Ruby/model/rehearsal.py

27 lines
703 B
Python

import torch
from model.brain import model, tokenizer, DEVICE
from model.trainer import train_on_message
from model.dynamic_expand import expand_model_if_needed
from utils.unicleaner import clean_unicode
def simulate_conversation():
expand_model_if_needed()
model.eval()
seed = torch.randint(0, tokenizer.next_id, (1, 5), device=DEVICE)
seed = seed[:, -128:]
output = model(seed)
preds = torch.argmax(output, dim=-1).squeeze().tolist()
if isinstance(preds, int):
preds = [preds]
text = tokenizer.detokenize(preds)
# 🧹 Clean the generated text too
text = clean_unicode(text)
if text and len(text.split()) >= 3:
train_on_message(text)