跳转至

heads

heads 提供了一系列的模型预测头，用于处理不同的任务。

heads 接受 ModelOutupt、dict 或 tuple 作为输入。 它会自动查找预测所需的模型输出并相应地处理。

一些预测头可能需要额外的信息，例如 attention_mask 或 input_ids，例如 ContactPredictionHead。 这些额外的参数可以作为参数/关键字参数传入。

请注意，heads 使用与 🤗 Transformers 相同的 ModelOutupt 约定。 如果模型输出是一个 tuple，我们将第一个元素视为 pooler_output，第二个元素视为 last_hidden_state，最后一个元素视为 attention_map。 用户有责任确保模型输出格式正确。

如果模型输出是一个 ModelOutupt 或一个 dict，heads 将从模型输出中查找 HeadConfig.output_name。 你可以在 HeadConfig 中指定 output_name，以确保 heads 可以正确定位所需的张量。

multimolecule.module.heads.config

HeadConfig dataclass

Bases: BaseHeadConfig

Configuration class for a prediction head.

Parameters:

Name Type Description Default

num_labels

int

Number of labels to use in the last layer added to the model, typically for a classification task.

Head should look for Config.num_labels if is None.

 None

problem_type

str

Problem type for XxxForYyyPrediction models. Can be one of "regression", "single_label_classification" or "multi_label_classification".

Head should look for Config.problem_type if is None.

 None

hidden_size

int | None

Dimensionality of the encoder layers and the pooler layer.

Head should look for Config.hidden_size if is None.

 None

dropout

float

The dropout ratio for the hidden states.

 0.0

transform

str | None

The transform operation applied to hidden states.

 None

transform_act

str | None

The activation function of transform applied to hidden states.

 'gelu'

bias

bool

Whether to apply bias to the final prediction layer.

 True

act

str | None

The activation function of the final prediction output.

 None

layer_norm_eps

float

The epsilon used by the layer normalization layers.

 1e-12

output_name

`str`, *optional*

The name of the tensor required in model outputs.

If is None, will use the default output name of the corresponding head.

 None
Source code in multimolecule/module/heads/config.py
Python
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
@dataclass
class HeadConfig(BaseHeadConfig):
    r"""
    Configuration class for a prediction head.

    Args:
        num_labels:
            Number of labels to use in the last layer added to the model, typically for a classification task.

            Head should look for [`Config.num_labels`][multimolecule.PreTrainedConfig] if is `None`.
        problem_type:
            Problem type for `XxxForYyyPrediction` models. Can be one of `"regression"`,
            `"single_label_classification"` or `"multi_label_classification"`.

            Head should look for [`Config.problem_type`][multimolecule.PreTrainedConfig] if is `None`.
        hidden_size:
            Dimensionality of the encoder layers and the pooler layer.

            Head should look for [`Config.hidden_size`][multimolecule.PreTrainedConfig] if is `None`.
        dropout:
            The dropout ratio for the hidden states.
        transform:
            The transform operation applied to hidden states.
        transform_act:
            The activation function of transform applied to hidden states.
        bias:
            Whether to apply bias to the final prediction layer.
        act:
            The activation function of the final prediction output.
        layer_norm_eps:
            The epsilon used by the layer normalization layers.
        output_name (`str`, *optional*):
            The name of the tensor required in model outputs.

            If is `None`, will use the default output name of the corresponding head.
    """

    num_labels: int = None  # type: ignore[assignment]
    problem_type: str = None  # type: ignore[assignment]
    hidden_size: int | None = None
    dropout: float = 0.0
    transform: str | None = None
    transform_act: str | None = "gelu"
    bias: bool = True
    act: str | None = None
    layer_norm_eps: float = 1e-12
    output_name: str | None = None

MaskedLMHeadConfig dataclass

Bases: BaseHeadConfig

Configuration class for a Masked Language Modeling head.

Parameters:

Name Type Description Default

hidden_size

int | None

Dimensionality of the encoder layers and the pooler layer.

Head should look for Config.hidden_size if is None.

 None

dropout

float

The dropout ratio for the hidden states.

 0.0

transform

str | None

The transform operation applied to hidden states.

 'nonlinear'

transform_act

str | None

The activation function of transform applied to hidden states.

 'gelu'

bias

bool

Whether to apply bias to the final prediction layer.

 True

act

str | None

The activation function of the final prediction output.

 None

layer_norm_eps

float

The epsilon used by the layer normalization layers.

 1e-12

output_name

`str`, *optional*

The name of the tensor required in model outputs.

If is None, will use the default output name of the corresponding head.

 None
Source code in multimolecule/module/heads/config.py
Python
 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
@dataclass
class MaskedLMHeadConfig(BaseHeadConfig):
    r"""
    Configuration class for a Masked Language Modeling head.

    Args:
        hidden_size:
            Dimensionality of the encoder layers and the pooler layer.

