-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path_train.py
165 lines (134 loc) · 7.88 KB
/
_train.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
import torch
import random
from tqdm import tqdm
import torch.optim as optim
from torch.utils.data import DataLoader
import numpy as np
from _models import *
from _utils import *
def enc_logs(batch_size, enc_dim, eras, eras_dict, encodings, epoch, count):
random_sample = random.randint(0, batch_size - 1)
for era in eras_dict.keys():
if eras[random_sample] == era:
eras_dict[era] = encodings[random_sample].squeeze().cpu().detach().numpy()
if np.all([v is not None for v in eras_dict.values()]):
eras_dict = save_encodings(eras_dict, enc_dim, epoch, count)
count += 1
return eras_dict, count
def gan_logs(batch_size, epoch, style_images, pair_images):
style_images = transpose_image(style_images, range_min=0, range_max=1)
pair_images = transpose_image(pair_images, range_min=0, range_max=1)
random_sample = random.randint(0, batch_size - 1)
save_images(style_images[random_sample], pair_images[random_sample], epoch)
def train_encoder(device, data, enc_dim=10, batch_size=1, num_epochs=15, visualize=True):
data_loader = DataLoader(data, batch_size=batch_size, shuffle=True, num_workers=2)
losses, batches = [], [batch for batch in batch_generator(data_loader)]
encoder = Encoder().to(device)
enc_optimizer = optim.Adam(encoder.parameters(), lr=0.00005, betas=(0.5, 0.999))
encoder.train()
for epoch in range(num_epochs):
enc_epoch_loss, steps, count = 0, 0, 0
eras_dict = {"renaissance": None, "baroque": None, "classical": None, "romantic": None, "modern": None}
pbar = tqdm(data_loader, desc="Epoch {}".format(epoch), total=len(data_loader))
for _, dl in enumerate(pbar):
melspecs, eras, _ = dl
melspecs = torch.transpose(torch.stack(melspecs, dim=0), 0, 1)
pos_melspecs, neg_melspecs = [], []
for sample in range(batch_size):
pos, neg = False, False
while not pos or not neg:
rbID = random.randint(0, len(data_loader) - 1)
rand_melspecs, rand_eras = torch.transpose(torch.stack(batches[rbID][0], dim=0), 0, 1), batches[rbID][1]
rsID = random.randint(0, batch_size - 1)
rand_melspec, rand_era = rand_melspecs[rsID], rand_eras[rsID]
if not pos and rand_era == eras[sample] and not torch.all(torch.eq(rand_melspec, melspecs[rsID])):
pos_melspecs.append(rand_melspec)
pos = True
elif not neg and rand_era != eras[sample]:
neg_melspecs.append(rand_melspec)
neg = True
pos_melspecs, neg_melspecs = torch.stack(pos_melspecs, dim=0), torch.stack(neg_melspecs, dim=0)
EA, EPA, ENA = encoder(melspecs.to(device)), encoder(pos_melspecs.to(device)), encoder(neg_melspecs.to(device))
enc_optimizer.zero_grad()
enc_loss = triplet_margin_loss(EA.view(EA.size(0), -1), EPA.view(EPA.size(0), -1), ENA.view(ENA.size(0), -1))
enc_loss.backward()
enc_optimizer.step()
enc_epoch_loss += enc_loss.item()
steps += 1
pbar.set_postfix({'Encoder Loss': '{0:.3f}'.format(enc_epoch_loss/steps)})
eras_dict, count = enc_logs(batch_size, enc_dim, eras, eras_dict, EA, epoch, count) if visualize and count < 5 else (eras_dict, count)
losses.append(enc_epoch_loss / steps)
torch.save(encoder.state_dict(), "/content/drive/MyDrive/MODELS/encoder.pth")
monitor_encoder(losses, num_epochs)
def train_gan(device, data, enc_dim=10, batch_size=15, num_epochs=100, gen_iters=1, dis_iters=1, noise_dim=100, visualize=True):
gan_losses = {'gen': [], 'dis': []}
