训练需要8组数据,所以没有通用性
class Generator(nn.Module) :
def __init__(self, noise_dim, num_classes) :
super(Generator, self).__init__()
self.path_left_eye = LocalPathway()
self.path_right_eye = LocalPathway()
self.path_nose = LocalPathway()
self.path_mouth = LocalPathway()
self.globalpath = GlobalPathway(noise_dim)
self.fuser = LocalFuser()
self.feature_predict = FeaturePredict(num_classes)
def forward(self, img128, img64, img32, left_eye, right_eye, nose, mouth, noise):
tf平台:
开源地址:https://github.com/HRLTY/TP-GAN
TP GAN ;全名 TWO pathway Geanerator Network
TP-GAN 根据不同姿势合成的结果。从左到右:90°、75°、45°、60°、30°和 15°。最后一栏是真实相片。
。这两条路径,一条专注于推理全局结构,另一条则推理局部的纹理,分别得到两个特征地图。这两个特征图会融合在一起,用于接下来的最终合成。