def main():
# parse options
parser = TrainSketchOptions()
opts = parser.parse()
# create model
print('--- create model ---')
netSketch = SketchModule(opts.GB_nlayers, opts.DB_nlayers, opts.GB_nf, opts.DB_nf, opts.gpu)
if opts.gpu:
netSketch.cuda()
netSketch.init_networks(weights_init)
netSketch.train()
print('--- training ---')
for epoch in range(opts.epochs):
itr = 0
fnames = load_train_batchfnames(opts.text_path, opts.batchsize,
opts.text_datasize, trainnum=opts.Btraining_num)
fnames2 = load_train_batchfnames(opts.augment_text_path, opts.batchsize,
opts.augment_text_datasize, trainnum=opts.Btraining_num)
for ii in range(len(fnames)):
fnames[ii][0:opts.batchsize//2-1] = fnames2[ii][0:opts.batchsize//2-1]
for fname in fnames:
itr += 1
t = prepare_text_batch(fname, anglejitter=True)
t = to_var(t) if opts.gpu else t
losses = netSketch.one_pass(t, [l/4.-1. for l in range(0,9)])
print('Epoch [%d/%d][%03d/%03d]' %(epoch+1, opts.epochs,itr,len(fnames)), end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f'%(losses[0], losses[1], losses[2]))
print('--- save ---')
# directory
torch.save(netSketch.state_dict(), opts.save_GB_name)
def main():
# parse options
parser = TrainShapeMatchingOptions()
opts = parser.parse()
# create model
print('--- create model ---')
netShapeM = ShapeMatchingGAN(opts.GS_nlayers, opts.DS_nlayers, opts.GS_nf, opts.DS_nf,
opts.GT_nlayers, opts.DT_nlayers, opts.GT_nf, opts.DT_nf, opts.gpu)
netSketch = SketchModule(opts.GB_nlayers, opts.DB_nlayers, opts.GB_nf, opts.DB_nf, opts.gpu)
if opts.gpu:
netShapeM.cuda()
netSketch.cuda()
netShapeM.init_networks(weights_init)
netShapeM.train()
netSketch.load_state_dict(torch.load(opts.load_GB_name))
netSketch.eval()
print('--- training ---')
# load image pair
scales = [l*2.0/(opts.scale_num-1)-1 for l in range(opts.scale_num)]
Xl, X, _, Noise = load_style_image_pair(opts.style_name, scales, netSketch, opts.gpu)
Xl = [to_var(a) for a in Xl] if opts.gpu else Xl
X = to_var(X) if opts.gpu else X
Noise = to_var(Noise) if opts.gpu else Noise
for epoch in range(opts.step1_epochs):
for i in range(opts.Straining_num//opts.batchsize):
idx = opts.scale_num-1
xl, x = cropping_training_batches(Xl[idx], X, Noise, opts.batchsize,
opts.Sanglejitter, opts.subimg_size, opts.subimg_size)
losses = netShapeM.structure_one_pass(x, xl, scales[idx])
print('Step1, Epoch [%02d/%02d][%03d/%03d]' %(epoch+1, opts.step1_epochs, i+1,
opts.Straining_num//opts.batchsize), end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f'%(losses[0], losses[1], losses[2], losses[3]))
netShapeM.G_S.myCopy()
for epoch in range(opts.step2_epochs):
for i in range(opts.Straining_num//opts.batchsize):
idx = random.choice([0, opts.scale_num-1])
xl, x = cropping_training_batches(Xl[idx], X, Noise, opts.batchsize,
opts.Sanglejitter, opts.subimg_size, opts.subimg_size)
losses = netShapeM.structure_one_pass(x, xl, scales[idx])
print('Step2, Epoch [%02d/%02d][%03d/%03d]' %(epoch+1, opts.step2_epochs, i+1,
opts.Straining_num//opts.batchsize), end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f'%(losses[0], losses[1], losses[2], losses[3]))
for epoch in range(opts.step3_epochs):
for i in range(opts.Straining_num//opts.batchsize):
idx = random.choice(range(opts.scale_num))