conditional_noise_dim = noise_dim + enc_dim
data_loader = DataLoader(data, batch_size=batch_size, shuffle=True, num_workers=2)
encoder = Encoder().to(device)
encoder.load_state_dict(torch.load("/content/drive/MyDrive/MODELS/encoder.pth"))
encoder.eval()
generator = Generator(in_dim=conditional_noise_dim).to(device)
gen_optimizer = optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))
generator.apply(xavier_weights)
generator.train()
discriminator = Discriminator().to(device)
dis_optimizer = optim.Adam(discriminator.parameters(), lr=0.00005, betas=(0.5, 0.999))
discriminator.apply(xavier_weights)
discriminator.train()
for epoch in range(num_epochs):
gen_epoch_loss, dis_epoch_loss, steps = 0, 0, 0
random_batch = random.randint(0, len(data_loader) - 1)
pbar = tqdm(data_loader, desc="Epoch {}".format(epoch), total=len(data_loader))
for i, dl in enumerate(pbar):
melspecs, _, pairs = dl
melspecs = torch.transpose(torch.stack(melspecs, dim=0), 0, 1)
noise = torch.randn(batch_size, noise_dim).to(device)
encoding = encoder(melspecs.to(device))
input = torch.cat((noise, encoding), dim=1).to(device)
SI = generator(input).to(device)
PI = torch.stack([pair for pair in pairs]).requires_grad_(True).to(device)
PI = transpose_image(PI, range_min=-1, range_max=1)
for _ in range(dis_iters):
dis_optimizer.zero_grad()
logSI, sigSI = discriminator(SI.detach())
logPI, sigPI = discriminator(PI)
logPI.requires_grad_(); sigPI.requires_grad_()
dis_loss = stDis_loss(logSI, sigSI, logPI, sigPI)
dis_loss.backward(retain_graph=False)
dis_optimizer.step()
dis_epoch_loss += dis_loss.item()
for _ in range(gen_iters):
gen_optimizer.zero_grad()
logSI, sigSI = discriminator(SI)
gen_loss = stGen_loss(logSI, sigSI)
gen_loss.backward(retain_graph=False)
gen_optimizer.step()
gen_epoch_loss += gen_loss.item()
steps += 1
pbar.set_postfix({'Generator Loss': '{0:.3f}'.format(gen_epoch_loss / (steps * gen_iters)),
'Discriminator Loss': '{0:.3f}'.format(dis_epoch_loss / (steps * dis_iters))})
gan_logs(batch_size, epoch, SI, PI) if visualize and random_batch == i else None
"""if i == random_batch:
TSI = transpose_image(SI, range_min=0, range_max=1)
show_image(TSI[0])
show_image(TSI[1])
show_image(TSI[2])"""
gan_losses['gen'].append(gen_epoch_loss / (steps * gen_iters))
gan_losses['dis'].append(dis_epoch_loss / (steps * dis_iters))
monitor_gan(gan_losses, epoch+1)
torch.save(generator.state_dict(), "/content/drive/MyDrive/MODELS/generator.pth")
torch.save(discriminator.state_dict(), "/content/drive/MyDrive/MODELS/discriminator.pth")
def train(device):
dataset_path = "/content/drive/MyDrive/DATASETS/musart-dataset/training"
art_eras = ["renaissance", "baroque", "classical", "romantic", "modern"]
prepared = input("<training_data.npy> already exists in ~/content/drive/MyDrive/DATA (y/n): ")
if prepared == "n":
data = prepare_data(dataset_path, art_eras, sr=22050, n_fft=1024, hop_length=256, n_mels=128, display=False)
np.save('/content/drive/MyDrive/DATA/training_data.npy', np.array(data))
elif prepared == "y":
data = np.load('/content/drive/MyDrive/DATA/training_data.npy', allow_pickle=True)
data = data.tolist()
else:
print("Invalid choice. Please select 'y' or 'n'.")
train_encoder(device, data, enc_dim=10, batch_size=1, num_epochs=15, visualize=True)
train_gan(device, data, enc_dim=10, batch_size=15, num_epochs=100, gen_iters=1, dis_iters=1, noise_dim=100, visualize=True)