-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathstudent.py
325 lines (294 loc) · 10.6 KB
/
student.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
from adapters.load_mcl import ModularMixin
from train_utils import (
load_optimizer,
evaluate_model,
train_epoch,
get_hparams,
get_model,
)
import torch.nn as nn
import torch
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from transformers import (
get_scheduler,
)
from utils import (
setup_basics,
EarlyStopper,
neptune_log,
set_seeds,
)
import copy
from task import (
get_task,
)
from metrics import Metric
import pdb
import time
import numpy as np
logger = get_logger(__name__)
LOG_TRAIN = False
class student:
def __init__(self, args, task, run, accelerator):
self.cache = []
self.task_name = args.task_name
self.seed = args.seed
self.target = args.target
self.args = get_hparams(args, self.task_name)
self.init_model()
self.test = task.data["test_dataloader"]
self.test_wrong = task.data["test_wrong_dataloader"]
self.run = run
self.seed = args.seed
self.accelerator = accelerator
self.iteration = 0
self.save_checkpoint = args.save_checkpoint
self.soft_labels = args.soft_labels
self.test_scores_gold = [0, 0]
self.test_scores_llm = [0, 0]
self.suffixes = [""]
if task.is_classification:
self.dic_classes = list(task.classes_dict_gold.values())
else:
self.dic_classes = None
self.metric = Metric(self.args, soft=self.args.is_classification)
self.metric_test = Metric(self.args, soft=self.args.is_classification)
def init_model(self):
set_seeds(self.seed)
model = get_model(self.args)
self.model = ModularMixin(
model,
freeze=True,
ac_kwargs={
"r": self.args.r,
"lora_scaling": self.args.lora_scaling,
"seed": self.seed,
},
)
return
def init_checkpoint(self, PATH):
self.model.load_state_dict(torch.load(PATH))
self.model.cuda()
return
def query(self, input):
torch.cuda.empty_cache()
self.model.eval()
self.model.cuda()
with torch.no_grad():
if self.soft_labels:
predictions = self.model.generate(
**{
"input_ids": input["input_ids"].cuda(),
"attention_mask": input["attention_mask"].cuda(),
},
max_new_tokens=1,
output_scores=True,
return_dict_in_generate=True,
)
predictions = [
torch.tensor(
list(np.array(predictions[1][0].cpu())[0][self.dic_classes])
)
]
else:
predictions = self.model.generate(
**{
"input_ids": input["input_ids"].cuda(),
"attention_mask": input["attention_mask"].cuda(),
},
num_beams=self.args.num_beams,
max_length=self.args.max_out_length,
decoder_start_token_id=self.model.model.config.bos_token_id,
)
return predictions
def evaluate(self):
self.metric_test.reset()
test_metric_gold = evaluate_model(
model=self.model,
accelerator=self.accelerator,
eval_dataloader=self.test,
metric=self.metric_test,
args=self.args,
dic_classes=self.dic_classes,
target="gold",
)
self.metric_test.reset()
test_metric_wrong_gold = evaluate_model(
model=self.model,
accelerator=self.accelerator,
eval_dataloader=self.test_wrong,
metric=self.metric_test,
args=self.args,
dic_classes=self.dic_classes,
target="gold",
)
self.metric_test.reset()
test_metric_llm = evaluate_model(
model=self.model,
accelerator=self.accelerator,
eval_dataloader=self.test,
metric=self.metric_test,
args=self.args,
dic_classes=self.dic_classes,
target="llm",
)
test_metric_wrong_llm = evaluate_model(
model=self.model,
accelerator=self.accelerator,
eval_dataloader=self.test_wrong,
metric=self.metric_test,
args=self.args,
dic_classes=self.dic_classes,
target="llm",
)
if self.run is not None:
stats = {
"test_gold_acc": test_metric_gold[0],
"test_llm_acc": test_metric_llm[0],
"test_wrong_gold_acc": test_metric_wrong_gold[0],
"test_wrong_llm_acc": test_metric_wrong_llm[0],
"data amount": self.data_amount,
}
for suffix in self.suffixes:
neptune_log(
run=self.run,
pref=f"test/" + suffix,
stats=stats,
epoch=self.iteration,
)
self.test_scores_gold = [self.test_scores_gold[1], test_metric_gold[0]]
self.test_scores_llm = [self.test_scores_llm[1], test_metric_llm[0]]
self.suffixes = [""]
def train(self, train_dataloader, eval_dataloader):
torch.cuda.empty_cache()
self.early_stopper = EarlyStopper(self.args.early_stop)
self.iteration += 1
if self.seed is not None:
set_seed(self.args.seed)
self.metric.reset()
# for every retraining, we train from scratch
self.init_model()
logger.info(f" Running task {self.task_name}")
logger.info(f" Num examples = {len(train_dataloader.dataset)}")
self.data_amount = len(train_dataloader.dataset) + len(eval_dataloader.dataset)
# Re-initialise lr_scheduler + optimized
optimizer = load_optimizer(self.model, self.args)
lr_scheduler = get_scheduler(
name=self.args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=int(
self.args.warmup
* self.args.num_train_epochs
* len(
train_dataloader.dataset
)
),
num_training_steps=self.args.num_train_epochs
* len(
train_dataloader.dataset
),
)
# Move to the device
self.model, optimizer, lr_scheduler = self.accelerator.prepare(
self.model, optimizer, lr_scheduler
)
for epoch in range(0, self.args.num_train_epochs):
total_loss = train_epoch(
model=self.model,
train_dataloader=train_dataloader,
accelerator=self.accelerator,
lr_scheduler=lr_scheduler,
optimizer=optimizer,
args=self.args,
dic_classes=self.dic_classes,
)
if (
epoch % self.args.eval_every_epochs == 0
or epoch == self.args.num_train_epochs - 1
):
eval_metrics = evaluate_model(
model=self.model,
accelerator=self.accelerator,
eval_dataloader=eval_dataloader,
metric=self.metric,
args=self.args,
dic_classes=self.dic_classes,
target=self.target,
)
self.model.cpu()
self.early_stopper.update(eval_metrics[0], self.model)
self.model.cuda()
log_msg = f"Epoch {epoch} -----> Average_Train_loss: {total_loss / len(train_dataloader.dataset)} ===== Eval_metric: {eval_metrics[0]}"
logger.info(log_msg)
if self.run is not None and LOG_TRAIN:
self.run[f"{self.iteration}-eval"].log(eval_metrics[0], step=epoch)
# log metrics are desactivated
if self.run is not None and LOG_TRAIN:
stats = {
"loss": total_loss / len(train_dataloader.dataset),
"main_lr": optimizer.param_groups[0]["lr"],
}
if self.early_stopper.should_finish():
break
# copying from a cpu
self.model.cpu()
self.model = copy.deepcopy(self.early_stopper.get_best())
self.model = self.early_stopper.get_best().cuda()
del self.early_stopper.best_model
self.evaluate()
if self.save_checkpoint != "no":
PATH_DEST = (
"checkpoints/"
+ self.task_name
+ "/"
+ str(self.seed)
+ "_"
+ str(len(train_dataloader.dataset) + len(eval_dataloader.dataset))
+ ".pt"
)
torch.save(self.model.state_dict(), PATH_DEST)
class aux_student:
def __init__(self, model, args, task):
self.model = model
self.task_name = args.task_name
self.args = args
self.task = task
self.soft_labels = args.soft_labels
if task.is_classification:
self.dic_classes = list(task.classes_dict_gold.values())
else:
self.dic_classes = None
def query(self, input):
torch.cuda.empty_cache()
self.model.eval()
with torch.no_grad():
if self.soft_labels:
predictions = self.model.generate(
**{
"input_ids": input["input_ids"].cpu(),
"attention_mask": input["attention_mask"].cpu(),
},
max_new_tokens=1,
output_scores=True,
return_dict_in_generate=True,
)
predictions = [
torch.tensor(
list(np.array(predictions[1][0].cpu())[0][self.dic_classes])
)
]
else:
predictions = self.model.generate(
**{
"input_ids": input["input_ids"],
"attention_mask": input["attention_mask"],
},
num_beams=self.args.num_beams,
max_length=self.args.max_out_length,
decoder_start_token_id=self.model.model.config.bos_token_id,
)
self.model.cpu()
input["input_ids"].cuda()
input["attention_mask"].cuda()
return predictions