File size: 1,705 Bytes
a600684 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 |
import numpy as np
from typing import Dict
# https://stackoverflow.com/a/50425683
def softmax(x: np.ndarray, axis: int):
x -= x.max(axis=axis, keepdims=True)
e: np.ndarray = np.exp(x)
return e / e.sum(axis=axis, keepdims=True)
def sample_logits(out, temperature: float = 1.0, top_p: float = 0.8, logit_bias: Dict[int, float] = None) -> int:
if hasattr(out, '__module__') and out.__module__ == 'torch':
out = out.cpu().numpy()
probs: np.ndarray = softmax(out, axis=-1)
return sample_probs(probs, temperature, top_p, logit_bias)
def sample_probs(probs: np.ndarray, temperature: float = 1.0, top_p: float = 0.8, logit_bias: Dict[int, float] = None) -> int:
if not (0.0 <= temperature):
raise ValueError('temperature')
if not (0.0 <= top_p <= 1.0):
raise ValueError('top_p')
if top_p == 0.0:
top_p = 1.0
if logit_bias is not None and len(logit_bias) > 0:
logits: np.ndarray = np.log(probs)
ids, values = zip(*logit_bias.items())
logits[list(ids)] += values
# Makes calculation more numerically stable, does not change the result
logits -= logits.max(axis=-1, keepdims=True)
probs = np.exp(logits) / np.sum(np.exp(logits))
if temperature == 0.0:
return np.argmax(probs).item()
if top_p < 1.0:
sorted_probs = np.sort(probs)[::-1]
cumulative_probs = np.cumsum(sorted_probs)
cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
probs[probs < cutoff] = 0
if temperature != 1.0:
probs = np.power(probs, 1.0 / temperature)
probs = probs / np.sum(probs)
return np.random.choice(a=len(probs), p=probs)
|