evo_prot_grad

get_expert

evo_prot_grad.get_expert(expert_name: str, scoring_strategy: str, temperature: float = 1.0, model: Optional[nn.Module] = None, tokenizer: Optional[Union[ExpertTokenizer, PreTrainedTokenizerBase]] = None, device: str = 'cpu') -> Expert

Current supported expert types (to pass to argument expert_name):

- `bert`
- `causallm`
- `esm`
- `evcouplings`
- `onehot_downstream_regression`

Customize the expert by specifying the model and tokenizer. For example:

from evo_prot_grad.experts import get_expert
from transformers import AutoTokenizer, EsmForMaskedLM

expert = get_expert(
    expert_name = 'esm',
    model = EsmForMaskedLM.from_pretrained("facebook/esm2_t36_3B_UR50D"),
    tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t36_3B_UR50D"),
    scoring_strategy = 'mutant_marginal',
    temperature = 1.0,
    device = 'cuda'
)   

Parameters:

Name	Type	Description	Default
expert_name	str	Name of the expert to be used.	required
scoring_strategy	str	Approach for scoring variants that the expert will use.	required
temperature	float	Temperature for the expert. Defaults to 1.0.	1.0
model	Optional[nn.Module]	Model to be used for the expert. Defaults to None.	None
tokenizer	Optional[Union[ExpertTokenizer, PreTrainedTokenizerBase]]	Tokenizer to be used for the expert. Defaults to None.	None
device	str	Device to be used for the expert. Defaults to 'cpu'.	'cpu'

Raises:

Type	Description
ValueError	If the expert name is not found.

Returns:

Name	Type	Description
expert	Expert	An instance of the expert.