multimodalart HF staff commited on
Commit
be190eb
1 Parent(s): f2725ef

Upload lora.py

Browse files
Files changed (1) hide show
  1. lora.py +1222 -0
lora.py ADDED
@@ -0,0 +1,1222 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # LoRA network module
2
+ # reference:
3
+ # https://github.com/microsoft/LoRA/blob/main/loralib/layers.py
4
+ # https://github.com/cloneofsimo/lora/blob/master/lora_diffusion/lora.py
5
+
6
+ import math
7
+ import os
8
+ from typing import Dict, List, Optional, Tuple, Type, Union
9
+ from diffusers import AutoencoderKL
10
+ from transformers import CLIPTextModel
11
+ import numpy as np
12
+ import torch
13
+ import re
14
+
15
+
16
+ RE_UPDOWN = re.compile(r"(up|down)_blocks_(\d+)_(resnets|upsamplers|downsamplers|attentions)_(\d+)_")
17
+
18
+ RE_UPDOWN = re.compile(r"(up|down)_blocks_(\d+)_(resnets|upsamplers|downsamplers|attentions)_(\d+)_")
19
+
20
+
21
+ class LoRAModule(torch.nn.Module):
22
+ """
23
+ replaces forward method of the original Linear, instead of replacing the original Linear module.
24
+ """
25
+
26
+ def __init__(
27
+ self,
28
+ lora_name,
29
+ org_module: torch.nn.Module,
30
+ multiplier=1.0,
31
+ lora_dim=4,
32
+ alpha=1,
33
+ dropout=None,
34
+ rank_dropout=None,
35
+ module_dropout=None,
36
+ ):
37
+ """if alpha == 0 or None, alpha is rank (no scaling)."""
38
+ super().__init__()
39
+ self.lora_name = lora_name
40
+
41
+ if org_module.__class__.__name__ == "Conv2d":
42
+ in_dim = org_module.in_channels
43
+ out_dim = org_module.out_channels
44
+ else:
45
+ in_dim = org_module.in_features
46
+ out_dim = org_module.out_features
47
+
48
+ # if limit_rank:
49
+ # self.lora_dim = min(lora_dim, in_dim, out_dim)
50
+ # if self.lora_dim != lora_dim:
51
+ # print(f"{lora_name} dim (rank) is changed to: {self.lora_dim}")
52
+ # else:
53
+ self.lora_dim = lora_dim
54
+
55
+ if org_module.__class__.__name__ == "Conv2d":
56
+ kernel_size = org_module.kernel_size
57
+ stride = org_module.stride
58
+ padding = org_module.padding
59
+ self.lora_down = torch.nn.Conv2d(in_dim, self.lora_dim, kernel_size, stride, padding, bias=False)
60
+ self.lora_up = torch.nn.Conv2d(self.lora_dim, out_dim, (1, 1), (1, 1), bias=False)
61
+ else:
62
+ self.lora_down = torch.nn.Linear(in_dim, self.lora_dim, bias=False)
63
+ self.lora_up = torch.nn.Linear(self.lora_dim, out_dim, bias=False)
64
+
65
+ if type(alpha) == torch.Tensor:
66
+ alpha = alpha.detach().float().numpy() # without casting, bf16 causes error
67
+ alpha = self.lora_dim if alpha is None or alpha == 0 else alpha
68
+ self.scale = alpha / self.lora_dim
69
+ self.register_buffer("alpha", torch.tensor(alpha)) # 定数として扱える
70
+
71
+ # same as microsoft's
72
+ torch.nn.init.kaiming_uniform_(self.lora_down.weight, a=math.sqrt(5))
73
+ torch.nn.init.zeros_(self.lora_up.weight)
74
+
75
+ self.multiplier = multiplier
76
+ self.org_module = org_module # remove in applying
77
+ self.dropout = dropout
78
+ self.rank_dropout = rank_dropout
79
+ self.module_dropout = module_dropout
80
+
81
+ def apply_to(self):
82
+ self.org_forward = self.org_module.forward
83
+ self.org_module.forward = self.forward
84
+ del self.org_module
85
+
86
+ def forward(self, x):
87
+ org_forwarded = self.org_forward(x)
88
+
89
+ # module dropout
90
+ if self.module_dropout is not None and self.training:
91
+ if torch.rand(1) < self.module_dropout:
92
+ return org_forwarded
93
+
94
+ lx = self.lora_down(x)
95
+
96
+ # normal dropout
97
+ if self.dropout is not None and self.training:
98
+ lx = torch.nn.functional.dropout(lx, p=self.dropout)
99
+
100
+ # rank dropout
101
+ if self.rank_dropout is not None and self.training:
102
+ mask = torch.rand((lx.size(0), self.lora_dim), device=lx.device) > self.rank_dropout
103
+ if len(lx.size()) == 3:
104
+ mask = mask.unsqueeze(1) # for Text Encoder
105
+ elif len(lx.size()) == 4:
106
+ mask = mask.unsqueeze(-1).unsqueeze(-1) # for Conv2d
107
+ lx = lx * mask
108
+
109
+ # scaling for rank dropout: treat as if the rank is changed
110
+ # maskから計算することも考えられるが、augmentation的な効果を期待してrank_dropoutを用いる
111
+ scale = self.scale * (1.0 / (1.0 - self.rank_dropout)) # redundant for readability
112
+ else:
113
+ scale = self.scale
114
+
115
+ lx = self.lora_up(lx)
116
+
117
+ return org_forwarded + lx * self.multiplier * scale
118
+
119
+
120
+ class LoRAInfModule(LoRAModule):
121
+ def __init__(
122
+ self,
123
+ lora_name,
124
+ org_module: torch.nn.Module,
125
+ multiplier=1.0,
126
+ lora_dim=4,
127
+ alpha=1,
128
+ **kwargs,
129
+ ):
130
+ # no dropout for inference
131
+ super().__init__(lora_name, org_module, multiplier, lora_dim, alpha)
132
+
133
+ self.org_module_ref = [org_module] # 後から参照できるように
134
+ self.enabled = True
135
+
136
+ # check regional or not by lora_name
137
+ self.text_encoder = False
138
+ if lora_name.startswith("lora_te_"):
139
+ self.regional = False
140
+ self.use_sub_prompt = True
141
+ self.text_encoder = True
142
+ elif "attn2_to_k" in lora_name or "attn2_to_v" in lora_name:
143
+ self.regional = False
144
+ self.use_sub_prompt = True
145
+ elif "time_emb" in lora_name:
146
+ self.regional = False
147
+ self.use_sub_prompt = False
148
+ else:
149
+ self.regional = True
150
+ self.use_sub_prompt = False
151
+
152
+ self.network: LoRANetwork = None
153
+
154
+ def set_network(self, network):
155
+ self.network = network
156
+
157
+ # freezeしてマージする
158
+ def merge_to(self, sd, dtype, device):
159
+ # get up/down weight
160
+ up_weight = sd["lora_up.weight"].to(torch.float).to(device)
161
+ down_weight = sd["lora_down.weight"].to(torch.float).to(device)
162
+
163
+ # extract weight from org_module
164
+ org_sd = self.org_module.state_dict()
