-
Notifications
You must be signed in to change notification settings - Fork 1
/
midi_glue.py
311 lines (248 loc) · 12.2 KB
/
midi_glue.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
# This file is a modification of MIDI-BERT/MidiBERT/CP/finetunetrainer.py
# from https://github.com/wazenmai/MIDI-BERT.git
# which is used to finetune MidiBERT-Piano on symbolic music understanding tasks
# like composer classification, mood prediction, etc.
# We have modified it to finetune on GLUE tasks
import argparse
import numpy as np
import pickle
import os
import random
from torch.utils.data import DataLoader
import torch
from transformers import BertConfig
from datasets import load_dataset
import pandas as pd
from transformers import BertTokenizer
from model import MidiBert
from finetune_trainer import FinetuneTrainer
from finetune_dataset import FinetuneDataset
from finetune_model import TokenClassification, SequenceClassification
from matplotlib import pyplot as plt
def get_args():
parser = argparse.ArgumentParser(description='')
### mode ###
parser.add_argument('--task', type=str, required=True)
### path setup ###
parser.add_argument('--dict_file', type=str, default='../../dict/CP.pkl')
parser.add_argument('--name', type=str, default='')
parser.add_argument('--ckpt', default='result/pretrain/test/model_best.ckpt')
### parameter setting ###
parser.add_argument('--num_workers', type=int, default=2)
parser.add_argument('--class_num', type=int, default=2)
parser.add_argument('--batch_size', type=int, default=16)
parser.add_argument('--max_seq_len', type=int, default=512, 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=3, 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") # default: false
parser.add_argument('--resume', action="store_true") # default: false
### cuda ###
parser.add_argument("--cpu", action="store_true") # default=False
parser.add_argument("--cuda_devices", type=int, nargs='+', default=[0,1,2,3], help="CUDA device ids")
args = parser.parse_args()
return args
def load_data(task):
#!pip install datasets==1.11.0
args = get_args()
max_seq_length = args.max_seq_len
actual_task = "mnli" if task == "mnli-mm" else task
dataset = load_dataset("glue", actual_task)
print (actual_task)
task_to_keys = {
"mnli": ["premise", "hypothesis"],
"mrpc": ["sentence1", "sentence2"],
"qnli": ["question", "sentence"],
"qqp": ["question1", "question2"],
"rte": ["sentence1", "sentence2"],
"stsb": ["sentence1", "sentence2"],
"wnli": ["sentence1", "sentence2"],
}
# Load the training dataset
df = pd.DataFrame(dataset["train"][:])
labels = df.label.values
print('Loading BERT tokenizer...')
tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
input_ids = []
if actual_task=="cola" or actual_task=="sst2":
sentences = df.sentence.values
for sent in sentences:
encoded_dict = tokenizer.encode(
sent,
max_length = max_seq_length , # Pad
add_special_tokens = False,
truncation=True,
pad_to_max_length = True,
return_tensors = 'pt', # Return pytorch tensors
)
input_ids.append(encoded_dict)
else:
sentences = df[task_to_keys[actual_task]].values.astype("str")
for sent in sentences:
encoded_dict = tokenizer.encode(
sent[0], sent[1],
max_length = max_seq_length , # Pad
add_special_tokens = False,
truncation=True,
pad_to_max_length = True,
return_tensors = 'pt', # Return pytorch tensors
)
input_ids.append(encoded_dict)
input_ids = torch.cat(input_ids, dim=0)
labels = torch.tensor(labels)
x = input_ids.numpy()
y_train = labels.numpy()
X_train = np.empty((x.shape[0], max_seq_length, 4))
X_train[:,:,0] = 1 # map to 1 in the Bar column
X_train[:,:,1] = x%16 # map to 0-16 in the Position column
X_train[:,:,2] = ((x/16).astype(int))%86 # map to 0-86 in the Pitch column
X_train[:,:,3] = ((x/(16*86)).astype(int))%32 # map to 0-32 in the Duration column
# the pad token is initially mapped to [1 0 0 0]
# but we want it mapped to [ 2 16 86 64]
for i in range (0,x.shape[0]):
for j in range (0,max_seq_length):
X_train[i,j,:] = np.array([2, 16, 86, 64]) if(np.sum(X_train[i,j,:]) == 1) else X_train[i,j,:]
X_train = X_train.astype(int)
# Load validation dataset
df = pd.DataFrame(dataset["validation"][:])
labels = df.label.values
print('Loading BERT tokenizer...')
