search

PaddlePaddle / PaddleGAN

PaddlePaddle GAN library, including lots of interesting applications like First-Order motion transfer, Wav2Lip, picture repair, image editing, photo2cartoon, image style transfer, GPEN, and so on.
Apache License 2.0
7.75k stars 1.23k forks source link

stylegan_v2_256是否可以训练生成512或者1024尺寸的图片，如何配置yaml? #837

Open zhw511006 opened 4 months ago

zhw511006 commented 4 months ago

在config文件夹里面找到了stylegan_v2_256_ffhq.yaml配置， model: name: StyleGAN2Model generator: name: StyleGANv2Generator size: 256 style_dim: 512 n_mlp: 8 discriminator: name: StyleGANv2Discriminator size: 256

这样训练完的模型可以生成尺寸为256的图片，简单修改其中size为512或者1024，进行训练的时候直接报错，如下：

ValueError: (InvalidArgument) Input(Y) has error dim.Y'dims[0] must be equal to 12800But received Y'dims[0] is 8192 [Hint: Expected y_dims[y_ndim - 2] == K, but received y_dims[y_ndim - 2]:8192 != K:12800.] (at /paddle/paddle/phi/kernels/impl/matmul_kernel_impl.h:315) [operator < matmul_v2 > error]

请问如何修改yaml配置可以生成512或者1024尺寸的图片，谢谢！

jerrywgz commented 3 months ago

在组网阶段一些输入shape也需要检查下

1196228233 commented 4 weeks ago

有没有一种可能把大图切成多个256小图，生成后再拼接回去。就是不知道会不会有接痕？