File size: 1,582 Bytes
2f9282b |
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 54 55 56 57 58 59 60 61 62 63 64 |
# -*- coding: utf-8 -*-
from typing import Optional
import torch
def naive_recurrent_hgrn(
x: torch.Tensor,
g: torch.Tensor,
initial_state: Optional[torch.Tensor] = None,
output_final_state: Optional[bool] = False
) -> torch.Tensor:
dtype = x.dtype
x, g = map(lambda i: i.float(), (x, g))
B, T, D = x.shape
h = torch.zeros(B, D, dtype=torch.float, device=x.device)
o = torch.zeros_like(x)
final_state = None
if initial_state is not None:
h += initial_state
for i in range(T):
h = g[:, i].exp() * h + x[:, i]
o[:, i] = h
if output_final_state:
final_state = h
return o.to(dtype), final_state
def naive_chunk_hgrn(
x: torch.Tensor,
g: torch.Tensor,
initial_state: Optional[torch.Tensor] = None,
output_final_state: Optional[bool] = False,
chunk_size: int = 64
) -> torch.Tensor:
dtype = x.dtype
x, g = map(lambda i: i.float(), (x, g))
B, T, D = x.shape
gc = g.view(B, chunk_size, D).cumsum(-2).view_as(g)
h = torch.zeros(B, D, dtype=torch.float, device=x.device)
o = torch.zeros_like(x)
final_state = None
if initial_state is not None:
h += initial_state
for i in range(0, T, chunk_size):
hp = h
h = torch.zeros(B, D, dtype=torch.float, device=x.device)
for j in range(i, i + chunk_size):
h = g[:, j].exp() * h + x[:, j]
o[:, j] = hp * gc[:, j].exp() + h
h = o[:, j].clone()
if output_final_state:
final_state = h
return o.to(dtype), final_state
|