xl, x = cropping_training_batches(Xl[idx], X, Noise, opts.batchsize,
opts.Sanglejitter, opts.subimg_size, opts.subimg_size)
losses = netShapeM.structure_one_pass(x, xl, scales[idx])
print('Step3, Epoch [%02d/%02d][%03d/%03d]' %(epoch+1, opts.step3_epochs, i+1,
opts.Straining_num//opts.batchsize), end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f'%(losses[0], losses[1], losses[2], losses[3]))
if opts.glyph_preserve:
fnames = load_train_batchfnames(opts.text_path, opts.batchsize,
opts.text_datasize, opts.Straining_num)
for epoch in range(opts.step4_epochs):
itr = 0
for fname in fnames:
itr += 1
t = prepare_text_batch(fname, anglejitter=False)
idx = random.choice(range(opts.scale_num))
xl, x = cropping_training_batches(Xl[idx], X, Noise, opts.batchsize,
opts.Sanglejitter, opts.subimg_size, opts.subimg_size)
t = to_var(x) if opts.gpu else t
losses = netShapeM.structure_one_pass(x, xl, scales[idx], t)
print('Step4, Epoch [%02d/%02d][%03d/%03d]' %(epoch+1, opts.step4_epochs, itr+1,
len(fnames)), end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f'%(losses[0], losses[1], losses[2], losses[3]))
print('--- save ---')
# directory
netShapeM.save_structure_model(opts.save_path, opts.save_name)
def main():
# parse options
parser = TrainShapeMatchingOptions()
opts = parser.parse()
# create model
print('--- create model ---')
# 6,4,32,32，6,4,32,32，False
netShape_M_GAN = ShapeMatchingGAN(opts.GS_nlayers, opts.DS_nlayers,
opts.GS_nf, opts.DS_nf, opts.GT_nlayers,
opts.DT_nlayers, opts.GT_nf, opts.DT_nf,
opts.gpu)
# 6, 5, 32, 32
netSketch = SketchModule(opts.GB_nlayers, opts.DB_nlayers, opts.GB_nf,
opts.DB_nf, opts.gpu)
if opts.gpu:
netShape_M_GAN.cuda()
netSketch.cuda()
netShape_M_GAN.init_networks(weights_init)
netShape_M_GAN.train()
netSketch.load_state_dict(torch.load(opts.load_GB_name))
netSketch.eval() # 不训练SketchModule
print('--- training ---')
# load image pair
scales = [
l * 2.0 / (opts.scale_num - 1) - 1 for l in range(opts.scale_num)
] # opts.scale_num默认值 4
# [-1.0, -0.33333333333333337, 0.33333333333333326, 1.0]
# 使用已训练的 netSketch！！！
Xl, X, _, Noise = load_style_image_pair(opts.style_name, scales, netSketch,
opts.gpu)
"""
Xl: 经过SketchModule的不同模糊程度的 4 个（scale_num默认值 4）距离图像 X
Xl[0] -- scales[0] -- -1.0
Xl[3] -- scales[3] -- 1.0
X：风格图像的距离图像 shape [1, 3, 740图像高度, 1000图像宽度]
Noise： 噪声？均值.0，方差.2，shape [1, 3, 740图像高度, 1000图像宽度]
"""
Xl = [to_var(a) for a in Xl] if opts.gpu else Xl
X = to_var(X) if opts.gpu else X
Noise = to_var(Noise) if opts.gpu else Noise
for epoch in range(opts.step1_epochs): # 默认 30
for i in range(opts.Straining_num //
opts.batchsize): # 2560 // 32 = 80
# 论文5.1 ：首先以固定的 l=1 训练G_S ，以学习最大变形。。。。。。。。
idx = opts.scale_num - 1 # 3
xl, x = cropping_training_batches(Xl[idx], X, Noise,
opts.batchsize,
opts.Sanglejitter,
opts.subimg_size,
opts.subimg_size)
# xl与x裁剪的坐标是相同的。
# xl是加入了一些噪声的自Xl[idx]随机裁剪出的 32 个 大小为 256x256 的xl图像 [32, 3, 256, 256]
# x就是输入的Output的随机裁剪/选择后的结果，也就是原距离图像随机裁剪/选择后的，与 xl shape 相同 [32, 3, 256, 256]
losses = netShape_M_GAN.structure_one_pass(x, xl, scales[idx])
print('Step1, Epoch [%02d/%02d][%03d/%03d]' %
(epoch + 1, opts.step1_epochs, i + 1,
opts.Straining_num // opts.batchsize),
end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f' %
