SpliceBERT
Pre-trained model on messenger RNA precursor (pre-mRNA) using a masked language modeling (MLM) objective.
Disclaimer
This is an UNOFFICIAL implementation of the Self-supervised learning on millions of pre-mRNA sequences improves sequence-based RNA splicing prediction by Ken Chen, et al.
The OFFICIAL repository of SpliceBERT is at chenkenbio/SpliceBERT.
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 SpliceBERT did not write this model card for this model so this model card has been written by the MultiMolecule team.
Model Details
SpliceBERT is a bert-style model pre-trained on a large corpus of messenger RNA precursor sequences in a self-supervised fashion. This means that the model was trained on the raw nucleotides of RNA sequences only, with an automatic process to generate inputs and labels from those texts. Please refer to the Training Details section for more information on the training process.
Variations
Model Specification
| Variants |
Num Layers |
Hidden Size |
Num Heads |
Intermediate Size |
Num Parameters (M) |
FLOPs (G) |
MACs (G) |
Max Num Tokens |
| splicebert |
6 |
512 |
16 |
2048 |
19.72 |
5.04 |
2.52 |
1024 |
| splicebert.510nt |
19.45 |
510 |
| splicebert-human.510nt |
Links
Usage
The model file depends on the multimolecule library. You can install it using pip:
| Bash |
|---|
| pip install multimolecule
|
Direct Use
You can use this model directly with a pipeline for masked language modeling:
| Python |
|---|
| >>> import multimolecule # you must import multimolecule to register models
>>> from transformers import pipeline
>>> unmasker = pipeline("fill-mask", model="multimolecule/splicebert")
>>> unmasker("gguc<mask>cucugguuagaccagaucugagccu")
[{'score': 0.340412974357605,
'token': 9,
'token_str': 'U',
'sequence': 'G G U C U C U C U G G U U A G A C C A G A U C U G A G C C U'},
{'score': 0.13882005214691162,
'token': 12,
'token_str': 'Y',
'sequence': 'G G U C Y C U C U G G U U A G A C C A G A U C U G A G C C U'},
{'score': 0.056610625237226486,
'token': 7,
'token_str': 'C',
'sequence': 'G G U C C C U C U G G U U A G A C C A G A U C U G A G C C U'},
{'score': 0.05455885827541351,
'token': 19,
'token_str': 'H',
'sequence': 'G G U C H C U C U G G U U A G A C C A G A U C U G A G C C U'},
{'score': 0.05356108024716377,
'token': 14,
'token_str': 'W',
'sequence': 'G G U C W C U C U G G U U A G A C C A G A U C U G A G C C U'}]
|
Downstream Use
Here is how to use this model to get the features of a given sequence in PyTorch:
| Python |
|---|
| from multimolecule import RnaTokenizer, SpliceBertModel
tokenizer = RnaTokenizer.from_pretrained("multimolecule/splicebert")
model = SpliceBertModel.from_pretrained("multimolecule/splicebert")
text = "UAGCUUAUCAGACUGAUGUUGA"
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 |
|---|
| import torch
from multimolecule import RnaTokenizer, SpliceBertForSequencePrediction
tokenizer = RnaTokenizer.from_pretrained("multimolecule/splicebert")
model = SpliceBertForSequencePrediction.from_pretrained("multimolecule/splicebert")
text = "UAGCUUAUCAGACUGAUGUUGA"
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 |
|---|
| import torch
from multimolecule import RnaTokenizer, SpliceBertForTokenPrediction
tokenizer = RnaTokenizer.from_pretrained("multimolecule/splicebert")
model = SpliceBertForTokenPrediction.from_pretrained("multimolecule/splicebert")
text = "UAGCUUAUCAGACUGAUGUUGA"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), ))
output = model(**input, labels=label)
|
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 |
|---|
| import torch
from multimolecule import RnaTokenizer, SpliceBertForContactPrediction
tokenizer = RnaTokenizer.from_pretrained("multimolecule/splicebert")
model = SpliceBertForContactPrediction.from_pretrained("multimolecule/splicebert")
text = "UAGCUUAUCAGACUGAUGUUGA"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), len(text)))
output = model(**input, labels=label)
|
Training Details
SpliceBERT 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 SpliceBERT model was pre-trained on messenger RNA precursor sequences from UCSC Genome Browser.
UCSC Genome Browser provides visualization, analysis, and download of comprehensive vertebrate genome data with aligned annotation tracks (known genes, predicted genes, ESTs, mRNAs, CpG islands, etc.).
SpliceBERT collected reference genomes and gene annotations from the UCSC Genome Browser for 72 vertebrate species. It applied bedtools getfasta to extract pre-mRNA sequences from the reference genomes based on the gene annotations. The pre-mRNA sequences are then used to pre-train SpliceBERT. The pre-training data contains 2 million pre-mRNA sequences with a total length of 65 billion nucleotides.
Note 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
SpliceBERT 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 8 NVIDIA V100 GPUs.
- Learning rate: 1e-4
- Learning rate scheduler: ReduceLROnPlateau(patience=3)
- Optimizer: AdamW
SpliceBERT trained model in a two-stage training process:
- Pre-train with sequences of a fixed length of 510 nucleotides.
- Pre-train with sequences of a variable length between 64 and 1024 nucleotides.
The intermediate model after the first stage is available as multimolecule/splicebert.510nt.
SpliceBERT also pre-trained a model on human data only to validate the contribution of multi-species pre-training. The intermediate model after the first stage is available as multimolecule/splicebert-human.510nt.
Citation
BibTeX:
| BibTeX |
|---|
| @article {chen2023self,
author = {Chen, Ken and Zhou, Yue and Ding, Maolin and Wang, Yu and Ren, Zhixiang and Yang, Yuedong},
title = {Self-supervised learning on millions of pre-mRNA sequences improves sequence-based RNA splicing prediction},
elocation-id = {2023.01.31.526427},
year = {2023},
doi = {10.1101/2023.01.31.526427},
publisher = {Cold Spring Harbor Laboratory},
abstract = {RNA splicing is an important post-transcriptional process of gene expression in eukaryotic cells. Predicting RNA splicing from primary sequences can facilitate the interpretation of genomic variants. In this study, we developed a novel self-supervised pre-trained language model, SpliceBERT, to improve sequence-based RNA splicing prediction. Pre-training on pre-mRNA sequences from vertebrates enables SpliceBERT to capture evolutionary conservation information and characterize the unique property of splice sites. SpliceBERT also improves zero-shot prediction of variant effects on splicing by considering sequence context information, and achieves superior performance for predicting branchpoint in the human genome and splice sites across species. Our study highlighted the importance of pre-training genomic language models on a diverse range of species and suggested that pre-trained language models were promising for deciphering the sequence logic of RNA splicing.Competing Interest StatementThe authors have declared no competing interest.},
URL = {https://www.biorxiv.org/content/early/2023/05/09/2023.01.31.526427},
eprint = {https://www.biorxiv.org/content/early/2023/05/09/2023.01.31.526427.full.pdf},
journal = {bioRxiv}
}
|
Please use GitHub issues of MultiMolecule for any questions or comments on the model card.
Please contact the authors of the SpliceBERT paper for questions or comments on the paper/model.
License
This model is licensed under the AGPL-3.0 License.
| Text Only |
|---|
| SPDX-License-Identifier: AGPL-3.0-or-later
|
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
standardextendedstreamlinenucleobase
- 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>>> 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 |
|---|
| 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.condon = 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.condon:
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)
|
Bases: PreTrainedConfig
This is the configuration class to store the configuration of a
SpliceBertModel. It is used to instantiate a SpliceBert 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 SpliceBert
biomed-AI/SpliceBERT 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 SpliceBert model. Defines the number of different tokens that can be represented by
the inputs_ids passed when calling [SpliceBertModel].
|
26
|
hidden_size
|
int
|
Dimensionality of the encoder layers and the pooler layer.
|
512
|
num_hidden_layers
|
int
|
Number of hidden layers in the Transformer encoder.
|
6
|
num_attention_heads
|
int
|
Number of attention heads for each attention layer in the Transformer encoder.
|
16
|
|
|
int
|
Dimensionality of the “intermediate” (often named feed-forward) layer in the Transformer encoder.
|
2048
|
hidden_dropout
|
float
|
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
0.1
|
attention_dropout
|
float
|
The dropout ratio for the attention probabilities.
|
0.1
|
max_position_embeddings
|
int
|
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).
|
1026
|
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
|
Examples:
Python Console Session>>> from multimolecule import SpliceBertModel, SpliceBertConfig
>>> # Initializing a SpliceBERT multimolecule/splicebert style configuration
>>> configuration = SpliceBertConfig()
>>> # Initializing a model (with random weights) from the multimolecule/splicebert style configuration
>>> model = SpliceBertModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
Source code in multimolecule/models/splicebert/configuration_splicebert.py
| Python |
|---|
| class SpliceBertConfig(PreTrainedConfig):
r"""
This is the configuration class to store the configuration of a
[`SpliceBertModel`][multimolecule.models.SpliceBertModel]. It is used to instantiate a SpliceBert 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 SpliceBert
[biomed-AI/SpliceBERT](https://github.com/biomed-AI/SpliceBERT) 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 SpliceBert model. Defines the number of different tokens that can be represented by
the `inputs_ids` passed when calling [`SpliceBertModel`].
