-
Notifications
You must be signed in to change notification settings - Fork 2
/
new_distill_spikformer.py
278 lines (238 loc) · 12.4 KB
/
new_distill_spikformer.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
import torch
import os
import torch.nn as nn
import pickle
import argparse
import torch.nn.functional as F
import torch.optim as optim
from model import new_spikformer
from torch.utils.data import DataLoader
from tqdm import tqdm
import numpy as np
import os
import time
from dataset import TxtDataset
from transformers import BertTokenizer, BertForSequenceClassification, AutoTokenizer
from spikingjelly.activation_based import encoding
from spikingjelly.activation_based import functional
import math
from utils.public import set_seed
from torchmetrics.classification import MatthewsCorrCoef
os.environ["CUDA_VISIBLE_DEVICES"] = "1,0,2,3"
def to_device(x, device):
for key in x:
x[key] = x[key].to(device)
def args():
parser = argparse.ArgumentParser()
parser.add_argument("--seed", default=42, type=int)
parser.add_argument("--dataset_name",default="sst2",type=str)
parser.add_argument("--data_augment", default="True", type=str)
parser.add_argument("--batch_size",default=4,type=int)
parser.add_argument("--fine_tune_lr",default=1e-2,type=float)
parser.add_argument("--epochs",default=100,type=int)
parser.add_argument("--teacher_model_path",default="", type=str)
parser.add_argument("--label_num",default=2,type=int)
parser.add_argument("--depths",default=6,type=int)
parser.add_argument("--max_length",default=64,type=int)
parser.add_argument("--dim",default=768,type=int)
parser.add_argument("--ce_weight", default=0.0, type=float)
parser.add_argument("--emb_weight", default=1.0, type=float)
parser.add_argument("--logit_weight", default=1.0, type=float)
parser.add_argument("--rep_weight", default=5.0, type=float)
parser.add_argument("--num_step", default=32, type=int)
parser.add_argument("--tau", default=10.0, type=float)
parser.add_argument("--common_thr", default=1.0, type=float)
parser.add_argument("--predistill_model_path", default="", type=str)
# parser.add_argument("--predistill_requires_grad", default="True", type=str)
parser.add_argument("--ignored_layers", default=1, type=int)
parser.add_argument("--metric", default="acc", type=str)
args = parser.parse_args()
return args
def distill(args):
device=torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = BertTokenizer.from_pretrained(args.teacher_model_path)
teacher_model = BertForSequenceClassification.from_pretrained(args.teacher_model_path, num_labels=args.label_num, output_hidden_states=True).to(device)
for param in teacher_model.parameters():
param.requires_grad = False
teacher_model.eval()
student_model = new_spikformer(depths=args.depths, length=args.max_length, T=args.num_step, \
tau=args.tau, common_thr=args.common_thr, vocab_size = len(tokenizer), dim=args.dim, num_classes=args.label_num, mode="distill")
# load embedding layer
# student_model.emb.weight = teacher_model.bert.embeddings.word_embeddings.weight
# student_model.emb.weight.requires_grad = True
if args.predistill_model_path != "":
weights = torch.load(args.predistill_model_path)
# for key in weights.keys():
# if "module.transforms" not in key:
# weights[key] = weights[key].float()
# if args.predistill_requires_grad == "False":
# weights[key].requires_grad = False
# elif args.predistill_requires_grad == "True":
# weights[key].requires_grad = True
student_model.load_state_dict(weights, strict=False)
print("load predistill model finish!")