            Head should look for [`Config.hidden_size`][multimolecule.PreTrainedConfig] if is `None`.
        dropout:
            The dropout ratio for the hidden states.
        transform:
            The transform operation applied to hidden states.
        transform_act:
            The activation function of transform applied to hidden states.
        bias:
            Whether to apply bias to the final prediction layer.
        act:
            The activation function of the final prediction output.
        layer_norm_eps:
            The epsilon used by the layer normalization layers.
        output_name (`str`, *optional*):
            The name of the tensor required in model outputs.

            If is `None`, will use the default output name of the corresponding head.
    """

    hidden_size: int | None = None
    dropout: float = 0.0
    transform: str | None = "nonlinear"
    transform_act: str | None = "gelu"
    bias: bool = True
    act: str | None = None
    layer_norm_eps: float = 1e-12
    output_name: str | None = None

multimolecule.module.heads.sequence

SequencePredictionHead

Bases: PredictionHead

Head for tasks in sequence-level.

Parameters:

Name Type Description Default

config

PreTrainedConfig

The configuration object for the model.

 required

head_config

HeadConfig | None

The configuration object for the head. If None, will use configuration from the config.

 None
Source code in multimolecule/module/heads/sequence.py
Python
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
@HeadRegistry.register("sequence")
class SequencePredictionHead(PredictionHead):
    r"""
    Head for tasks in sequence-level.

    Args:
        config: The configuration object for the model.
        head_config: The configuration object for the head.
            If None, will use configuration from the `config`.
    """

    output_name: str = "pooler_output"
    r"""The default output to use for the head."""

    def __init__(self, config: PreTrainedConfig, head_config: HeadConfig | None = None):
        super().__init__(config, head_config)
        if head_config is not None and head_config.output_name is not None:
            self.output_name = head_config.output_name

    def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
        self,
        outputs: ModelOutput | Tuple[Tensor, ...],
        labels: Tensor | None = None,
        output_name: str | None = None,
        **kwargs,
    ) -> HeadOutput:
        r"""
        Forward pass of the SequencePredictionHead.

        Args:
            outputs: The outputs of the model.
            labels: The labels for the head.
            output_name: The name of the output to use.
                Defaults to `self.output_name`.
        """
        if isinstance(outputs, (Mapping, ModelOutput)):
            output = outputs[output_name or self.output_name]
        elif isinstance(outputs, tuple):
            output = outputs[1]
        return super().forward(output, labels, **kwargs)

output_name class-attribute instance-attribute

Python
1
output_name: str = 'pooler_output'

The default output to use for the head.

forward

Python
1
forward(outputs: ModelOutput | Tuple[Tensor, ...], labels: Tensor | None = None, output_name: str | None = None, **kwargs) -> HeadOutput

Forward pass of the SequencePredictionHead.

Parameters:

Name Type Description Default
outputs
ModelOutput | Tuple[Tensor, ...]

The outputs of the model.

 required
labels
Tensor | None

The labels for the head.

 None
output_name
str | None

The name of the output to use. Defaults to self.output_name.

 None
Source code in multimolecule/module/heads/sequence.py
Python
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
    self,
    outputs: ModelOutput | Tuple[Tensor, ...],
    labels: Tensor | None = None,
    output_name: str | None = None,
    **kwargs,
) -> HeadOutput:
    r"""
    Forward pass of the SequencePredictionHead.

    Args:
        outputs: The outputs of the model.
        labels: The labels for the head.
        output_name: The name of the output to use.
            Defaults to `self.output_name`.
    """
    if isinstance(outputs, (Mapping, ModelOutput)):
        output = outputs[output_name or self.output_name]
    elif isinstance(outputs, tuple):
        output = outputs[1]
    return super().forward(output, labels, **kwargs)

multimolecule.module.heads.token

TokenPredictionHead

Bases: PredictionHead

Head for tasks in token-level.

Parameters:

Name Type Description Default

config

PreTrainedConfig

The configuration object for the model.

 required

head_config

HeadConfig | None

The configuration object for the head. If None, will use configuration from the config.

 None
Source code in multimolecule/module/heads/token.py
Python
 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
@HeadRegistry.token.register("single", default=True)
@TokenHeadRegistryHF.register("single", default=True)
class TokenPredictionHead(PredictionHead):
    r"""
    Head for tasks in token-level.

    Args:
        config: The configuration object for the model.
        head_config: The configuration object for the head.
            If None, will use configuration from the `config`.
    """

    output_name: str = "last_hidden_state"
    r"""The default output to use for the head."""

    def __init__(self, config: PreTrainedConfig, head_config: HeadConfig | None = None):
        super().__init__(config, head_config)
        self.bos_token_id = config.bos_token_id
        self.eos_token_id = config.eos_token_id
        self.pad_token_id = config.pad_token_id
        if head_config is not None and head_config.output_name is not None:
            self.output_name = head_config.output_name

    def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
        self,
        outputs: ModelOutput | Tuple[Tensor, ...],
        attention_mask: Tensor | None = None,
        input_ids: NestedTensor | Tensor | None = None,
        labels: Tensor | None = None,
        output_name: str | None = None,
        **kwargs,
    ) -> HeadOutput:
        r"""
        Forward pass of the TokenPredictionHead.

