unitorch.cli.models.visualbert

VisualBertProcessor

Tip

core/process/visualbert is the section for configuration of VisualBertProcessor.

Bases: VisualBertProcessor

VisualBERT Processor for text and image inputs.

Initialize the VisualBertProcessor.

Parameters:

Name	Type	Description	Default
vocab_path	str	The path to the vocabulary file.	required
max_seq_length	Optional[int]	The maximum sequence length. Defaults to 128.	128
special_input_ids	Optional[Dict]	A dictionary containing special input IDs. Defaults to an empty dictionary.	dict()
do_lower_case	Optional[bool]	Whether to convert the text to lowercase. Defaults to True.	True
do_basic_tokenize	Optional[bool]	Whether to perform basic tokenization. Defaults to True.	True
do_whole_word_mask	Optional[bool]	Whether to use whole-word masking. Defaults to True.	True
masked_lm_prob	Optional[float]	The probability of masked language model masking. Defaults to 0.15.	0.15
max_predictions_per_seq	Optional[int]	The maximum number of masked language model predictions per sequence. Defaults to 20.	20

Source code in src/unitorch/cli/models/visualbert/processing.py

def __init__(
    self,
    vocab_path,
    max_seq_length: Optional[int] = 128,
    special_input_ids: Optional[Dict] = dict(),
    do_lower_case: Optional[bool] = True,
    do_basic_tokenize: Optional[bool] = True,
    do_whole_word_mask: Optional[bool] = True,
    masked_lm_prob: Optional[float] = 0.15,
    max_predictions_per_seq: Optional[int] = 20,
):
    """
    Initialize the VisualBertProcessor.

    Args:
        vocab_path (str): The path to the vocabulary file.
        max_seq_length (Optional[int]): The maximum sequence length. Defaults to 128.
        special_input_ids (Optional[Dict]): A dictionary containing special input IDs. Defaults to an empty dictionary.
        do_lower_case (Optional[bool]): Whether to convert the text to lowercase. Defaults to True.
        do_basic_tokenize (Optional[bool]): Whether to perform basic tokenization. Defaults to True.
        do_whole_word_mask (Optional[bool]): Whether to use whole-word masking. Defaults to True.
        masked_lm_prob (Optional[float]): The probability of masked language model masking. Defaults to 0.15.
        max_predictions_per_seq (Optional[int]): The maximum number of masked language model predictions per sequence. Defaults to 20.
    """
    super().__init__(
        vocab_path=vocab_path,
        max_seq_length=max_seq_length,
        special_input_ids=special_input_ids,
        do_lower_case=do_lower_case,
        do_basic_tokenize=do_basic_tokenize,
        do_whole_word_mask=do_whole_word_mask,
        masked_lm_prob=masked_lm_prob,
        max_predictions_per_seq=max_predictions_per_seq,
    )

from_core_configure classmethod

from_core_configure(config, **kwargs)

Create an instance of VisualBertProcessor from a core configuration.

Parameters:

Name	Type	Description	Default
config		The core configuration.	required
**kwargs		Additional keyword arguments.	{}

Returns:

Name	Type	Description
dict		A dictionary containing the processor's configuration.

Source code in src/unitorch/cli/models/visualbert/processing.py

@classmethod
@add_default_section_for_init("core/process/visualbert")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of VisualBertProcessor from a core configuration.

    Args:
        config: The core configuration.
        **kwargs: Additional keyword arguments.

    Returns:
        dict: A dictionary containing the processor's configuration.
    """
    config.set_default_section("core/process/visualbert")
    pretrained_name = config.getoption("pretrained_name", "visualbert-vqa-coco-pre")
    vocab_path = config.getoption("vocab_path", None)
    vocab_path = pop_value(
        vocab_path,
        nested_dict_value(pretrained_visualbert_infos, pretrained_name, "vocab"),
    )
    vocab_path = cached_path(vocab_path)

    return {
        "vocab_path": vocab_path,
    }

VisualBertForClassification

Tip

core/model/classification/visualbert is the section for configuration of VisualBertForClassification.

Bases: VisualBertForClassification

VisualBERT for Classification model.

Initialize VisualBertForClassification.

Parameters:

Name	Type	Description	Default
config_path	str	The path to the model's configuration file.	required
num_classes	Optional[int]	The number of classes for classification. Defaults to 1.	1
gradient_checkpointing	Optional[bool]	Whether to use gradient checkpointing to save memory during training. Defaults to False.	False

Source code in src/unitorch/cli/models/visualbert/modeling.py

def __init__(
    self,
    config_path: str,
    num_classes: Optional[int] = 1,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initialize VisualBertForClassification.

    Args:
        config_path (str): The path to the model's configuration file.
        num_classes (Optional[int]): The number of classes for classification.
            Defaults to 1.
        gradient_checkpointing (Optional[bool]): Whether to use gradient checkpointing
            to save memory during training. Defaults to False.
    """
    super().__init__(
        config_path=config_path,
        num_classes=num_classes,
        gradient_checkpointing=gradient_checkpointing,
    )

forward

forward(
    input_ids: Tensor,
    attention_mask: Tensor,
    token_type_ids: Tensor,
    position_ids: Tensor,
    visual_embeds: Tensor,
    visual_attention_mask: Tensor,
    visual_token_type_ids: Tensor,
)

Forward pass of the VisualBertForClassification model.

Parameters:

Name	Type	Description	Default
input_ids	Tensor	The input token IDs.	required
attention_mask	Tensor	The attention mask.	required
token_type_ids	Tensor	The token type IDs.	required
position_ids	Tensor	The position IDs.	required
visual_embeds	Tensor	The visual embeddings.	required
visual_attention_mask	Tensor	The visual attention mask.	required
visual_token_type_ids	Tensor	The visual token type IDs.	required

Returns:

Name	Type	Description
ClassificationOutputs		The output logits of the model.

Source code in src/unitorch/cli/models/visualbert/modeling.py

@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: torch.Tensor,
    token_type_ids: torch.Tensor,
    position_ids: torch.Tensor,
    visual_embeds: torch.Tensor,
    visual_attention_mask: torch.Tensor,
    visual_token_type_ids: torch.Tensor,
):
    """
    Forward pass of the VisualBertForClassification model.

    Args:
        input_ids (torch.Tensor): The input token IDs.
        attention_mask (torch.Tensor): The attention mask.
        token_type_ids (torch.Tensor): The token type IDs.
        position_ids (torch.Tensor): The position IDs.
        visual_embeds (torch.Tensor): The visual embeddings.
        visual_attention_mask (torch.Tensor): The visual attention mask.
        visual_token_type_ids (torch.Tensor): The visual token type IDs.

    Returns:
        ClassificationOutputs: The output logits of the model.
    """
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        token_type_ids=token_type_ids,
        position_ids=position_ids,
        visual_embeds=visual_embeds,
        visual_attention_mask=visual_attention_mask,
        visual_token_type_ids=visual_token_type_ids,
    )
    return ClassificationOutputs(outputs=outputs)

from_core_configure classmethod

from_core_configure(config, **kwargs)

Create an instance of VisualBertForClassification from a core configuration.

Parameters:

Name	Type	Description	Default
config		The core configuration.	required
**kwargs		Additional keyword arguments.	{}

Returns:

Name	Type	Description
VisualBertForClassification		The initialized VisualBertForClassification instance.

Source code in src/unitorch/cli/models/visualbert/modeling.py

@classmethod
@add_default_section_for_init("core/model/classification/visualbert")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of VisualBertForClassification from a core configuration.

    Args:
        config: The core configuration.
        **kwargs: Additional keyword arguments.

    Returns:
        VisualBertForClassification: The initialized VisualBertForClassification instance.
    """
    config.set_default_section("core/model/classification/visualbert")
    pretrained_name = config.getoption("pretrained_name", "visualbert-vqa-coco-pre")
    config_path = config.getoption("config_path", None)

    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_visualbert_infos, pretrained_name, "config"),
    )

    config_path = cached_path(config_path)
    gradient_checkpointing = config.getoption("gradient_checkpointing", False)
    num_classes = config.getoption("num_classes", 1)

    inst = cls(config_path, num_classes, gradient_checkpointing)
    pretrained_weight_path = config.getoption("pretrained_weight_path", None)
    weight_path = pop_value(
        pretrained_weight_path,
        nested_dict_value(pretrained_visualbert_infos, pretrained_name, "weight"),
        check_none=False,
    )
    if weight_path is not None:
        inst.from_pretrained(weight_path)

    return inst

VisualBertForPretrain

Tip

core/model/pretrain/visualbert is the section for configuration of VisualBertForPretrain.

Bases: VisualBertForPretrain

VisualBERT for Pretraining model.

Initialize VisualBertForPretrain.

Parameters:

Name	Type	Description	Default
config_path	str	The path to the model's configuration file.	required
gradient_checkpointing	Optional[bool]	Whether to use gradient checkpointing to save memory during training. Defaults to False.	False

Source code in src/unitorch/cli/models/visualbert/modeling.py

def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initialize VisualBertForPretrain.

