modules/generate_binders.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import random

# Load ProtGPT2 or equivalent model
tokenizer = AutoTokenizer.from_pretrained("nferruz/ProtGPT2")
model = AutoModelForCausalLM.from_pretrained("nferruz/ProtGPT2")

# Set pad token to a different value than eos token
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    model.config.pad_token_id = model.config.eos_token_id

def generate_binders(fusion_context, strategy='low_shot', num_candidates=10):
    seed_sequence = fusion_context['embedding_vector'][:10]
    seed = ''.join([chr(int(65 + abs(int(x * 10)) % 20)) for x in seed_sequence])
    
    # Create input tensors with attention mask
    inputs = tokenizer(seed, return_tensors="pt", padding=True)
    input_ids = inputs["input_ids"]
    attention_mask = inputs["attention_mask"]

    outputs = model.generate(
        input_ids,
        attention_mask=attention_mask,
        do_sample=True,
        top_k=950,
        top_p=0.96,
        temperature=1.0,
        max_length=200,
        num_return_sequences=num_candidates,
        pad_token_id=tokenizer.pad_token_id
    )

    binders = []
    for output in outputs:
        sequence = tokenizer.decode(output, skip_special_tokens=True)
        sequence = ''.join([aa for aa in sequence if aa in "ACDEFGHIKLMNPQRSTVWY"])
        if len(sequence) > 30:
            binders.append({"sequence": sequence})

    return {"generated_binders": binders}