scaler = torch.cuda.amp.GradScaler()
optimer = torch.optim.AdamW(params=student_model.parameters(), lr=args.fine_tune_lr, betas=(0.9, 0.999), weight_decay=5e-3)
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimer, T_max=args.epochs, eta_min=0)
# scheduler_warmup = GradualWarmupScheduler(optimer, multiplier=1, total_epoch=5, after_scheduler=scheduler)
if args.data_augment == "True":
print("With Augmentation")
train_dataset = TxtDataset(data_path=f"data/{args.dataset_name}/train_augment.txt")
else:
print("Without Augmentation")
train_dataset = TxtDataset(data_path=f"data/{args.dataset_name}/train.txt")
train_data_loader = DataLoader(dataset=train_dataset, batch_size= args.batch_size, shuffle=True, drop_last=False)
test_dataset = TxtDataset(data_path=f"data/{args.dataset_name}/test.txt")
test_data_loader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False,drop_last=False)
valid_dataset = TxtDataset(data_path=f"data/{args.dataset_name}/validation.txt")
valid_data_loader = DataLoader(dataset=test_dataset, batch_size=args.batch_size, shuffle=False,drop_last=False)
device_ids = [i for i in range(torch.cuda.device_count())]
print(device_ids)
if len(device_ids) > 1:
student_model = nn.DataParallel(student_model, device_ids=device_ids).to(device)
student_model = student_model.to(device)
metric_list = []
for epoch in tqdm(range(args.epochs)):
# if epoch == 5:
# args.rep_weight == 0
total_loss_list = []
embeddings_loss_list = []
ce_loss_list = []
logit_loss_list = []
rep_loss_list = []
for batch in tqdm(train_data_loader):
student_model.train()
batch_size = len(batch[0])
labels = batch[1].to(device)
inputs = tokenizer(batch[0], padding="max_length", truncation=True, \
return_tensors="pt", max_length=args.max_length)
# inputs = tokenizer(batch[0], padding=True, truncation=True, \
# return_tensors="pt", max_length=args.max_length)
to_device(inputs, device)
with torch.no_grad():
teacher_outputs = teacher_model(**inputs)
# [1:] means 12 outputs of each layer; [::x] means get value every x layers
tea_embeddings = teacher_model.bert.embeddings.word_embeddings.weight
if len(device_ids) > 1:
stu_embeddings = student_model.module.emb.weight
else:
stu_embeddings = student_model.emb.weight
embeddings_loss = F.mse_loss(stu_embeddings, tea_embeddings)
embeddings_loss_list.append(embeddings_loss.item())
tea_rep = teacher_outputs.hidden_states[1:][::int(12/args.depths)] # layers output
# len(stu_rep) = depth
# stu_rep[0] shape: B L D
stu_rep, student_outputs = student_model(inputs['input_ids'])
student_outputs = student_outputs.reshape(-1 , args.num_step, args.label_num)
# Before transpose: B T Label_num
# After transpose: T B Label_num
student_outputs = student_outputs.transpose(0, 1)
student_logits = torch.mean(student_outputs, dim=0) # B Label_num
# last step
# student_logits = student_outputs[-1,:,:]# B Label_num
# print("student_logits.shape: ", student_logits.shape)
# print("labels.shape: ", labels.shape)
ce_loss = F.cross_entropy(student_logits, labels)
ce_loss_list.append(ce_loss.item())
# print("ce_loss: ", ce_loss, ce_loss.dtype)
# print("student_logits.shape: ", student_logits.shape)
# print("teacher_outputs.logits.shape: ", teacher_outputs.logits.shape)
logit_loss = F.kl_div(F.log_softmax(student_logits, dim=1), F.softmax(teacher_outputs.logits, dim=1), reduction='batchmean')
logit_loss_list.append(logit_loss.item())
# print("logit_loss: ", logit_loss, logit_loss.dtype)
tea_rep = torch.tensor(np.array([item.cpu().detach().numpy() for item in tea_rep]), dtype=torch.float32)
