跳转至

RibonanzaNet

The original implementation of RibonanzaNet applied dropout-residual-norm path twice to the output of the Self-Attention layer.

By default, the MultiMolecule follows the original implementation.

You can set fix_attention_norm=True in the model configuration to apply the dropout-residual-norm path once.

See more at issue #3

Caution

The MultiMolecule team is aware of a potential risk in reproducing the results of RibonanzaNet.

The original implementation of RibonanzaNet does not apply attention mask correctly.

By default, the MultiMolecule follows the original implementation.

You can set fix_attention_mask=True in the model configuration to apply the correct attention mask.

See more at issue #4, issue #5, and issue #7

Caution

The MultiMolecule team is aware of a potential risk in reproducing the results of RibonanzaNet.

The original implementation of RibonanzaNet applies dropout in an axis different from the one described in the paper.

By default, the MultiMolecule follows the original implementation.

You can set fix_pairwise_dropout=True in the model configuration to follow the description in the paper.

See more at issue #6

Tip

The MultiMolecule team has confirmed that the provided model and checkpoints are producing the same intermediate representations as the original implementation.

The team releasing RibonanzaNet did not write this model card for this model so this model card has been written by the MultiMolecule team.

Model Details

RibonanzaNet is a bert-style model pre-trained on a large corpus of RNA sequences. Please refer to the Training Details section for more information on the training process.

Model Specification

Num Layers Hidden Size Num Heads Intermediate Size Num Parameters (M) FLOPs (G) MACs (G) Max Num Tokens
9 256 8 1024 11.37 107.31 53.32 inf
  • Code: multimolecule.ribonanzanet
  • Weights: multimolecule/ribonanzanet
  • Data: Ribonanza
  • Paper: Ribonanza: deep learning of RNA structure through dual crowdsourcing
  • Developed by: Shujun He, Rui Huang, Jill Townley, Rachael C. Kretsch, Thomas G. Karagianes, David B.T. Cox, Hamish Blair, Dmitry Penzar, Valeriy Vyaltsev, Elizaveta Aristova, Arsenii Zinkevich, Artemy Bakulin, Hoyeol Sohn, Daniel Krstevski, Takaaki Fukui, Fumiya Tatematsu, Yusuke Uchida, Donghoon Jang, Jun Seong Lee, Roger Shieh, Tom Ma, Eduard Martynov, Maxim V. Shugaev, Habib S.T. Bukhari, Kazuki Fujikawa, Kazuki Onodera, Christof Henkel, Shlomo Ron, Jonathan Romano, John J. Nicol, Grace P. Nye, Yuan Wu, Christian Choe, Walter Reade, Eterna participants, Rhiju Das
  • Model type: BERT
  • Original Repository: Shujun-He/RibonanzaNet

Usage

The model file depends on the multimolecule library. You can install it using pip:

Bash
1
pip install multimolecule

Direct Use

You can use this model directly to predict chemical mapping:

Python
1
2
3
>>> from multimolecule import RnaTokenizer, RibonanzaNetForPreTraining
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/ribonanzanet")
>>> model = RibonanzaNetForPreTraining.from_pretrained("multimolecule/ribonanzanet")

Downstream Use

Extract Features

Here is how to use this model to get the features of a given sequence in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
from multimolecule import RnaTokenizer, RibonanzaNetModel


tokenizer = RnaTokenizer.from_pretrained("multimolecule/ribonanzanet")
model = RibonanzaNetModel.from_pretrained("multimolecule/ribonanzanet")

text = "UAGCUUAUCAGACUGAUGUUG"
input = tokenizer(text, return_tensors="pt")

output = model(**input)

Sequence Classification / Regression

Note: This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for sequence classification or regression.

Here is how to use this model as backbone to fine-tune for a sequence-level task in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
import torch
from multimolecule import RnaTokenizer, RibonanzaNetForSequencePrediction


tokenizer = RnaTokenizer.from_pretrained("multimolecule/ribonanzanet")
model = RibonanzaNetForSequencePrediction.from_pretrained("multimolecule/ribonanzanet")

text = "UAGCUUAUCAGACUGAUGUUG"
input = tokenizer(text, return_tensors="pt")
label = torch.tensor([1])

output = model(**input, labels=label)

Token Classification / Regression

Note: This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for nucleotide classification or regression.

Here is how to use this model as backbone to fine-tune for a nucleotide-level task in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
import torch
from multimolecule import RnaTokenizer, RibonanzaNetForTokenPrediction


tokenizer = RnaTokenizer.from_pretrained("multimolecule/ribonanzanet")
model = RibonanzaNetForTokenPrediction.from_pretrained("multimolecule/ribonanzanet")

text = "UAGCUUAUCAGACUGAUGUUG"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), ))

output = model(**input, labels=label)

Contact Classification / Regression

Note: This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for contact classification or regression.