    Args:
        config_path (str): The path to the model's configuration file.
        gradient_checkpointing (Optional[bool]): Whether to use gradient checkpointing
            to save memory during training. Defaults to False.
    """
    super().__init__(
        config_path=config_path,
        gradient_checkpointing=gradient_checkpointing,
    )

forward

forward(
    input_ids: Tensor,
    attention_mask: Tensor,
    token_type_ids: Tensor,
    position_ids: Tensor,
    visual_embeds: Tensor,
    visual_attention_mask: Tensor,
    visual_token_type_ids: Tensor,
    nsp_label: Tensor,
    mlm_label: Tensor,
    mlm_label_mask: Tensor,
)

Forward pass of the VisualBertForPretrain model.

Parameters:

Name	Type	Description	Default
input_ids	Tensor	The input token IDs.	required
attention_mask	Tensor	The attention mask.	required
token_type_ids	Tensor	The token type IDs.	required
position_ids	Tensor	The position IDs.	required
visual_embeds	Tensor	The visual embeddings.	required
visual_attention_mask	Tensor	The visual attention mask.	required
visual_token_type_ids	Tensor	The visual token type IDs.	required
nsp_label	Tensor	The next sentence prediction label.	required
mlm_label	Tensor	The masked language model label.	required
mlm_label_mask	Tensor	The masked language model label mask.	required

Returns:

Name	Type	Description
LossOutputs		The output loss of the model.

Source code in src/unitorch/cli/models/visualbert/modeling.py

@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: torch.Tensor,
    token_type_ids: torch.Tensor,
    position_ids: torch.Tensor,
    visual_embeds: torch.Tensor,
    visual_attention_mask: torch.Tensor,
    visual_token_type_ids: torch.Tensor,
    nsp_label: torch.Tensor,
    mlm_label: torch.Tensor,
    mlm_label_mask: torch.Tensor,
):
    """
    Forward pass of the VisualBertForPretrain model.

    Args:
        input_ids (torch.Tensor): The input token IDs.
        attention_mask (torch.Tensor): The attention mask.
        token_type_ids (torch.Tensor): The token type IDs.
        position_ids (torch.Tensor): The position IDs.
        visual_embeds (torch.Tensor): The visual embeddings.
        visual_attention_mask (torch.Tensor): The visual attention mask.
        visual_token_type_ids (torch.Tensor): The visual token type IDs.
        nsp_label (torch.Tensor): The next sentence prediction label.
        mlm_label (torch.Tensor): The masked language model label.
        mlm_label_mask (torch.Tensor): The masked language model label mask.

    Returns:
        LossOutputs: The output loss of the model.
    """
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        token_type_ids=token_type_ids,
        position_ids=position_ids,
        visual_embeds=visual_embeds,
        visual_attention_mask=visual_attention_mask,
        visual_token_type_ids=visual_token_type_ids,
        nsp_label=nsp_label,
        mlm_label=mlm_label,
        mlm_label_mask=mlm_label_mask,
    )
    return LossOutputs(loss=outputs)

from_core_configure classmethod

from_core_configure(config, **kwargs)

Create an instance of VisualBertForPretrain from a core configuration.

Parameters:

Name	Type	Description	Default
config		The core configuration.	required
**kwargs		Additional keyword arguments.	{}

Returns:

Name	Type	Description
VisualBertForPretrain		The initialized VisualBertForPretrain instance.

Source code in src/unitorch/cli/models/visualbert/modeling.py

@classmethod
@add_default_section_for_init("core/model/pretrain/visualbert")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of VisualBertForPretrain from a core configuration.

    Args:
        config: The core configuration.
        **kwargs: Additional keyword arguments.

    Returns:
        VisualBertForPretrain: The initialized VisualBertForPretrain instance.
    """
    config.set_default_section("core/model/pretrain/visualbert")
    pretrained_name = config.getoption("pretrained_name", "visualbert-vqa-coco-pre")
    config_path = config.getoption("config_path", None)

    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_visualbert_infos, pretrained_name, "config"),
    )

    config_path = cached_path(config_path)
    gradient_checkpointing = config.getoption("gradient_checkpointing", False)

    inst = cls(config_path, gradient_checkpointing)
    pretrained_weight_path = config.getoption("pretrained_weight_path", None)
    weight_path = pop_value(
        pretrained_weight_path,
        nested_dict_value(pretrained_visualbert_infos, pretrained_name, "weight"),
        check_none=False,
    )
    if weight_path is not None:
        inst.from_pretrained(weight_path)

    return inst