Views
No views yet
1from transformers import (
2 T5Config,
3 T5EncoderModel,
4 T5Tokenizer,
5 PreTrainedModel,
6 TrainingArguments,
7 Trainer,
8 DataCollatorWithPadding,
9)
10class T5RegressionModel(PreTrainedModel):
11
12 config_class = T5Config
13
14 def __init__(self, config, d_model=None):
15 super().__init__(config)
16
17 self.encoder = T5EncoderModel.from_pretrained("Rostlab/prot_t5_xl_uniref50")
18
19
20 hidden_dim = d_model if d_model is not None else config.d_model
21 self.regression_head = nn.Linear(hidden_dim, 1)
22
23 def forward(
24 self,
25 input_ids=None,
26 attention_mask=None,
27 labels=None,
28 **kwargs
29 ):
30 encoder_outputs = self.encoder(input_ids=input_ids, attention_mask=attention_mask)
31 hidden_states = encoder_outputs.last_hidden_state
32
33 mask = attention_mask.unsqueeze(-1)
34 pooled_output = (hidden_states * mask).sum(dim=1) / mask.sum(dim=1)
35 logits = self.regression_head(pooled_output).squeeze(-1) # [batch_size]
36
37
38 loss = None
39 if labels is not None:
40 labels = labels.float()
41 loss = nn.MSELoss()(logits, labels)
42
43 return {
44 "loss": loss,
45 "logits": logits
46 }1tokenizer = T5Tokenizer.from_pretrained("jiaxie/DeepProtT5-Beta", do_lower_case=False)
2model = T5RegressionModel.from_pretrained("jiaxie/DeepProtT5-Beta", torch_dtype=torch.bfloat16).to("cuda")