        Args:
            outputs: The outputs of the model.
            attention_mask: The attention mask for the inputs.
            input_ids: The input ids for the inputs.
            labels: The labels for the head.
            output_name: The name of the output to use.
                Defaults to `self.output_name`.
        """
        if attention_mask is None:
            if isinstance(input_ids, NestedTensor):
                input_ids, attention_mask = input_ids.tensor, input_ids.mask
            else:
                if input_ids is None:
                    raise ValueError(
                        f"Either attention_mask or input_ids must be provided for {self.__class__.__name__} to work."
                    )
                if self.pad_token_id is None:
                    raise ValueError(
                        f"pad_token_id must be provided when attention_mask is not passed to {self.__class__.__name__}."
                    )
                attention_mask = input_ids.ne(self.pad_token_id)

        if isinstance(outputs, (Mapping, ModelOutput)):
            output = outputs[output_name or self.output_name]
        elif isinstance(outputs, tuple):
            output = outputs[0]
        else:
            raise ValueError(f"Unsupported type for outputs: {type(outputs)}")
        output = output * attention_mask.unsqueeze(-1)

        # remove cls token embeddings
        if self.bos_token_id is not None:
            output = output[..., 1:, :]
            # process attention_mask and input_ids to make removal of eos token happy
            attention_mask = attention_mask[..., 1:]
            if input_ids is not None:
                input_ids = input_ids[..., 1:]
        # remove eos token embeddings
        if self.eos_token_id is not None:
            if input_ids is not None:
                eos_mask = input_ids.ne(self.eos_token_id).to(output)
            else:
                last_valid_indices = attention_mask.sum(dim=-1)
                seq_length = attention_mask.size(-1)
                eos_mask = torch.arange(seq_length, device=output.device) == last_valid_indices.unsqueeze(1)
            output = output * eos_mask[:, :, None]
            output = output[..., :-1, :]

        return super().forward(output, labels, **kwargs)

output_name class-attribute instance-attribute

Python
1
output_name: str = 'last_hidden_state'

The default output to use for the head.

forward

Python
1
forward(outputs: ModelOutput | Tuple[Tensor, ...], attention_mask: Tensor | None = None, input_ids: NestedTensor | Tensor | None = None, labels: Tensor | None = None, output_name: str | None = None, **kwargs) -> HeadOutput

Forward pass of the TokenPredictionHead.

Parameters:

Name Type Description Default
outputs
ModelOutput | Tuple[Tensor, ...]

The outputs of the model.

 required
attention_mask
Tensor | None

The attention mask for the inputs.

 None
input_ids
NestedTensor | Tensor | None

The input ids for the inputs.

 None
labels
Tensor | None

The labels for the head.

 None
output_name
str | None

The name of the output to use. Defaults to self.output_name.

 None
Source code in multimolecule/module/heads/token.py
Python
 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
def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
    self,
    outputs: ModelOutput | Tuple[Tensor, ...],
    attention_mask: Tensor | None = None,
    input_ids: NestedTensor | Tensor | None = None,
    labels: Tensor | None = None,
    output_name: str | None = None,
    **kwargs,
) -> HeadOutput:
    r"""
    Forward pass of the TokenPredictionHead.

    Args:
        outputs: The outputs of the model.
        attention_mask: The attention mask for the inputs.
        input_ids: The input ids for the inputs.
        labels: The labels for the head.
        output_name: The name of the output to use.
            Defaults to `self.output_name`.
    """
    if attention_mask is None:
        if isinstance(input_ids, NestedTensor):
            input_ids, attention_mask = input_ids.tensor, input_ids.mask
        else:
            if input_ids is None:
                raise ValueError(
                    f"Either attention_mask or input_ids must be provided for {self.__class__.__name__} to work."
                )
            if self.pad_token_id is None:
                raise ValueError(
                    f"pad_token_id must be provided when attention_mask is not passed to {self.__class__.__name__}."
                )
            attention_mask = input_ids.ne(self.pad_token_id)

    if isinstance(outputs, (Mapping, ModelOutput)):
        output = outputs[output_name or self.output_name]
    elif isinstance(outputs, tuple):
        output = outputs[0]
    else:
        raise ValueError(f"Unsupported type for outputs: {type(outputs)}")
    output = output * attention_mask.unsqueeze(-1)

    # remove cls token embeddings
    if self.bos_token_id is not None:
        output = output[..., 1:, :]
        # process attention_mask and input_ids to make removal of eos token happy
        attention_mask = attention_mask[..., 1:]
        if input_ids is not None:
            input_ids = input_ids[..., 1:]
    # remove eos token embeddings
    if self.eos_token_id is not None:
        if input_ids is not None:
            eos_mask = input_ids.ne(self.eos_token_id).to(output)
        else:
            last_valid_indices = attention_mask.sum(dim=-1)
            seq_length = attention_mask.size(-1)
            eos_mask = torch.arange(seq_length, device=output.device) == last_valid_indices.unsqueeze(1)
        output = output * eos_mask[:, :, None]
        output = output[..., :-1, :]

    return super().forward(output, labels, **kwargs)

TokenKMerHead

Bases: PredictionHead

Head for tasks in token-level with kmer inputs.

Parameters:

Name Type Description Default

config

PreTrainedConfig

The configuration object for the model.

 required

head_config

HeadConfig | None

The configuration object for the head. If None, will use configuration from the config.

 None
Source code in multimolecule/module/heads/token.py
Python
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
@HeadRegistry.register("token.kmer")
@TokenHeadRegistryHF.register("kmer")
class TokenKMerHead(PredictionHead):
    r"""
    Head for tasks in token-level with kmer inputs.

    Args:
        config: The configuration object for the model.
        head_config: The configuration object for the head.
            If None, will use configuration from the `config`.
    """

    output_name: str = "last_hidden_state"
    r"""The default output to use for the head."""

    def __init__(self, config: PreTrainedConfig, head_config: HeadConfig | None = None):
        super().__init__(config, head_config)
        self.nmers = config.nmers
        self.bos_token_id = config.bos_token_id
        self.eos_token_id = config.eos_token_id
        self.pad_token_id = config.pad_token_id
        if head_config is not None and head_config.output_name is not None:
            self.output_name = head_config.output_name
        # Do not pass bos_token_id and eos_token_id to unfold_kmer_embeddings
        # As they will be removed in preprocess
        self.unfold_kmer_embeddings = partial(unfold_kmer_embeddings, nmers=self.nmers)

    def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
        self,
        outputs: ModelOutput | Tuple[Tensor, ...],
        attention_mask: Tensor | None = None,
        input_ids: NestedTensor | Tensor | None = None,
        labels: Tensor | None = None,
        output_name: str | None = None,
        **kwargs,
    ) -> HeadOutput:
        r"""
        Forward pass of the TokenKMerHead.

        Args:
            outputs: The outputs of the model.
            attention_mask: The attention mask for the inputs.
            input_ids: The input ids for the inputs.
            labels: The labels for the head.
            output_name: The name of the output to use.
                Defaults to `self.output_name`.
        """
        if attention_mask is None:
            if isinstance(input_ids, NestedTensor):
                input_ids, attention_mask = input_ids.tensor, input_ids.mask
            else:
                if input_ids is None:
                    raise ValueError(
                        f"Either attention_mask or input_ids must be provided for {self.__class__.__name__} to work."
                    )
                if self.pad_token_id is None:
                    raise ValueError(
                        f"pad_token_id must be provided when attention_mask is not passed to {self.__class__.__name__}."
                    )
                attention_mask = input_ids.ne(self.pad_token_id)

        if isinstance(outputs, (Mapping, ModelOutput)):
            output = outputs[output_name or self.output_name]
        elif isinstance(outputs, tuple):
            output = outputs[0]
        else:
            raise ValueError(f"Unsupported type for outputs: {type(outputs)}")
        output = output * attention_mask.unsqueeze(-1)

        # remove cls token embeddings
        if self.bos_token_id is not None:
            output = output[..., 1:, :]
            attention_mask = attention_mask[..., 1:]
            if input_ids is not None:
                input_ids = input_ids[..., 1:]
        # remove eos token embeddings
        if self.eos_token_id is not None:
            if input_ids is not None:
                eos_mask = input_ids.ne(self.eos_token_id).to(output)
                input_ids = input_ids[..., :-1]
            else:
                last_valid_indices = attention_mask.sum(dim=-1)
                seq_length = attention_mask.size(-1)
                eos_mask = torch.arange(seq_length, device=output.device) == last_valid_indices.unsqueeze(1)
            output = output * eos_mask[:, :, None]
            output = output[..., :-1, :]
            attention_mask = attention_mask[..., 1:]

        output = self.unfold_kmer_embeddings(output, attention_mask)
        return super().forward(output, labels, **kwargs)

output_name class-attribute instance-attribute

Python
1
output_name: str = 'last_hidden_state'

The default output to use for the head.

forward

Python
1
forward(outputs: ModelOutput | Tuple[Tensor, ...], attention_mask: Tensor | None = None, input_ids: NestedTensor | Tensor | None = None, labels: Tensor | None = None, output_name: str | None = None, **kwargs) -> HeadOutput

Forward pass of the TokenKMerHead.

Parameters:

Name Type Description Default
outputs
ModelOutput | Tuple[Tensor, ...]

The outputs of the model.

 required
attention_mask
Tensor | None

The attention mask for the inputs.

 None
input_ids
NestedTensor | Tensor | None

The input ids for the inputs.

 None
labels
Tensor | None

The labels for the head.

 None
output_name
str | None

The name of the output to use. Defaults to self.output_name.