(losses[0], losses[1], losses[2], losses[3]))
netShape_M_GAN.G_S.myCopy()
for epoch in range(opts.step2_epochs): # 40
for i in range(opts.Straining_num //
opts.batchsize): # 2560 // 32 = 80
idx = random.choice([0, opts.scale_num - 1]) # 0 或 3
xl, x = cropping_training_batches(Xl[idx], X, Noise,
opts.batchsize,
opts.Sanglejitter,
opts.subimg_size,
opts.subimg_size)
losses = netShape_M_GAN.structure_one_pass(x, xl, scales[idx])
print('Step2, Epoch [%02d/%02d][%03d/%03d]' %
(epoch + 1, opts.step2_epochs, i + 1,
opts.Straining_num // opts.batchsize),
end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f' %
(losses[0], losses[1], losses[2], losses[3]))
for epoch in range(opts.step3_epochs):
for i in range(opts.Straining_num // opts.batchsize):
idx = random.choice(range(opts.scale_num)) # 0,1,2,3
xl, x = cropping_training_batches(Xl[idx], X, Noise,
opts.batchsize,
opts.Sanglejitter,
opts.subimg_size,
opts.subimg_size)
losses = netShape_M_GAN.structure_one_pass(x, xl, scales[idx])
print('Step3, Epoch [%02d/%02d][%03d/%03d]' %
(epoch + 1, opts.step3_epochs, i + 1,
opts.Straining_num // opts.batchsize),
end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f' %
(losses[0], losses[1], losses[2], losses[3]))
# glyph_preserve 默认False,如果是True那么复杂结构字的论文效果会比不加更好吗？
if opts.glyph_preserve:
fnames = load_train_batchfnames(opts.text_path, opts.batchsize,
opts.text_datasize, opts.Straining_num)
for epoch in range(opts.step4_epochs):
itr = 0
for fname in fnames:
itr += 1
t = prepare_text_batch(fname, anglejitter=False)
idx = random.choice(range(opts.scale_num))
xl, x = cropping_training_batches(Xl[idx], X, Noise,
opts.batchsize,
opts.Sanglejitter,
opts.subimg_size,
opts.subimg_size)
t = to_var(x) if opts.gpu else t
losses = netShape_M_GAN.structure_one_pass(
x, xl, scales[idx], t)
print('Step4, Epoch [%02d/%02d][%03d/%03d]' %
(epoch + 1, opts.step4_epochs, itr + 1, len(fnames)),
end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lgly: %+.3f' %
(losses[0], losses[1], losses[2], losses[3]))
print('--- save ---')
# directory
netShape_M_GAN.save_structure_model(opts.save_path, opts.save_name)
def main():
# parse options
parser = TrainSketchOptions()
opts = parser.parse()
"""
opts.GB_nlayers = 6 生成器的层数
opts.DB_nlayers = 5 判别器的层数
opts.GB_nf = 32 生成器第一层中的特征数量
opts.DB_nf = 32 判别器第一层中的特征数量
opts.epochs = 3
"""
# create model
print('--- create model ---')
netSketch = SketchModule(opts.GB_nlayers, opts.DB_nlayers, opts.GB_nf,
opts.DB_nf, opts.gpu)
if opts.gpu:
netSketch.cuda()
netSketch.init_networks(weights_init)
netSketch.train()
# print(netSketch)
print('--- training ---')
for epoch in range(opts.epochs):
itr = 0
"""
text_path = ../data/rawtext/yaheiB/train
batchsize = 16
text_datasize = 708
Btraining_num = 12800
"""
# 按批次划分的文件名，没有的文件也不会报错，(trainnum // batch_size) X (batch_size) 矩阵
f_names = load_train_batchfnames(opts.text_path,
opts.batchsize,
opts.text_datasize,
trainnum=opts.Btraining_num)
"""
augment_text_path = ../data/rawtext/augment/
batchsize = 16
augment_text_datasize = 5
"""
f_names2 = load_train_batchfnames(opts.augment_text_path,
opts.batchsize,
opts.augment_text_datasize,
trainnum=opts.Btraining_num)
# fnames的每个数组里（即每个批次的）前 (opts.batchsize // 2 - 1) 为增强数据
for ii in range(len(f_names)):
