-
Notifications
You must be signed in to change notification settings - Fork 178
/
softmax.py
51 lines (35 loc) · 1.82 KB
/
softmax.py
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
from typing import Optional
import torch
from torch_scatter import scatter_sum, scatter_max
from torch_scatter.utils import broadcast
def scatter_softmax(src: torch.Tensor, index: torch.Tensor,
dim: int = -1,
dim_size: Optional[int] = None) -> torch.Tensor:
if not torch.is_floating_point(src):
raise ValueError('`scatter_softmax` can only be computed over tensors '
'with floating point data types.')
index = broadcast(index, src, dim)
max_value_per_index = scatter_max(
src, index, dim=dim, dim_size=dim_size)[0]
max_per_src_element = max_value_per_index.gather(dim, index)
recentered_scores = src - max_per_src_element
recentered_scores_exp = recentered_scores.exp_()
sum_per_index = scatter_sum(
recentered_scores_exp, index, dim, dim_size=dim_size)
normalizing_constants = sum_per_index.gather(dim, index)
return recentered_scores_exp.div(normalizing_constants)
def scatter_log_softmax(src: torch.Tensor, index: torch.Tensor, dim: int = -1,
eps: float = 1e-12,
dim_size: Optional[int] = None) -> torch.Tensor:
if not torch.is_floating_point(src):
raise ValueError('`scatter_log_softmax` can only be computed over '
'tensors with floating point data types.')
index = broadcast(index, src, dim)
max_value_per_index = scatter_max(
src, index, dim=dim, dim_size=dim_size)[0]
max_per_src_element = max_value_per_index.gather(dim, index)
recentered_scores = src - max_per_src_element
sum_per_index = scatter_sum(
recentered_scores.exp(), index, dim, dim_size=dim_size)
normalizing_constants = sum_per_index.add_(eps).log_().gather(dim, index)
return recentered_scores.sub_(normalizing_constants)