 None
Source code in multimolecule/module/heads/token.py
Python
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
def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
    self,
    outputs: ModelOutput | Tuple[Tensor, ...],
    attention_mask: Tensor | None = None,
    input_ids: NestedTensor | Tensor | None = None,
    labels: Tensor | None = None,
    output_name: str | None = None,
    **kwargs,
) -> HeadOutput:
    r"""
    Forward pass of the TokenKMerHead.

    Args:
        outputs: The outputs of the model.
        attention_mask: The attention mask for the inputs.
        input_ids: The input ids for the inputs.
        labels: The labels for the head.
        output_name: The name of the output to use.
            Defaults to `self.output_name`.
    """
    if attention_mask is None:
        if isinstance(input_ids, NestedTensor):
            input_ids, attention_mask = input_ids.tensor, input_ids.mask
        else:
            if input_ids is None:
                raise ValueError(
                    f"Either attention_mask or input_ids must be provided for {self.__class__.__name__} to work."
                )
            if self.pad_token_id is None:
                raise ValueError(
                    f"pad_token_id must be provided when attention_mask is not passed to {self.__class__.__name__}."
                )
            attention_mask = input_ids.ne(self.pad_token_id)

    if isinstance(outputs, (Mapping, ModelOutput)):
        output = outputs[output_name or self.output_name]
    elif isinstance(outputs, tuple):
        output = outputs[0]
    else:
        raise ValueError(f"Unsupported type for outputs: {type(outputs)}")
    output = output * attention_mask.unsqueeze(-1)

    # remove cls token embeddings
    if self.bos_token_id is not None:
        output = output[..., 1:, :]
        attention_mask = attention_mask[..., 1:]
        if input_ids is not None:
            input_ids = input_ids[..., 1:]
    # remove eos token embeddings
    if self.eos_token_id is not None:
        if input_ids is not None:
            eos_mask = input_ids.ne(self.eos_token_id).to(output)
            input_ids = input_ids[..., :-1]
        else:
            last_valid_indices = attention_mask.sum(dim=-1)
            seq_length = attention_mask.size(-1)
            eos_mask = torch.arange(seq_length, device=output.device) == last_valid_indices.unsqueeze(1)
        output = output * eos_mask[:, :, None]
        output = output[..., :-1, :]
        attention_mask = attention_mask[..., 1:]

    output = self.unfold_kmer_embeddings(output, attention_mask)
    return super().forward(output, labels, **kwargs)

multimolecule.module.heads.contact

ContactPredictionHead

Bases: PredictionHead

Head for tasks in contact-level.

Performs symmetrization, and average product correct.

Parameters:

Name Type Description Default

config

PreTrainedConfig

The configuration object for the model.

 required

head_config

HeadConfig | None

The configuration object for the head. If None, will use configuration from the config.

 None
Source code in multimolecule/module/heads/contact.py
Python
 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
@HeadRegistry.register("contact")
class ContactPredictionHead(PredictionHead):
    r"""
    Head for tasks in contact-level.

    Performs symmetrization, and average product correct.

    Args:
        config: The configuration object for the model.
        head_config: The configuration object for the head.
            If None, will use configuration from the `config`.
    """

    output_name: str = "attentions"
    r"""The default output to use for the head."""

    def __init__(self, config: PreTrainedConfig, head_config: HeadConfig | None = None):
        super().__init__(config, head_config)
        self.bos_token_id = config.bos_token_id
        self.eos_token_id = config.eos_token_id
        self.pad_token_id = config.pad_token_id
        self.decoder = nn.Linear(
            config.num_hidden_layers * config.num_attention_heads, self.num_labels, bias=self.config.bias
        )
        if head_config is not None and head_config.output_name is not None:
            self.output_name = head_config.output_name

    def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
        self,
        outputs: ModelOutput | Mapping | Tuple[Tensor, ...],
        attention_mask: Tensor | None = None,
        input_ids: NestedTensor | Tensor | None = None,
        labels: Tensor | None = None,
        output_name: str | None = None,
        **kwargs,
    ) -> HeadOutput:
        r"""
        Forward pass of the ContactPredictionHead.

        Args:
            outputs: The outputs of the model.
            attention_mask: The attention mask for the inputs.
            input_ids: The input ids for the inputs.
            labels: The labels for the head.
            output_name: The name of the output to use.
                Defaults to `self.output_name`.
        """
        if attention_mask is None:
            if isinstance(input_ids, NestedTensor):
                input_ids, attention_mask = input_ids.tensor, input_ids.mask
            else:
                if input_ids is None:
                    raise ValueError(
                        f"Either attention_mask or input_ids must be provided for {self.__class__.__name__} to work."
                    )
                if self.pad_token_id is None:
                    raise ValueError(
                        f"pad_token_id must be provided when attention_mask is not passed to {self.__class__.__name__}."
                    )
                attention_mask = input_ids.ne(self.pad_token_id)

        if isinstance(outputs, (Mapping, ModelOutput)):
            output = outputs[output_name or self.output_name]
        elif isinstance(outputs, tuple):
            output = outputs[-1]
        attentions = torch.stack(output, 1)

