-
Notifications
You must be signed in to change notification settings - Fork 21
/
train.py
181 lines (141 loc) · 6.21 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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
import argparse
import logging
import os
import time
import matplotlib.pyplot as plt
import numpy as np
import torch
from torch.utils.data import DataLoader
from nvae.dataset import ImageFolderDataset
from nvae.utils import add_sn
from nvae.vae_celeba import NVAE
class WarmupKLLoss:
def __init__(self, init_weights, steps,
M_N=0.005,
eta_M_N=1e-5,
M_N_decay_step=3000):
"""
预热KL损失，先对各级别的KL损失进行预热，预热完成后，对M_N的值进行衰减,所有衰减策略采用线性衰减
:param init_weights: 各级别 KL 损失的初始权重
:param steps: 各级别KL损失从初始权重增加到1所需的步数
:param M_N: 初始M_N值
:param eta_M_N: 最小M_N值
:param M_N_decay_step: 从初始M_N值到最小M_N值所需的衰减步数
"""
self.init_weights = init_weights
self.M_N = M_N
self.eta_M_N = eta_M_N
self.M_N_decay_step = M_N_decay_step
self.speeds = [(1. - w) / s for w, s in zip(init_weights, steps)]
self.steps = np.cumsum(steps)
self.stage = 0
self._ready_start_step = 0
self._ready_for_M_N = False
self._M_N_decay_speed = (self.M_N - self.eta_M_N) / self.M_N_decay_step
def _get_stage(self, step):
while True:
if self.stage > len(self.steps) - 1:
break
if step <= self.steps[self.stage]:
return self.stage
else:
self.stage += 1
return self.stage
def get_loss(self, step, losses):
loss = 0.
stage = self._get_stage(step)
for i, l in enumerate(losses):
# Update weights
if i == stage:
speed = self.speeds[stage]
t = step if stage == 0 else step - self.steps[stage - 1]
w = min(self.init_weights[i] + speed * t, 1.)
elif i < stage:
w = 1.
else:
w = self.init_weights[i]
# 如果所有级别的KL损失的预热都已完成
if self._ready_for_M_N == False and i == len(losses) - 1 and w == 1.:
# 准备M_N的衰减
self._ready_for_M_N = True
self._ready_start_step = step
l = losses[i] * w
loss += l
if self._ready_for_M_N:
M_N = max(self.M_N - self._M_N_decay_speed *
(step - self._ready_start_step), self.eta_M_N)
else:
M_N = self.M_N
return M_N * loss
if __name__ == '__main__':
LOG_FORMAT = "%(asctime)s - %(levelname)s - %(message)s"
logging.basicConfig(level=logging.INFO, format=LOG_FORMAT)
parser = argparse.ArgumentParser(description="Trainer for state AutoEncoder model.")
parser.add_argument("--epochs", type=int, default=500, help="number of epochs.")
parser.add_argument("--batch_size", type=int, default=128, help="size of each sample batch")
parser.add_argument("--dataset_path", type=str, required=True, help="dataset path")
parser.add_argument("--pretrained_weights", type=str, help="if specified starts from checkpoint model")
parser.add_argument("--n_cpu", type=int, default=16, help="number of cpu threads to use during batch generation")
opt = parser.parse_args()
epochs = opt.epochs
batch_size = opt.batch_size
dataset_path = opt.dataset_path
train_ds = ImageFolderDataset(dataset_path, img_dim=64)
train_dataloader = DataLoader(train_ds, batch_size=batch_size, shuffle=True, num_workers=opt.n_cpu)
os.makedirs("checkpoints", exist_ok=True)
os.makedirs("output", exist_ok=True)
device = "cuda:0" if torch.cuda.is_available() else "cpu"
model = NVAE(z_dim=512, img_dim=(64, 64))
# apply Spectral Normalization
model.apply(add_sn)
model.to(device)
if opt.pretrained_weights:
model.load_state_dict(torch.load(opt.pretrained_weights, map_location=device), strict=False)
warmup_kl = WarmupKLLoss(init_weights=[1., 1. / 2, 1. / 8],
steps=[4500, 3000, 1500],
M_N=opt.batch_size / len(train_ds),
eta_M_N=5e-6,
M_N_decay_step=36000)
print('M_N=', warmup_kl.M_N, 'ETA_M_N=', warmup_kl.eta_M_N)
optimizer = torch.optim.Adamax(model.parameters(), lr=0.01)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=15, eta_min=1e-4)
step = 0
for epoch in range(epochs):
model.train()
n_true = 0
total_size = 0
total_loss = 0
for i, image in enumerate(train_dataloader):
optimizer.zero_grad()
image = image.to(device)
image_recon, recon_loss, kl_losses = model(image)
kl_loss = warmup_kl.get_loss(step, kl_losses)
loss = recon_loss + kl_loss
log_str = "\r---- [Epoch %d/%d, Step %d/%d] loss: %.6f----" % (
epoch, epochs, i, len(train_dataloader), loss.item())
logging.info(log_str)
loss.backward()
optimizer.step()
step += 1
if step != 0 and step % 100 == 0:
with torch.no_grad():
z = torch.randn((1, 512, 2, 2)).to(device)
gen_img, _ = model.decoder(z)
gen_img = gen_img.permute(0, 2, 3, 1)
gen_img = gen_img[0].cpu().numpy() * 255
gen_img = gen_img.astype(np.uint8)
plt.imshow(gen_img)
# plt.savefig(f"output/ae_ckpt_%d_%.6f.png" % (epoch, total_loss))
plt.show()
scheduler.step()
torch.save(model.state_dict(), f"checkpoints/ae_ckpt_%d_%.6f.pth" % (epoch, loss.item()))
model.eval()
with torch.no_grad():
z = torch.randn((1, 512, 2, 2)).to(device)
gen_img, _ = model.decoder(z)
gen_img = gen_img.permute(0, 2, 3, 1)
gen_img = gen_img[0].cpu().numpy() * 255
gen_img = gen_img.astype(np.uint8)
plt.imshow(gen_img)
# plt.savefig(f"output/ae_ckpt_%d_%.6f.png" % (epoch, total_loss))
plt.show()