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
65
66
67
68
69
70
71
72
|
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import torch
from typing import Optional
from torch import Tensor
import torch.nn.functional as F
from bitsandbytes.optim import GlobalOptimManager
class StableEmbedding(torch.nn.Embedding):
def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None,
max_norm: Optional[float] = None, norm_type: float = 2., scale_grad_by_freq: bool = False,
sparse: bool = False, _weight: Optional[Tensor] = None) -> None:
super(StableEmbedding, self).__init__(num_embeddings, embedding_dim, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse, _weight)
self.norm = torch.nn.LayerNorm(embedding_dim)
GlobalOptimManager.get_instance().register_module_override(self, 'weight', {'optim_bits': 32})
def reset_parameters(self) -> None:
torch.nn.init.xavier_uniform_(self.weight)
self._fill_padding_idx_with_zero()
''' !!! This is a redefinition of _fill_padding_idx_with_zero in torch.nn.Embedding
to make the Layer compatible with Pytorch < 1.9.
This means that if this changes in future PyTorch releases this need to change too
which is cumbersome. However, with this we can ensure compatibility with previous
PyTorch releases.
'''
def _fill_padding_idx_with_zero(self) -> None:
if self.padding_idx is not None:
with torch.no_grad():
self.weight[self.padding_idx].fill_(0)
def forward(self, input: Tensor) -> Tensor:
emb = F.embedding(
input, self.weight, self.padding_idx, self.max_norm,
self.norm_type, self.scale_grad_by_freq, self.sparse)
return self.norm(emb)
class Embedding(torch.nn.Embedding):
def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: Optional[int] = None,
max_norm: Optional[float] = None, norm_type: float = 2., scale_grad_by_freq: bool = False,
sparse: bool = False, _weight: Optional[Tensor] = None) -> None:
super(Embedding, self).__init__(num_embeddings, embedding_dim, padding_idx, max_norm, norm_type, scale_grad_by_freq, sparse, _weight)
GlobalOptimManager.get_instance().register_module_override(self, 'weight', {'optim_bits': 32})
def reset_parameters(self) -> None:
torch.nn.init.xavier_uniform_(self.weight)
self._fill_padding_idx_with_zero()
''' !!! This is a redefinition of _fill_padding_idx_with_zero in torch.nn.Embedding
to make the Layer compatible with Pytorch < 1.9.
This means that if this changes in future PyTorch releases this need to change too
which is cumbersome. However, with this we can ensure compatibility with previous
PyTorch releases.
'''
def _fill_padding_idx_with_zero(self) -> None:
if self.padding_idx is not None:
with torch.no_grad():
self.weight[self.padding_idx].fill_(0)
def forward(self, input: Tensor) -> Tensor:
emb = F.embedding(
input, self.weight, self.padding_idx, self.max_norm,
self.norm_type, self.scale_grad_by_freq, self.sparse)
return emb
|
