File size: 863 Bytes
4d5aff4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# This code is licensed under a non-commercial license.

import torch
from torch.autograd import Variable

def to_var(x, requires_grad=False, volatile=False):
    if torch.cuda.is_available():
        x = x.cuda()
    return Variable(x, requires_grad=requires_grad, volatile=volatile)

def top_k_logits(logits, k, probs=False):
    """
    Masks everything but the k top entries as -infinity (1e10).
    Used to mask logits such that e^-infinity -> 0 won't contribute to the
    sum of the denominator.
    """
    if k == 0:
        return logits
    else:
        values = torch.topk(logits, k)[0]
        batch_mins = values[:, -1].view(-1, 1).expand_as(logits)
        if probs:
            return torch.where(logits < batch_mins, torch.ones_like(logits) * 0.0, logits)
        return torch.where(logits < batch_mins, torch.ones_like(logits) * -1e10, logits)