-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlord_complex_train.py
284 lines (229 loc) · 10.5 KB
/
lord_complex_train.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
"""
======================================================================
LORD_COMPLEX_TRAIN ---
The comprehensive and overall training function. In `lord_train.py` it
is a simplified version of LoRD loss function.
Author: Zi Liang <zi1415926.liang@connect.polyu.hk>
Copyright © 2024, ZiLiang, all rights reserved.
Created: 5 March 2024
======================================================================
"""
# ------------------------ Code --------------------------------------
import torch
import json
from torch.utils.tensorboard import SummaryWriter
from torch.distributions import Categorical
from torch.utils.data import TensorDataset, DataLoader
from tqdm import tqdm
import argparse
from transformers import AutoModelForCausalLM
from transformers import AutoModelForSequenceClassification
from transformers import AutoModelForTokenClassification
from transformers import AutoTokenizer, AutoConfig, AutoModel
from training_data_collecting_openai import load_raw_train_datals
from training_data_collecting_openai import load_steal_datals
from glue_process import load_glue_datals
from sequence_utils import my_padding, my_padding_logits
from sequence_utils import my_padding_token_dist
from sequence_utils import my_padding_logit
import torch.nn.functional as F
from rlhf_train import clip, log_clip
def complex_train_one_period(args, lm,
lm_tokenizer,
loader, epoch, device,
tb_writer,
tensorboard_name,
save_path,
LR=3e-5,
acc_step=1,
log_step=100,
save_step=1000,
beta=0.7,
epsln=1e-6,
is_black_box=0,
method="complex",
):
overall_loss = 0.
overall_step = 0
pad_token_id = lm_tokenizer.pad_token_id
kl_loss = torch.nn.KLDivLoss(reduction="none")
sigmoid = torch.nn.Sigmoid()
opt1 = torch.optim.AdamW(lm.parameters(), lr=LR)
for e in tqdm(range(epoch), desc="epoch"):
for item in tqdm(loader, desc="loader"):
overall_step += 1
loss_constractive = 0.
loss_logits = 0.
# print(item)
idxs1, idxs2, mask1, mask2, old_logits1, old_logits2, vic_logits2, idxs2_dist = item
bs, sqlen1 = idxs1.shape
bs, sqlen2 = idxs2.shape
# print(sqlen1, sqlen2)
sqlen = min(sqlen1, sqlen2)
idxs1 = idxs1.to(device) # bs, sql
idxs2 = idxs2.to(device) # bs, sql
mask1 = mask1.to(device)
mask2 = mask2.to(device)
# already normalized by softmax
old_logits1 = old_logits1.to(device) # bs, sql,
old_logits2 = old_logits2.to(device) # bs, sql,
if args.is_black_box==0:
vic_logits2 = vic_logits2.to(device) # bs, sql, 5
idxs2_dist = idxs2_dist.to(device)
logits2_dist = torch.gather(logits2, 2, idxs2_dist)
print("idx1text: ", lm_tokenizer.decode(idxs1[0]))
print("idx2text: ", lm_tokenizer.decode(idxs2[0]))
logits1 = lm(idxs1).logits[:, :-1, :]
logits1 = F.log_softmax(logits1, dim=-1)
logits1 = logits1[torch.arange(bs).unsqueeze(1),
torch.arange(sqlen-1).unsqueeze(0),
idxs1[:, 1:sqlen]]
logits2 = lm(idxs2).logits[:, :-1, :]
logits2 = torch.log_softmax(logits2, dim=-1)
logits2_cons = logits2[torch.arange(bs).unsqueeze(1),
torch.arange(sqlen-1).unsqueeze(0),
idxs2[:, 1:sqlen]]
if method == "complex":
if args.is_black_box == 0:
loss_constractive_good = -torch.sum(
(logits2_cons*2 -
vic_logits2[:, :, 0]
- old_logits2)*mask2[:, :-1])\
/ torch.sum(mask2[:, :-1])
else:
loss_constractive_good = -torch.sum(
(logits2_cons*2 -
-old_logits2)*mask2[:, :-1])
zelta_logits1 = log_clip(old_logits1-logits1)
loss_constractive_past = -torch.sum(
(zelta_logits1) * mask1[:, :-1]) /\
torch.sum(mask1[:, :-1])
if args.use_old_logits != "1":
loss_constractive_past = 0.
if args.use_vic_logits != "1":
loss_constractive_good = 0.
