def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
# self.adain1 = ADAIN_NORM(ngf * 2)
self.adain2 = ADAIN_NORM(ngf)
self.residual = self._make_layer(ResBlock, ngf * 2)
self.upsample = upBlock(ngf * 2, ngf)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
self.jointConv = Block3x3_relu(ngf + 100, ngf) # FIXME del
self.residual = self._make_layer(ResBlock, ngf) # FIXME ngf * 2
# self.upsample = upBlock(ngf * 2, ngf) # FIXME
self.upsample = upBlock(ngf , ngf)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
self.att_sent = ATT_NET_sent(ngf, self.ef_dim)
self.residual1 = ResBlock(ngf)
self.residual2 = ResBlock(ngf)
self.residual3 = ResBlock(ngf)
self.upsample = upBlock(ngf, ngf)
def define_module(self):
ngf = self.gf_dim
nef = self.ef_dim
nef2 = self.ef_dim2
self.att = ATT_NET(ngf, nef)
self.bt_att = BT_ATT_NET(ngf, nef)
self.residual = self._make_layer(HmapResBlock, ngf*3+nef2)
self.upsample = upBlock(ngf*3+nef2, ngf)
def define_module(self):
self.att = ATT_NET(self.gen_feat_dim, self.text_emb_dim)
self.residual = self._make_layer(ResBlock, self.gen_feat_dim * 2)
self.upsample = upBlock(self.gen_feat_dim * 3, self.gen_feat_dim)
# local pathway
label_input_dim = cfg.GAN.TEXT_CONDITION_DIM + cfg.TEXT.CLASSES_NUM # no noise anymore
self.label = nn.Sequential(
nn.Linear(label_input_dim, self.label_dim, bias=False),
nn.BatchNorm1d(self.label_dim), nn.ReLU(True))
self.local1 = upBlock(self.label_dim + self.gen_feat_dim,
self.gen_feat_dim * 2)
self.local2 = upBlock(self.gen_feat_dim * 2, self.gen_feat_dim)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
self.residual = self._make_layer(ResBlock, ngf * 2)
self.upsample = upBlock(ngf * 3, ngf)
# local pathway
linput = cfg.GAN.Z_DIM + 81
self.label = nn.Sequential(
nn.Linear(linput, self.ef_dim // 2, bias=False),
nn.BatchNorm1d(self.ef_dim // 2), nn.ReLU(True))
self.local1 = upBlock(self.ef_dim // 2 + ngf, ngf * 2)
self.local2 = upBlock(ngf * 2, ngf)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
self.residual = self._make_layer(ResBlock, ngf * 2)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
# GlobalAttentionGeneral模块，作为输入
self.residual = self._make_layer(ResBlock, ngf * 2) # 两块残差块
self.upsample = upBlock(ngf * 2, ngf) # 上采样
def __init__(self):
super(generator, self).__init__()
self.image_size = 64
self.num_channels = 3
self.noise_dim = 100
self.embed_dim = 2400 # compatible with skip thought (1024)
self.projected_embed_dim = 128
self.latent_dim = self.noise_dim + self.projected_embed_dim
self.ngf = 64
self.projection = nn.Sequential(
nn.Linear(in_features=self.embed_dim,
out_features=self.projected_embed_dim),
nn.BatchNorm1d(num_features=self.projected_embed_dim),
nn.LeakyReLU(negative_slope=0.2, inplace=True))
self.att = ATT_NET(self.ngf, self.ef_dim)
# based on: https://github.com/pytorch/examples/blob/master/dcgan/main.py
self.netG = nn.Sequential(
nn.ConvTranspose2d(self.latent_dim,
self.ngf * 8,
4,
1,
0,
bias=True),
nn.BatchNorm2d(self.ngf * 8),
nn.ReLU(True),
# adding extra convs will give output (ngf*8) x 4 x 4
nn.Conv2d(self.ngf * 8, self.ngf * 2, 1, 1, 0),
nn.BatchNorm2d(self.ngf * 2),
nn.ReLU(True),
nn.Conv2d(self.ngf * 2, self.ngf * 2, 3, 1, 1),
nn.BatchNorm2d(self.ngf * 2),
nn.ReLU(True),
nn.Conv2d(self.ngf * 2, self.ngf * 8, 3, 1, 1),
nn.BatchNorm2d(self.ngf * 8),
nn.ReLU(True),
# state size. (ngf*8) x 4 x 4
nn.ConvTranspose2d(self.ngf * 8, self.ngf * 4, 4, 2, 1, bias=True),
nn.BatchNorm2d(self.ngf * 4),
nn.ReLU(True),
# adding extra convs will give output (ngf*4) x 4 x 4
nn.Conv2d(self.ngf * 4, self.ngf, 1, 1, 0),
nn.BatchNorm2d(self.ngf),
nn.ReLU(True),
nn.Conv2d(self.ngf, self.ngf, 3, 1, 1),
nn.BatchNorm2d(self.ngf),
nn.ReLU(True),
nn.Conv2d(self.ngf, self.ngf * 4, 3, 1, 1),
nn.BatchNorm2d(self.ngf * 4),
nn.ReLU(True),
# state size. (ngf*4) x 8 x 8
nn.ConvTranspose2d(self.ngf * 4, self.ngf * 2, 4, 2, 1, bias=True),
nn.BatchNorm2d(self.ngf * 2),
nn.ReLU(True),
# state size. (ngf*2) x 16 x 16
nn.ConvTranspose2d(self.ngf * 2, self.ngf, 4, 2, 1, bias=True),
nn.BatchNorm2d(self.ngf),
nn.ReLU(True),
# state size. (ngf) x 32 x 32
nn.ConvTranspose2d(self.ngf, self.num_channels, 4, 2, 1,
bias=True),
nn.Tanh()
# state size. (num_channels) x 64 x 64
)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim, nhw=self.hw_size)
ngf_counts = 3
self.residual = self._make_layer(ResBlock, ngf * ngf_counts)
self.upsample = upBlock(ngf * ngf_counts, ngf)
def define_module(self):
ngf = self.gf_dim
self.att = ATT_NET(ngf, self.ef_dim)
self.ca_net = CA_NET()
self.residual = self._make_layer(ResBlock, ngf * 2)
self.upsample = upBlock(ngf * 2, ngf)
