├── Stackelberg GAN
├── images
│ ├── exp1.png
│ ├── exp2.png
│ ├── exp3.png
│ ├── mnist.png
│ ├── converge.gif
│ ├── architecture.png
│ ├── fashion_mnist.png
│ └── cifar10_imagenet.png
├── CIFAR-10
│ ├── utils.py
│ ├── ops.py
│ ├── main.py
│ └── models.py
├── Gaussian mixture
│ ├── gan_branch_mG.py
│ └── gan_stackelberg_mG.py
├── fashion_MNIST
│ └── gan_mnist_fashion_classifier.py
└── MNIST
│ └── gan_mnist_classifier.py
├── README.md
└── LICENSE
/Stackelberg GAN/images/exp1.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/exp1.png
--------------------------------------------------------------------------------
/Stackelberg GAN/images/exp2.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/exp2.png
--------------------------------------------------------------------------------
/Stackelberg GAN/images/exp3.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/exp3.png
--------------------------------------------------------------------------------
/Stackelberg GAN/images/mnist.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/mnist.png
--------------------------------------------------------------------------------
/Stackelberg GAN/images/converge.gif:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/converge.gif
--------------------------------------------------------------------------------
/Stackelberg GAN/images/architecture.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/architecture.png
--------------------------------------------------------------------------------
/Stackelberg GAN/images/fashion_mnist.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/fashion_mnist.png
--------------------------------------------------------------------------------
/Stackelberg GAN/images/cifar10_imagenet.png:
--------------------------------------------------------------------------------
https://raw.githubusercontent.com/hongyanz/Stackelberg-GAN/HEAD/Stackelberg GAN/images/cifar10_imagenet.png
--------------------------------------------------------------------------------
/Stackelberg GAN/CIFAR-10/utils.py:
--------------------------------------------------------------------------------
1 | from __future__ import division
2 | from __future__ import print_function
3 | from __future__ import absolute_import
4 |
5 | import numpy as np
6 | import math
7 | import matplotlib
8 | matplotlib.use('TkAgg')
9 | import matplotlib.pyplot as plt
10 | plt.switch_backend('agg')
11 | plt.style.use('ggplot')
12 |
13 | class Prior(object):
14 | def __init__(self, type):
15 | self.type = type
16 |
17 | def sample(self, shape):
18 | if self.type == "uniform":
19 | return np.random.uniform(-1.0, 1.0, shape)
20 | else:
21 | return np.random.normal(0, 1, shape)
22 |
23 | def conv_out_size_same(size, stride):
24 | return int(math.ceil(float(size) / float(stride)))
25 |
26 | def make_batches(size, batch_size):
27 | '''Returns a list of batch indices (tuples of indices).
28 | '''
29 | return [(i, min(size, i + batch_size)) for i in range(0, size, batch_size)]
30 |
31 | def create_image_grid(x, img_size, tile_shape):
32 | assert (x.shape[0] == tile_shape[0] * tile_shape[1])
33 | assert (x[0].shape == img_size)
34 |
35 | img = np.zeros((img_size[0] * tile_shape[0] + tile_shape[0] - 1,
36 | img_size[1] * tile_shape[1] + tile_shape[1] - 1,
37 | 3))
38 |
39 | for t in range(x.shape[0]):
40 | i, j = t // tile_shape[1], t % tile_shape[1]
41 | img[i * img_size[0] + i : (i + 1) * img_size[0] + i, j * img_size[1] + j : (j + 1) * img_size[1] + j] = x[t]
42 |
43 | return img
44 |
45 | def merge(images, size):
46 | h, w = images.shape[1], images.shape[2]
47 | img = np.zeros((h * size[0], w * size[1], 3))
48 |
49 | for idx, image in enumerate(images):
50 | i = idx % size[1]
51 | j = idx // size[1]
52 | img[j*h:j*h+h, i*w:i*w+w, :] = image
53 |
54 | return img
55 |
56 | def disp_scatter(x, g, gen, num_gens, fig=None, ax=None):
57 | colors = ['darkblue', 'yellow', 'indigo', 'darkgreen', 'purple',
58 | 'dodgerblue', 'lime', 'brown', 'darkcyan', 'deeppink']