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_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.
Examples:
>>> from multimolecule import SpliceBertModel, SpliceBertConfig
>>> # Initializing a SpliceBERT multimolecule/splicebert style configuration
>>> configuration = SpliceBertConfig()
>>> # Initializing a model (with random weights) from the multimolecule/splicebert style configuration
>>> model = SpliceBertModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
"""
model_type = "splicebert"
def __init__(
self,
vocab_size: int = 26,
hidden_size: int = 512,
num_hidden_layers: int = 6,
num_attention_heads: int = 16,
intermediate_size: int = 2048,
hidden_act: str = "gelu",
hidden_dropout: float = 0.1,
attention_dropout: float = 0.1,
max_position_embeddings: int = 1026,
initializer_range: float = 0.02,
layer_norm_eps: float = 1e-12,
position_embedding_type: str = "absolute",
is_decoder: bool = False,
use_cache: bool = True,
head: HeadConfig | None = None,
lm_head: MaskedLMHeadConfig | None = None,
**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.hidden_act = hidden_act
self.hidden_dropout = hidden_dropout
self.attention_dropout = attention_dropout
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.position_embedding_type = position_embedding_type
self.is_decoder = is_decoder
self.use_cache = use_cache
self.head = HeadConfig(**head if head is not None else {})
self.lm_head = MaskedLMHeadConfig(**lm_head if lm_head is not None else {})
|
Bases: SpliceBertPreTrainedModel
Examples:
Python Console Session>>> from multimolecule import SpliceBertConfig, SpliceBertForContactPrediction, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForContactPrediction(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, 2])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| class SpliceBertForContactPrediction(SpliceBertPreTrainedModel):
"""
Examples:
>>> from multimolecule import SpliceBertConfig, SpliceBertForContactPrediction, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForContactPrediction(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, 2])
>>> output["loss"] # doctest:+ELLIPSIS
tensor(..., grad_fn=<NllLossBackward0>)
"""
def __init__(self, config: SpliceBertConfig):
super().__init__(config)
self.splicebert = SpliceBertModel(config, add_pooling_layer=True)
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,
position_ids: 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,
return_dict: bool | None = None,
**kwargs,
) -> Tuple[Tensor, ...] | ContactPredictorOutput:
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.splicebert(
input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
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 ContactPredictorOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
|
Bases: SpliceBertPreTrainedModel
Examples:
Python Console Session>>> from multimolecule import SpliceBertConfig, SpliceBertForMaskedLM, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForMaskedLM(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=input["input_ids"])
>>> output["logits"].shape
torch.Size([1, 7, 26])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| class SpliceBertForMaskedLM(SpliceBertPreTrainedModel):
"""
Examples:
>>> from multimolecule import SpliceBertConfig, SpliceBertForMaskedLM, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForMaskedLM(config)
>>> tokenizer = RnaTokenizer.from_pretrained("multimolecule/rna")
>>> input = tokenizer("ACGUN", return_tensors="pt")
>>> output = model(**input, labels=input["input_ids"])
>>> output["logits"].shape
torch.Size([1, 7, 26])
>>> output["loss"] # doctest:+ELLIPSIS
tensor(..., grad_fn=<NllLossBackward0>)
"""
_tied_weights_keys = ["lm_head.decoder.bias", "lm_head.decoder.weight"]
def __init__(self, config: SpliceBertConfig):
super().__init__(config)
if config.is_decoder:
logger.warning(
"If you want to use `SpliceBertForMaskedLM` make sure `config.is_decoder=False` for "
"bi-directional self-attention."