Here is how to use this model as backbone to fine-tune for a contact-level task in PyTorch:

Python
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
import torch
from multimolecule import RnaTokenizer, RibonanzaNetForContactPrediction


tokenizer = RnaTokenizer.from_pretrained("multimolecule/ribonanzanet")
model = RibonanzaNetForContactPrediction.from_pretrained("multimolecule/ribonanzanet")

text = "UAGCUUAUCAGACUGAUGUUG"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), len(text)))

output = model(**input, labels=label)

Training Details

RibonanzaNet used Masked Language Modeling (MLM) as the pre-training objective: taking a sequence, the model randomly masks 15% of the tokens in the input then runs the entire masked sentence through the model and has to predict the masked tokens. This is comparable to the Cloze task in language modeling.

Training Data

The RibonanzaNet model was pre-trained on Ribonanza. RNAcentral is a free, public resource that offers integrated access to a comprehensive and up-to-date set of non-coding RNA sequences provided by a collaborating group of Expert Databases representing a broad range of organisms and RNA types.

RibonanzaNet applied CD-HIT (CD-HIT-EST) with a cut-off at 100% sequence identity to remove redundancy from the RNAcentral. The final dataset contains 23.7 million non-redundant RNA sequences.

RibonanzaNet preprocessed all tokens by replacing “U”s with “T”s.

Note that during model conversions, “T” is replaced with “U”. RnaTokenizer will convert “T”s to “U”s for you, you may disable this behaviour by passing replace_T_with_U=False.

Training Procedure

Preprocessing

RibonanzaNet used masked language modeling (MLM) as the pre-training objective. The masking procedure is similar to the one used in BERT:

  • 15% of the tokens are masked.
  • In 80% of the cases, the masked tokens are replaced by <mask>.
  • In 10% of the cases, the masked tokens are replaced by a random token (different) from the one they replace.
  • In the 10% remaining cases, the masked tokens are left as is.

PreTraining

The model was trained on 10 NVIDIA L40S GPUs with 48GiB memories.

  • Learning rate: 1e-4
  • Weight decay: 0.01

Citation

BibTeX:

BibTeX
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
@article{He2024.02.24.581671,
  author       = {He, Shujun and Huang, Rui and Townley, Jill and Kretsch, Rachael C. and Karagianes, Thomas G. and Cox, David B.T. and Blair, Hamish and Penzar, Dmitry and Vyaltsev, Valeriy and Aristova, Elizaveta and Zinkevich, Arsenii and Bakulin, Artemy and Sohn, Hoyeol and Krstevski, Daniel and Fukui, Takaaki and Tatematsu, Fumiya and Uchida, Yusuke and Jang, Donghoon and Lee, Jun Seong and Shieh, Roger and Ma, Tom and Martynov, Eduard and Shugaev, Maxim V. and Bukhari, Habib S.T. and Fujikawa, Kazuki and Onodera, Kazuki and Henkel, Christof and Ron, Shlomo and Romano, Jonathan and Nicol, John J. and Nye, Grace P. and Wu, Yuan and Choe, Christian and Reade, Walter and Eterna participants and Das, Rhiju},
  title        = {Ribonanza: deep learning of RNA structure through dual crowdsourcing},
  elocation-id = {2024.02.24.581671},
  year         = {2024},
  doi          = {10.1101/2024.02.24.581671},
  publisher    = {Cold Spring Harbor Laboratory},
  abstract     = {Prediction of RNA structure from sequence remains an unsolved problem, and progress has been slowed by a paucity of experimental data. Here, we present Ribonanza, a dataset of chemical mapping measurements on two million diverse RNA sequences collected through Eterna and other crowdsourced initiatives. Ribonanza measurements enabled solicitation, training, and prospective evaluation of diverse deep neural networks through a Kaggle challenge, followed by distillation into a single, self-contained model called RibonanzaNet. When fine tuned on auxiliary datasets, RibonanzaNet achieves state-of-the-art performance in modeling experimental sequence dropout, RNA hydrolytic degradation, and RNA secondary structure, with implications for modeling RNA tertiary structure.Competing Interest StatementStanford University is filing patent applications based on concepts described in this paper. R.D. is a cofounder of Inceptive.},
  url          = {https://www.biorxiv.org/content/early/2024/06/11/2024.02.24.581671},
  eprint       = {https://www.biorxiv.org/content/early/2024/06/11/2024.02.24.581671.full.pdf},
  journal      = {bioRxiv}
}

Contact

Please use GitHub issues of MultiMolecule for any questions or comments on the model card.

Please contact the authors of the RibonanzaNet paper for questions or comments on the paper/model.

License

This model is licensed under the AGPL-3.0 License.

Text Only
1
SPDX-License-Identifier: AGPL-3.0-or-later

multimolecule.models.ribonanzanet

RnaTokenizer

Bases: Tokenizer

Tokenizer for RNA sequences.

