znxlwm / pytorch-mnist-celeba-cgan-cdcgan

Pytorch implementation of conditional Generative Adversarial Networks (cGAN) and conditional Deep Convolutional Generative Adversarial Networks (cDCGAN) for MNIST dataset

pytorch-MNIST-CelebA-cGAN-cDCGAN

Pytorch implementation of conditional Generative Adversarial Networks (cGAN) [1] and conditional Generative Adversarial Networks (cDCGAN) for MNIST [2] and CelebA [3] datasets.

• The network architecture (number of layer, layer size and activation function etc.) of this code differs from the paper.

• CelebA dataset used gender lable as condition.

• If you want to train using cropped CelebA dataset, you have to change isCrop = False to isCrop = True.

• you can download

  • MNIST dataset: http://yann.lecun.com/exdb/mnist/
  • CelebA dataset: http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Implementation details

• cGAN

• cDCGAN

Resutls

MNIST

• Generate using fixed noise (fixed_z_)

cGAN	cDCGAN

• MNIST vs Generated images

MNIST	cGAN after 50 epochs	cDCGAN after 20 epochs

• Learning Time
  • MNIST cGAN - Avg. per epoch: 9.13 sec; Total 50 epochs: 937.06 sec
  • MNIST cDCGAN - Avg. per epoch: 47.16 sec; Total 20 epochs: 1024.26 sec

CelebA

• Generate using fixed noise (fixed_z_; odd line - female (y: 0) & even line - male (y: 1); each two lines have the same style (1-2) & (3-4).)

cDCGAN	cDCGAN crop

• CelebA vs Generated images

CelebA	cDCGAN after 20 epochs	cDCGAN crop after 30 epochs

• CelebA cDCGAN morphing (noise interpolation)

cDCGAN	cDCGAN crop

• Learning Time
  • CelebA cDCGAN - Avg. per epoch: 826.69 sec; total 20 epochs ptime: 16564.10 sec

Development Environment

• Ubuntu 14.04 LTS
• NVIDIA GTX 1080 ti
• cuda 8.0
• Python 2.7.6
• pytorch 0.1.12
• torchvision 0.1.8
• matplotlib 1.3.1
• imageio 2.2.0

Reference

[1] Mirza, Mehdi, and Simon Osindero. "Conditional generative adversarial nets." arXiv preprint arXiv:1411.1784 (2014).

(Full paper: https://arxiv.org/pdf/1411.1784.pdf)

[2] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. "Gradient-based learning applied to document recognition." Proceedings of the IEEE, 86(11):2278-2324, November 1998.

[3] Liu, Ziwei, et al. "Deep learning face attributes in the wild." Proceedings of the IEEE International Conference on Computer Vision. 2015.

pytorch-mnist-celeba-cgan-cdcgan's Issues

niu bi

代码直接cuda，不考虑cpu的情况。而且还一堆错误，这种代码为啥

Question about mnist-cDCGAN

I notice that you use "y_ = (torch.rand(mini_batch, 1) * 10).type(torch.LongTensor).squeeze()" in line 235,why not use the real y_ from input image?

LeakyReLU before Sigmoid

pytorch-MNIST-CelebA-cGAN-cDCGAN/pytorch_MNIST_cDCGAN.py

Line 69 in bc8be78

x = F.leaky_relu(self.conv3_bn(self.conv3(x)), 0.2)

I see in so much other codes what uses only Conv2D and Sigmod, but never a Conv2D, LeakyReLU and Sigmoid.

miss of files

There are no files named "Fixed_results", it seems that some files are missed

ERROR

ValueError: Using a target size (torch.Size([128])) that is different to the input size (torch.Size([128, 6, 6])) is deprecated. Please ensure they have the same size.

Anyone able to get this working on collab?

Would much appreciate a link if so, been trying for days.

