forked from zhaojw1998/AccoMontage
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtransition_model_train_contrastive.py
187 lines (165 loc) · 6.79 KB
/
transition_model_train_contrastive.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
import pretty_midi as pyd
import numpy as np
import os
from torch.utils.tensorboard.summary import text
from tqdm import tqdm
#import converter
import time
import platform
import torch
from torch.utils.tensorboard import SummaryWriter
from torch.optim.lr_scheduler import ExponentialLR
from torch import optim
from transition_model_data_loader import two_bar_dataset
import sys
sys.path.append('./models')
from two_bar_contrastive_model import contrastive_model
from ptvae import TextureEncoder
args={
"Linux_batch_size": 8,
"batch_size": 8,
"Linux_data_path": "./data files/song_data.npz",
"data_path": "./data files/song_data.npz",
'Linux_weight_path': "./data files/model_master_final.pt",
'weight_path': "./data files/model_master_final.pt",
"Linux_embed_size": 256,
"embed_size": 256,
"Linux_hidden_dim": 1024,
"hidden_dim": 1024,
"time_step": 32,
"n_epochs": 100,
"lr": 1e-3,
"decay": 0.99991,
"Linux_log_save": "./demo/demo_generate/log",
"log_save": "./demo/demo_generate/log",
}
# 10个epoch contrastive optimizer稳定
class MinExponentialLR(ExponentialLR):
def __init__(self, optimizer, gamma, minimum, last_epoch=-1):
self.min = minimum
super(MinExponentialLR, self).__init__(optimizer, gamma, last_epoch=last_epoch)
def get_lr(self):
return [
max(base_lr * self.gamma**self.last_epoch, self.min)
for base_lr in self.base_lrs
]
class AverageMeter(object):
"""Computes and stores the average and current value"""
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = self.sum / self.count
def train(contrastive_model, texture_model, dataset, optimizer, scheduler, loss_recorder, writer):
batch = dataset.get_batch('train') #8 * 6 * 8 * 32 * 128
batch = torch.from_numpy(batch).float().cuda()
bs, pos_neg, time, roll = batch.shape
optimizer_2.zero_grad()
_, batch = texture_model(batch.view(-1, time, roll))
batch = batch.view(bs, pos_neg, -1)
optimizer.zero_grad()
similarity = contrastive_model(batch)
model_loss = contrastive_model.contrastive_loss(similarity)
model_loss.backward()
torch.nn.utils.clip_grad_norm_(contrastive_model.parameters(), 1)
optimizer_2.step()
optimizer.step()
loss_recorder.update(model_loss.item())
scheduler_2.step()
scheduler.step()
n_epoch = dataset.get_epoch()
total_batch = dataset.get_batch_volumn('train')
current_batch = dataset.train_batch_anchor
step = current_batch + n_epoch * total_batch
print('---------------------------Training VAE----------------------------')
for param in optimizer.param_groups:
print('lr1: ', param['lr'])
print('Epoch: [{0}][{1}/{2}]'.format(n_epoch, current_batch, total_batch))
print('loss: {loss:.5f}'.format(loss=model_loss.item()))
writer.add_scalar('train_vae/1-loss_total-epoch', loss_recorder.avg, step)
writer.add_scalar('train_vae/5-learning-rate', param['lr'], step)
def val(contrastive_model, texture_model, dataset, writer, val_loss_recoder):
loss = val_loss_recoder
step = 1
for i in range(dataset.get_batch_volumn('val')):
batch = dataset.get_batch('val')
batch = torch.from_numpy(batch).float().cuda()
bs, pos_neg, time, roll = batch.shape
_, batch = texture_model(batch.view(-1, time, roll))
batch = batch.view(bs, pos_neg, -1)
with torch.no_grad():
similarity = contrastive_model(batch)
model_loss = contrastive_model.contrastive_loss(similarity)
loss.update(model_loss.item())
n_epoch = dataset.get_epoch()
total_batch = dataset.get_batch_volumn('val')
print('----validation----')
print('Epoch: [{0}][{1}/{2}]'.format(n_epoch, step, total_batch))
print('loss: {loss:.5f}'.format(loss=model_loss.item()))
step += 1
writer.add_scalar('val/loss_total-epoch', loss.avg, n_epoch)
if platform.system() == 'Linux':
embed_size = args["Linux_embed_size"]
hidden_dim = args["Linux_hidden_dim"]
weight_path = args["Linux_weight_path"]
else:
embed_size = args["embed_size"]
hidden_dim = args["hidden_dim"]
weight_path = args["weight_path"]
contrastive_model = contrastive_model(emb_size=embed_size, hidden_dim=hidden_dim).cuda()
texture_model = TextureEncoder(emb_size=256, hidden_dim=1024, z_dim=256, num_channel=10, for_contrastive=True)
checkpoint = torch.load(weight_path)
from collections import OrderedDict
rhy_checkpoint = OrderedDict()
for k, v in checkpoint.items():
part = k.split('.')[0]
name = '.'.join(k.split('.')[1:])
if part == 'rhy_encoder':
rhy_checkpoint[name] = v
texture_model.load_state_dict(rhy_checkpoint)
texture_model.cuda()
run_time = time.asctime(time.localtime(time.time())).replace(':', '-')
logdir = 'log/' + run_time[4:]
save_dir = 'params/' + run_time[4:]
if platform.system() == 'Linux':
logdir = os.path.join(args["Linux_log_save"], logdir)
save_dir = os.path.join(args["Linux_log_save"], save_dir)
batch_size = args['Linux_batch_size']
else:
logdir = os.path.join(args["log_save"], logdir)
save_dir = os.path.join(args["log_save"], save_dir)
batch_size = args['batch_size']
if not os.path.exists(logdir):
os.makedirs(logdir)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
writer = SummaryWriter(logdir)
training_loss_recoder = AverageMeter()
val_loss_recoder = AverageMeter()
dataset = two_bar_dataset(args['data_path'], batch_size, args['time_step'])
dataset.make_batch(batch_size)
optimizer = optim.Adam(contrastive_model.parameters(), lr=args['lr'])
optimizer_2 = optim.Adam(texture_model.parameters(), lr=1e-4)
scheduler = MinExponentialLR(optimizer, gamma=args['decay'], minimum=1e-5,)
scheduler_2 = MinExponentialLR(optimizer_2, gamma=0.999995, minimum=5e-6,)
while dataset.get_epoch() < args['n_epochs']:
if dataset.train_batch_anchor == 0:
contrastive_model.eval()
val(contrastive_model, texture_model, dataset, writer, val_loss_recoder)
if (dataset.get_epoch()) % 1 == 0:
checkpoint = save_dir + '/contrastive_model_params' + str(dataset.get_epoch()).zfill(3) + '.pt'
torch.save(contrastive_model.cpu().state_dict(), checkpoint)
contrastive_model.cuda()
checkpoint = save_dir + '/texture_model_params' + str(dataset.get_epoch()).zfill(3) + '.pt'
torch.save(texture_model.cpu().state_dict(), checkpoint)
texture_model.cuda()
print('Model saved!')
contrastive_model.train()
train(contrastive_model, texture_model, dataset, optimizer, scheduler, training_loss_recoder, writer)