Parameters:

Name Type Description Default

alphabet

Alphabet | str | List[str] | None

alphabet to use for tokenization.

  • If is None, the standard RNA alphabet will be used.
  • If is a string, it should correspond to the name of a predefined alphabet. The options include
    • standard
    • extended
    • streamline
    • nucleobase
  • If is an alphabet or a list of characters, that specific alphabet will be used.
 None

nmers

int

Size of kmer to tokenize.

 1

codon

bool

Whether to tokenize into codons.

 False

replace_T_with_U

bool

Whether to replace T with U.

 True

do_upper_case

bool

Whether to convert input to uppercase.

 True

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
>>> from multimolecule import RnaTokenizer
>>> tokenizer = RnaTokenizer()
>>> tokenizer('<pad><cls><eos><unk><mask><null>ACGUNRYSWKMBDHV.X*-I')["input_ids"]
[1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 2]
>>> tokenizer('acgu')["input_ids"]
[1, 6, 7, 8, 9, 2]
>>> tokenizer('acgt')["input_ids"]
[1, 6, 7, 8, 9, 2]
>>> tokenizer = RnaTokenizer(replace_T_with_U=False)
>>> tokenizer('acgt')["input_ids"]
[1, 6, 7, 8, 3, 2]
>>> tokenizer = RnaTokenizer(nmers=3)
>>> tokenizer('uagcuuauc')["input_ids"]
[1, 83, 17, 64, 49, 96, 84, 22, 2]
>>> tokenizer = RnaTokenizer(codon=True)
>>> tokenizer('uagcuuauc')["input_ids"]
[1, 83, 49, 22, 2]
>>> tokenizer('uagcuuauca')["input_ids"]
Traceback (most recent call last):
ValueError: length of input sequence must be a multiple of 3 for codon tokenization, but got 10
Source code in multimolecule/tokenisers/rna/tokenization_rna.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
 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
class RnaTokenizer(Tokenizer):
    """
    Tokenizer for RNA sequences.

    Args:
        alphabet: alphabet to use for tokenization.

            - If is `None`, the standard RNA alphabet will be used.
            - If is a `string`, it should correspond to the name of a predefined alphabet. The options include
                + `standard`
                + `extended`
                + `streamline`
                + `nucleobase`
            - If is an alphabet or a list of characters, that specific alphabet will be used.
        nmers: Size of kmer to tokenize.
        codon: Whether to tokenize into codons.
        replace_T_with_U: Whether to replace T with U.
        do_upper_case: Whether to convert input to uppercase.

    Examples:
        >>> from multimolecule import RnaTokenizer
        >>> tokenizer = RnaTokenizer()
        >>> tokenizer('<pad><cls><eos><unk><mask><null>ACGUNRYSWKMBDHV.X*-I')["input_ids"]
        [1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 2]
        >>> tokenizer('acgu')["input_ids"]
        [1, 6, 7, 8, 9, 2]
        >>> tokenizer('acgt')["input_ids"]
        [1, 6, 7, 8, 9, 2]
        >>> tokenizer = RnaTokenizer(replace_T_with_U=False)
        >>> tokenizer('acgt')["input_ids"]
        [1, 6, 7, 8, 3, 2]
        >>> tokenizer = RnaTokenizer(nmers=3)
        >>> tokenizer('uagcuuauc')["input_ids"]
        [1, 83, 17, 64, 49, 96, 84, 22, 2]
        >>> tokenizer = RnaTokenizer(codon=True)
        >>> tokenizer('uagcuuauc')["input_ids"]
        [1, 83, 49, 22, 2]
        >>> tokenizer('uagcuuauca')["input_ids"]
        Traceback (most recent call last):
        ValueError: length of input sequence must be a multiple of 3 for codon tokenization, but got 10
    """

    model_input_names = ["input_ids", "attention_mask"]

    def __init__(
        self,
        alphabet: Alphabet | str | List[str] | None = None,
        nmers: int = 1,
        codon: bool = False,
        replace_T_with_U: bool = True,
        do_upper_case: bool = True,
        additional_special_tokens: List | Tuple | None = None,
        **kwargs,
    ):
        if codon and (nmers > 1 and nmers != 3):
            raise ValueError("Codon and nmers cannot be used together.")
        if codon:
            nmers = 3  # set to 3 to get correct vocab
        if not isinstance(alphabet, Alphabet):
            alphabet = get_alphabet(alphabet, nmers=nmers)
        super().__init__(
            alphabet=alphabet,
            nmers=nmers,
            codon=codon,
            replace_T_with_U=replace_T_with_U,
            do_upper_case=do_upper_case,
            additional_special_tokens=additional_special_tokens,
            **kwargs,
        )
        self.replace_T_with_U = replace_T_with_U
        self.nmers = nmers
        self.codon = codon

    def _tokenize(self, text: str, **kwargs):
        if self.do_upper_case:
            text = text.upper()
        if self.replace_T_with_U:
            text = text.replace("T", "U")
        if self.codon:
            if len(text) % 3 != 0:
                raise ValueError(
                    f"length of input sequence must be a multiple of 3 for codon tokenization, but got {len(text)}"
                )
            return [text[i : i + 3] for i in range(0, len(text), 3)]
        if self.nmers > 1:
            return [text[i : i + self.nmers] for i in range(len(text) - self.nmers + 1)]  # noqa: E203
        return list(text)

