-
Notifications
You must be signed in to change notification settings - Fork 203
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Request for conditional CIFAR10 codes #40
Comments
What does "conditional SNGAN results" refer to? Is it the scores shown in Table 3 of https://arxiv.org/pdf/1802.05637.pdf? |
@takerum Oh, yes. I should have indicated that table in the first place. |
The hyper-parameters are the same as the ones used in the ImageNet experiments, except that:
The followings are the models used in that paper. Discriminator:
Generator:
|
@takerum Thank you! I will try ASAP. |
I added some changes to the code written by you @mbsariyildiz and I hope you can now reproduce the results. |
@takerum, I saw the changes and run an experiment right away 😄 I got same results. Now I can focus on "why Pytorch version does not reproduce the results". |
Great, and thanks again for your contribution! |
@takerum, I tried to reproduce unconditional SNGAN results on the paper with the following config file: batchsize: 64
iteration: 50000
iteration_decay_start: 0
seed: 0
display_interval: 100
progressbar_interval: 100
snapshot_interval: 10000
evaluation_interval: 1000
models:
generator:
fn: gen_models/resnet_32.py
name: ResNetGenerator
args:
dim_z: 128
bottom_width: 4
ch: 256
n_classes: 1
discriminator:
fn: dis_models/snresnet_32.py
name: SNResNetProjectionDiscriminator
args:
ch: 128
n_classes: 1
dataset:
dataset_fn: datasets/cifar10.py
dataset_name: CIFAR10Dataset
args:
test: False
adam:
alpha: 0.0002
beta1: 0.5
beta2: 0.999
updater:
fn: updater.py
name: Updater
args:
n_dis: 1
n_gen_samples: 128
conditional: False
loss_type: hinge Everything else was same as in the conditional GAN setup. However, IS oscillates between 4.3 and 4.6. Do you see what I am doing wrong here? |
Which results do you want to reproduce? If it's the score with standard CNNs (7.42 in the paper), you should use different architectures from those specified in the config files. If the score with ResNet, you can reproduce the results with the same config file as https://github.com/pfnet-research/sngan_projection/blob/master/configs/sn_cifar10.yml, but please change parameters not to use label supervision. I will make the config for the latter one, just moment |
I made and uploaded the code just now, please check it. |
Thank you. I'm running an experiment now. |
With the code that you provided above, I do get an Inception score around 8.7 as you mentioned but I get the intra-FID score to be around 23. Is there something I am missing ? @takerum Thank you for your help! |
sorry for the late reply. |
Thanks a lot.
By any chance do you have the saved models for CIFAR-10, CIFAR-100 ? @takerum
|
@takerum In table 8 of the paper, (where the result of 8.6 IS on CIFAR10 is obtained) it says that the model used a gradient penalty if I am reading correctly. However I can't find where this is done in the code — could you clarify whether this result indeed includes gradient penalty and if so, where this computation happens in this repo? Thank you so much for your time! Edit: To add to this, can anyone show how the gradient penalty computation works for the class-conditional GAN? I have not been able to find a good resource on this |
@amartya18x I get the intra-FID score 23 like you, but my conventional FID is 14.5. What's your conventional FID? |
Hello! I have been really trying hard to reproduce conditional SNGAN results on CIFAR10, recently. But I am far from the ones stated in the paper. I started coding in Pytorch, but somehow couldn't manage to achieve IS score above 7. While looking around, I saw you mentioning this link in anaother issue, which reproduces DCGAN results (with standard CNNs) on CIFAR10. So I basically replaced conv2ds in common/net/DownResBlockXs with snconv2ds, and linear layer with snlinear layer in common/net/ResnetDiscriminator. I also modified wgan_gp updater to optimize hinge loss and used the optimal hyperparameters. But neither dcgan nor the updated wgan_gp (with lam=0.) achieved relatively competitive results. Could you please provide the scripts that reproduce conditional SNGAN results?
The text was updated successfully, but these errors were encountered: