File size: 14,803 Bytes
faa56b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
#  ------------------------------------------------------------------------------------------
#  Copyright (c) Microsoft Corporation. All rights reserved.
#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
#  ------------------------------------------------------------------------------------------
import math
from typing import List, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F

from .layers import LoRALayer


class SVDLinear(nn.Linear, LoRALayer):
    # SVD-based adaptation implemented in a dense layer
    def __init__(
        self, 
        in_features: int, 
        out_features: int, 
        r: int = 0, 
        lora_alpha: int = 1, 
        lora_dropout: float = 0.,
        fan_in_fan_out: bool = False, 
        merge_weights: bool = True,
        **kwargs
    ):
        nn.Linear.__init__(self, in_features, out_features, **kwargs)
        LoRALayer.__init__(self, r=r, lora_alpha=lora_alpha, lora_dropout=lora_dropout,
                           merge_weights=merge_weights)

        self.fan_in_fan_out = fan_in_fan_out
        # Actual trainable parameters
        if r > 0:
            self.lora_A = nn.Parameter(
                self.weight.new_zeros((r, in_features))
            )
            self.lora_E = nn.Parameter(
                self.weight.new_zeros(r, 1)
            ) 
            self.lora_B = nn.Parameter(
                self.weight.new_zeros((out_features, r))
            )
            self.ranknum = nn.Parameter(
                self.weight.new_zeros(1), requires_grad=False
            )
            self.ranknum.data.fill_(float(self.r))
            self.scaling = self.lora_alpha if self.lora_alpha>0 else float(self.r)   
            # Freezing the pre-trained weight matrix
            self.weight.requires_grad = False
            self.ranknum.requires_grad = False
        self.reset_parameters()
        if fan_in_fan_out:
            self.weight.data = self.weight.data.T

    def reset_parameters(self):
        nn.Linear.reset_parameters(self)
        if hasattr(self, 'lora_A'):
            # initialize A,B the same way as the default for nn.Linear 
            # and E (singular values) for zero 
            nn.init.zeros_(self.lora_E)
            nn.init.normal_(self.lora_A, mean=0.0, std=0.02)
            nn.init.normal_(self.lora_B, mean=0.0, std=0.02)

    def train(self, mode: bool = True):
        def T(w):
            return w.T if self.fan_in_fan_out else w
        nn.Linear.train(self, mode)
        if self.merge_weights and self.merged:
            # Make sure that the weights are not merged
            if self.r > 0:
                self.weight.data -= T(
                    self.lora_B @ (self.lora_A*self.lora_E)
                ) * self.scaling / (self.ranknum+1e-5)
            self.merged = False
    
    def eval(self):
        def T(w):
            return w.T if self.fan_in_fan_out else w
        nn.Linear.eval(self)
        if self.merge_weights and not self.merged:
            # Merge the weights and mark it
            if self.r > 0:
                self.weight.data += T(
                    self.lora_B @ (self.lora_A * self.lora_E)
                ) * self.scaling / (self.ranknum+1e-5)
            self.merged = True

    def forward(self, x: torch.Tensor):
        def T(w):
            return w.T if self.fan_in_fan_out else w
        if self.r > 0 and not self.merged:
            result = F.linear(x, T(self.weight), bias=self.bias)
            if self.r > 0:
                result += (
                    self.lora_dropout(x) @ (self.lora_A * self.lora_E).T @ self.lora_B.T
                ) * self.scaling / (self.ranknum+1e-5)
            return result
        else:
            return F.linear(x, T(self.weight), bias=self.bias)


class RankAllocator(object):
    """
    The RankAllocator for AdaLoRA Model that will be called every training step. 
    Paper: https://openreview.net/pdf?id=lq62uWRJjiY

    Args:
        model: the model that we apply AdaLoRA to.
        lora_r (`int`): The initial rank for each incremental matrix.
        target_rank (`int`): The target average rank of incremental matrix.
        init_warmup (`int`): The steps of initial fine-tuning warmup.
        final_warmup (`int`): The step of final fine-tuning.
        mask_interval (`int`): The time internval between two budget allocations.
        beta1 (`float`): The hyperparameter of EMA for sensitivity smoothing.
        beta2 (`float`): The hyperparameter of EMA for undertainty quantification.
        total_step (`int`): The total training steps, correctly configured before training.
        target_total_rank (`Optinal[int]`): The speficified final total rank. 
        tb_writter (`SummaryWriter`): Tensorboard SummaryWriter. 
        tb_writter_loginterval (`int`): The logging interval of SummaryWriter. 
    """
    def __init__(
        self, model, 
        lora_r:int,
        target_rank:int, 
        init_warmup:int, 
        final_warmup:int,
        mask_interval:int,
        beta1:float, 
        beta2:float, 
        total_step:Optional[int]=None, 
        target_total_rank:Optional[int]=None,
        tb_writter=None,
        tb_writter_loginterval:int=500, 
    ):
        self.ave_target_rank = target_rank 
        self.target_rank = target_total_rank
        self.lora_init_rank = lora_r 
        self.initial_warmup = init_warmup
        self.final_warmup = final_warmup 
        self.mask_interval = mask_interval
        self.beta1 = beta1
        self.beta2 = beta2
        self.total_step = total_step

