forked from catniplab/ML-music-analysis
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_model.py
300 lines (235 loc) · 10.8 KB
/
train_model.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
import json
import string
import random
import pickle as pkl
import simmanager
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from utils.models import MusicRNN
from utils.metrics import FrameAccuracy, MaskedBCE
from utils.data_loader import get_datasets
from utils.initialization import initialize
from utils.plotting import plot_note_comparison, plot_phase_portrait
from absl import flags, app
from datetime import datetime
from os.path import join as opj
FLAGS = flags.FLAGS
# system
flags.DEFINE_bool(
'use_gpu', False, 'Whether or not to use the GPU. Fails if True and CUDA is not available.')
flags.DEFINE_string(
'model_name', '', 'If non-empty works as a special name for this model.')
flags.DEFINE_string('results_path', 'models',
'Name of the directory to save all results within.')
flags.DEFINE_integer(
'random_seed', -1, 'If not -1, set the random seed to this value. Otherwise the random seed will be the current microsecond.')
# training
flags.DEFINE_enum('dataset', 'JSB_Chorales', [
'JSB_Chorales', 'Nottingham', 'Piano_midi', 'MuseData'], 'Which dataset to train the model on.')
flags.DEFINE_integer('n_steps', 10000,
'How many training batches to show the network.')
flags.DEFINE_integer('batch_size', 50, 'Batch size.')
flags.DEFINE_float('lr', 0.001, 'Learning rate.')
flags.DEFINE_integer(
'decay_every', 1000, 'Shrink the learning rate after this many training steps.')
flags.DEFINE_float(
'lr_decay', 0.95, 'Shrink the learning rate by this factor.')
flags.DEFINE_enum('optimizer', 'RMSprop', [
'Adam', 'SGD', 'Adagrad', 'RMSprop'], 'Which optimizer to use.')
flags.DEFINE_float('reg_coeff', 0.0001,
'Coefficient for L2 regularization of weights.')
flags.DEFINE_bool(
'use_grad_clip', False, 'Whether or not to clip the backward gradients by their magnitude.')
flags.DEFINE_float(
'grad_clip', 1, 'Maximum magnitude of gradients if gradient clipping is used.')
flags.DEFINE_integer(
'validate_every', 500, 'Validate the model at this many training steps.')
flags.DEFINE_integer(
'save_every', 1000, 'Save the model at this many training steps.')
flags.DEFINE_boolean('plot', False, 'Plot note comparison and phase portrait every validation step.')
# model
flags.DEFINE_enum('architecture', 'TANH', [
'TANH', 'LSTM', 'GRU'], 'Which recurrent architecture to use.')
flags.DEFINE_integer(
'n_rec', 400, 'How many recurrent neurons to use.')
flags.DEFINE_enum('initialization', 'default', ['default', 'orthogonal', 'limit_cycle'],
'Which initialization to use for the recurrent weight matrices. Default is uniform Xavier. Limit cycles only apply to TANH and GRU.')
flags.DEFINE_string(
'restore_from', '', 'If non-empty, restore the previous model from this directory and train it using the new flags.')
def train_loop(sm, FLAGS, model, train_iter, valid_iter, test_iter):
# if we are on cuda we construct the device and run everything on it
device = torch.device('cpu')
if FLAGS.use_gpu:
dev_name = 'cuda:0'
device = torch.device(dev_name)
model = model.to(device)
train_loss = []
train_reg = []
train_acc = []
valid_loss = []
valid_acc = []
# construct the optimizer
if FLAGS.optimizer == 'SGD':
optimizer = optim.SGD(model.parameters(), lr=FLAGS.lr)
elif FLAGS.optimizer == 'Adam':
optimizer = optim.Adam(model.parameters(), lr=FLAGS.lr)
elif FLAGS.optimizer == 'RMSprop':
optimizer = optim.RMSprop(model.parameters(), lr=FLAGS.lr)
elif FLAGS.optimizer == 'Adagrad':
optimizer = optim.Adagrad(model.parameters(), lr=FLAGS.lr)
else:
raise ValueError(f'Optimizer {FLAGS.optimizier} not recognized.')