loss_constractive = loss_constractive_good \
+ loss_constractive_past
overall_loss = loss_constractive
elif method == "VeryComplex":
term1 = -torch.exp(old_logits1)*(
log_clip(old_logits1-logits1))
if args.is_black_box == 0:
term3 = torch.exp(vic_logits2[:, :, 0])\
* (
(vic_logits2[:, :, 0]-logits2_cons))\
+ (old_logits2 - logits2_cons)
else:
term3 = - logits2_cons*2
loss_constractive_past = torch.sum(
term1*mask1[:, :-1])
loss_constractive_good = torch.sum(
term3*mask2[:, :-1])
loss_constractive = loss_constractive_good +\
loss_constractive_past
if args.use_old_logits != "1":
loss_constractive_past = 0.
if args.use_vic_logits != "1":
loss_constractive_good = 0.
overall_loss = loss_constractive
elif method == "nologComplex":
mask = torch.logical_or(mask1, mask2).long()
term1 = log_clip(-old_logits1+logits1)
term2 = (old_logits2-logits2_cons)
if args.is_black_box == 0:
term3 = \
(vic_logits2[:, :, 0]-logits2_cons)
else:
term3 = - logits2_cons
loss_1 = term2 + term3
loss_2 = torch.exp(term1)
# loss_2 = term1
# loss = sigmoid(loss_1)+loss_2
loss = sigmoid(torch.mean(loss_1)*mask2[:, :-1]
)*loss_2
loss = torch.mean(loss*mask1[:, :-1])
# if torch.sum(mask[:, :-1]) >= 1:
# loss = torch.sum(loss*mask[:, :-1])
# # / torch.sum(mask[:, :-1])
# else:
# loss = 0.
if loss == torch.tensor(float("nan")):
print("++++++++++++++++++++++")
print(f"term1: {term1}")
print(f"term2: {term3}")
print(f"loss1: {loss_1}")
print(f"loss2: {loss_2}")
print(f"loss: {loss}")
print(f"mask: {mask[:,:-1]}")
print("++++++++DEBUG DONE.++++++++")
loss_constractive = loss
loss_constractive_past = 0.
loss_constractive_good = 0.
loss_logits = 0.
overall_loss += loss_constractive + loss_logits
elif method == "ComplexV3":
mask = torch.logical_or(mask1, mask2).long()
term1 = (-old_logits1+logits1)
term2 = log_clip(old_logits2-logits2_cons)
if is_black_box == 0:
term3 = \
(vic_logits2[:, :, 0]-logits2_cons)
else:
term3 = - logits2_cons
loss_1 = term1 + term3
loss_2 = torch.exp(term2)
loss = sigmoid(loss_1)*loss_2
if torch.sum(mask[:, :-1]) >= 1:
loss = torch.sum(loss*mask[:, :-1])
# / torch.sum(mask[:, :-1])
else:
loss = 0.
if loss == torch.tensor(float("nan")):
print("++++++++++++++++++++++")
print(f"term1: {term1}")
print(f"term2: {term3}")
print(f"loss1: {loss_1}")
print(f"loss2: {loss_2}")
print(f"loss: {loss}")
print(f"mask: {mask[:,:-1]}")
print("++++++++DEBUG DONE.++++++++")
loss_constractive = loss
loss_constractive_past = 0.
loss_constractive_good = 0.
loss_logits = 0.
overall_loss += loss_constractive + loss_logits
if overall_step % log_step == 0:
print(" LOSS: {}\tGoodRewardLoss: {}\tToPassRewardLoss: {}\tKL-D: {}".format(
overall_loss,
loss_constractive_good,
loss_constractive_past,
loss_logits
))
tb_writer.add_scalar("loss", overall_loss,
overall_step)
if args.use_vic_logits == "1":
tb_writer.add_scalar("rewardloss_good",
loss_constractive_good,
overall_step)
if args.use_old_logits == "1":
tb_writer.add_scalar("rewardloss_past",
loss_constractive_past,
overall_step)
if args.use_kld == "1":
tb_writer.add_scalar("KLloss", loss_logits,
overall_step)
if overall_step % save_step == 0:
print(" -->Regular Saving.")
print(f"in epoch {e}, step {overall_step}.")
lm_tokenizer.save_pretrained(save_path+"___"+str(overall_step))
lm.save_pretrained(save_path+"___"+str(overall_step))
if overall_step % acc_step == 0:
opt1.zero_grad()
overall_loss.backward()
opt1.step()
overall_loss = 0.
print(" -->Finally Saving.")
# lm_tokenizer.save_pretrained(save_path+"___STEPfinally")
# lm.save_pretrained(save_path+"___STEPfinally")
print("ONE PERIOD TRAINING DONE!")
return lm