165
+ weight = org_sd["weight"].to(torch.float)
166
+
167
+ # merge weight
168
+ if len(weight.size()) == 2:
169
+ # linear
170
+ weight = weight + self.multiplier * (up_weight @ down_weight) * self.scale
171
+ elif down_weight.size()[2:4] == (1, 1):
172
+ # conv2d 1x1
173
+ weight = (
174
+ weight
175
+ + self.multiplier
176
+ * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3)
177
+ * self.scale
178
+ )
179
+ else:
180
+ # conv2d 3x3
181
+ conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3)
182
+ # print(conved.size(), weight.size(), module.stride, module.padding)
183
+ weight = weight + self.multiplier * conved * self.scale
184
+
185
+ # set weight to org_module
186
+ org_sd["weight"] = weight.to(dtype)
187
+ self.org_module.load_state_dict(org_sd)
188
+
189
+ # 復元できるマージのため、このモジュールのweightを返す
190
+ def get_weight(self, multiplier=None):
191
+ if multiplier is None:
192
+ multiplier = self.multiplier
193
+
194
+ # get up/down weight from module
195
+ up_weight = self.lora_up.weight.to(torch.float)
196
+ down_weight = self.lora_down.weight.to(torch.float)
197
+
198
+ # pre-calculated weight
199
+ if len(down_weight.size()) == 2:
200
+ # linear
201
+ weight = self.multiplier * (up_weight @ down_weight) * self.scale
202
+ elif down_weight.size()[2:4] == (1, 1):
203
+ # conv2d 1x1
204
+ weight = (
205
+ self.multiplier
206
+ * (up_weight.squeeze(3).squeeze(2) @ down_weight.squeeze(3).squeeze(2)).unsqueeze(2).unsqueeze(3)
207
+ * self.scale
208
+ )
209
+ else:
210
+ # conv2d 3x3
211
+ conved = torch.nn.functional.conv2d(down_weight.permute(1, 0, 2, 3), up_weight).permute(1, 0, 2, 3)
212
+ weight = self.multiplier * conved * self.scale
213
+
214
+ return weight
215
+
216
+ def set_region(self, region):
217
+ self.region = region
218
+ self.region_mask = None
219
+
220
+ def default_forward(self, x):
221
+ # print("default_forward", self.lora_name, x.size())
222
+ return self.org_forward(x) + self.lora_up(self.lora_down(x)) * self.multiplier * self.scale
223
+
224
+ def forward(self, x):
225
+ if not self.enabled:
226
+ return self.org_forward(x)
227
+
228
+ if self.network is None or self.network.sub_prompt_index is None:
229
+ return self.default_forward(x)
230
+ if not self.regional and not self.use_sub_prompt:
231
+ return self.default_forward(x)
232
+
233
+ if self.regional:
234
+ return self.regional_forward(x)
235
+ else:
236
+ return self.sub_prompt_forward(x)
237
+
238
+ def get_mask_for_x(self, x):
239
+ # calculate size from shape of x
240
+ if len(x.size()) == 4:
241
+ h, w = x.size()[2:4]
242
+ area = h * w
243
+ else:
244
+ area = x.size()[1]
245
+
246
+ mask = self.network.mask_dic[area]
247
+ if mask is None:
248
+ raise ValueError(f"mask is None for resolution {area}")
249
+ if len(x.size()) != 4:
250
+ mask = torch.reshape(mask, (1, -1, 1))
251
+ return mask
252
+
253
+ def regional_forward(self, x):
254
+ if "attn2_to_out" in self.lora_name:
255
+ return self.to_out_forward(x)
256
+
257
+ if self.network.mask_dic is None: # sub_prompt_index >= 3
258
+ return self.default_forward(x)
259
+
260
+ # apply mask for LoRA result
261
+ lx = self.lora_up(self.lora_down(x)) * self.multiplier * self.scale
262
+ mask = self.get_mask_for_x(lx)
263
+ # print("regional", self.lora_name, self.network.sub_prompt_index, lx.size(), mask.size())
264
+ lx = lx * mask
265
+
266
+ x = self.org_forward(x)
267
+ x = x + lx
268
+
269
+ if "attn2_to_q" in self.lora_name and self.network.is_last_network:
270
+ x = self.postp_to_q(x)
271
+
272
+ return x
273
+
274
+ def postp_to_q(self, x):
275
+ # repeat x to num_sub_prompts
276
+ has_real_uncond = x.size()[0] // self.network.batch_size == 3
277
+ qc = self.network.batch_size # uncond
278
+ qc += self.network.batch_size * self.network.num_sub_prompts # cond
279
+ if has_real_uncond:
280
+ qc += self.network.batch_size # real_uncond
281
+
282
+ query = torch.zeros((qc, x.size()[1], x.size()[2]), device=x.device, dtype=x.dtype)
283
+ query[: self.network.batch_size] = x[: self.network.batch_size]
284
+
285
+ for i in range(self.network.batch_size):
286
+ qi = self.network.batch_size + i * self.network.num_sub_prompts
287
+ query[qi : qi + self.network.num_sub_prompts] = x[self.network.batch_size + i]
288
+
289
+ if has_real_uncond:
290
+ query[-self.network.batch_size :] = x[-self.network.batch_size :]
291
+
292
+ # print("postp_to_q", self.lora_name, x.size(), query.size(), self.network.num_sub_prompts)
293
+ return query
294
+
295
+ def sub_prompt_forward(self, x):
296
+ if x.size()[0] == self.network.batch_size: # if uncond in text_encoder, do not apply LoRA
297
+ return self.org_forward(x)
298
+
299
+ emb_idx = self.network.sub_prompt_index
300
+ if not self.text_encoder:
301
+ emb_idx += self.network.batch_size
302
+
303
+ # apply sub prompt of X
304
+ lx = x[emb_idx :: self.network.num_sub_prompts]
305
+ lx = self.lora_up(self.lora_down(lx)) * self.multiplier * self.scale
306
+
307
+ # print("sub_prompt_forward", self.lora_name, x.size(), lx.size(), emb_idx)
308
+
309
+ x = self.org_forward(x)
310
+ x[emb_idx :: self.network.num_sub_prompts] += lx
311
+
312
+ return x
313
+
314
+ def to_out_forward(self, x):
315
+ # print("to_out_forward", self.lora_name, x.size(), self.network.is_last_network)
316
+
317
+ if self.network.is_last_network:
318
+ masks = [None] * self.network.num_sub_prompts
319
+ self.network.shared[self.lora_name] = (None, masks)
320
+ else:
321
+ lx, masks = self.network.shared[self.lora_name]
322
+
323
+ # call own LoRA
324
+ x1 = x[self.network.batch_size + self.network.sub_prompt_index :: self.network.num_sub_prompts]
325
