-
Notifications
You must be signed in to change notification settings - Fork 43
/
weight_permutation.py
48 lines (41 loc) · 2.12 KB
/
weight_permutation.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
import torch
@torch.jit.script
def find_greedy_nearest_indices(weight: torch.Tensor, use_abs: bool = False):
print("use_abs", use_abs)
ordered_unit_weight_t = weight.detach().t().clone()
ordered_unit_weight_t /= ordered_unit_weight_t.norm(p=2, dim=-1, keepdim=True)
distance_matrix = ordered_unit_weight_t @ ordered_unit_weight_t.T
if use_abs:
distance_matrix = abs(distance_matrix)
permutation = torch.arange(len(ordered_unit_weight_t), device=weight.device)
for dim_i in range(len(ordered_unit_weight_t) - 2):
nearest_dim_i = (dim_i + 1) + distance_matrix[dim_i, dim_i + 1 :].argmax()
next_dim_i = torch.full_like(nearest_dim_i, dim_i + 1)
index_pair = torch.stack([next_dim_i, nearest_dim_i])
swapped_index_pair = torch.stack([nearest_dim_i, next_dim_i])
ordered_unit_weight_t[index_pair] = ordered_unit_weight_t[swapped_index_pair]
distance_matrix[index_pair] = distance_matrix[swapped_index_pair]
distance_matrix[:, index_pair] = distance_matrix[:, swapped_index_pair]
permutation[index_pair] = permutation[swapped_index_pair]
return permutation
def get_permutation_order(H: torch.Tensor, W: torch.Tensor, permutation_order: str = "identity", use_abs: bool = False):
"""
Permutation order for layer weights.
:param H: Hessian of Weights
:param W: Layer weights
:param permutation_order: which permutation order to use default: identity, act_order,nearest
:return: permutation order 1d int tensor
"""
if permutation_order == "spearman":
w_rank = W.argsort(dim=0).argsort(dim=0).float()
w_rank = w_rank - w_rank.mean(dim=0, keepdim=True)
perm = find_greedy_nearest_indices(w_rank, use_abs)
elif permutation_order == "act_order":
perm = torch.argsort(torch.diag(H), descending=True)
elif permutation_order == "identity":
perm = torch.arange(H.shape[0], device=H.device)
elif isinstance(permutation_order, torch.Tensor):
return permutation_order # user-defined
else:
raise ValueError(f"Unknown permutation order name: {permutation_order}")
return perm