59 |
60 | if ax is None:
61 | fig, ax = plt.subplots()
62 |
63 | ax.cla()
64 | ax.scatter(x[:, 0], x[:, 1], s=10, marker='+', color='r', alpha=0.8)
65 | for i in range(num_gens):
66 | ax.scatter(g[gen == i, 0], g[gen == i, 1], s=10, marker='o',
67 | color=colors[i], alpha=0.8)
68 | ax.legend(["real data"] + ['gen {}'.format(i) for i in range(num_gens)])
69 | ax.set_xlim(-3, 3)
70 | ax.set_ylim(-3, 3)
71 | return fig, ax
--------------------------------------------------------------------------------
/Stackelberg GAN/CIFAR-10/ops.py:
--------------------------------------------------------------------------------
1 | from __future__ import division
2 | from __future__ import print_function
3 | from __future__ import absolute_import
4 |
5 | import numpy as np
6 | import tensorflow as tf
7 |
8 | def lrelu(x, alpha=0.2):
9 | return tf.maximum(x, alpha * x)
10 |
11 | def linear(input, output_dim, scope='linear', stddev=0.01):
12 | norm = tf.random_normal_initializer(stddev=stddev)
13 | const = tf.constant_initializer(0.0)
14 | with tf.variable_scope(scope):
15 | w = tf.get_variable('weights', [input.get_shape()[1], output_dim], initializer=norm)
16 | b = tf.get_variable('biases', [output_dim], initializer=const)
17 | return tf.matmul(input, w) + b
18 |
19 | def conv2d(input_, output_dim,
20 | k_h=5, k_w=5, d_h=2, d_w=2, stddev=0.02,
21 | name="conv2d"):
22 | with tf.variable_scope(name):
23 | w = tf.get_variable('weights', [k_h, k_w, input_.get_shape()[-1], output_dim],
24 | initializer=tf.truncated_normal_initializer(stddev=stddev))
25 | conv = tf.nn.conv2d(input_, w, strides=[1, d_h, d_w, 1], padding='SAME')
26 |
27 | biases = tf.get_variable('biases', [output_dim], initializer=tf.constant_initializer(0.0))
28 | # conv = tf.reshape(tf.nn.bias_add(conv, biases), conv.get_shape())
29 |
30 | return tf.nn.bias_add(conv, biases)
31 |
32 |
33 | def deconv2d(input_, output_shape,
34 | k_h=5, k_w=5, d_h=2, d_w=2, stddev=0.02,
35 | name="deconv2d", with_w=False):
36 | with tf.variable_scope(name):
37 | # filter : [height, width, output_channels, in_channels]
38 | w = tf.get_variable('weights', [k_h, k_w, output_shape[-1], input_.get_shape()[-1]],
39 | initializer=tf.random_normal_initializer(stddev=stddev))
40 |
41 | try:
42 | deconv = tf.nn.conv2d_transpose(input_, w, output_shape=output_shape,
43 | strides=[1, d_h, d_w, 1])
44 |
45 | # Support for versions of TensorFlow before 0.7.0
46 | except AttributeError:
47 | deconv = tf.nn.deconv2d(input_, w, output_shape=output_shape,
48 | strides=[1, d_h, d_w, 1])
49 |
50 | biases = tf.get_variable('biases', [output_shape[-1]],
51 | initializer=tf.constant_initializer(0.0))
52 | deconv = tf.reshape(tf.nn.bias_add(deconv, biases), deconv.get_shape())
53 |
54 | if with_w:
55 | return deconv, w, biases
56 | else:
57 | return deconv
58 |
59 | def gmm_sample(num_samples, mix_coeffs, mean, cov):
60 | z = np.random.multinomial(num_samples, mix_coeffs)
61 | samples = np.zeros(shape=[num_samples, len(mean[0])])
62 | i_start = 0
63 | for i in range(len(mix_coeffs)):
64 | i_end = i_start + z[i]
65 | samples[i_start:i_end, :] = np.random.multivariate_normal(
66 | mean=np.array(mean)[i, :],
67 | cov=np.diag(np.array(cov)[i, :]),
68 | size=z[i])
69 | i_start = i_end
70 | return samples
71 |
--------------------------------------------------------------------------------
/Stackelberg GAN/CIFAR-10/main.py:
--------------------------------------------------------------------------------
1 | from __future__ import division
2 | from __future__ import print_function
3 | from __future__ import absolute_import
4 |
5 | import sys
6 | import pickle
7 | import argparse
8 | import numpy as np
9 | import tensorflow as tf
10 | import os
11 | from models import SGAN
12 |
13 |
14 | FLAGS = None
15 | # os.environ['CUDA_VISIBLE_DEVICES'] = '0'
16 |
17 | def main(_):
18 | tmp = pickle.load(open("data/cifar10_train.pkl", "rb"))
19 | x_train = tmp['data'].astype(np.float32).reshape([-1, 32, 32, 3]) / 127.5 - 1.