        self.model = model
        self.ipt = {} 
        self.exp_avg_ipt = {}
        self.exp_avg_unc = {}
        self.cat_ipt = {}
        self.rank_pattern = {} 
        self.get_lora_param_name()

        self.tb_writter = tb_writter
        self.log_interval = tb_writter_loginterval 

        assert (self.beta1<1 and self.beta1>0)
        assert (self.beta2<1 and self.beta2>0)

    def set_total_step(self, total_step:int): 
        # Set total step number 
        self.total_step = total_step
        assert self.total_step>self.initial_warmup+self.final_warmup

    def get_rank_pattern(self):
        # Return rank pattern 
        return self.rank_pattern

    def get_lora_param_name(self):
        # Prepare the budget scheduler 
        self.name_set = set() 
        self.total_rank = 0 
        self.shape_dict = {}
        for n,p in self.model.named_parameters():
            if "lora_A" in n: 
                name_mat = n.replace("lora_A", "%s")
                self.name_set.add(name_mat)
                self.total_rank += p.size(0) 
                self.shape_dict[n] = p.shape
            if "lora_B" in n:
                self.shape_dict[n] = p.shape
        self.name_set = list(sorted(self.name_set)) 
        if self.target_rank is None:
            self.target_rank = self.ave_target_rank * len(self.name_set) 

    def schedule_threshold(self, step:int):
        # Global budget schedule
        mask_ind = False 
        target_rank = self.target_rank 
        initial_warmup = self.initial_warmup 
        final_warmup = self.final_warmup 
        total_step = self.total_step 
        self.global_step = step
        if step <= initial_warmup: 
            # Initial warmup 
            curr_rank = self.total_rank 
            mask_ind = False 
        elif step > total_step - final_warmup: 
            # Final fine-tuning 
            curr_rank = self.target_rank 
            # Fix the rank pattern by 
            # always masking the same unimportant singluar values 
            mask_ind = True 
        else: 
            # Budget decreasing 
            mul_coeff = 1-(step-initial_warmup)/(total_step-final_warmup-initial_warmup)
            curr_rank = target_rank + (self.total_rank-target_rank)*(mul_coeff**3)
            curr_rank = int(curr_rank)
            mask_ind = True if step % self.mask_interval == 0 else False 
        return curr_rank, mask_ind 


    def update_ipt(self, model): 
        for n,p in model.named_parameters():
            if "lora_" in n: 
                if n not in self.ipt:
                    self.ipt[n] = torch.zeros_like(p)
                    self.exp_avg_ipt[n] = torch.zeros_like(p) 
                    self.exp_avg_unc[n] = torch.zeros_like(p) 
                with torch.no_grad():
                    # Calculate sensitivity 
                    self.ipt[n] = (p * p.grad).abs().detach()
                    # Update sensitivity 
                    self.exp_avg_ipt[n] = self.beta1 * self.exp_avg_ipt[n] + \
                                        (1-self.beta1)*self.ipt[n]
                    # Update uncertainty 
                    self.exp_avg_unc[n] = self.beta2 * self.exp_avg_unc[n] + \
                                        (1-self.beta2)*(self.ipt[n]-self.exp_avg_ipt[n]).abs()

    def calculate_score(self, n, p=None, metric="ipt"):
        if metric == "ipt":
            # Combine the senstivity and uncertainty 
            ipt_score = self.exp_avg_ipt[n] * self.exp_avg_unc[n]
        elif metric == "mag":
            ipt_score = p.abs().detach().clone() 
        else:
            raise ValueError("Unexcptected Metric: %s"%metric)
        return ipt_score 

    def _combine_ipt(self, ipt_E, ipt_AB):
        ipt_AB = ipt_AB.sum(dim=1, keepdim=False)
        sum_ipt = ipt_E.view(-1) + ipt_AB.view(-1)
        return sum_ipt

