Closed tianweiy closed 4 years ago
Hi, Sorry for the delay.
There isn't a pretained model for CIFAR yet. But I can have a look.
I've found a bug explaining partly the low score on CIFAR so far, after https://github.com/facebookresearch/pytorch_GAN_zoo/pull/72 you should easily get a 6.1 score. I'm working on the parameters to go up to 8.
Thanks. No hurry, I am using gan in TensorFlow in the meantime.
Thanks. No hurry, I am using gan in TensorFlow in the meantime.
How much time do you cost in training in cifar10 with 5.2 inception score ?
For the 5.2, I am trying to say that the loss diverges in the end so that current implementation is far from optimal. I suspect you can get a little higher inception score by training less time. It took about 2~3 days using 2 GPUs if I remember correctly.
God, it 's too much time for me
@Molugan Can the current version reproduce the results (fid=8.80) for cifar-10?
@Molugan Can the current version reproduce the results (fid=8.80) for cifar-10?
@Molugan Having same question here, any insight would be greatly appreciated!
Is there any pretrained model for cifar 10 dataset? I tried to train using the code here but the loss diverges in the end and I can only get 5.2 inception score.