RibonanzaNetConfig

Bases: PreTrainedConfig

This is the configuration class to store the configuration of a RibonanzaNetModel. It is used to instantiate a RibonanzaNet model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the RibonanzaNet Shujun-He/RibonanzaNet architecture.

Configuration objects inherit from PreTrainedConfig and can be used to control the model outputs. Read the documentation from PreTrainedConfig for more information.

Parameters:

Name Type Description Default

vocab_size

int

Vocabulary size of the RibonanzaNet model. Defines the number of different tokens that can be represented by the inputs_ids passed when calling [RibonanzaNetModel].

 26

hidden_size

int

Dimensionality of the encoder layers and the pooler layer.

 256

num_hidden_layers

int

Number of hidden layers in the Transformer encoder.

 9

num_attention_heads

int

Number of attention heads for each attention layer in the Transformer encoder.

 8

intermediate_size

int

Dimensionality of the “intermediate” (often named feed-forward) layer in the Transformer encoder.

 1024

hidden_act

str

The non-linear activation function (function or string) in the encoder and pooler. If string, "gelu", "relu", "silu" and "gelu_new" are supported.

 'gelu'

hidden_dropout

float

The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.

 0.05

attention_dropout

float

The dropout ratio for the attention probabilities.

 0.05

max_position_embeddings

The maximum sequence length that this model might ever be used with. Typically set this to something large just in case (e.g., 512 or 1024 or 2048).

 required

initializer_range

float

The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

 0.02

layer_norm_eps

float

The epsilon used by the layer normalization layers.

 1e-12

position_embedding_type

Type of position embedding. Choose one of "absolute", "relative_key", "relative_key_query". For positional embeddings use "absolute". For more information on "relative_key", please refer to Self-Attention with Relative Position Representations (Shaw et al.). For more information on "relative_key_query", please refer to Method 4 in Improve Transformer Models with Better Relative Position Embeddings (Huang et al.).

 required

is_decoder

Whether the model is used as a decoder or not. If False, the model is used as an encoder.

 required

use_cache

Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if config.is_decoder=True.

 required

emb_layer_norm_before

Whether to apply layer normalization after embeddings but before the main stem of the network.

 required

token_dropout

When this is enabled, masked tokens are treated as if they had been dropped out by input dropout.

 required

head

HeadConfig | None

The configuration of the head.

 None

lm_head

MaskedLMHeadConfig | None

The configuration of the masked language model head.

 None

Examples:

Python Console Session
1
2
3
4
5
6
7
>>> from multimolecule import RibonanzaNetConfig, RibonanzaNetModel
>>> # Initializing a RibonanzaNet multimolecule/ribonanzanet style configuration
>>> configuration = RibonanzaNetConfig()
>>> # Initializing a model (with random weights) from the multimolecule/ribonanzanet style configuration
>>> model = RibonanzaNetModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
Source code in multimolecule/models/ribonanzanet/configuration_ribonanzanet.py
Python
 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
class RibonanzaNetConfig(PreTrainedConfig):
    r"""
    This is the configuration class to store the configuration of a
    [`RibonanzaNetModel`][multimolecule.models.RibonanzaNetModel].
    It is used to instantiate a RibonanzaNet model according to the specified arguments, defining the model
    architecture.
    Instantiating a configuration with the defaults will yield a similar configuration to that of the RibonanzaNet
    [Shujun-He/RibonanzaNet](https://github.com/Shujun-He/RibonanzaNet) architecture.

    Configuration objects inherit from [`PreTrainedConfig`][multimolecule.models.PreTrainedConfig] and can be used to
    control the model outputs. Read the documentation from [`PreTrainedConfig`][multimolecule.models.PreTrainedConfig]
    for more information.