        # In the original model, attentions for padding tokens are completely zeroed out.
        # This makes no difference most of the time because the other tokens won't attend to them,
        # but it does for the contact prediction task, which takes attentions as input,
        # so we have to mimic that here.
        attention_mask = attention_mask.unsqueeze(1) * attention_mask.unsqueeze(2)
        attentions *= attention_mask[:, None, None, :, :]

        # remove cls token attentions
        if self.bos_token_id is not None:
            attentions = attentions[..., 1:, 1:]
            # process attention_mask and input_ids to make removal of eos token happy
            attention_mask = attention_mask[..., 1:]
            if input_ids is not None:
                input_ids = input_ids[..., 1:]
        # remove eos token attentions
        if self.eos_token_id is not None:
            if input_ids is not None:
                eos_mask = input_ids.ne(self.eos_token_id).to(attentions)
            else:
                last_valid_indices = attention_mask.sum(dim=-1)
                seq_length = attention_mask.size(-1)
                eos_mask = torch.arange(seq_length, device=attentions.device).unsqueeze(0) == last_valid_indices
            eos_mask = eos_mask.unsqueeze(1) * eos_mask.unsqueeze(2)
            attentions *= eos_mask[:, None, None, :, :]
            attentions = attentions[..., :-1, :-1]

        # features: batch x channels x input_ids x input_ids (symmetric)
        batch_size, layers, heads, seqlen, _ = attentions.size()
        attentions = attentions.view(batch_size, layers * heads, seqlen, seqlen)
        attentions = attentions.to(self.decoder.weight.device)
        attentions = average_product_correct(symmetrize(attentions))
        attentions = attentions.permute(0, 2, 3, 1).squeeze(3)

        return super().forward(attentions, labels, **kwargs)

output_name class-attribute instance-attribute

Python
1
output_name: str = 'attentions'

The default output to use for the head.

forward

Python
1
forward(outputs: ModelOutput | Mapping | Tuple[Tensor, ...], attention_mask: Tensor | None = None, input_ids: NestedTensor | Tensor | None = None, labels: Tensor | None = None, output_name: str | None = None, **kwargs) -> HeadOutput

Forward pass of the ContactPredictionHead.

Parameters:

Name Type Description Default
outputs
ModelOutput | Mapping | Tuple[Tensor, ...]

The outputs of the model.

 required
attention_mask
Tensor | None

The attention mask for the inputs.

 None
input_ids
NestedTensor | Tensor | None

The input ids for the inputs.

 None
labels
Tensor | None

The labels for the head.

 None
output_name
str | None

The name of the output to use. Defaults to self.output_name.

 None
Source code in multimolecule/module/heads/contact.py
Python
 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
def forward(  # type: ignore[override]  # pylint: disable=arguments-renamed
    self,
    outputs: ModelOutput | Mapping | Tuple[Tensor, ...],
    attention_mask: Tensor | None = None,
    input_ids: NestedTensor | Tensor | None = None,
    labels: Tensor | None = None,
    output_name: str | None = None,
    **kwargs,
) -> HeadOutput:
    r"""
    Forward pass of the ContactPredictionHead.

    Args:
        outputs: The outputs of the model.
        attention_mask: The attention mask for the inputs.
        input_ids: The input ids for the inputs.
        labels: The labels for the head.
        output_name: The name of the output to use.
            Defaults to `self.output_name`.
    """
    if attention_mask is None:
        if isinstance(input_ids, NestedTensor):
            input_ids, attention_mask = input_ids.tensor, input_ids.mask
        else:
            if input_ids is None:
                raise ValueError(
                    f"Either attention_mask or input_ids must be provided for {self.__class__.__name__} to work."
                )
            if self.pad_token_id is None:
                raise ValueError(
                    f"pad_token_id must be provided when attention_mask is not passed to {self.__class__.__name__}."
                )
            attention_mask = input_ids.ne(self.pad_token_id)

    if isinstance(outputs, (Mapping, ModelOutput)):
        output = outputs[output_name or self.output_name]
    elif isinstance(outputs, tuple):
        output = outputs[-1]
    attentions = torch.stack(output, 1)

    # In the original model, attentions for padding tokens are completely zeroed out.
    # This makes no difference most of the time because the other tokens won't attend to them,
    # but it does for the contact prediction task, which takes attentions as input,
    # so we have to mimic that here.
    attention_mask = attention_mask.unsqueeze(1) * attention_mask.unsqueeze(2)
    attentions *= attention_mask[:, None, None, :, :]

    # remove cls token attentions
    if self.bos_token_id is not None:
        attentions = attentions[..., 1:, 1:]
        # process attention_mask and input_ids to make removal of eos token happy
        attention_mask = attention_mask[..., 1:]
        if input_ids is not None:
            input_ids = input_ids[..., 1:]
    # remove eos token attentions
    if self.eos_token_id is not None:
        if input_ids is not None:
            eos_mask = input_ids.ne(self.eos_token_id).to(attentions)
        else:
            last_valid_indices = attention_mask.sum(dim=-1)
            seq_length = attention_mask.size(-1)
            eos_mask = torch.arange(seq_length, device=attentions.device).unsqueeze(0) == last_valid_indices
        eos_mask = eos_mask.unsqueeze(1) * eos_mask.unsqueeze(2)
        attentions *= eos_mask[:, None, None, :, :]
        attentions = attentions[..., :-1, :-1]

