True,
            eye_token_type_ids=eye_token_type_ids,
        )

        logits = output.logits
        if self.task == TaskTypes.REGRESSION:
            logits = logits.squeeze()

        loss = self.loss(logits, labels)

        return labels, loss, logits

    def configure_optimizers(self) -> tuple[list, list]:
        """
        Configure the optimizer and learning rate scheduler.

        Returns:
            tuple: Optimizer and learning rate scheduler.
        """
        # Define the optimizer
        assert self.warmup_proportion is not None

        # Copied from bert
        param_optimizer = list(self.named_parameters())

        # hack to remove pooler, which is not used
        # thus it produce None grad that break apex
        param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]

        no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
        optimizer_grouped_parameters = [
            {
                'params': [
                    p for n, p in param_optimizer if not any(nd in n for nd in no_decay)
                ],
                'weight_decay': 0.1,
            },
            {
                'params': [
                    p for n, p in param_optimizer if any(nd in n for nd in no_decay)
                ],
                'weight_decay': 0.0,
            },
        ]
        optimizer = torch.optim.AdamW(
            optimizer_grouped_parameters,
            lr=self.learning_rate,
            betas=(0.9, 0.98),
            eps=1e-6,
        )

        stepping_batches = self.trainer.estimated_stepping_batches
        scheduler = get_linear_schedule_with_warmup(
            optimizer,
            num_warmup_steps=int(stepping_batches * self.warmup_proportion),
            num_training_steps=stepping_batches,
        )
        return [optimizer], [{'scheduler': scheduler, 'interval': 'step'}]

configure_optimizers()

Configure the optimizer and learning rate scheduler.

Returns:

Name Type Description
tuple tuple[list, list]

Optimizer and learning rate scheduler.
Source code in src/models/base_roberta.py 
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
def configure_optimizers(self) -> tuple[list, list]:
    """
    Configure the optimizer and learning rate scheduler.

    Returns:
        tuple: Optimizer and learning rate scheduler.
    """
    # Define the optimizer
    assert self.warmup_proportion is not None

    # Copied from bert
    param_optimizer = list(self.named_parameters())

    # hack to remove pooler, which is not used
    # thus it produce None grad that break apex
    param_optimizer = [n for n in param_optimizer if 'pooler' not in n[0]]

    no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
    optimizer_grouped_parameters = [
        {
            'params': [
                p for n, p in param_optimizer if not any(nd in n for nd in no_decay)
            ],
            'weight_decay': 0.1,
        },
        {
            'params': [
                p for n, p in param_optimizer if any(nd in n for nd in no_decay)
            ],
            'weight_decay': 0.0,
        },
    ]
    optimizer = torch.optim.AdamW(
        optimizer_grouped_parameters,
        lr=self.learning_rate,
        betas=(0.9, 0.98),
        eps=1e-6,
    )

    stepping_batches = self.trainer.estimated_stepping_batches
    scheduler = get_linear_schedule_with_warmup(
        optimizer,
        num_warmup_steps=int(stepping_batches * self.warmup_proportion),
        num_training_steps=stepping_batches,
    )
    return [optimizer], [{'scheduler': scheduler, 'interval': 'step'}]

finalize_eye_data(batch)

Set the final gaze_features, gaze_positions, and attention_mask based on whether we use fixation reports or not.

Parameters:

Name Type Description Default
batch BatchData

Batch data.

 required
actual_needed_eye_padding_len int | None

Actual needed eye padding length.

 required

Returns:

Name Type Description
tuple tuple[Tensor | None, Tensor | None, Tensor]

Gaze features, gaze positions, and attention mask.
Source code in src/models/base_roberta.py 
 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
def finalize_eye_data(
    self, batch
) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor]:
    """
    Set the final gaze_features, gaze_positions, and attention_mask
    based on whether we use fixation reports or not.

    Args:
        batch (BatchData): Batch data.
        actual_needed_eye_padding_len (int | None): Actual needed eye padding length.