the loss of Generator didn't decrease

I only changed one place
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
to transforms.Normalize((0.1307,), (0.3081,))
because it causes error.
following is trainning process:
[1/50] - ptime: 29.42, loss_d: 0.855, loss_g: 1.262
[2/50] - ptime: 29.97, loss_d: 0.288, loss_g: 2.511
[3/50] - ptime: 31.54, loss_d: 0.090, loss_g: 3.608
[4/50] - ptime: 32.44, loss_d: 0.086, loss_g: 4.054
[5/50] - ptime: 34.58, loss_d: 0.070, loss_g: 4.227
[6/50] - ptime: 32.97, loss_d: 0.033, loss_g: 4.767
[7/50] - ptime: 33.55, loss_d: 0.014, loss_g: 5.374
[8/50] - ptime: 34.24, loss_d: 0.029, loss_g: 5.416
[9/50] - ptime: 34.68, loss_d: 0.008, loss_g: 5.843
[10/50] - ptime: 35.56, loss_d: 0.016, loss_g: 5.878
[11/50] - ptime: 38.06, loss_d: 0.006, loss_g: 6.242
[12/50] - ptime: 34.45, loss_d: 0.026, loss_g: 6.199
[13/50] - ptime: 32.97, loss_d: 0.005, loss_g: 6.561
[14/50] - ptime: 32.59, loss_d: 0.015, loss_g: 6.542
[15/50] - ptime: 32.66, loss_d: 0.002, loss_g: 6.981
[16/50] - ptime: 32.96, loss_d: 0.025, loss_g: 6.400
[17/50] - ptime: 33.44, loss_d: 0.019, loss_g: 6.936
[18/50] - ptime: 33.80, loss_d: 0.008, loss_g: 6.489
[19/50] - ptime: 34.30, loss_d: 0.002, loss_g: 7.097
[20/50] - ptime: 34.98, loss_d: 0.040, loss_g: 6.319
[21/50] - ptime: 34.86, loss_d: 0.003, loss_g: 6.906
[22/50] - ptime: 34.65, loss_d: 0.003, loss_g: 7.332
[23/50] - ptime: 34.74, loss_d: 0.008, loss_g: 7.447
[24/50] - ptime: 34.87, loss_d: 0.003, loss_g: 7.552
[25/50] - ptime: 34.79, loss_d: 0.001, loss_g: 7.913
[26/50] - ptime: 35.62, loss_d: 0.009, loss_g: 7.999
[27/50] - ptime: 37.91, loss_d: 0.006, loss_g: 7.409
[28/50] - ptime: 33.79, loss_d: 0.001, loss_g: 8.188
[29/50] - ptime: 32.57, loss_d: 0.001, loss_g: 8.127
learning rate change!
[30/50] - ptime: 32.99, loss_d: 0.005, loss_g: 7.739
[31/50] - ptime: 33.69, loss_d: 0.001, loss_g: 7.975
[32/50] - ptime: 34.34, loss_d: 0.001, loss_g: 8.114
[33/50] - ptime: 35.43, loss_d: 0.001, loss_g: 8.140
[34/50] - ptime: 36.85, loss_d: 0.002, loss_g: 7.763
[35/50] - ptime: 40.14, loss_d: 0.001, loss_g: 8.218
[36/50] - ptime: 44.25, loss_d: 0.001, loss_g: 8.467
[37/50] - ptime: 45.44, loss_d: 0.001, loss_g: 8.708
[38/50] - ptime: 40.60, loss_d: 0.001, loss_g: 8.666
[39/50] - ptime: 40.33, loss_d: 0.001, loss_g: 8.824
learning rate change!
[40/50] - ptime: 39.38, loss_d: 0.000, loss_g: 8.775
[41/50] - ptime: 37.45, loss_d: 0.000, loss_g: 8.842
[42/50] - ptime: 37.45, loss_d: 0.001, loss_g: 8.971
[43/50] - ptime: 38.80, loss_d: 0.001, loss_g: 8.718
[44/50] - ptime: 38.23, loss_d: 0.000, loss_g: 8.672
[45/50] - ptime: 40.54, loss_d: 0.000, loss_g: 8.708
[46/50] - ptime: 36.43, loss_d: 0.000, loss_g: 8.707
[47/50] - ptime: 34.01, loss_d: 0.000, loss_g: 8.800
[48/50] - ptime: 36.30, loss_d: 0.000, loss_g: 8.910
[49/50] - ptime: 36.61, loss_d: 0.000, loss_g: 8.990
[50/50] - ptime: 35.31, loss_d: 0.000, loss_g: 9.075
Avg one epoch ptime: 35.55, total 50 epochs ptime: 1881.31
Training finish!... save training results

[solved] wrong MNIST format

error:
RuntimeError: output with shape [1, 28, 28] doesn't match the broadcast shape [3, 28, 28]

MNIST has [1, 28, 28] and [3. 28. 28] two format. In this code,
change
transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
to
transforms.Normalize([0.5], [0.5])

how to plot the figure about the cGAN and cDCGan you show?

@znxlwm good job you did. And also I think the figure about the framework of cGAN and cDCGan looks very clean and beautiful, what's the software you use? thanks.

How to pass custom dataset?

Can Anybody please help how to pass custom datasets instead of MNIST?

train_loader = torch.utils.data.DataLoader( datasets.MNIST('data', train=True, download=False, transform=transform), batch_size=batch_size, shuffle=True)

Discriminator Over-fitting Issue

Hello,

Thanks for sharing your implementation of cDCGAN.

I have tried implementing it for MNIST data set following the same steps you did in 'pytorch_MNIST_cDCGAN.py' with same parameter settings except for the batch size (I used 32 instead of 128). Unfortunately, I get unstable results for epochs larger than 5.

Based on your implementation, I cannot find something like label smoothing or arbitrary Gaussian noise addition to the discriminator input images in order to fix the over fitting issues as described here. Have you already used such stabilizing tricks and I couldn't observe in your implementation ?

With many thanks in advance

Best