训练gan需要多少个假示例？
最近，我训练了一个带有约1000张真实图像和64张假图像的gan。这显然还不够虚假的图像，因此gan中的大多数“潜在空间”都创建了相同的图像。通常对gan进行多少个假图像，以使其潜在空间可用？
I recently trained a GAN with ~1000 real images and 64 fake images. This obviously wasn't enough fake images, so most of the "latent space" in the GAN just creates the same image. How many fake images are GANs usually trained with in order to make their latent space usable?
如果你对这篇内容有疑问，欢迎到本站社区发帖提问 参与讨论，获取更多帮助，或者扫码二维码加入 Web 技术交流群。
绑定邮箱获取回复消息
由于您还没有绑定你的真实邮箱，如果其他用户或者作者回复了您的评论，将不能在第一时间通知您！
发布评论
评论（1）
您生成的假图像的数量取决于您采样的噪声向量的大小。理想情况下，在训练的单个步骤中，您将采样dimension
batch_size x Noings_dim的噪声向量，然后使用生成器生成batch_sizeface>伪造图像。同样，您的歧视者会看到batch_size真实图像的数量。这样，在每个步骤中，您的歧视者都会看到相等数量的假和真实图像，而歧视者看到的假图像总数将等于训练集中的图像总数本身。
The number of fake images that you generate depends on the size of the noise vector that you sample. Ideally, in a single step of training, you would sample a noise vector of dimension
batch_size x noise_dimand use the generator to generatebatch_sizefake images. Similarly, your discriminator would see thebatch_sizenumber of real images.This way, at every step, your discriminator sees an equal number of fake and real images, and the total number of fake images seen by your discriminator would be equal to the total number of images in your training set itself.