    Returns:
        tuple: Gaze features, gaze positions, and attention mask.
    """
    if self.model_args.use_fixation_report:
        assert batch.fixation_features is not None, (
            'fixation_features must be present if using fixation_report'
        )

        gaze_positions = batch.grouped_inversions
        gaze_features = batch.fixation_features

        attention_mask = batch.input_masks

    else:
        # No fixation report => we must have eyes instead
        assert batch.eyes is not None
        assert batch.input_ids is not None

        gaze_features = batch.eyes
        gaze_positions = batch.grouped_inversions
        attention_mask = batch.input_masks

    # If we do NOT want to prepend_eye_features_to_text for ROBERTEYE_MODEL, set them to None
    if (
        not self.model_args.prepend_eye_features_to_text
        and self.model_args.base_model_name == DLModelNames.ROBERTEYE_MODEL
    ):
        gaze_features = None
        gaze_positions = None

    return gaze_features, gaze_positions, attention_mask

forward(input_ids, attention_mask, labels, gaze_features, gaze_positions, eye_token_type_ids=None, **kwargs)

Forward pass of the model.

Parameters:

Name Type Description Default
input_ids Tensor

Input IDs.

 required
attention_mask Tensor

Attention mask.

 required
labels Tensor

Labels.

 required
gaze_features Tensor

Gaze features.

 required
gaze_positions Tensor

Gaze positions.

 required
eye_token_type_ids Tensor

Eye token type IDs. Defaults to None.

 None

Returns:

Type Description
Tensor

torch.Tensor: Model output.
Source code in src/models/base_roberta.py 
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
def forward(
    self,
    input_ids,
    attention_mask,
    labels,
    gaze_features,
    gaze_positions,
    eye_token_type_ids=None,
    **kwargs,
) -> torch.Tensor:
    """
    Forward pass of the model.

    Args:
        input_ids (torch.Tensor): Input IDs.
        attention_mask (torch.Tensor): Attention mask.
        labels (torch.Tensor): Labels.
        gaze_features (torch.Tensor): Gaze features.
        gaze_positions (torch.Tensor): Gaze positions.
        eye_token_type_ids (torch.Tensor, optional): Eye token type IDs. Defaults to None.

    Returns:
        torch.Tensor: Model output.
    """
    return self.model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        labels=labels,
        gaze_features=gaze_features,
        gaze_positions=gaze_positions,
        eye_token_type_ids=eye_token_type_ids,
        **kwargs,
    )

shared_step(batch)

Shared step for training, validation, and testing.

Parameters:

Name Type Description Default
batch list

Batch data.

 required

Returns:

Name Type Description
tuple tuple[Tensor, Tensor, Tensor]

Ordered labels, loss, ordered logits, labels, and logits.
Source code in src/models/base_roberta.py 
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
def shared_step(
    self, batch: list
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
    """
    Shared step for training, validation, and testing.

    Args:
        batch (list): Batch data.

    Returns:
        tuple: Ordered labels, loss, ordered logits, labels, and logits.
    """
    # 1. Unpack the batch
    batch_data = self.unpack_batch(batch)

    # 2. Process different modes that affect input_masks/grouped_inversions
    eye_token_type_ids = (
        None  # if it will stay None, it will be created automatically later
    )

    # Safety checks
    assert batch_data.input_masks is not None, 'input_masks cannot be None'
    assert batch_data.grouped_inversions is not None, (
        'grouped_inversions cannot be None'
    )

    # 3. Finalize eye/gaze data
    gaze_features, gaze_positions, attention_mask = self.finalize_eye_data(
        batch_data
    )

    labels = batch_data.labels

    if self.task == TaskTypes.REGRESSION:
        labels = labels.squeeze().float()

    output = self(
        input_ids=batch_data.input_ids,
        attention_mask=attention_mask,
        labels=labels,
        gaze_features=gaze_features,
        gaze_positions=gaze_positions,
        output_hidden_states=True,
        eye_token_type_ids=eye_token_type_ids,
    )

    logits = output.logits
    if self.task == TaskTypes.REGRESSION:
        logits = logits.squeeze()

    loss = self.loss(logits, labels)

    return labels, loss, logits