    Args:
        vocab_size:
            Vocabulary size of the RibonanzaNet model. Defines the number of different tokens that can be represented by
            the `inputs_ids` passed when calling [`RibonanzaNetModel`].
        hidden_size:
            Dimensionality of the encoder layers and the pooler layer.
        num_hidden_layers:
            Number of hidden layers in the Transformer encoder.
        num_attention_heads:
            Number of attention heads for each attention layer in the Transformer encoder.
        intermediate_size:
            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
        hidden_act:
            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
            `"relu"`, `"silu"` and `"gelu_new"` are supported.
        hidden_dropout:
            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
        attention_dropout:
            The dropout ratio for the attention probabilities.
        max_position_embeddings:
            The maximum sequence length that this model might ever be used with. Typically set this to something large
            just in case (e.g., 512 or 1024 or 2048).
        initializer_range:
            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
        layer_norm_eps:
            The epsilon used by the layer normalization layers.
        position_embedding_type:
            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
        is_decoder:
            Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
        use_cache:
            Whether or not the model should return the last key/values attentions (not used by all models). Only
            relevant if `config.is_decoder=True`.
        emb_layer_norm_before:
            Whether to apply layer normalization after embeddings but before the main stem of the network.
        token_dropout:
            When this is enabled, masked tokens are treated as if they had been dropped out by input dropout.
        head:
            The configuration of the head.
        lm_head:
            The configuration of the masked language model head.

    Examples:
        >>> from multimolecule import RibonanzaNetConfig, RibonanzaNetModel
        >>> # Initializing a RibonanzaNet multimolecule/ribonanzanet style configuration
        >>> configuration = RibonanzaNetConfig()
        >>> # Initializing a model (with random weights) from the multimolecule/ribonanzanet style configuration
        >>> model = RibonanzaNetModel(configuration)
        >>> # Accessing the model configuration
        >>> configuration = model.config
    """

    model_type = "ribonanzanet"

    def __init__(
        self,
        vocab_size: int = 26,
        hidden_size: int = 256,
        num_hidden_layers: int = 9,
        num_attention_heads: int = 8,
        intermediate_size: int = 1024,
        pairwise_size: int = 64,
        pairwise_attention_size: int = 32,
        pairwise_intermediate_size: int = 256,
        pairwise_num_attention_heads: int = 4,
        kernel_size: int = 5,
        use_triangular_attention: bool = False,
        hidden_act: str = "gelu",
        pairwise_hidden_act: str = "relu",
        hidden_dropout: float = 0.05,
        attention_dropout: float = 0.05,
        output_pairwise_states: bool = False,
        initializer_range: float = 0.02,
        layer_norm_eps: float = 1e-12,
        head: HeadConfig | None = None,
        lm_head: MaskedLMHeadConfig | None = None,
        fix_attention_mask: bool = False,
        fix_attention_norm: bool = False,
        fix_pairwise_dropout: bool = False,
        **kwargs,
    ):
        super().__init__(**kwargs)
        self.vocab_size = vocab_size
        self.type_vocab_size = 2
        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.pairwise_size = pairwise_size
        self.pairwise_attention_size = pairwise_attention_size
        self.pairwise_intermediate_size = pairwise_intermediate_size
        self.pairwise_num_attention_heads = pairwise_num_attention_heads
        self.kernel_size = kernel_size
        self.use_triangular_attention = use_triangular_attention
        self.hidden_act = hidden_act
        self.pairwise_hidden_act = pairwise_hidden_act
        self.hidden_dropout = hidden_dropout
        self.attention_dropout = attention_dropout
        self.output_pairwise_states = output_pairwise_states
        self.initializer_range = initializer_range
        self.layer_norm_eps = layer_norm_eps
        self.head = HeadConfig(**head) if head is not None else None
        self.lm_head = MaskedLMHeadConfig(**lm_head) if lm_head is not None else None
        self.fix_attention_mask = fix_attention_mask
        self.fix_attention_norm = fix_attention_norm
        self.fix_pairwise_dropout = fix_pairwise_dropout

RibonanzaNetForContactPrediction

Bases: RibonanzaNetPreTrainedModel

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForContactPrediction, RnaTokenizer
>>> config = RibonanzaNetConfig()
>>> model = RibonanzaNetForContactPrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.randint(2, (1, 5, 5)))
>>> output["logits"].shape
torch.Size([1, 5, 5, 1])
>>> output["loss"]
tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
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
class RibonanzaNetForContactPrediction(RibonanzaNetPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForContactPrediction, RnaTokenizer
        >>> config = RibonanzaNetConfig()
        >>> model = RibonanzaNetForContactPrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.randint(2, (1, 5, 5)))
        >>> output["logits"].shape
        torch.Size([1, 5, 5, 1])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
    """

    def __init__(self, config: RibonanzaNetConfig):
        super().__init__(config)
        self.ribonanzanet = RibonanzaNetModel(config)
        self.contact_head = ContactPredictionHead(config)
        self.head_config = self.contact_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        head_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        output_pairwise_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RibonanzaNetContactPredictorOutput:
        if output_attentions is False:
            warn("output_attentions must be True for contact classification and will be ignored.")
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.ribonanzanet(
            input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=True,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.contact_head(outputs, attention_mask, input_ids, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RibonanzaNetContactPredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            pairwise_states=outputs.pairwise_states,
            attentions=outputs.attentions,
        )