+ lx1 = self.lora_up(self.lora_down(x1)) * self.multiplier * self.scale
326
+
327
+ if self.network.is_last_network:
328
+ lx = torch.zeros(
329
+ (self.network.num_sub_prompts * self.network.batch_size, *lx1.size()[1:]), device=lx1.device, dtype=lx1.dtype
330
+ )
331
+ self.network.shared[self.lora_name] = (lx, masks)
332
+
333
+ # print("to_out_forward", lx.size(), lx1.size(), self.network.sub_prompt_index, self.network.num_sub_prompts)
334
+ lx[self.network.sub_prompt_index :: self.network.num_sub_prompts] += lx1
335
+ masks[self.network.sub_prompt_index] = self.get_mask_for_x(lx1)
336
+
337
+ # if not last network, return x and masks
338
+ x = self.org_forward(x)
339
+ if not self.network.is_last_network:
340
+ return x
341
+
342
+ lx, masks = self.network.shared.pop(self.lora_name)
343
+
344
+ # if last network, combine separated x with mask weighted sum
345
+ has_real_uncond = x.size()[0] // self.network.batch_size == self.network.num_sub_prompts + 2
346
+
347
+ out = torch.zeros((self.network.batch_size * (3 if has_real_uncond else 2), *x.size()[1:]), device=x.device, dtype=x.dtype)
348
+ out[: self.network.batch_size] = x[: self.network.batch_size] # uncond
349
+ if has_real_uncond:
350
+ out[-self.network.batch_size :] = x[-self.network.batch_size :] # real_uncond
351
+
352
+ # print("to_out_forward", self.lora_name, self.network.sub_prompt_index, self.network.num_sub_prompts)
353
+ # for i in range(len(masks)):
354
+ # if masks[i] is None:
355
+ # masks[i] = torch.zeros_like(masks[-1])
356
+
357
+ mask = torch.cat(masks)
358
+ mask_sum = torch.sum(mask, dim=0) + 1e-4
359
+ for i in range(self.network.batch_size):
360
+ # 1枚の画像ごとに処理する
361
+ lx1 = lx[i * self.network.num_sub_prompts : (i + 1) * self.network.num_sub_prompts]
362
+ lx1 = lx1 * mask
363
+ lx1 = torch.sum(lx1, dim=0)
364
+
365
+ xi = self.network.batch_size + i * self.network.num_sub_prompts
366
+ x1 = x[xi : xi + self.network.num_sub_prompts]
367
+ x1 = x1 * mask
368
+ x1 = torch.sum(x1, dim=0)
369
+ x1 = x1 / mask_sum
370
+
371
+ x1 = x1 + lx1
372
+ out[self.network.batch_size + i] = x1
373
+
374
+ # print("to_out_forward", x.size(), out.size(), has_real_uncond)
375
+ return out
376
+
377
+
378
+ def parse_block_lr_kwargs(nw_kwargs):
379
+ down_lr_weight = nw_kwargs.get("down_lr_weight", None)
380
+ mid_lr_weight = nw_kwargs.get("mid_lr_weight", None)
381
+ up_lr_weight = nw_kwargs.get("up_lr_weight", None)
382
+
383
+ # 以上のいずれにも設定がない場合は無効としてNoneを返す
384
+ if down_lr_weight is None and mid_lr_weight is None and up_lr_weight is None:
385
+ return None, None, None
386
+
387
+ # extract learning rate weight for each block
388
+ if down_lr_weight is not None:
389
+ # if some parameters are not set, use zero
390
+ if "," in down_lr_weight:
391
+ down_lr_weight = [(float(s) if s else 0.0) for s in down_lr_weight.split(",")]
392
+
393
+ if mid_lr_weight is not None:
394
+ mid_lr_weight = float(mid_lr_weight)
395
+
396
+ if up_lr_weight is not None:
397
+ if "," in up_lr_weight:
398
+ up_lr_weight = [(float(s) if s else 0.0) for s in up_lr_weight.split(",")]
399
+
400
+ down_lr_weight, mid_lr_weight, up_lr_weight = get_block_lr_weight(
401
+ down_lr_weight, mid_lr_weight, up_lr_weight, float(nw_kwargs.get("block_lr_zero_threshold", 0.0))
402
+ )
403
+
404
+ return down_lr_weight, mid_lr_weight, up_lr_weight
405
+
406
+
407
+ def create_network(
408
+ multiplier: float,
409
+ network_dim: Optional[int],
410
+ network_alpha: Optional[float],
411
+ vae: AutoencoderKL,
412
+ text_encoder: Union[CLIPTextModel, List[CLIPTextModel]],
413
+ unet,
414
+ neuron_dropout: Optional[float] = None,
415
+ **kwargs,
416
+ ):
417
+ if network_dim is None:
418
+ network_dim = 4 # default
419
+ if network_alpha is None:
420
+ network_alpha = 1.0
421
+
422
+ # extract dim/alpha for conv2d, and block dim
423
+ conv_dim = kwargs.get("conv_dim", None)
424
+ conv_alpha = kwargs.get("conv_alpha", None)
425
+ if conv_dim is not None:
426
+ conv_dim = int(conv_dim)
427
+ if conv_alpha is None:
428
+ conv_alpha = 1.0
429
+ else:
430
+ conv_alpha = float(conv_alpha)
431
+
432
+ # block dim/alpha/lr
433
+ block_dims = kwargs.get("block_dims", None)
434
+ down_lr_weight, mid_lr_weight, up_lr_weight = parse_block_lr_kwargs(kwargs)
435
+
436
+ # 以上のいずれかに指定があればblockごとのdim(rank)を有効にする
437
+ if block_dims is not None or down_lr_weight is not None or mid_lr_weight is not None or up_lr_weight is not None:
438
+ block_alphas = kwargs.get("block_alphas", None)
439
+ conv_block_dims = kwargs.get("conv_block_dims", None)
440
+ conv_block_alphas = kwargs.get("conv_block_alphas", None)
441
+
442
+ block_dims, block_alphas, conv_block_dims, conv_block_alphas = get_block_dims_and_alphas(
443
+ block_dims, block_alphas, network_dim, network_alpha, conv_block_dims, conv_block_alphas, conv_dim, conv_alpha
444
+ )
445
+
446
+ # remove block dim/alpha without learning rate
447
+ block_dims, block_alphas, conv_block_dims, conv_block_alphas = remove_block_dims_and_alphas(
448
+ block_dims, block_alphas, conv_block_dims, conv_block_alphas, down_lr_weight, mid_lr_weight, up_lr_weight
449
+ )
450
+
451
+ else:
452
+ block_alphas = None
453
+ conv_block_dims = None
454
+ conv_block_alphas = None
455
+
456
+ # rank/module dropout
457
+ rank_dropout = kwargs.get("rank_dropout", None)
458
+ if rank_dropout is not None:
459
+ rank_dropout = float(rank_dropout)
460
+ module_dropout = kwargs.get("module_dropout", None)
461
+ if module_dropout is not None:
462
+ module_dropout = float(module_dropout)
463
+
464
+ # すごく引数が多いな ( ^ω^)・・・
465
+ network = LoRANetwork(
466
+ text_encoder,
467
+ unet,
468
+ multiplier=multiplier,
469
+ lora_dim=network_dim,
470
