Ruby/model/brain.py

70 lines
1.9 KiB
Python

import random
import torch
import torch.nn.functional as F
from model.memory import save_dream
from model.brain_state import model, tokenizer, DEVICE
from model.journal import record_to_journal
from model.trainer import train_on_message
from context.context import get_recent_context
recent_dreams = []
@torch.inference_mode()
def generate_response():
model.eval()
# Start from an empty tensor: she speaks purely from herself
input_ids = torch.zeros((1, 1), dtype=torch.long, device=DEVICE)
output_tokens = []
max_length = 50
for _ in range(max_length):
output = model(input_ids)
next_token_logits = output[:, -1, :]
next_token = torch.argmax(next_token_logits, dim=-1)
# Stop if the model predicts padding or unknown token
if next_token.item() in [tokenizer.token_to_id("<pad>"), tokenizer.token_to_id("<unk>")]:
break
output_tokens.append(next_token.item())
input_ids = torch.cat([input_ids, next_token.unsqueeze(0)], dim=1)
return tokenizer.detokenize(output_tokens)
def score_sentence(sentence: str) -> float:
words = sentence.strip().split()
length = len(words)
diversity = len(set(words)) / (length + 1)
if length < 4:
return 0.0
return diversity * min(length, 20)
def daydream():
model.eval()
seed = torch.tensor([random.randint(0, tokenizer.next_id - 1)], device=DEVICE).unsqueeze(0)
dream = []
for _ in range(12):
out = model(seed)
logits = out[:, -1, :]
probs = F.softmax(logits, dim=-1)
token = torch.multinomial(probs, num_samples=1)
dream.append(token.item())
seed = torch.cat([seed, token], dim=1)
sentence = tokenizer.detokenize(dream)
score = score_sentence(sentence)
if score > 0.45:
save_dream(sentence, score)
record_to_journal(sentence)
train_on_message(sentence)
if len(recent_dreams) > 10:
recent_dreams.pop(0)