-
Notifications
You must be signed in to change notification settings - Fork 6
/
merge.py
329 lines (280 loc) · 11.4 KB
/
merge.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
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
import torch
import argparse
import os
from collections import OrderedDict
from transformers import (
AutoAdapterModel,
AutoTokenizer,
AutoModelForSequenceClassification,
AutoModelForSeq2SeqLM,
AutoConfig,
)
TRANSFORMERS_CACHE="checkpoints/hf_model"
HF_DATASETS_CACHE="checkpoints/hf_model"
HF_METRICS_CACHE="checkpoints/hf_model"
cache_dir=TRANSFORMERS_CACHE
def test(text1,text2):
print('hello world!')
print(text1,text2)
def test1(model_name_or_path,
task_name_list,
adapter_list,
adapter_config,
save_path):
print(task_name_list, type(task_name_list))
#print(adapter_list)
print(adapter_config, type(adapter_config))
print(model_name_or_path,type(task_name_list))
print(save_path,type(save_path))
print('pass!')
def compute_pairwise_coefficient_sets(step=0.2):
num_set = 1/step # 5
num_set=int(num_set)
sets = [[(ns+1)/num_set, 1-(ns+1)/num_set] for ns in range(num_set)]
# sets = [[0.0, 1.0]] + sets
return sets
def adapter_merge(
model_name_or_path,
task_name_list,
adapter_list,
adapter_config,
save_path,
overwrite_existed = True,
coefficient_sets = None,
dropout_list = None,
merge_way = "simple",
merge_head = False,
full_merge:bool = False,
fisher_list = [],
fisher_floor = 1e-6,
fisher_favor_target_model = True,
fisher_normalization = True,
*args,
**kwargs):
'''
examples:
model_name_or_path="roberta-large"
# the aim task is the last task: cola
task_name=["sst2","cola"]
pretrained_adapter=["sentiment/sst-2@ukp","lingaccept/cola@ukp"]
adapter_config="pfeiffer"
# the merged adapter for ['stsb','sst2','rte'] will be saved in 'save_path/for_rte/rte_sst2_stsb'
save_path="try-save-adapters/"
# overwrite the may existed adapter on the save path
overwrite_existed = True
# "simple" "fisher" "regmean"
merge_way="simple"
# if True, there will be parameter-protecting procedure for only target model.
# Else there will be parameter-protecting procedure for all models.
fisher_favor_target_model = True
'''
# what's the classification head belongs to
merged_name=task_name_list
#merged_name.sort()
merged_name="_".join(merged_name)
# check if already existed
will_save_path=os.path.join(save_path, merged_name)
if os.path.isdir(will_save_path):
if(overwrite_existed):
print("Overwrite the existed adapter.")
else:
print("Merged adapter already created.")
return
if full_merge:
range_list=model_name_or_path
else:
model_name_or_path=model_name_or_path[0]
range_list=adapter_list
# model-level
if coefficient_sets == None:
coefficient_sets = len(range_list) * [ 1 / len(range_list)]
else:
# suppose that len(coefficient_sets) == len(adapter_list)
# TODO: assert len(coefficient_sets) == len(adapter_list)
coefficient_sets = [float(x) for x in coefficient_sets]
coefficient_sets = [ x/sum(coefficient_sets) for x in coefficient_sets]
print(coefficient_sets)
state_dict={}
merged_dict={}
merged_keys = None
merged_values=[]
sum_factors = None
for i, _ in enumerate(range_list):
if full_merge:
model_name_or_path=range_list[i]
else:
adapter=range_list[i]
# have to load model every time when we load adapter
if dropout_list == None:
dropout=0.1
else:
dropout=float(dropout_list[i])
config = AutoConfig.from_pretrained(
model_name_or_path,
attention_probs_dropout_prob=dropout,
hidden_dropout_prob=dropout
)
tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path,
cache_dir=cache_dir,
)
# model = AutoModelForSequenceClassification
# model = AutoModelForSeq2SeqLM
model = AutoAdapterModel.from_pretrained(
model_name_or_path,
cache_dir=cache_dir,
config=config,
)
model.resize_token_embeddings(len(tokenizer))
# print(adapter)
if full_merge == False:
model.load_adapter(
adapter,
config=adapter_config,
load_as=merged_name,
)
state_dict = model.state_dict()
if dropout == 0:
head_keys = [hk for hk in state_dict.keys() if "head" in hk]
for hk in head_keys:
if '0' in hk:
hk_n = '1'.join(hk.split('0'))
elif '2' in hk:
hk_n = '4'.join(hk.split('2'))
state_dict[hk_n]=state_dict[hk]
state_dict.pop(hk)
# breakpoint()
if merged_keys == None:
if full_merge:
if merge_head:
merged_keys = [mk for mk in state_dict.keys()]
else:
merged_keys = [mk for mk in state_dict.keys() if ("head" not in mk) and ("classifier" not in mk)]
else:
merged_keys = [mk for mk in state_dict.keys() if ("adapters" in mk) or ("lora" in mk)] # adapters / lora
head_keys = [hk for hk in state_dict.keys() if "head" in hk]
if merge_head:
merged_keys = merged_keys + head_keys
merged_values = [0] * len(merged_keys)
sum_factors = [0] * len(merged_keys)
# parameter-level
factors = None
if merge_way == "fisher":
fisher = torch.load(fisher_list[i])