RibonanzaNetForSecondaryStructurePrediction

Bases: RibonanzaNetPreTrainedModel

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForSecondaryStructurePrediction, RnaTokenizer
>>> config = RibonanzaNetConfig()
>>> model = RibonanzaNetForSecondaryStructurePrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.randint(2, (1, 5, 5)))
>>> output["logits"].shape
torch.Size([1, 5, 5, 1])
>>> output["loss"]
tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
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
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
class RibonanzaNetForSecondaryStructurePrediction(RibonanzaNetPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForSecondaryStructurePrediction, RnaTokenizer
        >>> config = RibonanzaNetConfig()
        >>> model = RibonanzaNetForSecondaryStructurePrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.randint(2, (1, 5, 5)))
        >>> output["logits"].shape
        torch.Size([1, 5, 5, 1])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
    """

    def __init__(self, config: RibonanzaNetConfig):
        super().__init__(config)
        self.ribonanzanet = RibonanzaNetModel(config, add_pooling_layer=False)
        self.ss_head = RibonanzaNetForSecondaryStructurePredictionHead(config)
        self.decoder = nn.Linear(config.hidden_size, config.num_labels)
        self.head_config = self.ss_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        head_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        output_pairwise_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RibonanzaNetForSecondaryStructurePredictorOutput:
        if not output_pairwise_states:
            warn("output_pairwise_states must be True since prediction head requires pairwise states.")
            output_pairwise_states = True
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.ribonanzanet(
            input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_pairwise_states=output_pairwise_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.ss_head(outputs, attention_mask, input_ids, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RibonanzaNetForSecondaryStructurePredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            pairwise_states=outputs.pairwise_states,
            attentions=outputs.attentions,
        )

RibonanzaNetForSequencePrediction

Bases: RibonanzaNetPreTrainedModel

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForSequencePrediction, RnaTokenizer
>>> config = RibonanzaNetConfig()
>>> model = RibonanzaNetForSequencePrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.tensor([[1]]))
>>> output["logits"].shape
torch.Size([1, 1])
>>> output["loss"]
tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
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
class RibonanzaNetForSequencePrediction(RibonanzaNetPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForSequencePrediction, RnaTokenizer
        >>> config = RibonanzaNetConfig()
        >>> model = RibonanzaNetForSequencePrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.tensor([[1]]))
        >>> output["logits"].shape
        torch.Size([1, 1])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
    """

    def __init__(self, config: RibonanzaNetConfig):
        super().__init__(config)
        self.ribonanzanet = RibonanzaNetModel(config)
        self.sequence_head = SequencePredictionHead(config)
        self.head_config = self.sequence_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        head_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        output_pairwise_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RibonanzaNetSequencePredictorOutput:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.ribonanzanet(
            input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_pairwise_states=output_pairwise_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.sequence_head(outputs, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RibonanzaNetSequencePredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            pairwise_states=outputs.pairwise_states,
            attentions=outputs.attentions,
        )

RibonanzaNetForTokenPrediction

Bases: RibonanzaNetPreTrainedModel

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForTokenPrediction, RnaTokenizer
>>> config = RibonanzaNetConfig()
>>> model = RibonanzaNetForTokenPrediction(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=torch.randint(2, (1, 5)))
>>> output["logits"].shape
torch.Size([1, 5, 1])
>>> output["loss"]
tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
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
class RibonanzaNetForTokenPrediction(RibonanzaNetPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RibonanzaNetConfig, RibonanzaNetForTokenPrediction, RnaTokenizer
        >>> config = RibonanzaNetConfig()
        >>> model = RibonanzaNetForTokenPrediction(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input, labels=torch.randint(2, (1, 5)))
        >>> output["logits"].shape
        torch.Size([1, 5, 1])
        >>> output["loss"]  # doctest:+ELLIPSIS
        tensor(..., grad_fn=<BinaryCrossEntropyWithLogitsBackward0>)
    """

    def __init__(self, config: RibonanzaNetConfig):
        super().__init__(config)
        self.ribonanzanet = RibonanzaNetModel(config)
        self.token_head = TokenPredictionHead(config)
        self.head_config = self.token_head.config

        # Initialize weights and apply final processing
        self.post_init()

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        head_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        labels: Tensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        output_pairwise_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RibonanzaNetTokenPredictorOutput:
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        outputs = self.ribonanzanet(
            input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_pairwise_states=output_pairwise_states,
            return_dict=return_dict,
            **kwargs,
        )
        output = self.token_head(outputs, attention_mask, input_ids, labels)
        logits, loss = output.logits, output.loss

        if not return_dict:
            output = (logits,) + outputs[2:]
            return ((loss,) + output) if loss is not None else output

        return RibonanzaNetTokenPredictorOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            pairwise_states=outputs.pairwise_states,
            attentions=outputs.attentions,
        )

RibonanzaNetModel

Bases: RibonanzaNetPreTrainedModel

Examples:

Python Console Session
 1
 2
 3
 4
 5
 6
 7
 8
 9
10
>>> from multimolecule import RibonanzaNetConfig, RibonanzaNetModel, RnaTokenizer
>>> config = RibonanzaNetConfig()
>>> model = RibonanzaNetModel(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input)
>>> output["last_hidden_state"].shape
torch.Size([1, 7, 256])
>>> output["pooler_output"].shape
torch.Size([1, 256])
Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
 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
class RibonanzaNetModel(RibonanzaNetPreTrainedModel):
    """
    Examples:
        >>> from multimolecule import RibonanzaNetConfig, RibonanzaNetModel, RnaTokenizer
        >>> config = RibonanzaNetConfig()
        >>> model = RibonanzaNetModel(config)
        >>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
        >>> input = tokenizer("ACGUN", return_tensors="pt")
        >>> output = model(**input)
        >>> output["last_hidden_state"].shape
        torch.Size([1, 7, 256])
        >>> output["pooler_output"].shape
        torch.Size([1, 256])
    """

    def __init__(self, config: RibonanzaNetConfig, add_pooling_layer: bool = True):
        super().__init__(config)
        self.pad_token_id = config.pad_token_id
        self.embeddings = RibonanzaNetEmbeddings(config)
        self.encoder = RibonanzaNetEncoder(config)
        self.pooler = RibonanzaNetPooler(config) if add_pooling_layer else None
        self.fix_attention_mask = config.fix_attention_mask

        # Initialize weights and apply final processing
        self.post_init()

    def get_input_embeddings(self):
        return self.embeddings.word_embeddings

    def set_input_embeddings(self, value):
        self.embeddings.word_embeddings = value

    def _prune_heads(self, heads_to_prune):
        """
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        """
        for layer, heads in heads_to_prune.items():
            self.encoder.layer[layer].attention.prune_heads(heads)

    def forward(
        self,
        input_ids: Tensor | NestedTensor,
        attention_mask: Tensor | None = None,
        head_mask: Tensor | None = None,
        inputs_embeds: Tensor | NestedTensor | None = None,
        output_attentions: bool | None = None,
        output_hidden_states: bool | None = None,
        output_pairwise_states: bool | None = None,
        return_dict: bool | None = None,
        **kwargs,
    ) -> Tuple[Tensor, ...] | RibonanzaNetModelOutputWithPooling:
        if kwargs:
            warn(
                f"Additional keyword arguments `{', '.join(kwargs)}` are detected in "
                f"`{self.__class__.__name__}.forward`, they will be ignored.\n"
                "This is provided for backward compatibility and may lead to unexpected behavior."
            )
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        output_pairwise_states = (
            output_pairwise_states if output_pairwise_states is not None else self.config.output_pairwise_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        if isinstance(input_ids, NestedTensor):
            input_ids, attention_mask = input_ids.tensor, input_ids.mask
        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        if input_ids is not None:
            self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
            input_shape = input_ids.size()
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds")

        batch_size, seq_length = input_shape
        device = input_ids.device if input_ids is not None else inputs_embeds.device  # type: ignore[union-attr]

        if attention_mask is None:
            attention_mask = (
                input_ids.ne(self.pad_token_id)
                if self.pad_token_id is not None
                else torch.ones(((batch_size, seq_length)), device=device)
            )
        else:
            # Must make a clone here because the attention mask might be reused in other modules
            # and we need to process it to mimic the behavior of the original implementation.
            # See more in https://github.com/Shujun-He/RibonanzaNet/issues/4
            attention_mask = attention_mask.clone()

        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        extended_attention_mask: Tensor = self.get_extended_attention_mask(attention_mask, input_shape)
        attention_mask = attention_mask.float().unsqueeze(-1)

        # Prepare head mask if needed
        # 1.0 in head_mask indicate we keep the head
        # attention_probs has shape bsz x n_heads x N x N
        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)

        embedding_output = self.embeddings(
            input_ids=input_ids,
            inputs_embeds=inputs_embeds,
        )
        encoder_outputs = self.encoder(
            embedding_output,
            attention_mask=attention_mask,
            extended_attention_mask=extended_attention_mask,
            head_mask=head_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            output_pairwise_states=output_pairwise_states,
            return_dict=return_dict,
        )
        sequence_output = encoder_outputs[0]
        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None

        if not return_dict:
            return (sequence_output, pooled_output) + encoder_outputs[1:]

        return RibonanzaNetModelOutputWithPooling(
            last_hidden_state=sequence_output,
            pooler_output=pooled_output,
            hidden_states=encoder_outputs.hidden_states,
            pairwise_states=encoder_outputs.pairwise_states,
            attentions=encoder_outputs.attentions,
        )

    def get_extended_attention_mask(
        self, attention_mask: Tensor, input_shape: Tuple[int], dtype: torch.dtype | None = None
    ) -> Tensor:
        """
        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

        Arguments:
            attention_mask (`torch.Tensor`):
                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
            input_shape (`Tuple[int]`):
                The shape of the input to the model.