+ alpha=network_alpha,
471
+ dropout=neuron_dropout,
472
+ rank_dropout=rank_dropout,
473
+ module_dropout=module_dropout,
474
+ conv_lora_dim=conv_dim,
475
+ conv_alpha=conv_alpha,
476
+ block_dims=block_dims,
477
+ block_alphas=block_alphas,
478
+ conv_block_dims=conv_block_dims,
479
+ conv_block_alphas=conv_block_alphas,
480
+ varbose=True,
481
+ )
482
+
483
+ if up_lr_weight is not None or mid_lr_weight is not None or down_lr_weight is not None:
484
+ network.set_block_lr_weight(up_lr_weight, mid_lr_weight, down_lr_weight)
485
+
486
+ return network
487
+
488
+
489
+ # このメソッドは外部から呼び出される可能性を考慮しておく
490
+ # network_dim, network_alpha にはデフォルト値が入っている。
491
+ # block_dims, block_alphas は両方ともNoneまたは両方とも値が入っている
492
+ # conv_dim, conv_alpha は両方ともNoneまたは両方とも値が入っている
493
+ def get_block_dims_and_alphas(
494
+ block_dims, block_alphas, network_dim, network_alpha, conv_block_dims, conv_block_alphas, conv_dim, conv_alpha
495
+ ):
496
+ num_total_blocks = LoRANetwork.NUM_OF_BLOCKS * 2 + 1
497
+
498
+ def parse_ints(s):
499
+ return [int(i) for i in s.split(",")]
500
+
501
+ def parse_floats(s):
502
+ return [float(i) for i in s.split(",")]
503
+
504
+ # block_dimsとblock_alphasをパースする。必ず値が入る
505
+ if block_dims is not None:
506
+ block_dims = parse_ints(block_dims)
507
+ assert (
508
+ len(block_dims) == num_total_blocks
509
+ ), f"block_dims must have {num_total_blocks} elements / block_dimsは{num_total_blocks}個指定してください"
510
+ else:
511
+ print(f"block_dims is not specified. all dims are set to {network_dim} / block_dimsが指定されていません。すべてのdimは{network_dim}になります")
512
+ block_dims = [network_dim] * num_total_blocks
513
+
514
+ if block_alphas is not None:
515
+ block_alphas = parse_floats(block_alphas)
516
+ assert (
517
+ len(block_alphas) == num_total_blocks
518
+ ), f"block_alphas must have {num_total_blocks} elements / block_alphasは{num_total_blocks}個指定してください"
519
+ else:
520
+ print(
521
+ f"block_alphas is not specified. all alphas are set to {network_alpha} / block_alphasが指定されていません。すべてのalphaは{network_alpha}になります"
522
+ )
523
+ block_alphas = [network_alpha] * num_total_blocks
524
+
525
+ # conv_block_dimsとconv_block_alphasを、指定がある場合のみパースする。指定がなければconv_dimとconv_alphaを使う
526
+ if conv_block_dims is not None:
527
+ conv_block_dims = parse_ints(conv_block_dims)
528
+ assert (
529
+ len(conv_block_dims) == num_total_blocks
530
+ ), f"conv_block_dims must have {num_total_blocks} elements / conv_block_dimsは{num_total_blocks}個指定してください"
531
+
532
+ if conv_block_alphas is not None:
533
+ conv_block_alphas = parse_floats(conv_block_alphas)
534
+ assert (
535
+ len(conv_block_alphas) == num_total_blocks
536
+ ), f"conv_block_alphas must have {num_total_blocks} elements / conv_block_alphasは{num_total_blocks}個指定してください"
537
+ else:
538
+ if conv_alpha is None:
539
+ conv_alpha = 1.0
540
+ print(
541
+ f"conv_block_alphas is not specified. all alphas are set to {conv_alpha} / conv_block_alphasが指定されていません。すべてのalphaは{conv_alpha}になります"
542
+ )
543
+ conv_block_alphas = [conv_alpha] * num_total_blocks
544
+ else:
545
+ if conv_dim is not None:
546
+ print(
547
+ f"conv_dim/alpha for all blocks are set to {conv_dim} and {conv_alpha} / すべてのブロックのconv_dimとalphaは{conv_dim}および{conv_alpha}になります"
548
+ )
549
+ conv_block_dims = [conv_dim] * num_total_blocks
550
+ conv_block_alphas = [conv_alpha] * num_total_blocks
551
+ else:
552
+ conv_block_dims = None
553
+ conv_block_alphas = None
554
+
555
+ return block_dims, block_alphas, conv_block_dims, conv_block_alphas
556
+
557
+
558
+ # 層別学習率用に層ごとの学習率に対する倍率を定義する、外部から呼び出される可能性を考慮しておく
559
+ def get_block_lr_weight(
560
+ down_lr_weight, mid_lr_weight, up_lr_weight, zero_threshold
561
+ ) -> Tuple[List[float], List[float], List[float]]:
562
+ # パラメータ未指定時は何もせず、今までと同じ動作とする
563
+ if up_lr_weight is None and mid_lr_weight is None and down_lr_weight is None:
564
+ return None, None, None
565
+
566
+ max_len = LoRANetwork.NUM_OF_BLOCKS # フルモデル相当でのup,downの層の数
567
+
568
+ def get_list(name_with_suffix) -> List[float]:
569
+ import math
570
+
571
+ tokens = name_with_suffix.split("+")
572
+ name = tokens[0]
573
+ base_lr = float(tokens[1]) if len(tokens) > 1 else 0.0
574
+
575
+ if name == "cosine":
576
+ return [math.sin(math.pi * (i / (max_len - 1)) / 2) + base_lr for i in reversed(range(max_len))]
577
+ elif name == "sine":
578
+ return [math.sin(math.pi * (i / (max_len - 1)) / 2) + base_lr for i in range(max_len)]
579
+ elif name == "linear":
580
+ return [i / (max_len - 1) + base_lr for i in range(max_len)]
581
+ elif name == "reverse_linear":
582
+ return [i / (max_len - 1) + base_lr for i in reversed(range(max_len))]
583
+ elif name == "zeros":
584
+ return [0.0 + base_lr] * max_len
585
+ else:
586
+ print(
587
+ "Unknown lr_weight argument %s is used. Valid arguments: / 不明なlr_weightの引数 %s が使われました。有効な引数:\n\tcosine, sine, linear, reverse_linear, zeros"
588
+ % (name)
589
+ )
590
+ return None
591
+
592
+ if type(down_lr_weight) == str:
593
+ down_lr_weight = get_list(down_lr_weight)
594
+ if type(up_lr_weight) == str:
595
+ up_lr_weight = get_list(up_lr_weight)
596
+
597
+ if (up_lr_weight != None and len(up_lr_weight) > max_len) or (down_lr_weight != None and len(down_lr_weight) > max_len):
598
+ print("down_weight or up_weight is too long. Parameters after %d-th are ignored." % max_len)
599
+ print("down_weightもしくはup_weightが長すぎます。%d個目以降のパラメータは無視されます。" % max_len)
600
+ up_lr_weight = up_lr_weight[:max_len]
601
+ down_lr_weight = down_lr_weight[:max_len]
602
+
603
+ if (up_lr_weight != None and len(up_lr_weight) < max_len) or (down_lr_weight != None and len(down_lr_weight) < max_len):
604