# learning rate decay
scheduler = None
scheduler = optim.lr_scheduler.LambdaLR(
optimizer, lambda epoch: FLAGS.lr_decay**(epoch//FLAGS.decay_every))
acc_fcn = FrameAccuracy()
loss_fcn = MaskedBCE()
# begin training loop
for i in range(FLAGS.n_steps):
# get next training sample
x, y, mask = next(train_iter)
x, y, mask = x.to(device), y.to(device), mask.to(device)
# forward pass
output, hidden = model(x)
# binary cross entropy
bce_loss = loss_fcn(output, y, mask)
# weight regularization
l2_reg = torch.tensor(0, dtype=torch.float32, device=device)
for param in model.parameters():
l2_reg += FLAGS.reg_coeff*torch.norm(param)
# backward pass and optimization step
loss = bce_loss + l2_reg
optimizer.zero_grad()
loss.backward()
optimizer.step()
# learning rate decay
scheduler.step()
# compute accuracy
acc = acc_fcn(output, y, mask)
# append metrics
train_loss.append(bce_loss.cpu().item())
train_acc.append(acc.cpu().item())
train_reg.append(l2_reg.cpu().item())
if i > 0 and i % FLAGS.validate_every == 0:
timestamp = datetime.now().strftime('%H:%M:%S')
print(
f'{timestamp} Validating {sm.sim_name} at iteration {i}.\n Training loss: {train_loss[-1]:.3}\n Training accuracy: {100*train_acc[-1]:.3}%\n L2 regularization: {train_reg[-1]:.3}')
# get next validation sample
x, y, mask = next(valid_iter)
x, y, mask = x.to(device), y.to(device), mask.to(device)
# forward pass
output, hidden = model(x)
# binary cross entropy
bce_loss = loss_fcn(output, y, mask)
# compute accuracy
acc = acc_fcn(output, y, mask)
# append metrics
valid_loss.append(bce_loss.cpu().item())
valid_acc.append(acc.cpu().item())
print(
f' Validation loss: {valid_loss[-1]:.3}\n Validation accuracy: {100*valid_acc[-1]:.3}%\n')
if FLAGS.plot:
plot_note_comparison(sm, output, y, i)
if model.architecture in ['TANH', 'GRU']:
plot_phase_portrait(sm, model, i)
if i > 0 and i % FLAGS.save_every == 0:
timestamp = datetime.now().strftime('%H:%M:%S')
print(f'{timestamp} Saving {sm.sim_name} at iteration {i}.\n')
np.save(opj(sm.paths.results_path, 'training_loss'), train_loss)
np.save(opj(sm.paths.results_path,
'training_accuracy'), train_acc)
np.save(opj(sm.paths.results_path,
'training_regularization'), train_reg)
np.save(opj(sm.paths.results_path, 'validation_loss'), valid_loss)
np.save(opj(sm.paths.results_path,
'validation_accuracy'), valid_acc)
torch.save(model.state_dict(), opj(
sm.paths.results_path, 'model_checkpoint.pt'))
timestamp = datetime.now().strftime('%H:%M:%S')
print(f'{timestamp} Finished training. Entering testing phase.')
test_loss = []
test_acc = []
tot_test_samples = 0
# loop through entire testing set
for x, y, mask in test_iter:
x, y, mask = x.to(device), y.to(device), mask.to(device)
output, hidden = model(x)
bce_loss = loss_fcn(output, y, mask).cpu().item()
acc = acc_fcn(output, y, mask).cpu().item()
batch_size = x.shape[0]
tot_test_samples += batch_size
test_loss.append(batch_size*bce_loss)
test_acc.append(batch_size*acc)
final_test_loss = np.sum(test_loss)/tot_test_samples
final_test_acc = np.sum(test_acc)/tot_test_samples
print(
f' Testing loss: {final_test_loss:.3}\n Testing accuracy: {100*final_test_acc:.3}%')
print(f'Final save of {sm.sim_name}.')
np.save(opj(sm.paths.results_path, 'testing_loss'), final_test_loss)
np.save(opj(sm.paths.results_path, 'testing_accuracy'), final_test_acc)
torch.save(model.state_dict(), opj(
sm.paths.results_path, 'model_checkpoint.pt'))
date = datetime.now().strftime('%d-%m-%Y')
timestamp = datetime.now().strftime('%H:%M:%S')
print(f'Training {sm.sim_name} ended on {date} at {timestamp}.')
def main(_argv):
# construct simulation manager
if FLAGS.model_name == '':
identifier = ''.join(random.choice(
string.ascii_lowercase + string.digits) for _ in range(4))
sim_name = 'model_{}'.format(identifier)
else:
sim_name = FLAGS.model_name
date = datetime.now().strftime('%d-%m-%Y')
timestamp = datetime.now().strftime('%H:%M:%S')
print(f'\nBeginning to train {sim_name} on {date} at {timestamp}.\n')
sm = simmanager.SimManager(
sim_name, FLAGS.results_path, write_protect_dirs=False, tee_stdx_to='output.log')
with sm:
# check for cuda
if FLAGS.use_gpu and not torch.cuda.is_available():
raise OSError(
'CUDA is not available. Check your installation or set `use_gpu` to False.')
# dump FLAGS for this experiment
with open(opj(sm.paths.data_path, 'FLAGS.json'), 'w') as f:
flag_dict = {}
for k in FLAGS._flags().keys():
if k not in FLAGS.__dict__['__hiddenflags']:
flag_dict[k] = FLAGS.__getattr__(k)
json.dump(flag_dict, f)
# generate, save, and set random seed
if FLAGS.random_seed != -1:
random_seed = FLAGS.random_seed
else:
random_seed = datetime.now().microsecond
np.save(opj(sm.paths.data_path, 'random_seed'), random_seed)
torch.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)
# check for old model to restore from
if FLAGS.restore_from != '':
with open(opj(FLAGS.restore_from, 'data', 'FLAGS.json'), 'r') as f:
old_FLAGS = json.load(f)
architecture = old_FLAGS['architecture']
if architecture != FLAGS.architecture:
print(
'Warning: restored architecture does not agree with architecture specified in FLAGS.')
n_rec = old_FLAGS['n_rec']
if n_rec != FLAGS.n_rec:
print(
'Warning: restored number of recurrent units does not agree with number of recurrent units specified in FLAGS.')
model = MusicRNN(
architecture, n_rec, use_grad_clip=FLAGS.use_grad_clip, grad_clip=FLAGS.grad_clip)
model.load_state_dict(torch.load(
opj(FLAGS.restore_from, 'results', 'model_checkpoint.pt')))
else:
model = MusicRNN(FLAGS.architecture, FLAGS.n_rec,
use_grad_clip=FLAGS.use_grad_clip, grad_clip=FLAGS.grad_clip)
initialize(model, FLAGS)
train_iter, valid_iter, test_iter = get_datasets(FLAGS)
train_loop(sm, FLAGS, model, train_iter, valid_iter, test_iter)
if __name__ == '__main__':
app.run(main)