forked from huggingface/diffusers
-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathlora.py
More file actions
123 lines (92 loc) · 4.81 KB
/
lora.py
File metadata and controls
123 lines (92 loc) · 4.81 KB
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
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Optional
import torch.nn.functional as F
from torch import nn
class LoRALinearLayer(nn.Module):
def __init__(self, in_features, out_features, rank=4, network_alpha=None, device=None, dtype=None):
super().__init__()
if rank > min(in_features, out_features):
raise ValueError(f"LoRA rank {rank} must be less or equal than {min(in_features, out_features)}")
self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
# This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
# See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
self.network_alpha = network_alpha
self.rank = rank
nn.init.normal_(self.down.weight, std=1 / rank)
nn.init.zeros_(self.up.weight)
def forward(self, hidden_states):
orig_dtype = hidden_states.dtype
dtype = self.down.weight.dtype
down_hidden_states = self.down(hidden_states.to(dtype))
up_hidden_states = self.up(down_hidden_states)
if self.network_alpha is not None:
up_hidden_states *= self.network_alpha / self.rank
return up_hidden_states.to(orig_dtype)
class LoRAConv2dLayer(nn.Module):
def __init__(
self, in_features, out_features, rank=4, kernel_size=(1, 1), stride=(1, 1), padding=0, network_alpha=None
):
super().__init__()
if rank > min(in_features, out_features):
raise ValueError(f"LoRA rank {rank} must be less or equal than {min(in_features, out_features)}")
self.down = nn.Conv2d(in_features, rank, kernel_size=kernel_size, stride=stride, padding=padding, bias=False)
# according to the official kohya_ss trainer kernel_size are always fixed for the up layer
# # see: https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L129
self.up = nn.Conv2d(rank, out_features, kernel_size=(1, 1), stride=(1, 1), bias=False)
# This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
# See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
self.network_alpha = network_alpha
self.rank = rank
nn.init.normal_(self.down.weight, std=1 / rank)
nn.init.zeros_(self.up.weight)
def forward(self, hidden_states):
orig_dtype = hidden_states.dtype
dtype = self.down.weight.dtype
down_hidden_states = self.down(hidden_states.to(dtype))
up_hidden_states = self.up(down_hidden_states)
if self.network_alpha is not None:
up_hidden_states *= self.network_alpha / self.rank
return up_hidden_states.to(orig_dtype)
class LoRACompatibleConv(nn.Conv2d):
"""
A convolutional layer that can be used with LoRA.
"""
def __init__(self, *args, lora_layer: Optional[LoRAConv2dLayer] = None, **kwargs):
super().__init__(*args, **kwargs)
self.lora_layer = lora_layer
def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
self.lora_layer = lora_layer
def forward(self, x):
if self.lora_layer is None:
# make sure to the functional Conv2D function as otherwise torch.compile's graph will break
# see: https://github.com/huggingface/diffusers/pull/4315
return F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
else:
return super().forward(x) + self.lora_layer(x)
class LoRACompatibleLinear(nn.Linear):
"""
A Linear layer that can be used with LoRA.
"""
def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
super().__init__(*args, **kwargs)
self.lora_layer = lora_layer
def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
self.lora_layer = lora_layer
def forward(self, x):
if self.lora_layer is None:
return super().forward(x)
else:
return super().forward(x) + self.lora_layer(x)