    def mask_to_target_rank(self, model, curr_rank): 
        is_dict = {}
        combine_dict = {} 
        singular_dict = {}
        # Calculate the importance score for each sub matrix 
        for n,p in model.named_parameters(): 
            if "lora_A" in n: 
                rdim, hdim_a = p.shape
                ipt_score = self.calculate_score(n, metric="ipt")
                comb_ipt = torch.mean(ipt_score, dim=1, keepdim=True)
                name_mat = n.replace("lora_A", "%s")
                if name_mat not in combine_dict: 
                    combine_dict[name_mat] = [comb_ipt]
                else:
                    combine_dict[name_mat].append(comb_ipt)
            if "lora_B" in n: 
                hdim_b, rdim = p.shape 
                ipt_score = self.calculate_score(n, metric="ipt")
                comb_ipt = torch.mean(ipt_score, dim=0, keepdim=False).view(-1, 1)
                name_mat = n.replace("lora_B", "%s")
                if name_mat not in combine_dict: 
                    combine_dict[name_mat] = [comb_ipt]
                else:
                    combine_dict[name_mat].append(comb_ipt)
            if "lora_E" in n:
                ipt_score = self.calculate_score(n, p=p, metric="ipt")                
                name_mat = n.replace("lora_E", "%s")
                singular_dict[name_mat] = ipt_score

        # Combine the importance scores 
        all_is = []
        for name_mat in combine_dict: 
            ipt_E = singular_dict[name_mat] 
            ipt_AB = torch.cat(combine_dict[name_mat], dim=1)
            sum_ipt = self._combine_ipt(ipt_E, ipt_AB)
            name_E = name_mat%"lora_E"
            is_dict[name_E] = sum_ipt.view(-1, 1)
            all_is.append(sum_ipt.view(-1))

        # Calculate the masking threshold 
        mask_threshold = torch.kthvalue(torch.cat(all_is), (self.total_rank-curr_rank))[0].item()

        # Mask out unimportant singular values 
        with torch.no_grad():
            curr_sum_rank = 0
            sum_param = 0
            for n,p in model.named_parameters():
                if "lora_E" in n: 
                    p.data.masked_fill_(is_dict[n]<=mask_threshold, 0.0)
                    ranknum = (is_dict[n]>mask_threshold).sum().item() 

                    if self.tb_writter is not None and self.global_step%self.log_interval==0:
                        self.tb_writter.add_scalar("Ranknum/%s"%(n,), ranknum, self.global_step) 
                        self.rank_pattern[n] = ranknum 
                        curr_sum_rank += ranknum 
                        sum_param += ranknum*self.shape_dict[n.replace("lora_E", "lora_A")][1]  
                        sum_param += ranknum*self.shape_dict[n.replace("lora_E", "lora_B")][0]  

            if self.tb_writter is not None and self.global_step%self.log_interval==0:
                self.tb_writter.add_scalar("Budget/total_rank", curr_sum_rank, self.global_step)
                self.tb_writter.add_scalar("Budget/mask_threshold", mask_threshold, self.global_step)
                self.tb_writter.add_scalar("Budget/sum_param", sum_param, self.global_step)

        return mask_threshold


    def update_and_mask(self, model, global_step):
        if global_step<self.total_step-self.final_warmup:
            # Update importance scores element-wise 
            self.update_ipt(model)
            # do not update ipt during final fine-tuning 
        # Budget schedule
        curr_rank, mask_ind = self.schedule_threshold(global_step)
        if mask_ind:
            # Mask to target budget 
            mask_threshold = self.mask_to_target_rank(model, curr_rank) 
        else:
            mask_threshold = None 
        self._maybe_tb_writter_log(model)
        return curr_rank, mask_threshold

    def _maybe_tb_writter_log(self, model):
        if self.tb_writter is not None and self.global_step%self.log_interval==0:
            with torch.no_grad():
                regu_loss = []
                for n,p in model.named_parameters():
                    if "lora_A" in n or "lora_B" in n:
                        mat = p.data.detach().clone()
                        mat_cov = mat @ mat.T if "lora_A" in n else mat.T @ mat 
                        I = torch.eye(*mat_cov.size(), out=torch.empty_like(mat_cov))
                        I.requires_grad = False
                        orth_regu = torch.norm(mat_cov-I, p="fro")
                        regu_loss.append(orth_regu.item())
                        self.tb_writter.add_scalar(
                            "Orth_regu_loss/%s"%n, orth_regu.item(), self.global_step
                        )
                self.tb_writter.add_scalar(
                    "train/orth_regu_loss", sum(regu_loss)/len(regu_loss), self.global_step
                )


def compute_orth_regu(model, regu_weight=0.1):
    # The function to compute orthongonal regularization for SVDLinear in `model`. 
    regu_loss, num_param = 0., 0
    for n,p in model.named_parameters():
        if "lora_A" in n or "lora_B" in n:
            para_cov = p @ p.T if "lora_A" in n else p.T @ p 
            I = torch.eye(*para_cov.size(), out=torch.empty_like(para_cov))
            I.requires_grad = False
            regu_loss += torch.norm(para_cov-I, p="fro")
            num_param += 1
    return regu_weight*regu_loss/num_param