def __init__(self, hp):
super(GPPatchMcResDis, self).__init__()
assert hp['n_res_blks'] % 2 == 0, 'n_res_blk must be multiples of 2'
self.n_layers = hp['n_res_blks'] // 2
nf = hp['nf']
cnn_f = [Conv2dBlock(3, nf, 7, 1, 3,
pad_type='reflect',
norm='none',
activation='none')]
for i in range(self.n_layers - 1):
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_f += [nn.ReflectionPad2d(1)]
cnn_f += [nn.AvgPool2d(kernel_size=3, stride=2)]
nf = np.min([nf * 2, 1024])
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_c = [Conv2dBlock(nf_out, hp['num_classes'], 1, 1,
norm='none',
activation='lrelu',
activation_first=True)]
self.cnn_f = nn.Sequential(*cnn_f)
self.cnn_c = nn.Sequential(*cnn_c)
def __init__(self, num_writers):
super(WriterClaModel, self).__init__()
self.n_layers = 6
nf = 16
cnn_f = [Conv2dBlock(1, nf, 7, 1, 3,
pad_type='reflect',
norm='none',
activation='none')]
for i in range(self.n_layers - 1):
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_f += [nn.ReflectionPad2d(1)]
cnn_f += [nn.AvgPool2d(kernel_size=3, stride=2)]
nf = np.min([nf * 2, 1024])
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_c = [Conv2dBlock(nf_out, num_writers, IMG_HEIGHT//(2**(self.n_layers-1)), IMG_WIDTH//(2**(self.n_layers-1))+1,
norm='none',
activation='lrelu',
activation_first=True)]
self.cnn_f = nn.Sequential(*cnn_f)
self.cnn_c = nn.Sequential(*cnn_c)
self.cross_entropy = nn.CrossEntropyLoss()
def __init__(self):
super(DisModel, self).__init__()
self.n_layers = 6
self.final_size = 1024
nf = 16
cnn_f = [Conv2dBlock(1, nf, 7, 1, 3,
pad_type='reflect',
norm='none',
activation='none')]
for i in range(self.n_layers - 1):
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_f += [nn.ReflectionPad2d(1)]
cnn_f += [nn.AvgPool2d(kernel_size=3, stride=2)]
nf = np.min([nf * 2, 1024])
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_c = [Conv2dBlock(nf_out, self.final_size, IMG_HEIGHT//(2**(self.n_layers-1)), IMG_WIDTH//(2**(self.n_layers-1))+1,
norm='none',
activation='lrelu',
activation_first=True)]
self.cnn_f = nn.Sequential(*cnn_f)
self.cnn_c = nn.Sequential(*cnn_c)
self.bce = nn.BCEWithLogitsLoss()
def __init__(self, hp):
super(GPPatchMcResDis, self).__init__()
#InceptionBlock = Conv2dBlock
assert hp['n_res_blks'] % 2 == 0, 'n_res_blk must be multiples of 2'
self.n_layers = hp['n_res_blks'] // 2
nf = hp['nf']
input_channels = hp['input_nc']
cnn_f = [
Conv2dBlock(input_channels,
nf,
KERNEL_SIZE_7,
1,
3,
pad_type='reflect',
norm='none',
activation='none')
]
for i in range(self.n_layers - 1):
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_f += [nn.ReflectionPad2d(1)]
cnn_f += [nn.AvgPool2d(kernel_size=3, stride=2)]
nf = np.min([nf * 2, 1024])
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_c = [
Conv2dBlock(nf_out,
hp['num_classes'],
1,
1,
norm='none',
activation='lrelu',
activation_first=True)
]
self.cnn_f = nn.Sequential(*cnn_f)
self.cnn_c = nn.Sequential(*cnn_c)
#self.register_backward_hook(debug.printgradnorm)
self.debug = Debugger()
def __init__(self, hp):
super(GPPatchMcResDis, self).__init__()
assert hp['n_res_blks'] % 2 == 0, 'n_res_blk must be multiples of 2'
self.n_layers = hp['n_res_blks'] // 2 # 5
nf = hp['nf'] # 64
cnn_f = [
Conv2dBlock(3,
nf,
7,
1,
3,
pad_type='reflect',
norm='none',
activation='none')
] # 第一个卷积层提取特征，输出batchsize*128*128*64维度向量
for i in range(
self.n_layers -
1): # 128->256->512->1024,输出batchsize*8*8*1024维度向量，在原论文条件下
nf_out = np.min([nf * 2, 1024])
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu', 'none')]
cnn_f += [nn.ReflectionPad2d(1)]
cnn_f += [nn.AvgPool2d(kernel_size=3, stride=2)]
nf = np.min([nf * 2, 1024])
nf_out = np.min([nf * 2, 1024]) # 1024
cnn_f += [ActFirstResBlock(nf, nf, None, 'lrelu', 'none')]
cnn_f += [ActFirstResBlock(nf, nf_out, None, 'lrelu',
'none')] # 再经过两个ResBlock,得到1024深度
cnn_c = [
Conv2dBlock(nf_out,
hp['num_classes'],
1,
1,
norm='none',
activation='lrelu',
activation_first=True)
] # 最后再经过一个卷积层得到各个种类的概率：输出batchsize*patch*patch*num_class的四维向量
self.cnn_f = nn.Sequential(*cnn_f)
self.cnn_c = nn.Sequential(*cnn_c)