-
Notifications
You must be signed in to change notification settings - Fork 0
/
finetune.py
270 lines (240 loc) · 11.2 KB
/
finetune.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
import numpy as np
import tqdm
import torch
import torch.nn as nn
from model import MidiBert, TokenClassification, SequenceClassification
import argparse
from transformers import BertConfig,AdamW
import pickle
import os
from torch.utils.data import DataLoader
from dataset import FinetuneDataset
from peft import LoraConfig, get_peft_model
import copy
from pretrain import get_mask_ind
import random
def get_args():
parser = argparse.ArgumentParser(description='')
### mode ###
parser.add_argument('--task', type=str, required=True)
### dataset & data root ###
parser.add_argument('--dataset', type=str, required=True)
parser.add_argument('--dataroot', type=str, default=None)
### path setup ###
parser.add_argument('--dict_file', type=str,default='./Data/Octuple.pkl')
parser.add_argument('--model_path', default='./midibert_pretrain.pth')
### parameter setting ###
parser.add_argument('--num_workers', type=int, default=5)
parser.add_argument('--class_num', type=int, default=None)
parser.add_argument('--batch_size', type=int, default=8)
parser.add_argument('--max_seq_len', type=int, default=1024,
help='all sequences are padded to `max_seq_len`')
parser.add_argument('--hs', type=int, default=768)
parser.add_argument("--index_layer", type=int,
default=12, help="number of layers")
parser.add_argument('--epochs', type=int, default=10,
help='number of training epochs')
parser.add_argument('--lr', type=float, default=2e-5,
help='initial learning rate')
parser.add_argument('--nopretrain', action="store_true")
### cuda ###
parser.add_argument("--cpu", action="store_true")
parser.add_argument("--cuda_devices", type=int, nargs='+',
default=[5, 7], help="CUDA device ids")
parser.add_argument('--mask', action="store_true")
parser.add_argument('--aug', action="store_true")
args = parser.parse_args()
# check args
if args.class_num is None:
if args.task == 'melody':
args.class_num = 4
elif args.task == 'velocity':
args.class_num = 7
elif args.task == 'composer':
args.class_num = 8
elif args.task == 'emotion':
args.class_num = 4
return args
def load_data(dataset, data_root=None):
if data_root is None:
data_root = 'Data/finetune/others'
X_train = np.load(os.path.join(
data_root, f'{dataset}_train.npy'), allow_pickle=True)
X_val = np.load(os.path.join(
data_root, f'{dataset}_valid.npy'), allow_pickle=True)
X_test = np.load(os.path.join(
data_root, f'{dataset}_test.npy'), allow_pickle=True)
print('X_train: {}, X_valid: {}, X_test: {}'.format(
X_train.shape, X_val.shape, X_test.shape))
y_train = np.load(os.path.join(
data_root, f'{dataset}_train_ans.npy'), allow_pickle=True)
y_val = np.load(os.path.join(
data_root, f'{dataset}_valid_ans.npy'), allow_pickle=True)
y_test = np.load(os.path.join(
data_root, f'{dataset}_test_ans.npy'), allow_pickle=True)
print('y_train: {}, y_valid: {}, y_test: {}'.format(
y_train.shape, y_val.shape, y_test.shape))
return X_train, X_val, X_test, y_train, y_val, y_test
def iteration(model,midibert,optim,data_loader,task,device,train=True,mask=False,aug=False):
if train:
model.train()
torch.set_grad_enabled(True)
else:
model.eval()
torch.set_grad_enabled(False)
acc_list=[]
loss_list=[]
pbar = tqdm.tqdm(data_loader, disable=False)
for x,label in pbar:
x=x.to(device)
label=label.to(device)
x = x.long()
label = label.long()
attn_mask = (x[..., 0] != midibert.bar_pad_word).float().to(device)
batch, seq_len, _ = x.shape
input_ids = copy.deepcopy(x)
loss_mask = None
if aug:
for b in range(batch):
min = torch.min(x[b,:,3])
x[b,:,3][x[b,:,3]>127]=0
max = torch.max(x[b,:,3])
if min<11:
min = -min
else:
min = -11
if max>116:
max = 127 - max
else:
max = 11
rand = random.randint(min, max)
# rand = random.randint(-11, 11)
input_ids[b,:,3][input_ids[b,:,3]<128]+=rand
if mask:
for b in range(batch):
loss_mask = torch.zeros(batch, seq_len).to(device)
mask80, rand10, cur10 = get_mask_ind(mask_percent=0.15)
for i in mask80:
mask_word = torch.tensor(midibert.mask_word_np).to(device)
input_ids[b][i] = mask_word
loss_mask[b][i] = 1
# for i in rand10:
# rand_word = torch.tensor(midibert.get_rand_tok()).to(device)
# input_ids[b][i] = rand_word
# loss_mask[b][i] = 1
y=model(input_ids,attn=attn_mask,layer=-1)
output = torch.argmax(y,dim=-1)
if task=="emotion" or task=="composer":
loss_func = nn.CrossEntropyLoss()
loss=loss_func(y,label)
acc = torch.mean((output==label).float())
else:
if loss_mask is None:
# loss_func = nn.CrossEntropyLoss(reduction="none")
# loss = loss_func(y.reshape(label.shape[0]*label.shape[1],-1),label.reshape(label.shape[0]*label.shape[1])).reshape(label.shape[0],label.shape[1])
# loss = torch.sum(loss*attn_mask)/torch.sum(attn_mask)
# acc = torch.sum((output == label.reshape(label.shape[0],label.shape[1])).float()*attn_mask)/torch.sum(attn_mask)
loss_func = nn.CrossEntropyLoss()