    # features: batch x channels x input_ids x input_ids (symmetric)
    batch_size, layers, heads, seqlen, _ = attentions.size()
    attentions = attentions.view(batch_size, layers * heads, seqlen, seqlen)
    attentions = attentions.to(self.decoder.weight.device)
    attentions = average_product_correct(symmetrize(attentions))
    attentions = attentions.permute(0, 2, 3, 1).squeeze(3)

    return super().forward(attentions, labels, **kwargs)

multimolecule.module.heads.pretrain

MaskedLMHead

Bases: Module

Head for masked language modeling.

Parameters:

Name Type Description Default

config

PreTrainedConfig

The configuration object for the model.

 required

head_config

MaskedLMHeadConfig | None

The configuration object for the head. If None, will use configuration from the config.

 None
Source code in multimolecule/module/heads/pretrain.py
Python
 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
@HeadRegistry.register("masked_lm")
class MaskedLMHead(nn.Module):
    r"""
    Head for masked language modeling.

    Args:
        config: The configuration object for the model.
        head_config: The configuration object for the head.
            If None, will use configuration from the `config`.
    """

    output_name: str = "last_hidden_state"
    r"""The default output to use for the head."""

    def __init__(
        self, config: PreTrainedConfig, weight: Tensor | None = None, head_config: MaskedLMHeadConfig | None = None
    ):
        super().__init__()
        if head_config is None:
            head_config = config.lm_head if hasattr(config, "lm_head") else config.head  # type: ignore[assignment]
        self.config: MaskedLMHeadConfig = head_config  # type: ignore[assignment]
        if self.config.hidden_size is None:
            self.config.hidden_size = config.hidden_size
        self.num_labels = config.vocab_size
        self.dropout = nn.Dropout(self.config.dropout)
        self.transform = HeadTransformRegistryHF.build(self.config)
        self.decoder = nn.Linear(self.config.hidden_size, self.num_labels, bias=False)
        if weight is not None:
            self.decoder.weight = weight
        if self.config.bias:
            self.bias = nn.Parameter(torch.zeros(self.num_labels))
            self.decoder.bias = self.bias
        self.activation = ACT2FN[self.config.act] if self.config.act is not None else None
        if head_config is not None and head_config.output_name is not None:
            self.output_name = head_config.output_name

    def forward(
        self, outputs: ModelOutput | Tuple[Tensor, ...], labels: Tensor | None = None, output_name: str | None = None
    ) -> HeadOutput:
        r"""
        Forward pass of the MaskedLMHead.

        Args:
            outputs: The outputs of the model.
            labels: The labels for the head.
            output_name: The name of the output to use.
                Defaults to `self.output_name`.
        """
        if isinstance(outputs, (Mapping, ModelOutput)):
            output = outputs[output_name or self.output_name]
        elif isinstance(outputs, tuple):
            output = outputs[0]
        else:
            raise ValueError(f"Unsupported type for outputs: {type(outputs)}")
        output = self.dropout(output)
        output = self.transform(output)
        output = self.decoder(output)
        if self.activation is not None:
            output = self.activation(output)
        if labels is not None:
            if isinstance(labels, NestedTensor):
                if isinstance(output, Tensor):
                    output = labels.nested_like(output, strict=False)
                return HeadOutput(output, F.cross_entropy(torch.cat(output.storage()), torch.cat(labels.storage())))
            return HeadOutput(output, F.cross_entropy(output.view(-1, self.num_labels), labels.view(-1)))
        return HeadOutput(output)

output_name class-attribute instance-attribute

Python
1
output_name: str = 'last_hidden_state'

The default output to use for the head.

forward

Python
1
forward(outputs: ModelOutput | Tuple[Tensor, ...], labels: Tensor | None = None, output_name: str | None = None) -> HeadOutput

Forward pass of the MaskedLMHead.

Parameters:

Name Type Description Default
outputs
ModelOutput | Tuple[Tensor, ...]

The outputs of the model.

 required
labels
Tensor | None

The labels for the head.

 None
output_name
str | None

The name of the output to use. Defaults to self.output_name.

 None
Source code in multimolecule/module/heads/pretrain.py
Python
 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
def forward(
    self, outputs: ModelOutput | Tuple[Tensor, ...], labels: Tensor | None = None, output_name: str | None = None
) -> HeadOutput:
    r"""
    Forward pass of the MaskedLMHead.

