jiaxie/DeepProtT5-Beta

First we define a class T5RegressionModel:

from transformers import (
    T5Config,
    T5EncoderModel,
    T5Tokenizer,
    PreTrainedModel,
    TrainingArguments,
    Trainer,
    DataCollatorWithPadding,
)
class T5RegressionModel(PreTrainedModel):

    config_class = T5Config

    def __init__(self, config, d_model=None):
        super().__init__(config)

        self.encoder = T5EncoderModel.from_pretrained("Rostlab/prot_t5_xl_uniref50")


        hidden_dim = d_model if d_model is not None else config.d_model
        self.regression_head = nn.Linear(hidden_dim, 1)

    def forward(
        self,
        input_ids=None,
        attention_mask=None,
        labels=None,
        **kwargs
    ):
        encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
        hidden_states = encoder_outputs.last_hidden_state

        mask = attention_mask.unsqueeze(-1)
        pooled_output = (hidden_states * mask).sum(dim=1) / mask.sum(dim=1)
        logits = self.regression_head(pooled_output).squeeze(-1)  # [batch_size]


        loss = None
        if labels is not None:
            labels = labels.float()
            loss = nn.MSELoss()(logits, labels)

        return {
            "loss": loss,
            "logits": logits
        }

Then we load our pretrained model

tokenizer = T5Tokenizer.from_pretrained("jiaxie/DeepProtT5-Beta", do_lower_case=False)
model = T5RegressionModel.from_pretrained("jiaxie/DeepProtT5-Beta", torch_dtype=torch.bfloat16).to("cuda")