)
self.splicebert = SpliceBertModel(config, add_pooling_layer=False)
self.lm_head = MaskedLMHead(config)
# Initialize weights and apply final processing
self.post_init()
def get_output_embeddings(self):
return self.lm_head.decoder
def set_output_embeddings(self, new_embeddings):
self.lm_head.decoder = new_embeddings
def forward(
self,
input_ids: Tensor | NestedTensor,
attention_mask: Tensor | None = None,
position_ids: Tensor | None = None,
head_mask: Tensor | None = None,
inputs_embeds: Tensor | NestedTensor | None = None,
encoder_hidden_states: Tensor | None = None,
encoder_attention_mask: Tensor | None = None,
labels: Tensor | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
**kwargs,
) -> Tuple[Tensor, ...] | MaskedLMOutput:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.splicebert(
input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
**kwargs,
)
output = self.lm_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 MaskedLMOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
|
Bases: SpliceBertPreTrainedModel
Examples:
Python Console Session>>> from multimolecule import SpliceBertConfig, SpliceBertForSequencePrediction, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForSequencePrediction(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, 2])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| class SpliceBertForSequencePrediction(SpliceBertPreTrainedModel):
"""
Examples:
>>> from multimolecule import SpliceBertConfig, SpliceBertForSequencePrediction, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForSequencePrediction(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, 2])
>>> output["loss"] # doctest:+ELLIPSIS
tensor(..., grad_fn=<NllLossBackward0>)
"""
def __init__(self, config: SpliceBertConfig):
super().__init__(config)
self.splicebert = SpliceBertModel(config, add_pooling_layer=True)
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,
position_ids: 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,
return_dict: bool | None = None,
**kwargs,
) -> Tuple[Tensor, ...] | SequencePredictorOutput:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.splicebert(
input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_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 SequencePredictorOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
|
Bases: SpliceBertPreTrainedModel
Examples:
Python Console Session>>> from multimolecule import SpliceBertConfig, SpliceBertForTokenPrediction, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForTokenPrediction(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, 2])
>>> output["loss"]
tensor(..., grad_fn=<NllLossBackward0>)
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| class SpliceBertForTokenPrediction(SpliceBertPreTrainedModel):
"""
Examples:
>>> from multimolecule import SpliceBertConfig, SpliceBertForTokenPrediction, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertForTokenPrediction(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, 2])
>>> output["loss"] # doctest:+ELLIPSIS
tensor(..., grad_fn=<NllLossBackward0>)
"""
def __init__(self, config: SpliceBertConfig):
super().__init__(config)
self.splicebert = SpliceBertModel(config, add_pooling_layer=True)
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,
position_ids: 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,
return_dict: bool | None = None,
**kwargs,
) -> Tuple[Tensor, ...] | TokenPredictorOutput:
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.splicebert(
input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
head_mask=head_mask,
inputs_embeds=inputs_embeds,
output_attentions=output_attentions,
output_hidden_states=output_hidden_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 TokenPredictorOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
|
Bases: SpliceBertPreTrainedModel
Examples:
Python Console Session>>> from multimolecule import SpliceBertConfig, SpliceBertModel, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertModel(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, 512])
>>> output["pooler_output"].shape
torch.Size([1, 512])
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| class SpliceBertModel(SpliceBertPreTrainedModel):
"""
Examples:
>>> from multimolecule import SpliceBertConfig, SpliceBertModel, RnaTokenizer
>>> config = SpliceBertConfig()
>>> model = SpliceBertModel(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, 512])
>>> output["pooler_output"].shape
torch.Size([1, 512])
"""
def __init__(self, config: SpliceBertConfig, add_pooling_layer: bool = True):
super().__init__(config)
self.pad_token_id = config.pad_token_id
self.embeddings = SpliceBertEmbeddings(config)
self.encoder = SpliceBertEncoder(config)
self.pooler = SpliceBertPooler(config) if add_pooling_layer else None
# 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,
position_ids: Tensor | None = None,
head_mask: Tensor | None = None,
inputs_embeds: Tensor | NestedTensor | None = None,
encoder_hidden_states: Tensor | None = None,
encoder_attention_mask: Tensor | None = None,
past_key_values: Tuple[Tuple[Tensor, Tensor, Tensor, Tensor], ...] | None = None,
use_cache: bool | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
**kwargs,
) -> Tuple[Tensor, ...] | BaseModelOutputWithPoolingAndCrossAttentions:
r"""
Args:
encoder_hidden_states:
Shape: `(batch_size, sequence_length, hidden_size)`
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask:
Shape: `(batch_size, sequence_length)`
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used
in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values:
Tuple of length `config.n_layers` with each tuple having 4 tensors of shape
`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up
decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
use_cache:
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
"""
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
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if self.config.is_decoder:
use_cache = use_cache if use_cache is not None else self.config.use_cache
else:
use_cache = False
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]
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
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 + past_key_values_length)), device=device)
)
# 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)
# If a 2D or 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
if self.config.is_decoder and encoder_hidden_states is not None:
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
if encoder_attention_mask is None:
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = None
# 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,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
past_key_values_length=past_key_values_length,
)
encoder_outputs = self.encoder(
embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_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 BaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
past_key_values=encoder_outputs.past_key_values,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
cross_attentions=encoder_outputs.cross_attentions,
)
|
| Python |
|---|
| forward(input_ids: Tensor | NestedTensor, attention_mask: Tensor | None = None, position_ids: Tensor | None = None, head_mask: Tensor | None = None, inputs_embeds: Tensor | NestedTensor | None = None, encoder_hidden_states: Tensor | None = None, encoder_attention_mask: Tensor | None = None, past_key_values: Tuple[Tuple[Tensor, Tensor, Tensor, Tensor], ...] | None = None, use_cache: bool | None = None, output_attentions: bool | None = None, output_hidden_states: bool | None = None, return_dict: bool | None = None, **kwargs) -> Tuple[Tensor, ...] | BaseModelOutputWithPoolingAndCrossAttentions
|
Parameters:
| Name |
Type |
Description |
Default |
encoder_hidden_states
|
Tensor | None
|
Shape: (batch_size, sequence_length, hidden_size)
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
|
None
|
encoder_attention_mask
|
Tensor | None
|
Shape: (batch_size, sequence_length)
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used
in the cross-attention if the model is configured as a decoder. Mask values selected in [0, 1]:
- 1 for tokens that are not masked,
- 0 for tokens that are masked.