    Args:
        outputs: The outputs of the model.
        labels: The labels for the head.
        output_name: The name of the output to use.
            Defaults to `self.output_name`.
    """
    if isinstance(outputs, (Mapping, ModelOutput)):
        output = outputs[output_name or self.output_name]
    elif isinstance(outputs, tuple):
        output = outputs[0]
    else:
        raise ValueError(f"Unsupported type for outputs: {type(outputs)}")
    output = self.dropout(output)
    output = self.transform(output)
    output = self.decoder(output)
    if self.activation is not None:
        output = self.activation(output)
    if labels is not None:
        if isinstance(labels, NestedTensor):
            if isinstance(output, Tensor):
                output = labels.nested_like(output, strict=False)
            return HeadOutput(output, F.cross_entropy(torch.cat(output.storage()), torch.cat(labels.storage())))
        return HeadOutput(output, F.cross_entropy(output.view(-1, self.num_labels), labels.view(-1)))
    return HeadOutput(output)

multimolecule.module.heads.generic

PredictionHead

Bases: Module

Head for all-level of tasks.

Parameters:

Name Type Description Default

config

PreTrainedConfig

The configuration object for the model.

 required

head_config

HeadConfig | None

The configuration object for the head. If None, will use configuration from the config.

 None
Source code in multimolecule/module/heads/generic.py
Python
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
class PredictionHead(nn.Module):
    r"""
    Head for all-level of tasks.

    Args:
        config: The configuration object for the model.
        head_config: The configuration object for the head.
            If None, will use configuration from the `config`.
    """

    num_labels: int

    def __init__(self, config: PreTrainedConfig, head_config: HeadConfig | None = None):
        super().__init__()
        if head_config is None:
            head_config = config.head
        self.config = head_config
        if self.config.hidden_size is None:
            self.config.hidden_size = config.hidden_size
        if self.config.num_labels is None:
            self.config.num_labels = config.num_labels
        if self.config.problem_type is None:
            self.config.problem_type = config.problem_type
        self.num_labels = self.config.num_labels
        self.dropout = nn.Dropout(self.config.dropout)
        self.transform = HeadTransformRegistryHF.build(self.config)
        self.decoder = nn.Linear(config.hidden_size, self.num_labels, bias=self.config.bias)
        self.activation = ACT2FN[self.config.act] if self.config.act is not None else None
        self.criterion = Criterion(self.config)

    def forward(self, embeddings: Tensor, labels: Tensor | None, **kwargs) -> HeadOutput:
        r"""
        Forward pass of the PredictionHead.

        Args:
            embeddings: The embeddings to be passed through the head.
            labels: The labels for the head.
        """
        if kwargs:
            warn(
                f"The following arguments are not applicable to {self.__class__.__name__}"
                f"and will be ignored: {kwargs.keys()}"
            )
        output = self.dropout(embeddings)
        output = self.transform(output)
        output = self.decoder(output)
        if self.activation is not None:
            output = self.activation(output)
        if labels is not None:
            if isinstance(labels, NestedTensor):
                if isinstance(output, Tensor):
                    output = labels.nested_like(output, strict=False)
                return HeadOutput(output, self.criterion(torch.cat(output.storage()), torch.cat(labels.storage())))
            return HeadOutput(output, self.criterion(output, labels))
        return HeadOutput(output)

forward

Python
1
forward(embeddings: Tensor, labels: Tensor | None, **kwargs) -> HeadOutput

Forward pass of the PredictionHead.

Parameters:

Name Type Description Default
embeddings
Tensor

The embeddings to be passed through the head.

 required
labels
Tensor | None

The labels for the head.

 required
Source code in multimolecule/module/heads/generic.py
Python
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
def forward(self, embeddings: Tensor, labels: Tensor | None, **kwargs) -> HeadOutput:
    r"""
    Forward pass of the PredictionHead.

    Args:
        embeddings: The embeddings to be passed through the head.
        labels: The labels for the head.
    """
    if kwargs:
        warn(
            f"The following arguments are not applicable to {self.__class__.__name__}"
            f"and will be ignored: {kwargs.keys()}"
        )
    output = self.dropout(embeddings)
    output = self.transform(output)
    output = self.decoder(output)
    if self.activation is not None:
        output = self.activation(output)
    if labels is not None:
        if isinstance(labels, NestedTensor):
            if isinstance(output, Tensor):
                output = labels.nested_like(output, strict=False)
            return HeadOutput(output, self.criterion(torch.cat(output.storage()), torch.cat(labels.storage())))
        return HeadOutput(output, self.criterion(output, labels))
    return HeadOutput(output)

multimolecule.module.heads.output

HeadOutput dataclass

Bases: ModelOutput

Output of a prediction head.

Parameters:

Name Type Description Default

logits

FloatTensor

The prediction logits from the head.

 required

loss

FloatTensor | None

The loss from the head. Defaults to None.

 None
Source code in multimolecule/module/heads/output.py
Python
25
26
27
28
29
30
31
32
33
34
35
36
37
@dataclass
class HeadOutput(ModelOutput):
    r"""
    Output of a prediction head.

    Args:
        logits: The prediction logits from the head.
        loss: The loss from the head.
            Defaults to None.
    """

    logits: FloatTensor
    loss: FloatTensor | None = None