protify/base_models/esm2.py

"""
We use the FastESM2 implementation of ESM2.
"""
import sys
import os
import torch
import torch.nn as nn
from typing import Optional, Union, List, Dict

_FASTPLMS = os.path.join(os.path.dirname(os.path.dirname(os.path.abspath(__file__))), 'FastPLMs')
if _FASTPLMS not in sys.path:
    sys.path.insert(0, _FASTPLMS)

from esm2.modeling_fastesm import (
    FastEsmModel,
    FastEsmForMaskedLM,
    FastEsmForSequenceClassification,
    FastEsmForTokenClassification,
)
from transformers import EsmTokenizer
from .base_tokenizer import BaseSequenceTokenizer


presets = {
    'ESM2-8': 'Synthyra/ESM2-8M',
    'ESM2-35': 'Synthyra/ESM2-35M',
    'ESM2-150': 'Synthyra/ESM2-150M',
    'ESM2-650': 'Synthyra/ESM2-650M',
    'ESM2-3B': 'Synthyra/ESM2-3B',
    'DSM-150': 'GleghornLab/DSM_150',
    'DSM-650': 'GleghornLab/DSM_650',
    'DSM-PPI': 'Synthyra/DSM_ppi_full',
}


class ESM2TokenizerWrapper(BaseSequenceTokenizer):
    def __init__(self, tokenizer: EsmTokenizer):
        super().__init__(tokenizer)

    def __call__(self, sequences: Union[str, List[str]], **kwargs) -> Dict[str, torch.Tensor]:
        if isinstance(sequences, str):
            sequences = [sequences]
        kwargs.setdefault('return_tensors', 'pt')
        kwargs.setdefault('padding', 'longest')
        kwargs.setdefault('add_special_tokens', True)
        tokenized = self.tokenizer(sequences, **kwargs)
        return tokenized


class FastEsmForEmbedding(nn.Module):
    def __init__(self, model_path: str, dtype: torch.dtype = None):
        super().__init__()
        self.esm = FastEsmModel.from_pretrained(model_path, dtype=dtype)

    def forward(
            self,
            input_ids: torch.Tensor,
            attention_mask: Optional[torch.Tensor] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = False,
            **kwargs,
    ) -> torch.Tensor:
        if output_attentions:
            out = self.esm(input_ids=input_ids, attention_mask=attention_mask, output_attentions=output_attentions)
            return out.last_hidden_state, out.attentions
        else:
            return self.esm(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state


def get_esm2_tokenizer(preset: str, model_path: str = None):
    return ESM2TokenizerWrapper(EsmTokenizer.from_pretrained('facebook/esm2_t6_8M_UR50D'))


def build_esm2_model(preset: str, masked_lm: bool = False, dtype: torch.dtype = None, model_path: str = None, **kwargs):
    path = model_path or presets[preset]
    if masked_lm:
        model = FastEsmForMaskedLM.from_pretrained(path, dtype=dtype).eval()
    else:
        model = FastEsmForEmbedding(path, dtype=dtype).eval()
    tokenizer = get_esm2_tokenizer(preset)
    return model, tokenizer


def get_esm2_for_training(preset: str, tokenwise: bool = False, num_labels: int = None, hybrid: bool = False, dtype: torch.dtype = None, model_path: str = None):
    model_path = model_path or presets[preset]
    if hybrid:
        model = FastEsmModel.from_pretrained(model_path, dtype=dtype).eval()
    else:
        if tokenwise:
            model = FastEsmForTokenClassification.from_pretrained(model_path, num_labels=num_labels, dtype=dtype).eval()
        else:
            model = FastEsmForSequenceClassification.from_pretrained(model_path, num_labels=num_labels, dtype=dtype).eval()
    tokenizer = get_esm2_tokenizer(preset)
    return model, tokenizer


if __name__ == '__main__':
    # py -m src.protify.base_models.esm2
    model, tokenizer = build_esm2_model('ESM2-8')
    print(model)
    print(tokenizer)
    print(tokenizer('MEKVQYLTRSAIRRASTIEMPQQARQKLQNLFINFCLILICBBOLLICIIVMLL'))