loss = loss_func(y.reshape(label.shape[0]*label.shape[1],-1), label.reshape(label.shape[0]*label.shape[1]))
# acc = torch.mean((output == label.reshape(label.shape[0],label.shape[1])).float())
acc = torch.sum((output == label.reshape(label.shape[0],label.shape[1])).float()*attn_mask)/torch.sum(attn_mask)
else:
loss_func = nn.CrossEntropyLoss(reduction="none")
loss = loss_func(y.reshape(label.shape[0]*label.shape[1],-1),label.reshape(label.shape[0]*label.shape[1])).reshape(label.shape[0],label.shape[1])
# attn_mask = attn_mask * (1-loss_mask)
attn_mask = 1 - loss_mask
loss = torch.sum(loss*attn_mask)/torch.sum(attn_mask)
acc = torch.sum((output == label.reshape(label.shape[0],label.shape[1])).float()*attn_mask)/torch.sum(attn_mask)
if train:
# l2_regularization = 0
# for param in model.parameters():
# l2_regularization += torch.norm(param, 2)
# loss += 1e-4 * l2_regularization
model.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), 3.0)
optim.step()
acc_list.append(acc.item())
loss_list.append(loss.item())
return np.mean(loss_list),np.mean(acc_list)
def main():
args=get_args()
cuda_devices=args.cuda_devices
if not args.cpu and cuda_devices is not None and len(cuda_devices) >= 1:
device_name = "cuda:" + str(cuda_devices[0])
else:
device_name = "cpu"
device=torch.device(device_name)
with open(args.dict_file, 'rb') as f:
e2w, w2e = pickle.load(f)
configuration = BertConfig(max_position_embeddings=args.max_seq_len,
position_embedding_type='relative_key_query',
hidden_size=args.hs)
midibert = MidiBert(bertConfig=configuration, e2w=e2w, w2e=w2e).to(device)
if not args.nopretrain:
midibert.load_state_dict(torch.load(args.model_path,map_location ='cpu'))
# peft_config = LoraConfig(target_modules=['query', 'value', 'key'], r=8, lora_alpha=32, lora_dropout=0.1)
# midibert = get_peft_model(midibert, peft_config)
# midibert.print_trainable_parameters()
task = args.task
if task=="composer" or task=="emotion":
model = SequenceClassification(midibert, args.class_num, args.hs).to(device)
else:
model = TokenClassification(midibert, args.class_num+1, args.hs).to(device)
if len(cuda_devices) > 1 and not args.cpu:
model = nn.DataParallel(model, device_ids=cuda_devices)
X_train, X_val, X_test, y_train, y_val, y_test = load_data(args.dataset, args.dataroot)
trainset = FinetuneDataset(X=X_train, y=y_train)
validset = FinetuneDataset(X=X_val, y=y_val)
testset = FinetuneDataset(X=X_test, y=y_test)
train_loader = DataLoader(trainset, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=True)
print(" len of train_loader", len(train_loader))
valid_loader = DataLoader(validset, batch_size=args.batch_size, num_workers=args.num_workers)
print(" len of valid_loader", len(valid_loader))
test_loader = DataLoader(testset, batch_size=args.batch_size, num_workers=args.num_workers)
print(" len of valid_loader", len(test_loader))
optim = AdamW(model.parameters(), lr=args.lr, weight_decay=0.01)
best_acc=0
best_loss=1e8
acc_epoch=0
loss_epoch=0
result_acc=None
j = 0
while True:
j+=1
loss,acc=iteration(model, midibert, optim, train_loader, task, device, train=True, mask=args.mask,aug=args.aug)
# loss,acc=iteration(model, midibert, optim, train_loader, task, device, train=True, mask=args.mask,aug=False)
log = "Epoch {:} | Train Loss {:06f} Train Acc {:06f} | ".format(j,loss,acc)
with open(args.task+"_"+args.dataset+".txt",'a') as file:
file.write(log)
print(log)
loss,acc=iteration(model, midibert, optim, valid_loader, task, device, train=False, mask=args.mask,aug=args.aug)
# loss,acc=iteration(model, midibert, optim, valid_loader, task, device, train=False, mask=args.mask, aug=False)
log = "Valid Loss {:06f} Valid Acc {:06f} | ".format(loss,acc)
with open(args.task+"_"+args.dataset+".txt",'a') as file:
file.write(log)
print(log)
test_loss,test_acc=iteration(model, midibert, optim, test_loader, task, device, train=False, mask=args.mask,aug=args.aug)
# test_loss,test_acc=iteration(model, midibert, optim, test_loader, task, device, train=False, mask=args.mask, aug=False)
log = "Test Loss {:06f} Test Acc {:06f}".format(test_loss,test_acc)
with open(args.task+"_"+args.dataset+".txt",'a') as file:
file.write(log+"\n")
print(log)
if acc >= best_acc or loss <= best_loss:
# torch.save(model.state_dict(), args.task+"_"+args.dataset+".pth")
result_acc = test_acc
if acc >= best_acc:
best_acc = acc
acc_epoch = 0
else:
acc_epoch += 1
if loss < best_loss:
best_loss = loss
loss_epoch = 0
else:
loss_epoch += 1
if acc_epoch >= args.epochs and loss_epoch >= args.epochs:
break
print("Acc Epoch {:}, Loss Epcoh {:}, Result Acc {:}".format(acc_epoch, loss_epoch,result_acc))
if __name__ == '__main__':
main()