+ print("down_weight or up_weight is too short. Parameters after %d-th are filled with 1." % max_len)
605
+ print("down_weightもしくはup_weightが短すぎます。%d個目までの不足したパラメータは1で補われます。" % max_len)
606
+
607
+ if down_lr_weight != None and len(down_lr_weight) < max_len:
608
+ down_lr_weight = down_lr_weight + [1.0] * (max_len - len(down_lr_weight))
609
+ if up_lr_weight != None and len(up_lr_weight) < max_len:
610
+ up_lr_weight = up_lr_weight + [1.0] * (max_len - len(up_lr_weight))
611
+
612
+ if (up_lr_weight != None) or (mid_lr_weight != None) or (down_lr_weight != None):
613
+ print("apply block learning rate / 階層別学習率を適用します。")
614
+ if down_lr_weight != None:
615
+ down_lr_weight = [w if w > zero_threshold else 0 for w in down_lr_weight]
616
+ print("down_lr_weight (shallower -> deeper, 浅い層->深い層):", down_lr_weight)
617
+ else:
618
+ print("down_lr_weight: all 1.0, すべて1.0")
619
+
620
+ if mid_lr_weight != None:
621
+ mid_lr_weight = mid_lr_weight if mid_lr_weight > zero_threshold else 0
622
+ print("mid_lr_weight:", mid_lr_weight)
623
+ else:
624
+ print("mid_lr_weight: 1.0")
625
+
626
+ if up_lr_weight != None:
627
+ up_lr_weight = [w if w > zero_threshold else 0 for w in up_lr_weight]
628
+ print("up_lr_weight (deeper -> shallower, 深い層->浅い層):", up_lr_weight)
629
+ else:
630
+ print("up_lr_weight: all 1.0, すべて1.0")
631
+
632
+ return down_lr_weight, mid_lr_weight, up_lr_weight
633
+
634
+
635
+ # lr_weightが0のblockをblock_dimsから除外する、外部から呼び出す可能性を考慮しておく
636
+ def remove_block_dims_and_alphas(
637
+ block_dims, block_alphas, conv_block_dims, conv_block_alphas, down_lr_weight, mid_lr_weight, up_lr_weight
638
+ ):
639
+ # set 0 to block dim without learning rate to remove the block
640
+ if down_lr_weight != None:
641
+ for i, lr in enumerate(down_lr_weight):
642
+ if lr == 0:
643
+ block_dims[i] = 0
644
+ if conv_block_dims is not None:
645
+ conv_block_dims[i] = 0
646
+ if mid_lr_weight != None:
647
+ if mid_lr_weight == 0:
648
+ block_dims[LoRANetwork.NUM_OF_BLOCKS] = 0
649
+ if conv_block_dims is not None:
650
+ conv_block_dims[LoRANetwork.NUM_OF_BLOCKS] = 0
651
+ if up_lr_weight != None:
652
+ for i, lr in enumerate(up_lr_weight):
653
+ if lr == 0:
654
+ block_dims[LoRANetwork.NUM_OF_BLOCKS + 1 + i] = 0
655
+ if conv_block_dims is not None:
656
+ conv_block_dims[LoRANetwork.NUM_OF_BLOCKS + 1 + i] = 0
657
+
658
+ return block_dims, block_alphas, conv_block_dims, conv_block_alphas
659
+
660
+
661
+ # 外部から呼び出す可能性を考慮しておく
662
+ def get_block_index(lora_name: str) -> int:
663
+ block_idx = -1 # invalid lora name
664
+
665
+ m = RE_UPDOWN.search(lora_name)
666
+ if m:
667
+ g = m.groups()
668
+ i = int(g[1])
669
+ j = int(g[3])
670
+ if g[2] == "resnets":
671
+ idx = 3 * i + j
672
+ elif g[2] == "attentions":
673
+ idx = 3 * i + j
674
+ elif g[2] == "upsamplers" or g[2] == "downsamplers":
675
+ idx = 3 * i + 2
676
+
677
+ if g[0] == "down":
678
+ block_idx = 1 + idx # 0に該当するLoRAは存在しない
679
+ elif g[0] == "up":
680
+ block_idx = LoRANetwork.NUM_OF_BLOCKS + 1 + idx
681
+
682
+ elif "mid_block_" in lora_name:
683
+ block_idx = LoRANetwork.NUM_OF_BLOCKS # idx=12
684
+
685
+ return block_idx
686
+
687
+
688
+ # Create network from weights for inference, weights are not loaded here (because can be merged)
689
+ def create_network_from_weights(multiplier, file, vae, text_encoder, unet, weights_sd=None, for_inference=False, **kwargs):
690
+ if weights_sd is None:
691
+ if os.path.splitext(file)[1] == ".safetensors":
692
+ from safetensors.torch import load_file, safe_open
693
+
694
+ weights_sd = load_file(file)
695
+ else:
696
+ weights_sd = torch.load(file, map_location="cpu")
697
+
698
+ # get dim/alpha mapping
699
+ modules_dim = {}
700
+ modules_alpha = {}
701
+ for key, value in weights_sd.items():
702
+ if "." not in key:
703
+ continue
704
+
705
+ lora_name = key.split(".")[0]
706
+ if "alpha" in key:
707
+ modules_alpha[lora_name] = value
708
+ elif "lora_down" in key:
709
+ dim = value.size()[0]
710
+ modules_dim[lora_name] = dim
711
+ # print(lora_name, value.size(), dim)
712
+
713
+ # support old LoRA without alpha
714
+ for key in modules_dim.keys():
715
+ if key not in modules_alpha:
716
+ modules_alpha[key] = modules_dim[key]
717
+
718
+ module_class = LoRAInfModule if for_inference else LoRAModule
719
+
720
+ network = LoRANetwork(
721
+ text_encoder, unet, multiplier=multiplier, modules_dim=modules_dim, modules_alpha=modules_alpha, module_class=module_class
722
+ )
723
+
724
+ # block lr
725
+ down_lr_weight, mid_lr_weight, up_lr_weight = parse_block_lr_kwargs(kwargs)
726
+ if up_lr_weight is not None or mid_lr_weight is not None or down_lr_weight is not None:
727
+ network.set_block_lr_weight(up_lr_weight, mid_lr_weight, down_lr_weight)
728
+
729
+ return network, weights_sd
730
+
731
+
732
+ class LoRANetwork(torch.nn.Module):
733
+ NUM_OF_BLOCKS = 12 # フルモデル相当でのup,downの層の数
734
+
735
+ UNET_TARGET_REPLACE_MODULE = ["Transformer2DModel"]
736
+ UNET_TARGET_REPLACE_MODULE_CONV2D_3X3 = ["ResnetBlock2D", "Downsample2D", "Upsample2D"]
737
+ TEXT_ENCODER_TARGET_REPLACE_MODULE = ["CLIPAttention", "CLIPMLP"]
738
+ LORA_PREFIX_UNET = "lora_unet"
739
+ LORA_PREFIX_TEXT_ENCODER = "lora_te"
740
+
741
+ # SDXL: must starts with LORA_PREFIX_TEXT_ENCODER
742
+ LORA_PREFIX_TEXT_ENCODER1 = "lora_te1"
743
+ LORA_PREFIX_TEXT_ENCODER2 = "lora_te2"
744
+
745
+ def __init__(
746
+ self,
747
+ text_encoder: Union[List[CLIPTextModel], CLIPTextModel],
748
+ unet,
749
+ multiplier: float = 1.0,
750
+ lora_dim: int = 4,
751
+ alpha: float = 1,
752
+ dropout: Optional[float] = None,
753
+ rank_dropout: Optional[float] = None,
754
+ module_dropout: Optional[float] = None,
755
+ conv_lora_dim: Optional[int] = None,
756
