-
Notifications
You must be signed in to change notification settings - Fork 58
/
prefixtuning.py
300 lines (262 loc) · 12.8 KB
/
prefixtuning.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
import torch
from torch import nn
from transformers import AutoTokenizer
from .base import PushToHubFriendlyModel
from ..prompt.modeling_auto import AutoModelForSeq2SeqLM
class Model(PushToHubFriendlyModel):
def __init__(self, args):
super().__init__()
self.args = args
"""The prefix-tuning code"""
self.preseqlen = args.prefix_tuning.prefix_sequence_length
self.mid_dim = args.prefix_tuning.mid_dim
print("prefix-tuning sequence length is {}.".format(self.preseqlen))
# Load tokenizer and model.
self.tokenizer = AutoTokenizer.from_pretrained(args.bert.location, use_fast=False)
self.pretrain_model = AutoModelForSeq2SeqLM.from_pretrained(
args.bert.location
)
self.config = self.pretrain_model.config
from ..prompt.modeling_bart import BartForConditionalGeneration
from ..prompt.modeling_t5 import T5ForConditionalGeneration
if isinstance(self.pretrain_model, BartForConditionalGeneration):
self.match_n_layer = self.config.decoder_layers
self.match_n_head = self.config.decoder_attention_heads
self.n_embd = self.config.d_model
assert self.n_embd % self.match_n_head == 0
self.match_n_embd = self.n_embd // self.match_n_head # huggingface BART's dim of kv need to be calculated
elif isinstance(self.pretrain_model, (T5ForConditionalGeneration)):
self.match_n_layer = self.config.num_decoder_layers
self.match_n_head = self.config.num_heads
self.n_embd = self.config.d_model
self.match_n_embd = self.config.d_kv
else:
raise ValueError("Other models are not supported yet!")
if args.special_tokens:
self.tokenizer.add_tokens([v for k, v in args.special_tokens])
self.pretrain_model.resize_token_embeddings(len(self.tokenizer))
# Prefix related.
self.register_buffer('input_tokens', torch.arange(self.preseqlen).long())
self.wte = nn.Embedding(self.preseqlen, self.n_embd)
self.control_trans = nn.Sequential(
nn.Linear(self.n_embd, self.mid_dim),
nn.Tanh(),
nn.Linear(self.mid_dim, self.match_n_layer * 2 * self.match_n_head * self.match_n_embd),
)
if self.args.model.knowledge_usage == 'separate':
self.knowledge_trans = nn.Sequential(
nn.Linear(self.n_embd, self.mid_dim),
nn.Tanh(),
nn.Linear(self.mid_dim, self.match_n_layer * 2 * self.match_n_head * self.match_n_embd),
)
self.wte_enc = nn.Embedding(self.preseqlen, self.n_embd)
self.control_trans_enc = nn.Sequential(
nn.Linear(self.n_embd, self.mid_dim),
nn.Tanh(),
nn.Linear(self.mid_dim, self.match_n_layer * 2 * self.match_n_head * self.match_n_embd),
)
if self.args.model.knowledge_usage == 'separate':
self.knowledge_trans_enc = nn.Sequential(
nn.Linear(self.n_embd, self.mid_dim),
nn.Tanh(),
nn.Linear(self.mid_dim, self.match_n_layer * 2 * self.match_n_head * self.match_n_embd),
)
self.wte_dec = nn.Embedding(self.preseqlen, self.n_embd)
self.control_trans_dec = nn.Sequential(
nn.Linear(self.n_embd, self.mid_dim),
nn.Tanh(),
nn.Linear(self.mid_dim, self.match_n_layer * 2 * self.match_n_head * self.match_n_embd),
)
# Knowledge prompt.
if self.args.model.knowledge_usage == 'separate':
self.knowledge_trans_dec = nn.Sequential(
nn.Linear(self.n_embd, self.mid_dim),
nn.Tanh(),
nn.Linear(self.mid_dim, self.match_n_layer * 2 * self.match_n_head * self.match_n_embd),
)
self.dropout = nn.Dropout(args.prefix_tuning.prefix_dropout)
if self.args.model.freeze_plm:
for param in self.pretrain_model.parameters():
param.requires_grad = False
if self.args.model.freeze_prefix:
for param in self.wte.parameters():
param.requires_grad = False
for param in self.control_trans.parameters():
param.requires_grad = False
for param in self.wte_dec.parameters():
param.requires_grad = False
for param in self.control_trans_dec.parameters():
param.requires_grad = False
for param in self.wte_enc.parameters():
param.requires_grad = False
for param in self.control_trans_enc.parameters():
param.requires_grad = False
def get_prompt(self, bsz=None, sample_size=1, description=None, knowledge=None):
old_bsz = bsz
bsz = bsz * sample_size
input_tokens = self.input_tokens.unsqueeze(0).expand(bsz, -1)
temp_control = self.wte(input_tokens)
if description is not None:
temp_control = temp_control + description.repeat_interleave(sample_size, dim=0).unsqueeze(1)
past_key_values = self.control_trans(temp_control) # bsz, seqlen, layer*emb
if knowledge is not None:
past_key_values = torch.cat([past_key_values, self.knowledge_trans(knowledge.repeat_interleave(sample_size, dim=0))], dim=1)
bsz, seqlen, _ = past_key_values.shape
past_key_values = past_key_values.view(
bsz, seqlen, self.match_n_layer * 2, self.match_n_head, self.match_n_embd
)
past_key_values = self.dropout(past_key_values)
past_key_values = past_key_values.permute([2, 0, 3, 1, 4]).split(2)
# Cross prefix
temp_control_dec = self.wte_dec(input_tokens)
if description is not None:
temp_control_dec = temp_control_dec + description.repeat_interleave(sample_size, dim=0).unsqueeze(1)
past_key_values_dec = self.control_trans_dec(
temp_control_dec
) # bsz, seqlen, layer*emb
if knowledge is not None:
past_key_values_dec = torch.cat([past_key_values_dec, self.knowledge_trans_dec(knowledge.repeat_interleave(sample_size, dim=0))], dim=1)
bsz, seqlen, _ = past_key_values_dec.shape
past_key_values_dec = past_key_values_dec.view(
bsz, seqlen, self.match_n_layer * 2, self.match_n_head, self.match_n_embd
)
past_key_values_dec = self.dropout(past_key_values_dec)
past_key_values_dec = past_key_values_dec.permute([2, 0, 3, 1, 4]).split(2)
# Encoder prefix
input_tokens_enc = (
self.input_tokens.unsqueeze(0).expand(old_bsz, -1)
)
temp_control_enc = self.wte_enc(input_tokens_enc)
if description is not None:
temp_control_enc = temp_control_enc + description.unsqueeze(1)
past_key_values_enc = self.control_trans_enc(
temp_control_enc
) # bsz, seqlen, layer*emb
if knowledge is not None:
past_key_values_enc = torch.cat([past_key_values_enc, self.knowledge_trans_enc(knowledge)], dim=1)
bsz_enc, seqlen, _ = past_key_values_enc.shape
past_key_values_enc = past_key_values_enc.view(
bsz_enc,
seqlen,
self.match_n_layer * 2,
self.match_n_head,
self.match_n_embd,
)
past_key_values_enc = self.dropout(past_key_values_enc)
past_key_values_enc = past_key_values_enc.permute([2, 0, 3, 1, 4]).split(2)
result = []
for i, key_val in enumerate(past_key_values):
temp = dict()
temp["decoder_prompt"] = {
"prev_key": key_val[0].contiguous(),
"prev_value": key_val[1].contiguous(),
"prev_key_padding_mask": torch.zeros(bsz, seqlen)
.to(key_val.device)
.bool()
# bsz, preseqlen
}
key_val_dec = past_key_values_dec[i]
temp["cross_attention_prompt"] = {
"prev_key": key_val_dec[0].contiguous(),
"prev_value": key_val_dec[1].contiguous(),
"prev_key_padding_mask": torch.zeros(bsz, seqlen)
.to(key_val_dec.device)
.bool(),
}
key_val_enc = past_key_values_enc[i]
temp["encoder_prompt"] = {
"prev_key": key_val_enc[0].contiguous(),
"prev_value": key_val_enc[1].contiguous(),
"prev_key_padding_mask": torch.zeros(bsz_enc, seqlen)
.to(key_val_enc.device)
.bool(),
}
result.append(temp)
return result
def get_description_representation(self, kwargs):
if self.args.model.use_description and self.args.model.map_description:
description_input_ids = kwargs.pop("description_input_ids")
description_attention_mask = kwargs.pop("description_attention_mask")
if self.args.bert.location in ["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b"]:
description_outputs = self.pretrain_model.encoder(
input_ids=description_input_ids,
attention_mask=description_attention_mask,
)
description = description_outputs.last_hidden_state[:, 0] # TODO: the first token from the encoder.
elif self.args.bert.location in ["facebook/bart-base", "facebook/bart-large"]:
description_outputs = self.pretrain_model.model.encoder(
input_ids=description_input_ids,
attention_mask=description_attention_mask,
)
description = description_outputs.last_hidden_state[:, 0] # TODO: the first token from the encoder.
else:
raise ValueError()
else:
description = None
return description
def get_knowledge_representation(self, kwargs):
if self.args.model.knowledge_usage == 'separate':
knowledge_input_ids = kwargs.pop("knowledge_input_ids", None)
knowledge_attention_mask = kwargs.pop("knowledge_attention_mask", None)
if self.args.bert.location in ["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b"]:
knowledge_outputs = self.pretrain_model.encoder(
input_ids=knowledge_input_ids,
attention_mask=knowledge_attention_mask,
)
knowledge = knowledge_outputs.last_hidden_state
elif self.args.bert.location in ["facebook/bart-base", "facebook/bart-large"]:
knowledge_outputs = self.pretrain_model.model.encoder(
input_ids=knowledge_input_ids,
attention_mask=knowledge_attention_mask,
)
knowledge = knowledge_outputs.last_hidden_state
else:
raise ValueError()
elif self.args.model.knowledge_usage == 'concatenate':
knowledge = None
else:
raise ValueError()
return knowledge
def forward(self,
input_ids,
attention_mask,
labels,
**kwargs,
):
bsz = input_ids.shape[0]
# Encode description.
description_representation = self.get_description_representation(kwargs)
# Encode knowledge.
knowledge_representation = self.get_knowledge_representation(kwargs)
past_prompt = self.get_prompt(
bsz=bsz, description=description_representation, knowledge=knowledge_representation,
)
loss = self.pretrain_model(
input_ids=input_ids,
attention_mask=attention_mask,
labels=labels,
past_prompt=past_prompt,
).loss
return {'loss': loss}
def generate(self,
input_ids,
attention_mask,
**kwargs):
bsz = input_ids.shape[0]
# Encode description.
description_representation = self.get_description_representation(kwargs)
# Encode knowledge.
knowledge_representation = self.get_knowledge_representation(kwargs)
past_prompt = self.get_prompt(
bsz=bsz, sample_size=kwargs['num_beams'], description=description_representation, knowledge=knowledge_representation,
)
generated_ids = self.pretrain_model.generate(
input_ids=input_ids,
attention_mask=attention_mask,
past_prompt=past_prompt,
use_cache=True,
**kwargs,
)
return generated_ids