def __init__(self, code_size, weight_init=None, weight_norm=None):
super(Generator, self).__init__()
self.inplane = 8
self.input_linear = nn.Linear(code_size, 4 * 4 * self.inplane * 32)
self.model = nn.Sequential(
*conv2d_trans_bn_lrelu_block(self.inplane * 32,
self.inplane * 16,
4,
2,
1,
normalize=True,
bias=False,
weight_norm=weight_norm),
*conv2d_trans_bn_lrelu_block(self.inplane * 16,
self.inplane * 8,
4,
2,
1,
normalize=True,
bias=False,
weight_norm=weight_norm),
*conv2d_trans_bn_lrelu_block(self.inplane * 8,
self.inplane * 4,
4,
2,
1,
normalize=True,
bias=False,
weight_norm=weight_norm),
apply_weight_norm(
nn.ConvTranspose2d(self.inplane * 4, 1, 3, 1, 1, bias=False),
weight_norm), nn.Tanh())
if weight_init:
self.apply(weight_init)
def __init__(self, code_size, weight_init=None, weight_norm=None):
super(DCGenerator, self).__init__()
self.label_embedding = nn.Embedding(10, 10)
self.input_linear = nn.Linear(code_size + 10, 512 * 4 * 4)
self.model = nn.Sequential(
*conv2d_trans_bn_relu_block(512,
256,
4,
2,
1,
normalize=True,
bias=True,
weight_norm=weight_norm),
*conv2d_trans_bn_relu_block(256,
128,
4,
2,
1,
normalize=True,
bias=True,
weight_norm=weight_norm),
*conv2d_trans_bn_relu_block(128,
64,
4,
2,
1,
normalize=True,
bias=True,
weight_norm=weight_norm),
apply_weight_norm(nn.ConvTranspose2d(64, 3, 3, 1, 1), weight_norm),
nn.Tanh())
if weight_init:
self.apply(weight_init)
def __init__(self, latent_lst, code_size, weight_init=None, weight_norm=None):
super(Generator, self).__init__()
self.inplane = 8
self.label_embedding = nn.ModuleList()
input_size = code_size
for latent in latent_lst:
if isinstance(latent, int):
self.label_embedding.append(nn.Embedding(latent, latent))
input_size += latent
else:
self.label_embedding.append(None)
input_size += 1
self.input_linear = nn.Linear(input_size, 4 * 4 * self.inplane * 32)
self.model = nn.Sequential(
*conv2d_trans_bn_relu_block(self.inplane * 32, self.inplane * 16, 4, 2, 1, normalize=True, bias=False,
weight_norm=weight_norm),
*conv2d_trans_bn_relu_block(self.inplane * 16, self.inplane * 8, 4, 2, 1, normalize=True, bias=False,
weight_norm=weight_norm),
*conv2d_trans_bn_relu_block(self.inplane * 8, self.inplane * 4, 4, 2, 1, normalize=True, bias=False,
weight_norm=weight_norm),
apply_weight_norm(nn.ConvTranspose2d(self.inplane * 4, 1, 3, 1, 1, bias=False), weight_norm),
nn.Tanh()
)
if weight_init:
self.apply(weight_init)
def conv1d_bn_lrelu_block(in_channels, out_channels, kernel_size, stride, padding, alpha=0.2, normalize=True, bias=True,
weight_norm=None):
conv = apply_weight_norm(nn.Conv1d(in_channels, out_channels, kernel_size, stride, padding, bias=bias), weight_norm)
layers = [conv]
if normalize:
layers.append(nn.BatchNorm1d(out_channels))
layers.append(nn.LeakyReLU(alpha, inplace=True))
return layers
def conv2d_bn_relu_block(in_channels, out_channels, kernel_size, stride, padding, normalize=True, bias=True,
weight_norm=None):
""" conv2d + batchnorm (optional) + relu """
conv = apply_weight_norm(nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, bias=bias), weight_norm)
layers = [conv]
if normalize:
layers.append(nn.BatchNorm2d(out_channels))
layers.append(nn.ReLU(inplace=True))
return layers