+ conv_alpha: Optional[float] = None,
757
+ block_dims: Optional[List[int]] = None,
758
+ block_alphas: Optional[List[float]] = None,
759
+ conv_block_dims: Optional[List[int]] = None,
760
+ conv_block_alphas: Optional[List[float]] = None,
761
+ modules_dim: Optional[Dict[str, int]] = None,
762
+ modules_alpha: Optional[Dict[str, int]] = None,
763
+ module_class: Type[object] = LoRAModule,
764
+ varbose: Optional[bool] = False,
765
+ ) -> None:
766
+ """
767
+ LoRA network: すごく引数が多いが、パターンは以下の通り
768
+ 1. lora_dimとalphaを指定
769
+ 2. lora_dim、alpha、conv_lora_dim、conv_alphaを指定
770
+ 3. block_dimsとblock_alphasを指定 : Conv2d3x3には適用しない
771
+ 4. block_dims、block_alphas、conv_block_dims、conv_block_alphasを指定 : Conv2d3x3にも適用する
772
+ 5. modules_dimとmodules_alphaを指定 (推論用)
773
+ """
774
+ super().__init__()
775
+ self.multiplier = multiplier
776
+
777
+ self.lora_dim = lora_dim
778
+ self.alpha = alpha
779
+ self.conv_lora_dim = conv_lora_dim
780
+ self.conv_alpha = conv_alpha
781
+ self.dropout = dropout
782
+ self.rank_dropout = rank_dropout
783
+ self.module_dropout = module_dropout
784
+
785
+ if modules_dim is not None:
786
+ print(f"create LoRA network from weights")
787
+ elif block_dims is not None:
788
+ print(f"create LoRA network from block_dims")
789
+ print(f"neuron dropout: p={self.dropout}, rank dropout: p={self.rank_dropout}, module dropout: p={self.module_dropout}")
790
+ print(f"block_dims: {block_dims}")
791
+ print(f"block_alphas: {block_alphas}")
792
+ if conv_block_dims is not None:
793
+ print(f"conv_block_dims: {conv_block_dims}")
794
+ print(f"conv_block_alphas: {conv_block_alphas}")
795
+ else:
796
+ print(f"create LoRA network. base dim (rank): {lora_dim}, alpha: {alpha}")
797
+ print(f"neuron dropout: p={self.dropout}, rank dropout: p={self.rank_dropout}, module dropout: p={self.module_dropout}")
798
+ if self.conv_lora_dim is not None:
799
+ print(f"apply LoRA to Conv2d with kernel size (3,3). dim (rank): {self.conv_lora_dim}, alpha: {self.conv_alpha}")
800
+
801
+ # create module instances
802
+ def create_modules(
803
+ is_unet: bool,
804
+ text_encoder_idx: Optional[int], # None, 1, 2
805
+ root_module: torch.nn.Module,
806
+ target_replace_modules: List[torch.nn.Module],
807
+ ) -> List[LoRAModule]:
808
+ prefix = (
809
+ self.LORA_PREFIX_UNET
810
+ if is_unet
811
+ else (
812
+ self.LORA_PREFIX_TEXT_ENCODER
813
+ if text_encoder_idx is None
814
+ else (self.LORA_PREFIX_TEXT_ENCODER1 if text_encoder_idx == 1 else self.LORA_PREFIX_TEXT_ENCODER2)
815
+ )
816
+ )
817
+ loras = []
818
+ skipped = []
819
+ for name, module in root_module.named_modules():
820
+ if module.__class__.__name__ in target_replace_modules:
821
+ for child_name, child_module in module.named_modules():
822
+ is_linear = child_module.__class__.__name__ == "Linear"
823
+ is_conv2d = child_module.__class__.__name__ == "Conv2d"
824
+ is_conv2d_1x1 = is_conv2d and child_module.kernel_size == (1, 1)
825
+
826
+ if is_linear or is_conv2d:
827
+ lora_name = prefix + "." + name + "." + child_name
828
+ lora_name = lora_name.replace(".", "_")
829
+
830
+ dim = None
831
+ alpha = None
832
+
833
+ if modules_dim is not None:
834
+ # モジュール指定あり
835
+ if lora_name in modules_dim:
836
+ dim = modules_dim[lora_name]
837
+ alpha = modules_alpha[lora_name]
838
+ elif is_unet and block_dims is not None:
839
+ # U-Netでblock_dims指定あり
840
+ block_idx = get_block_index(lora_name)
841
+ if is_linear or is_conv2d_1x1:
842
+ dim = block_dims[block_idx]
843
+ alpha = block_alphas[block_idx]
844
+ elif conv_block_dims is not None:
845
+ dim = conv_block_dims[block_idx]
846
+ alpha = conv_block_alphas[block_idx]
847
+ else:
848
+ # 通常、すべて対象とする
849
+ if is_linear or is_conv2d_1x1:
850
+ dim = self.lora_dim
851
+ alpha = self.alpha
852
+ elif self.conv_lora_dim is not None:
853
+ dim = self.conv_lora_dim
854
+ alpha = self.conv_alpha
855
+
856
+ if dim is None or dim == 0:
857
+ # skipした情報を出力
858
+ if is_linear or is_conv2d_1x1 or (self.conv_lora_dim is not None or conv_block_dims is not None):
859
+ skipped.append(lora_name)
860
+ continue
861
+
862
+ lora = module_class(
863
+ lora_name,
864
+ child_module,
865
+ self.multiplier,
866
+ dim,
867
+ alpha,
868
+ dropout=dropout,
869
+ rank_dropout=rank_dropout,
870
+ module_dropout=module_dropout,
871
+ )
872
+ loras.append(lora)
873
+ return loras, skipped
874
+
875
+ text_encoders = text_encoder if type(text_encoder) == list else [text_encoder]
876
+ print(text_encoders)
877
+ # create LoRA for text encoder
878
+ # 毎回すべてのモジュールを作るのは無駄なので要検討
879
+ self.text_encoder_loras = []
880
+ skipped_te = []
881
+ for i, text_encoder in enumerate(text_encoders):
882
+ if len(text_encoders) > 1:
883
+ index = i + 1
884
+ print(f"create LoRA for Text Encoder {index}:")
885
+ else:
886
+ index = None
887
+ print(f"create LoRA for Text Encoder:")
888
+
889
+ print(text_encoder)
890
+ text_encoder_loras, skipped = create_modules(False, index, text_encoder, LoRANetwork.TEXT_ENCODER_TARGET_REPLACE_MODULE)
891
+ self.text_encoder_loras.extend(text_encoder_loras)
892
+ skipped_te += skipped
893
+ print(f"create LoRA for Text Encoder: {len(self.text_encoder_loras)} modules.")
894
+
895
+ # extend U-Net target modules if conv2d 3x3 is enabled, or load from weights
896
+ target_modules = LoRANetwork.UNET_TARGET_REPLACE_MODULE
897
+ if modules_dim is not None or self.conv_lora_dim is not None or conv_block_dims is not None:
898
+ target_modules += LoRANetwork.UNET_TARGET_REPLACE_MODULE_CONV2D_3X3
899
+
900
+ self.unet_loras, skipped_un = create_modules(True, None, unet, target_modules)
901
+ print(f"create LoRA for U-Net: {len(self.unet_loras)} modules.")
