דיסטילציה של ידע

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

import torch
import torch.nn as nn

def distillation_loss(student_logits, teacher_logits, temperature=2.0):
    T = temperature
    student_probs = nn.functional.softmax(student_logits / T, dim=1)
    teacher_probs = nn.functional.softmax(teacher_logits / T, dim=1)
    return nn.functional.kl_div(
        nn.functional.log_softmax(student_logits / T, dim=1),
        teacher_probs,
        reduction='batchmean'
    ) * (T * T)