|
None
|
past_key_values
|
Tuple[Tuple[Tensor, Tensor, Tensor, Tensor], ...] | None
|
Tuple of length config.n_layers with each tuple having 4 tensors of shape
`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up
decoding.
If past_key_values are used, the user can optionally input only the last decoder_input_ids (those
that don’t have their past key value states given to this model) of shape (batch_size, 1) instead of
all decoder_input_ids of shape (batch_size, sequence_length).
|
None
|
use_cache
|
bool | None
|
If set to True, past_key_values key value states are returned and can be used to speed up decoding
(see past_key_values).
|
None
|
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| def forward(
self,
input_ids: Tensor | NestedTensor,
attention_mask: Tensor | None = None,
position_ids: Tensor | None = None,
head_mask: Tensor | None = None,
inputs_embeds: Tensor | NestedTensor | None = None,
encoder_hidden_states: Tensor | None = None,
encoder_attention_mask: Tensor | None = None,
past_key_values: Tuple[Tuple[Tensor, Tensor, Tensor, Tensor], ...] | None = None,
use_cache: bool | None = None,
output_attentions: bool | None = None,
output_hidden_states: bool | None = None,
return_dict: bool | None = None,
**kwargs,
) -> Tuple[Tensor, ...] | BaseModelOutputWithPoolingAndCrossAttentions:
r"""
Args:
encoder_hidden_states:
Shape: `(batch_size, sequence_length, hidden_size)`
Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
the model is configured as a decoder.
encoder_attention_mask:
Shape: `(batch_size, sequence_length)`
Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used
in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
- 1 for tokens that are **not masked**,
- 0 for tokens that are **masked**.
past_key_values:
Tuple of length `config.n_layers` with each tuple having 4 tensors of shape
`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)
Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up
decoding.
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
all `decoder_input_ids` of shape `(batch_size, sequence_length)`.
use_cache:
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
(see `past_key_values`).
"""
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
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if self.config.is_decoder:
use_cache = use_cache if use_cache is not None else self.config.use_cache
else:
use_cache = False
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]
# past_key_values_length
past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
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 + past_key_values_length)), device=device)
)
# 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)
# If a 2D or 3D attention mask is provided for the cross-attention
# we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
if self.config.is_decoder and encoder_hidden_states is not None:
encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
if encoder_attention_mask is None:
encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
else:
encoder_extended_attention_mask = None
# 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,
position_ids=position_ids,
inputs_embeds=inputs_embeds,
past_key_values_length=past_key_values_length,
)
encoder_outputs = self.encoder(
embedding_output,
attention_mask=extended_attention_mask,
head_mask=head_mask,
encoder_hidden_states=encoder_hidden_states,
encoder_attention_mask=encoder_extended_attention_mask,
past_key_values=past_key_values,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_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 BaseModelOutputWithPoolingAndCrossAttentions(
last_hidden_state=sequence_output,
pooler_output=pooled_output,
past_key_values=encoder_outputs.past_key_values,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
cross_attentions=encoder_outputs.cross_attentions,
)
|
Bases: PreTrainedModel
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
Source code in multimolecule/models/splicebert/modeling_splicebert.py
| Python |
|---|
| class SpliceBertPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = SpliceBertConfig
base_model_prefix = "splicebert"
supports_gradient_checkpointing = True
_no_split_modules = ["SpliceBertLayer", "SpliceBertEmbeddings"]
# 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)
def _set_gradient_checkpointing(self, module, value=False):
if isinstance(module, SpliceBertEncoder):
module.gradient_checkpointing = value
|