902
+
903
+ skipped = skipped_te + skipped_un
904
+ if varbose and len(skipped) > 0:
905
+ print(
906
+ f"because block_lr_weight is 0 or dim (rank) is 0, {len(skipped)} LoRA modules are skipped / block_lr_weightまたはdim (rank)が0の為、次の{len(skipped)}個のLoRAモジュールはスキップされます:"
907
+ )
908
+ for name in skipped:
909
+ print(f"\t{name}")
910
+
911
+ self.up_lr_weight: List[float] = None
912
+ self.down_lr_weight: List[float] = None
913
+ self.mid_lr_weight: float = None
914
+ self.block_lr = False
915
+
916
+ # assertion
917
+ names = set()
918
+ for lora in self.text_encoder_loras + self.unet_loras:
919
+ assert lora.lora_name not in names, f"duplicated lora name: {lora.lora_name}"
920
+ names.add(lora.lora_name)
921
+
922
+ def set_multiplier(self, multiplier):
923
+ self.multiplier = multiplier
924
+ for lora in self.text_encoder_loras + self.unet_loras:
925
+ lora.multiplier = self.multiplier
926
+
927
+ def load_weights(self, file):
928
+ if os.path.splitext(file)[1] == ".safetensors":
929
+ from safetensors.torch import load_file
930
+
931
+ weights_sd = load_file(file)
932
+ else:
933
+ weights_sd = torch.load(file, map_location="cpu")
934
+ info = self.load_state_dict(weights_sd, False)
935
+ return info
936
+
937
+ def apply_to(self, text_encoder, unet, apply_text_encoder=True, apply_unet=True):
938
+ if apply_text_encoder:
939
+ print("enable LoRA for text encoder")
940
+ else:
941
+ self.text_encoder_loras = []
942
+
943
+ if apply_unet:
944
+ print("enable LoRA for U-Net")
945
+ else:
946
+ self.unet_loras = []
947
+
948
+ for lora in self.text_encoder_loras + self.unet_loras:
949
+ lora.apply_to()
950
+ self.add_module(lora.lora_name, lora)
951
+
952
+ # マージできるかどうかを返す
953
+ def is_mergeable(self):
954
+ return True
955
+
956
+ # TODO refactor to common function with apply_to
957
+ def merge_to(self, text_encoder, unet, weights_sd, dtype, device):
958
+ apply_text_encoder = apply_unet = False
959
+ for key in weights_sd.keys():
960
+ if key.startswith(LoRANetwork.LORA_PREFIX_TEXT_ENCODER):
961
+ apply_text_encoder = True
962
+ elif key.startswith(LoRANetwork.LORA_PREFIX_UNET):
963
+ apply_unet = True
964
+
965
+ if apply_text_encoder:
966
+ print("enable LoRA for text encoder")
967
+ else:
968
+ self.text_encoder_loras = []
969
+
970
+ if apply_unet:
971
+ print("enable LoRA for U-Net")
972
+ else:
973
+ self.unet_loras = []
974
+
975
+ for lora in self.text_encoder_loras + self.unet_loras:
976
+ sd_for_lora = {}
977
+ for key in weights_sd.keys():
978
+ if key.startswith(lora.lora_name):
979
+ sd_for_lora[key[len(lora.lora_name) + 1 :]] = weights_sd[key]
980
+ lora.merge_to(sd_for_lora, dtype, device)
981
+
982
+ print(f"weights are merged")
983
+
984
+ # 層別学習率用に層ごとの学習率に対する倍率を定義する 引数の順番が逆だがとりあえず気にしない
985
+ def set_block_lr_weight(
986
+ self,
987
+ up_lr_weight: List[float] = None,
988
+ mid_lr_weight: float = None,
989
+ down_lr_weight: List[float] = None,
990
+ ):
991
+ self.block_lr = True
992
+ self.down_lr_weight = down_lr_weight
993
+ self.mid_lr_weight = mid_lr_weight
994
+ self.up_lr_weight = up_lr_weight
995
+
996
+ def get_lr_weight(self, lora: LoRAModule) -> float:
997
+ lr_weight = 1.0
998
+ block_idx = get_block_index(lora.lora_name)
999
+ if block_idx < 0:
1000
+ return lr_weight
1001
+
1002
+ if block_idx < LoRANetwork.NUM_OF_BLOCKS:
1003
+ if self.down_lr_weight != None:
1004
+ lr_weight = self.down_lr_weight[block_idx]
1005
+ elif block_idx == LoRANetwork.NUM_OF_BLOCKS:
1006
+ if self.mid_lr_weight != None:
1007
+ lr_weight = self.mid_lr_weight
1008
+ elif block_idx > LoRANetwork.NUM_OF_BLOCKS:
1009
+ if self.up_lr_weight != None:
1010
+ lr_weight = self.up_lr_weight[block_idx - LoRANetwork.NUM_OF_BLOCKS - 1]
1011
+
1012
+ return lr_weight
1013
+
1014
+ # 二つのText Encoderに別々の学習率を設定できるようにするといいかも
1015
+ def prepare_optimizer_params(self, text_encoder_lr, unet_lr, default_lr):
1016
+ self.requires_grad_(True)
1017
+ all_params = []
1018
+
1019
+ def enumerate_params(loras):
1020
+ params = []
1021
+ for lora in loras:
1022
+ params.extend(lora.parameters())
1023
+ return params
1024
+
1025
+ if self.text_encoder_loras:
1026
+ param_data = {"params": enumerate_params(self.text_encoder_loras)}
1027
+ if text_encoder_lr is not None:
1028
+ param_data["lr"] = text_encoder_lr
1029
+ all_params.append(param_data)
1030
+
1031
+ if self.unet_loras:
1032
+ if self.block_lr:
1033
+ # 学習率のグラフをblockごとにしたいので、blockごとにloraを分類
1034
+ block_idx_to_lora = {}
1035
+ for lora in self.unet_loras:
1036
+ idx = get_block_index(lora.lora_name)
1037
+ if idx not in block_idx_to_lora:
1038
+ block_idx_to_lora[idx] = []
1039
+ block_idx_to_lora[idx].append(lora)
1040
+
1041
+ # blockごとにパラメータを設定する
1042
+ for idx, block_loras in block_idx_to_lora.items():
1043
+ param_data = {"params": enumerate_params(block_loras)}
1044
+
1045
+ if unet_lr is not None:
1046
+ param_data["lr"] = unet_lr * self.get_lr_weight(block_loras[0])
1047
+ elif default_lr is not None:
1048
+ param_data["lr"] = default_lr * self.get_lr_weight(block_loras[0])
1049
+ if ("lr" in param_data) and (param_data["lr"] == 0):
1050
+ continue
1051
+ all_params.append(param_data)
1052
+
1053
+ else:
1054
+ param_data = {"params": enumerate_params(self.unet_loras)}
1055
+ if unet_lr is not None:
1056
+ param_data["lr"] = unet_lr
1057
+ all_params.append(param_data)
1058
+
1059
+ return all_params
1060
+
1061
+ def enable_gradient_checkpointing(self):
1062
+ # not supported
1063
+ pass
1064
+
1065
+ def prepare_grad_etc(self, text_encoder, unet):
1066
+ self.requires_grad_(True)
1067
+
1068