        Returns:
            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
        """
        if dtype is None:
            dtype = self.dtype

        if attention_mask.dim() == 2:
            attention_mask = attention_mask.unsqueeze(-1)
            if not self.fix_attention_mask:
                attention_mask[attention_mask == 0] = -1
            attention_mask = torch.matmul(attention_mask, attention_mask.transpose(1, 2))
        elif attention_mask.shape != 3:
            raise ValueError(
                f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
            )

        extended_attention_mask = attention_mask[:, None, :, :]
        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
        # masked positions, this operation will create a tensor which is 0.0 for
        # positions we want to attend and the dtype's smallest value for masked positions.
        # Since we are adding it to the raw scores before the softmax, this is
        # effectively the same as removing these entirely.
        extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility
        if self.fix_attention_mask:
            extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
        return extended_attention_mask

_prune_heads

Python
_prune_heads(heads_to_prune)

Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel

Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
122
123
124
125
126
127
128
def _prune_heads(self, heads_to_prune):
    """
    Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
    class PreTrainedModel
    """
    for layer, heads in heads_to_prune.items():
        self.encoder.layer[layer].attention.prune_heads(heads)

get_extended_attention_mask

Python
get_extended_attention_mask(attention_mask: Tensor, input_shape: Tuple[int], dtype: dtype | None = None) -> Tensor

Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

Parameters:

Name Type Description Default
attention_mask
`torch.Tensor`

Mask with ones indicating tokens to attend to, zeros for tokens to ignore.

 required
input_shape
`Tuple[int]`

The shape of the input to the model.

 required

Returns:

Type Description
Tensor

torch.Tensor The extended attention mask, with a the same dtype as attention_mask.dtype.
Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
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
def get_extended_attention_mask(
    self, attention_mask: Tensor, input_shape: Tuple[int], dtype: torch.dtype | None = None
) -> Tensor:
    """
    Makes broadcastable attention and causal masks so that future and masked tokens are ignored.

    Arguments:
        attention_mask (`torch.Tensor`):
            Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
        input_shape (`Tuple[int]`):
            The shape of the input to the model.

    Returns:
        `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
    """
    if dtype is None:
        dtype = self.dtype

    if attention_mask.dim() == 2:
        attention_mask = attention_mask.unsqueeze(-1)
        if not self.fix_attention_mask:
            attention_mask[attention_mask == 0] = -1
        attention_mask = torch.matmul(attention_mask, attention_mask.transpose(1, 2))
    elif attention_mask.shape != 3:
        raise ValueError(
            f"Wrong shape for input_ids (shape {input_shape}) or attention_mask (shape {attention_mask.shape})"
        )

    extended_attention_mask = attention_mask[:, None, :, :]
    # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
    # masked positions, this operation will create a tensor which is 0.0 for
    # positions we want to attend and the dtype's smallest value for masked positions.
    # Since we are adding it to the raw scores before the softmax, this is
    # effectively the same as removing these entirely.
    extended_attention_mask = extended_attention_mask.to(dtype=dtype)  # fp16 compatibility
    if self.fix_attention_mask:
        extended_attention_mask = (1.0 - extended_attention_mask) * torch.finfo(dtype).min
    return extended_attention_mask

RibonanzaNetPreTrainedModel

Bases: PreTrainedModel

An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models.

Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
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
class RibonanzaNetPreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = RibonanzaNetConfig
    base_model_prefix = "ribonanzanet"
    supports_gradient_checkpointing = True
    _no_split_modules = ["RibonanzaNetLayer", "RibonanzaNetEmbeddings"]

    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
    def _init_weights(self, module: nn.Module):
        """Initialize the weights"""
        if isinstance(module, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()
        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)
        for n, m in module.named_modules():
            if "_gate" in n:
                m.weight.data.zero_()
                m.bias.data.fill_(1.0)

_init_weights

Python
_init_weights(module: Module)

Initialize the weights

Source code in multimolecule/models/ribonanzanet/modeling_ribonanzanet.py
Python
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
def _init_weights(self, module: nn.Module):
    """Initialize the weights"""
    if isinstance(module, nn.Linear):
        # Slightly different from the TF version which uses truncated_normal for initialization
        # cf https://github.com/pytorch/pytorch/pull/5617
        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        if module.bias is not None:
            module.bias.data.zero_()
    elif isinstance(module, nn.Embedding):
        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
        if module.padding_idx is not None:
            module.weight.data[module.padding_idx].zero_()
    elif isinstance(module, nn.LayerNorm):
        module.bias.data.zero_()
        module.weight.data.fill_(1.0)
    for n, m in module.named_modules():
        if "_gate" in n:
            m.weight.data.zero_()
            m.bias.data.fill_(1.0)