Skip to content

unitorch.cli.models.mt5¤

MT5Processor¤

Tip

core/process/mt5 is the section for configuration of MT5Processor.

Bases: MT5Processor

Processor for MT5 models.

Initialize the MT5Processor.

Parameters:

Name Type Description Default
vocab_path str

The path to the vocabulary file.

required
special_input_ids Dict

Special input IDs. Defaults to an empty dictionary.

dict()
max_seq_length int

The maximum sequence length. Defaults to 128.

128
max_gen_seq_length int

The maximum generation sequence length. Defaults to 48.

48
Source code in src/unitorch/cli/models/mt5/processing.py
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
def __init__(
    self,
    vocab_path: str,
    special_input_ids: Optional[Dict] = dict(),
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 48,
):
    """
    Initialize the MT5Processor.

    Args:
        vocab_path (str): The path to the vocabulary file.
        special_input_ids (Dict, optional): Special input IDs. Defaults to an empty dictionary.
        max_seq_length (int, optional): The maximum sequence length. Defaults to 128.
        max_gen_seq_length (int, optional): The maximum generation sequence length. Defaults to 48.
    """
    super().__init__(
        vocab_path=vocab_path,
        special_input_ids=special_input_ids,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

from_core_configure classmethod ¤

from_core_configure(config, **kwargs)

Create an instance of MT5Processor 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 arguments for initializing the MT5Processor.

Source code in src/unitorch/cli/models/mt5/processing.py
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
@classmethod
@add_default_section_for_init("core/process/mt5")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of MT5Processor from a core configuration.

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

    Returns:
        dict: A dictionary containing the arguments for initializing the MT5Processor.
    """
    config.set_default_section("core/process/mt5")
    pretrained_name = config.getoption("pretrained_name", "mt5-base")
    vocab_path = config.getoption("vocab_path", None)
    vocab_path = pop_value(
        vocab_path,
        nested_dict_value(pretrained_mt5_infos, pretrained_name, "vocab"),
    )
    vocab_path = cached_path(vocab_path)

    return {
        "vocab_path": vocab_path,
    }

MT5ForGeneration¤

Tip

core/model/generation/mt5 is the section for configuration of MT5ForGeneration.

Bases: MT5ForGeneration

MT5 model for generation tasks.

Initialize the MT5ForGeneration model.

Parameters:

Name Type Description Default
config_path str

The path to the model configuration file.

required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

False
Source code in src/unitorch/cli/models/mt5/modeling.py
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    """
    Initialize the MT5ForGeneration model.

    Args:
        config_path (str): The path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__(
        config_path=config_path, gradient_checkpointing=gradient_checkpointing
    )

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Tensor,
    decoder_input_ids: Tensor,
    decoder_attention_mask: Tensor,
)

Perform a forward pass through the model.

Parameters:

Name Type Description Default
input_ids Tensor

Input token IDs.

required
attention_mask Tensor

Attention mask.

required
decoder_input_ids Tensor

Decoder input token IDs.

required
decoder_attention_mask Tensor

Decoder attention mask.

required

Returns:

Name Type Description
GenerationOutputs

The generation outputs.

Source code in src/unitorch/cli/models/mt5/modeling.py
 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
@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: torch.Tensor,
    decoder_input_ids: torch.Tensor,
    decoder_attention_mask: torch.Tensor,
):
    """
    Perform a forward pass through the model.

    Args:
        input_ids (torch.Tensor): Input token IDs.
        attention_mask (torch.Tensor): Attention mask.
        decoder_input_ids (torch.Tensor): Decoder input token IDs.
        decoder_attention_mask (torch.Tensor): Decoder attention mask.

    Returns:
        GenerationOutputs: The generation outputs.
    """
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
        decoder_input_ids=decoder_input_ids,
        decoder_attention_mask=decoder_attention_mask,
    )
    return GenerationOutputs(sequences=outputs)

from_core_configure classmethod ¤

from_core_configure(config, **kwargs)

Create an instance of MT5ForGeneration from a core configuration.

Parameters:

Name Type Description Default
config

The core configuration.

required
**kwargs

Additional keyword arguments.

{}

Returns:

Name Type Description
MT5ForGeneration

An instance of MT5ForGeneration.

Source code in src/unitorch/cli/models/mt5/modeling.py
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
@classmethod
@add_default_section_for_init("core/model/generation/mt5")
def from_core_configure(cls, config, **kwargs):
    """
    Create an instance of MT5ForGeneration from a core configuration.

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

    Returns:
        MT5ForGeneration: An instance of MT5ForGeneration.
    """
    config.set_default_section("core/model/generation/mt5")
    pretrained_name = config.getoption("pretrained_name", "mt5-base")
    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_mt5_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_mt5_infos, pretrained_name, "weight"),
        check_none=False,
    )
    if weight_path is not None:
        inst.from_pretrained(weight_path)

    return inst

generate ¤

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 0,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 1,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
)

Generate sequences using the MT5 model.

Parameters:

Name Type Description Default
input_ids Tensor

Input token IDs.

required
num_beams int

Number of beams for beam search. Defaults to 5.

5
decoder_start_token_id int

Decoder start token ID. Defaults to 0.

0
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 1.

1
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

1
min_gen_seq_length int

Minimum generation sequence length. Defaults to 0.

0
max_gen_seq_length int

Maximum generation sequence length. Defaults to 48.

48
repetition_penalty float

Repetition penalty. Defaults to 1.0.

1.0
no_repeat_ngram_size int

Size of n-grams to prevent repetition. Defaults to 0.

0
early_stopping bool

Whether to perform early stopping. Defaults to True.

True
length_penalty float

Length penalty. Defaults to 1.0.

1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

1
diversity_penalty float

Diversity penalty for diverse beam search. Defaults to 0.0.

0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

False
temperature float

Sampling temperature. Defaults to 1.0.

1.0
top_k int

Top-k sampling parameter. Defaults to 50.

50
top_p float

Top-p sampling parameter. Defaults to 1.0.

1.0

Returns:

Name Type Description
GenerationOutputs

The generation outputs.

Source code in src/unitorch/cli/models/mt5/modeling.py
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
@add_default_section_for_function("core/model/generation/mt5")
@torch.no_grad()
@autocast(device_type=("cuda" if torch.cuda.is_available() else "cpu"))
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 0,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 1,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
):
    """
    Generate sequences using the MT5 model.

    Args:
        input_ids (torch.Tensor): Input token IDs.
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): Decoder start token ID. Defaults to 0.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 1.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum generation sequence length. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum generation sequence length. Defaults to 48.
        repetition_penalty (float, optional): Repetition penalty. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to prevent repetition. Defaults to 0.
        early_stopping (bool, optional): Whether to perform early stopping. Defaults to True.
        length_penalty (float, optional): Length penalty. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Diversity penalty for diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k sampling parameter. Defaults to 50.
        top_p (float, optional): Top-p sampling parameter. Defaults to 1.0.

    Returns:
        GenerationOutputs: The generation outputs.
    """
    outputs = super().generate(
        input_ids,
        num_beams=num_beams,
        decoder_start_token_id=decoder_start_token_id,
        decoder_end_token_id=decoder_end_token_id,
        num_return_sequences=num_return_sequences,
        min_gen_seq_length=min_gen_seq_length,
        max_gen_seq_length=max_gen_seq_length,
        repetition_penalty=repetition_penalty,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        do_sample=do_sample,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
    )

    return GenerationOutputs(
        sequences=outputs.sequences,
        sequences_scores=outputs.sequences_scores,
    )