Update the softmax issue with DNA encoder. #33
Description
The forward function of the DNA encoder is mistakenly using:
return self.base_dna_encoder(sequence).logits.softmax(dim=-1).mean(dim=1)
The correct way should be using:
return self.lora_barcode_bert(x).hidden_states[-1].mean(dim=1)
To fix this, the idea is to create a new branch, re-train all existed pre-trained models to update relevent checkpoints. Then merge to the main branch.