Hi, I am using these codes on my own project and I got a few confusions since I am new on cGANs.
the first one is how to set a suitable batch_size. Others said that batch_size=1 led to better perfomance. However, my training set is a bit larger (about 42k images, and another 8k images for test and validation). Should the batch_size kept 1 or setting a larger value, such as 128? For the first attemption, I used batch_size = 1 but the gen_loss increased fast within one epoch. However, when I set batch_size to 128, the gen_loss slowed down.
so, here is the second question, when to stop training. I learned from some blogs that an ideal stopping condition was that the discriminator could not tell real and generated images, which means predict_fake = predict_real = 0.5. If I understood correctly, the gen_loss (-log(predict_fake)) should be about ~0.7, while the discri_loss ( -(log(predict_fake) + log(predict_real))) should be ~1.4. Is this the correct condition to stop training?
Hi, I am using these codes on my own project and I got a few confusions since I am new on cGANs.
the first one is how to set a suitable batch_size. Others said that batch_size=1 led to better perfomance. However, my training set is a bit larger (about 42k images, and another 8k images for test and validation). Should the batch_size kept 1 or setting a larger value, such as 128? For the first attemption, I used batch_size = 1 but the gen_loss increased fast within one epoch. However, when I set batch_size to 128, the gen_loss slowed down.
so, here is the second question, when to stop training. I learned from some blogs that an ideal stopping condition was that the discriminator could not tell real and generated images, which means predict_fake = predict_real = 0.5. If I understood correctly, the gen_loss (-log(predict_fake)) should be about ~0.7, while the discri_loss ( -(log(predict_fake) + log(predict_real))) should be ~1.4. Is this the correct condition to stop training?
Hoping for your expert explanations. Thanks a lot