+ def on_epoch_start(self, text_encoder, unet):
1069
+ self.train()
1070
+
1071
+ def get_trainable_params(self):
1072
+ return self.parameters()
1073
+
1074
+ def save_weights(self, file, dtype, metadata):
1075
+ if metadata is not None and len(metadata) == 0:
1076
+ metadata = None
1077
+
1078
+ state_dict = self.state_dict()
1079
+
1080
+ if dtype is not None:
1081
+ for key in list(state_dict.keys()):
1082
+ v = state_dict[key]
1083
+ v = v.detach().clone().to("cpu").to(dtype)
1084
+ state_dict[key] = v
1085
+
1086
+ if os.path.splitext(file)[1] == ".safetensors":
1087
+ from safetensors.torch import save_file
1088
+ from library import train_util
1089
+
1090
+ # Precalculate model hashes to save time on indexing
1091
+ if metadata is None:
1092
+ metadata = {}
1093
+ model_hash, legacy_hash = train_util.precalculate_safetensors_hashes(state_dict, metadata)
1094
+ metadata["sshs_model_hash"] = model_hash
1095
+ metadata["sshs_legacy_hash"] = legacy_hash
1096
+
1097
+ save_file(state_dict, file, metadata)
1098
+ else:
1099
+ torch.save(state_dict, file)
1100
+
1101
+ # mask is a tensor with values from 0 to 1
1102
+ def set_region(self, sub_prompt_index, is_last_network, mask):
1103
+ if mask.max() == 0:
1104
+ mask = torch.ones_like(mask)
1105
+
1106
+ self.mask = mask
1107
+ self.sub_prompt_index = sub_prompt_index
1108
+ self.is_last_network = is_last_network
1109
+
1110
+ for lora in self.text_encoder_loras + self.unet_loras:
1111
+ lora.set_network(self)
1112
+
1113
+ def set_current_generation(self, batch_size, num_sub_prompts, width, height, shared):
1114
+ self.batch_size = batch_size
1115
+ self.num_sub_prompts = num_sub_prompts
1116
+ self.current_size = (height, width)
1117
+ self.shared = shared
1118
+
1119
+ # create masks
1120
+ mask = self.mask
1121
+ mask_dic = {}
1122
+ mask = mask.unsqueeze(0).unsqueeze(1) # b(1),c(1),h,w
1123
+ ref_weight = self.text_encoder_loras[0].lora_down.weight if self.text_encoder_loras else self.unet_loras[0].lora_down.weight
1124
+ dtype = ref_weight.dtype
1125
+ device = ref_weight.device
1126
+
1127
+ def resize_add(mh, mw):
1128
+ # print(mh, mw, mh * mw)
1129
+ m = torch.nn.functional.interpolate(mask, (mh, mw), mode="bilinear") # doesn't work in bf16
1130
+ m = m.to(device, dtype=dtype)
1131
+ mask_dic[mh * mw] = m
1132
+
1133
+ h = height // 8
1134
+ w = width // 8
1135
+ for _ in range(4):
1136
+ resize_add(h, w)
1137
+ if h % 2 == 1 or w % 2 == 1: # add extra shape if h/w is not divisible by 2
1138
+ resize_add(h + h % 2, w + w % 2)
1139
+ h = (h + 1) // 2
1140
+ w = (w + 1) // 2
1141
+
1142
+ self.mask_dic = mask_dic
1143
+
1144
+ def backup_weights(self):
1145
+ # 重みのバックアップを行う
1146
+ loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras
1147
+ for lora in loras:
1148
+ org_module = lora.org_module_ref[0]
1149
+ if not hasattr(org_module, "_lora_org_weight"):
1150
+ sd = org_module.state_dict()
1151
+ org_module._lora_org_weight = sd["weight"].detach().clone()
1152
+ org_module._lora_restored = True
1153
+
1154
+ def restore_weights(self):
1155
+ # 重みのリストアを行う
1156
+ loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras
1157
+ for lora in loras:
1158
+ org_module = lora.org_module_ref[0]
1159
+ if not org_module._lora_restored:
1160
+ sd = org_module.state_dict()
1161
+ sd["weight"] = org_module._lora_org_weight
1162
+ org_module.load_state_dict(sd)
1163
+ org_module._lora_restored = True
1164
+
1165
+ def pre_calculation(self):
1166
+ # 事前計算を行う
1167
+ loras: List[LoRAInfModule] = self.text_encoder_loras + self.unet_loras
1168
+ for lora in loras:
1169
+ org_module = lora.org_module_ref[0]
1170
+ sd = org_module.state_dict()
1171
+
1172
+ org_weight = sd["weight"]
1173
+ lora_weight = lora.get_weight().to(org_weight.device, dtype=org_weight.dtype)
1174
+ sd["weight"] = org_weight + lora_weight
1175
+ assert sd["weight"].shape == org_weight.shape
1176
+ org_module.load_state_dict(sd)
1177
+
1178
+ org_module._lora_restored = False
1179
+ lora.enabled = False
1180
+
1181
+ def apply_max_norm_regularization(self, max_norm_value, device):
1182
+ downkeys = []
1183
+ upkeys = []
1184
+ alphakeys = []
1185
+ norms = []
1186
+ keys_scaled = 0
1187
+
1188
+ state_dict = self.state_dict()
1189
+ for key in state_dict.keys():
1190
+ if "lora_down" in key and "weight" in key:
1191
+ downkeys.append(key)
1192
+ upkeys.append(key.replace("lora_down", "lora_up"))
1193
+ alphakeys.append(key.replace("lora_down.weight", "alpha"))
1194
+
1195
+ for i in range(len(downkeys)):
1196
+ down = state_dict[downkeys[i]].to(device)
1197
+ up = state_dict[upkeys[i]].to(device)
1198
+ alpha = state_dict[alphakeys[i]].to(device)
1199
+ dim = down.shape[0]
1200
+ scale = alpha / dim
1201
+
1202
+ if up.shape[2:] == (1, 1) and down.shape[2:] == (1, 1):
1203
+ updown = (up.squeeze(2).squeeze(2) @ down.squeeze(2).squeeze(2)).unsqueeze(2).unsqueeze(3)
1204
+ elif up.shape[2:] == (3, 3) or down.shape[2:] == (3, 3):
1205
+ updown = torch.nn.functional.conv2d(down.permute(1, 0, 2, 3), up).permute(1, 0, 2, 3)
1206
+ else:
1207
+ updown = up @ down
1208
+
1209
+ updown *= scale
1210
+
1211
+ norm = updown.norm().clamp(min=max_norm_value / 2)
1212
+ desired = torch.clamp(norm, max=max_norm_value)
1213
+ ratio = desired.cpu() / norm.cpu()
1214
+ sqrt_ratio = ratio**0.5
1215
+ if ratio != 1:
1216
+ keys_scaled += 1
1217
+ state_dict[upkeys[i]] *= sqrt_ratio
1218
+ state_dict[downkeys[i]] *= sqrt_ratio
1219
+ scalednorm = updown.norm() * ratio
1220
+ norms.append(scalednorm.item())
1221
+
1222
+ return keys_scaled, sum(norms) / len(norms), max(norms)