f_names[ii][0:opts.batchsize // 2 -
1] = f_names2[ii][0:opts.batchsize // 2 - 1]
# print(f'fnames2[1]: {fnames2[0]}') # OK
for fname in f_names: # 每次循环是一个批次大小，即fname有batch_size个元素
itr += 1
t = prepare_text_batch(fname, anglejitter=True)
t = to_var(t) if opts.gpu else t
control_l = [e / 4. - 1. for e in range(0, 9)]
# t的shape为[batch_size, 3, 256, 256])
# control_l为[-1.0, -0.75, -0.5, -0.25, 0.0, 0.25, 0.5, 0.75, 1.0]
losses = netSketch.one_pass(t, control_l)
print('Epoch [%d/%d][%03d/%03d]' %
(epoch + 1, opts.epochs, itr, len(f_names)),
end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f' %
(losses[0], losses[1], losses[2]))
print('--- save ---')
# directory
torch.save(netSketch.state_dict(), opts.save_GB_name)
def main():
# parse options
parser = TrainShapeMatchingOptions()
opts = parser.parse()
# create model
print('--- create model ---')
netShapeM = ShapeMatchingGAN(opts.GS_nlayers, opts.DS_nlayers, opts.GS_nf, opts.DS_nf,
opts.GT_nlayers, opts.DT_nlayers, opts.GT_nf, opts.DT_nf, opts.gpu)
if opts.gpu:
netShapeM.cuda()
netShapeM.init_networks(weights_init)
netShapeM.train()
# 默认值是 False
if opts.style_loss:
netShapeM.G_S.load_state_dict(torch.load(opts.load_GS_name))
netShapeM.G_S.eval()
VGGNet = models.vgg19(pretrained=True).features
VGGfeatures = VGGFeature(VGGNet, opts.gpu)
for param in VGGfeatures.parameters():
param.requires_grad = False
if opts.gpu:
VGGfeatures.cuda()
style_targets = get_GRAM(opts.style_name, VGGfeatures, opts.batchsize, opts.gpu)
print('--- training ---')
# load image pair
_, X, Y, Noise = load_style_image_pair(opts.style_name, gpu=opts.gpu)
# X Y 显然大小相同，为 [1, 3, H ,W]
Y = to_var(Y) if opts.gpu else Y # 风格图
X = to_var(X) if opts.gpu else X # 风格距离图
Noise = to_var(Noise) if opts.gpu else Noise # 与X，Y形状相同
for epoch in range(opts.texture_step1_epochs): # 40
for i in range(opts.Ttraining_num // opts.batchsize): # 800 // 32 = 25
# x 风格距离图加上了红色Noise，y 风格图,
# shape为 [batchsize,6,256,256]
x, y = cropping_training_batches(X, Y, Noise, opts.batchsize,
opts.Tanglejitter, opts.subimg_size, opts.subimg_size)
losses = netShapeM.texture_one_pass(x, y)
print('Step1, Epoch [%02d/%02d][%03d/%03d]' % (epoch + 1, opts.texture_step1_epochs, i + 1,
opts.Ttraining_num // opts.batchsize), end=': ')
print('LDistance: %+.3f, LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lsty: %+.3f' % (losses[0], losses[1], losses[2], losses[3], losses[4]))
# 默认值是 False
if opts.style_loss:
fnames = load_train_batchfnames(opts.text_path, opts.batchsize,
opts.text_datasize, trainnum=opts.Ttraining_num)
for epoch in range(opts.texture_step2_epochs):
itr = 0
for fname in fnames:
itr += 1
t = prepare_text_batch(fname, anglejitter=False)
# x 风格距离图，y 风格图
x, y = cropping_training_batches(X, Y, Noise, opts.batchsize,
opts.Tanglejitter, opts.subimg_size, opts.subimg_size)
t = to_var(t) if opts.gpu else t
losses = netShapeM.texture_one_pass(x, y, t, 0, VGGfeatures, style_targets)
print('Step2, Epoch [%02d/%02d][%03d/%03d]' % (epoch + 1, opts.texture_step2_epochs,
itr, len(fnames)), end=': ')
print('LDadv: %+.3f, LGadv: %+.3f, Lrec: %+.3f, Lsty: %+.3f' % (
losses[0], losses[1], losses[2], losses[3]))
print('--- save ---')
# directory
netShapeM.save_texture_model(opts.save_path, opts.save_name)