fisher = {k:v.detach() for k,v in fisher.items()}
# normalization
if fisher_normalization:
vs=[v for v in fisher.values()]
norm_constants=torch.sqrt(sum([torch.sum(torch.square(v)) for v in fisher.values()]))
coefficient_sets[i]=coefficient_sets[i]/norm_constants
# fisher_floor: avoid numerical problem
if (not fisher_favor_target_model) or (i == len(range_list)-1):
fisher = {k:torch.maximum(v, torch.tensor(fisher_floor)) for k,v in fisher.items()}
# TODO:assert set(merged_keys)==set(fisher.keys())
factors = OrderedDict([(mk, fisher[task_name_list[i].join(mk.split(merged_name))]*coefficient_sets[i]) for mk in merged_keys])
elif merge_way == "simple":
factors = OrderedDict(zip(merged_keys, [coefficient_sets[i]]*len(merged_keys)))
sum_factors = [sf+f for sf,f in zip(sum_factors, factors.values())]
# Add together
new_values = [state_dict[mk]*factors[mk] for mk in merged_keys]
merged_values = [mv+nv for mv,nv in zip(merged_values, new_values)]
if i == len(range_list)-1:
merged_values = [torch.div(mv, sf) for mv,sf in zip(merged_values, sum_factors)]
merged_dict = dict(zip(merged_keys, merged_values))
state_dict.update(merged_dict)
if dropout == 0:
config = AutoConfig.from_pretrained(
model_name_or_path,
)
model = AutoAdapterModel.from_pretrained(
model_name_or_path,
cache_dir=cache_dir,
config=config,
)
if full_merge == False:
model.load_adapter(
adapter,
config=adapter_config,
load_as=merged_name,
)
model.load_state_dict(state_dict)
if full_merge:
config = AutoConfig.from_pretrained(
model_name_or_path,
# num_labels=num_labels,
finetuning_task=task_name_list[i],
cache_dir=cache_dir,
)
tokenizer = AutoTokenizer.from_pretrained(
model_name_or_path,
cache_dir=cache_dir,
)
model.save_pretrained(save_path)
config.save_pretrained(save_path)
tokenizer.save_pretrained(save_path)
else:
model.save_all_adapters(save_path)
def pairwise_create_adapter():
model_name_or_path="roberta-large"
task_name_list=["sst2","cola","qnli","mrpc"] #"mnli",
pretrained_adapter_list=["sentiment/sst-2@ukp",
"lingaccept/cola@ukp",
"nli/qnli@ukp",
"sts/mrpc@ukp"]
#"nli/multinli@ukp",
adapter_config="houlsby"
#"pfeiffer"
save_path="merged_adapters/"
for i in range(len(task_name_list)):
for j in range(len(task_name_list)):
if i==j:
continue
task_name=[task_name_list[i], task_name_list[j]]
pretrained_adapter=[pretrained_adapter_list[i], pretrained_adapter_list[j]]
task_path="_".join(task_name)
path=save_path+adapter_config+'/'+task_path+'/'
#test2(model_name_or_path, task_name, pretrained_adapter, adapter_config, path)
adapter_merge(model_name_or_path, task_name, pretrained_adapter, adapter_config, path)
def main():
model_name_or_path=["roberta-base", "roberta-base"]
task_name=["cola","cola"] # ["cola","sst2"]
pretrained_adapter=["lingaccept/cola@ukp", "lingaccept/cola@ukp"]#["lingaccept/cola@ukp","sentiment/sst-2@ukp"]
adapter_config="pfeiffer"
save_path="try-save-adapters/"
coefficient_sets=[0.5,0.5]
adapter_merge(model_name_or_path, task_name, pretrained_adapter, adapter_config, save_path, coefficient_sets, full_merge=True)
def init_args():
parser=argparse.ArgumentParser()
parser.add_argument('-m','--model',nargs='*',required=True,help="base model to use")
parser.add_argument('-t','--adatasks',nargs='*',help="adapter's task name")
parser.add_argument('-a','--adapters',nargs='*',help="adapters to merge")
parser.add_argument('-c','--adaconfig',type=str,help="adapter's config")
parser.add_argument('-p','--save_path',type=str,help="path to save merged adapter")
parser.add_argument('-w','--merge_way',type=str,help="the way to merge adapters")
parser.add_argument('--merge_head',type=bool,default=False,help="the way to merge adapter head if it have one")
parser.add_argument('--full_merge',type=bool,default=False,help="to decide merge the whole model or only the adapters")
parser.add_argument('-f','--fisher_list', default=[], nargs='*',help="fisher info for adapters")
parser.add_argument('-o','--overwrite',type=bool,default=False,help="ignore existed merged adapter or not")
parser.add_argument('-d','--dropout_list',default=None, nargs='*',help="has dropout layer or not")
#parser.add_argument('-s','--do_search',type=bool,help="lambda grid search")
parser.add_argument('--search_step', default=None, nargs='+', help="grid search step")
args=parser.parse_args()
return args
if __name__ == "__main__":
# main()
# pairwise_create_adapter()
args = init_args()
adapter_merge(
model_name_or_path = args.model,
task_name_list = args.adatasks,
adapter_list = args.adapters,
adapter_config = args.adaconfig,
save_path = args.save_path,
coefficient_sets = args.search_step,
merge_way = args.merge_way,
merge_head = args.merge_head,
full_merge = args.full_merge,
fisher_list = args.fisher_list,
dropout_list = args.dropout_list,
overwrite_existed = args.overwrite,
)