tokenizer = BertTokenizer.from_pretrained('bert-base-cased')
input_ids = []
if actual_task=="cola" or actual_task=="sst2":
sentences = df.sentence.values
for sent in sentences:
encoded_dict = tokenizer.encode(
sent,
max_length = max_seq_length ,
add_special_tokens = False,
truncation=True,
pad_to_max_length = True,
return_tensors = 'pt',
)
input_ids.append(encoded_dict)
else:
sentences = df[task_to_keys[actual_task]].values.astype("str")
for sent in sentences:
encoded_dict = tokenizer.encode(
sent[0], sent[1],
max_length = max_seq_length ,
add_special_tokens = False,
truncation=True,
pad_to_max_length = True,
return_tensors = 'pt',
)
input_ids.append(encoded_dict)
input_ids = torch.cat(input_ids, dim=0)
labels = torch.tensor(labels)
x = input_ids.numpy()
y_val = labels.numpy()
X_val = np.empty((x.shape[0], max_seq_length, 4))
X_val[:,:,0] = 1 # map to 1 in the Bar column
X_val[:,:,1] = x%16 # map to 0-16 in the Position column
X_val[:,:,2] = ((x/16).astype(int))%86 # map to 0-86 in the Pitch column
X_val[:,:,3] = ((x/(16*86)).astype(int))%32 # map to 0-32 in the Duration column
# the pad token is initially mapped to [1 0 0 0]
# but we want it mapped to [ 2 16 86 64]
for i in range (0,x.shape[0]):
for j in range (0,max_seq_length):
X_val[i,j,:] = np.array([2, 16, 86, 64]) if(np.sum(X_val[i,j,:]) == 1) else X_val[i,j,:]
X_val = X_val.astype(int)
print('X_train: {}, X_valid: {}'.format(X_train.shape, X_val.shape))
print('y_train: {}, y_valid: {}'.format(y_train.shape, y_val.shape))
return X_train, X_val, y_train, y_val
def main():
# set seed
seed = 2021
torch.manual_seed(seed) # cpu
torch.cuda.manual_seed(seed) # current gpu
torch.cuda.manual_seed_all(seed) # all gpu
np.random.seed(seed)
random.seed(seed)
# argument
args = get_args()
print("Loading Dictionary")
with open(args.dict_file, 'rb') as f:
e2w, w2e = pickle.load(f)
print("\nLoading Dataset")
seq_class = True
args.class_num = 1 if args.task=="stsb" else 2
X_train, X_val, y_train, y_val = load_data(args.task)
trainset = FinetuneDataset(X=X_train, y=y_train)
validset = FinetuneDataset(X=X_val, y=y_val)
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))
###
X_test, y_test = X_val, y_val
test_loader = valid_loader
print("\nBuilding BERT model")
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)
best_mdl = ''
if args.resume:
model = SequenceClassification(midibert, args.class_num, args.hs)
best_mdl = args.ckpt
print(" Loading model from", best_mdl.split('/')[-1])
checkpoint = torch.load(best_mdl, map_location='cpu')
model.load_state_dict(checkpoint['state_dict'])
elif not args.nopretrain:
model = None
best_mdl = args.ckpt
print(" Loading pre-trained model from", best_mdl.split('/')[-1])
checkpoint = torch.load(best_mdl, map_location='cpu')
midibert.load_state_dict(checkpoint['state_dict'])
else:
model = None
index_layer = int(args.index_layer)-13
print("\nCreating Finetune Trainer using index layer", index_layer)
trainer = FinetuneTrainer(midibert, train_loader, valid_loader, test_loader, index_layer, args.lr, args.class_num, args.task,
args.hs, y_test.shape, args.cpu, args.cuda_devices, model, seq_class, args.max_seq_len)
print("\nTraining Start")
save_dir = os.path.join('result/finetune/', args.task)
os.makedirs(save_dir, exist_ok=True)
filename = os.path.join(save_dir, 'model.ckpt')
print(" save model at {}".format(filename))
best_acc, best_epoch = 0, 0
bad_cnt = 0
# train_accs, valid_accs = [], []
with open(os.path.join(save_dir, 'log'), 'a') as outfile:
outfile.write("Loading pre-trained model from " + best_mdl.split('/')[-1] + '\n')
valid_loss, valid_acc = trainer.valid()
print("initial valid loss & acc: ", valid_loss, valid_acc)
for epoch in range(args.epochs):
train_loss, train_acc = trainer.train()
valid_loss, valid_acc = trainer.valid()
test_loss, test_acc, _ = 0,0,0 #trainer.test()
is_best = valid_acc >= best_acc
best_acc = max(valid_acc, best_acc)
if is_best:
bad_cnt, best_epoch = 0, epoch
else:
bad_cnt += 1
print('epoch: {}/{} | Train Loss: {} | Train acc: {} | Valid Loss: {} | Valid acc: {} | Test loss: {} | Test acc: {}'.format(
epoch+1, args.epochs, train_loss, train_acc, valid_loss, valid_acc, test_loss, test_acc))
# train_accs.append(train_acc)
# valid_accs.append(valid_acc)
trainer.save_checkpoint(epoch, train_acc, valid_acc,
valid_loss, train_loss, is_best, filename)
outfile.write('Epoch {}: train_loss={}, valid_loss={}, test_loss={}, train_acc={}, valid_acc={}, test_acc={}\n'.format(
epoch+1, train_loss, valid_loss, test_loss, train_acc, valid_acc, test_acc))
if bad_cnt > 3:
print('valid acc not improving for 3 epochs')
break
# draw figure valid_acc & train_acc
'''plt.figure()
plt.plot(train_accs)
plt.plot(valid_accs)
plt.title(f'{args.task} task accuracy (w/o pre-training)')
plt.xlabel('epoch')
plt.ylabel('accuracy')
plt.legend(['train','valid'], loc='upper left')
plt.savefig(f'acc_{args.task}_scratch.jpg')'''
if __name__ == '__main__':
main()