sleepychord/improvedgan-pytorch

Semi-supervised GAN in "Improved Techniques for Training GANs"

improvedgan-pytorch's Issues

Relation with bad GANs

Hello
You mentioned this code is inspired from the paper called "Good Semi-supervised Learning
That Requires a Bad GAN".

Have you changed any part of "Improved Techniques for Training GANs" to close to bad GANs paper?
What are the exact changes?
It is not clear for me.

Thank you

accuracy?

98.5% is train acc or test acc?

AttributeError: 'BatchNorm1d' object has no attribute 'track_running_stats'

I am getting this error:

File "/home/elhamod/melhamodenv/AML/project2/ImprovedGAN-pytorch-master/Nets.py", line 80, in forward
x = F.softplus(self.bn1(self.fc1(x)) + self.bn1_b)
File "/home/elhamod/melhamodenv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 541, in call
result = self.forward(*input, **kwargs)
File "/home/elhamod/melhamodenv/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py", line 69, in forward
if self.training and self.track_running_stats:
File "/home/elhamod/melhamodenv/lib/python3.6/site-packages/torch/nn/modules/module.py", line 585, in getattr
type(self).name, name))
AttributeError: 'BatchNorm1d' object has no attribute 'track_running_stats'

Any ideas how to fix it?

i don't understand code: torch.mean(F.softplus(logz_unlabel) why add F.softplus activation?

hello, the question about the accuracy on the test data by using default configs

when I run the command, python ImprovedGAN.py --cuda, and set the 100 epochs, the accuracy on test data is only around 65%. that is not same with your mentened 98.5%.

change input data

Hello
I tried to feed cifar10, I change the input size to 3072, however in run time I still get
(0): LinearWeightNorm(in_features=784, out_features=1000, weight_scale=1)
(1): LinearWeightNorm(in_features=1000, out_features=500, weight_scale=1)
(2): LinearWeightNorm(in_features=500, out_features=250, weight_scale=1)
(3): LinearWeightNorm(in_features=250, out_features=250, weight_scale=1)
(4): LinearWeightNorm(in_features=250, out_features=250, weight_scale=1)

for first layer it is still 784

About 'self.labeled.data_tensor.clone()' in ImprovedGAN.py

ImprovedGAN-pytorch/ImprovedGAN.py

Line 77 in bc14e80

t1 = self.labeled.data_tensor.clone()

when i run ImprovedGAN.py. An error is raised:

AttributeError: 'TensorDataset' object has no attribute 'data_tensor'

how can i solve this probleam?

the unlabeled samples include the labeled samples

Dear author, thanks for sharing the improved GAN in pytorch. In readme.md, you said: Default configs can train models achieving 98.5% accuracy on test dataset with 100 labeled data(10 per class) and other 59,000 unlabeled data after 100 epochs.
However, according to your implementation, the unlabeled samples are selected as the full train samples of MNIST. It seems they are inconsistent. Which is correct?

hello. When the loss_unsupervised is nearly equal with 0, the loss_supervised is still pretty large, how should I do to get better result?

when I run this code, the partial results are as follows.
Iteration 8, loss_supervised = 0.5682, loss_unsupervised = 0.0012, loss_gen = 1.8700 train acc = 0.7541
Eval: correct 6277/ 10000
.......
Iteration 34, loss_supervised = 0.8746, loss_unsupervised = 0.0009, loss_gen = 8.1296 train acc = 0.5644
Eval: correct 8050/ 1800
.......
Iteration 48, loss_supervised = 0.7172, loss_unsupervised = 0.0002, loss_gen = 7.7574 train acc = 0.6869
Eval: correct 8056/ 1800
........
then the results have few change.