tea_rep = tea_rep.to(device=device)
rep_loss = 0
tea_rep = tea_rep[args.ignored_layers:]
stu_rep = stu_rep[args.ignored_layers:]
# print(len(stu_rep))
for i in range(len(stu_rep)):
# print("stu_rep[i]", stu_rep[i])
# print("tea_rep[i]", tea_rep[i])
rep_loss += F.mse_loss(stu_rep[i], tea_rep[i])
rep_loss = rep_loss / batch_size # batch mean
rep_loss_list.append(rep_loss.item())
# print("rep_loss: ",rep_loss)
total_loss = (args.emb_weight * embeddings_loss) \
+ (args.ce_weight * ce_loss) \
+ (args.logit_weight * logit_loss) \
+ (args.rep_weight * rep_loss)
# print("total_loss: ", total_loss.item())
# print("total_loss:{} | ce_loss {} | logit_loss {} | rep_loss {} ".format(total_loss.item(), \
# ce_loss.item(), logit_loss.item(), rep_loss))
total_loss_list.append(total_loss.item())
optimer.zero_grad()
scaler.scale(total_loss).backward()
scaler.step(optimer)
scaler.update()
functional.reset_net(student_model)
# print(
# f"In average, at epoch {epoch}, "
# + f"ce_loss: {np.mean(ce_loss_list)}, "
# + f"emb_loss: {np.mean(embeddings_loss_list)} "
# + f"logit_loss: {np.mean(logit_loss_list)}, "
# + f"rep_loss: {np.mean(rep_loss_list)}, "
# + f"total_loss: {np.mean(total_loss_list)}"
# )
scheduler.step()
# scheduler_warmup.step()
y_true = []
y_pred = []
student_model.eval()
with torch.no_grad():
for batch in tqdm(test_data_loader):
# for batch in tqdm(valid_data_loader):
batch_size = len(batch[0])
b_y = batch[1]
y_true.extend(b_y.to("cpu").tolist())
inputs = tokenizer(batch[0], padding="max_length", truncation=True, \
return_tensors='pt', max_length=args.max_length)
# inputs = tokenizer(batch[0], padding=True, truncation=True, \
# return_tensors="pt", max_length=args.max_length)
to_device(inputs, device)
_, outputs = student_model(inputs['input_ids'])
outputs = outputs.to("cpu")
outputs = outputs.reshape(-1, args.num_step, args.label_num)
# print("outputs.shape", outputs.shape)
# Before transpose: B T Label_num
# After transpose: T B Label_num
outputs = outputs.transpose(0, 1)
# print("After transpose, outputs.shape", outputs.shape)
logits = torch.mean(outputs, dim=0) # B Label_num
# print("logits.shape", logits.shape)
y_pred.extend(torch.max(logits,1)[1].tolist())
# for line in torch.softmax(logits, 1):
# result.append(line.numpy())
# correct += int(b_y.eq(torch.max(logits,1)[1]).sum())
functional.reset_net(student_model)
if args.metric == "acc":
correct = 0
for i in range(len(y_true)):
correct += 1 if y_true[i] == y_pred[i] else 0
acc = correct / len(y_pred)
print("acc", acc)
elif args.metric == "mcc":
print(y_true)
print(y_pred)
matthews_corrcoef = MatthewsCorrCoef(task='binary')
mcc = matthews_corrcoef(torch.tensor(y_true), torch.tensor(y_pred))
print("mcc, ", mcc)
record = acc if args.metric == "acc" else mcc
metric_list.append(record)
print(f"Epoch {epoch} {args.metric}: {record}")
if record >= np.max(metric_list):
torch.save(student_model.state_dict(),
f"saved_models/distilled_spikformer/{args.dataset_name}_epoch{epoch}_{record}" +
f"num_step_{args.num_step}_lr{args.fine_tune_lr}_seed{args.seed}" +
f"_batch_size{args.batch_size}_depths{args.depths}_max_length{args.max_length}" +
f"_ce_weight{args.ce_weight}_logit_weight{args.logit_weight}_rep_weight{args.rep_weight}" +
f"_tau{args.tau}_common_thr{args.common_thr}"
)
print("best: ", np.max(metric_list))
return
if __name__ == "__main__":
_args = args()
for arg in vars(_args):
print(arg, getattr(_args, arg))
set_seed(_args.seed)
distill(_args)