20 | print(x_train.shape)
21 | model = SGAN(
22 | num_z=FLAGS.num_z,
23 | beta=FLAGS.beta,
24 | num_gens=FLAGS.num_gens,
25 | d_batch_size=FLAGS.d_batch_size,
26 | g_batch_size=FLAGS.g_batch_size,
27 | z_prior=FLAGS.z_prior,
28 | learning_rate1=FLAGS.learning_rate1,
29 | learning_rate2=FLAGS.learning_rate2,
30 | img_size=(32, 32, 3),
31 | g_num_conv_layers=FLAGS.g_num_conv_layers,
32 | d_num_conv_layers=FLAGS.d_num_conv_layers,
33 | num_gen_feature_maps=FLAGS.num_gen_feature_maps,
34 | num_dis_feature_maps=FLAGS.num_dis_feature_maps,
35 | num_epochs=FLAGS.num_epochs,
36 | sample_fp="classifier_samples_g"+str(FLAGS.num_gens)+"_epoch_"+str(FLAGS.num_epochs)+"_layers_"+str(FLAGS.g_num_conv_layers)+"_lr1_"+str(FLAGS.learning_rate1)+"_lr2_"+str(FLAGS.learning_rate2)+"/samples_{epoch:04d}.png",
37 | sample_by_gen_fp="classifier_samples_by_gen_g"+str(FLAGS.num_gens)+"_epoch_"+str(FLAGS.num_epochs)+"_layers_"+str(FLAGS.g_num_conv_layers)+"_lr1_"+str(FLAGS.learning_rate1)+"_lr2_"+str(FLAGS.learning_rate2)+"_new",
38 | random_seed=6789,
39 | checkpoint_dir="classifier_checkpoint_g"+str(FLAGS.num_gens)+"_best")
40 | model.fit(x_train)
41 | # model.predict()
42 |
43 |
44 | if __name__ == '__main__':
45 | parser = argparse.ArgumentParser()
46 | parser.add_argument('--num_z', type=int, default=100,
47 | help='Number of latent units.')
48 | parser.add_argument('--beta', type=float, default=0.1,
49 | help='Diversity parameter beta.')
50 | parser.add_argument('--num_gens', type=int, default=10,
51 | help='Number of generators.')
52 | parser.add_argument('--d_batch_size', type=int, default=64,
53 | help='Minibatch size for the discriminator.')
54 | parser.add_argument('--g_batch_size', type=int, default=64,
55 | help='Minibatch size for the generators.')
56 | parser.add_argument('--z_prior', type=str, default="uniform",
57 | help='Prior distribution of the noise (uniform/gaussian).')
58 | parser.add_argument('--learning_rate1', type=float, default=0.000005,
59 | help='Learning rate1.')
60 | parser.add_argument('--learning_rate2', type=float, default=0.00001,
61 | help='Learning rate2.')
62 | parser.add_argument('--g_num_conv_layers', type=int, default=2,
63 | help='Number of G convolutional layers.')
64 | parser.add_argument('--d_num_conv_layers', type=int, default=3,
65 | help='Number of D convolutional layers.')
66 | parser.add_argument('--num_gen_feature_maps', type=int, default=128,
67 | help='Number of feature maps of Generator.')
68 | parser.add_argument('--num_dis_feature_maps', type=int, default=128,
69 | help='Number of feature maps of Discriminator.')
70 | parser.add_argument('--num_epochs', type=int, default=500,
71 | help='Number of epochs.')
72 | FLAGS, unparsed = parser.parse_known_args()
73 | tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)
74 |
--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
1 | # Stackelberg-GAN
2 | This is the code for [the paper](https://arxiv.org/abs/1811.08010) "Stackelberg GAN: Towards Provable Minimax Equilibrium via Multi-Generator Architectures".
3 |
4 | ## Install
5 | This code depends on python 3.6, pytorch 0.4.1 (for the experiemnts of mixture of Gaussians, MNIST, and Fashion MNIST) and Tensorflow (version>=1.4.1, for CIFAR-10 and Tiny Imagenet experiments). We suggest to install the dependencies using Anaconda or Miniconda. Here is an exemplary command:
6 | ```
7 | $ wget https://repo.anaconda.com/archive/Anaconda3-5.1.0-Linux-x86_64.sh
8 | $ bash Anaconda3-5.1.0-Linux-x86_64.sh
9 | $ source ~/.bashrc
10 | $ conda install pytorch=0.4.1
11 | ```
12 |
13 | ## Get started
14 | To get started, cd into the directory. Then runs the scripts:
15 | * gan_stackelberg_mG.py is a demo on the performance of Stackelberg GAN on Gaussian mixture dataset,
16 | * gan_branch_mG.py is a demo on the performance of multi-branch GAN (a baseline method) on Gaussian mixture dataset,
17 | * gan_mnist_classifier.py is a demo on the performance of Stackelberg GAN on MNIST dataset,
18 | * gan_mnist_fashion_classifier.py is a demo on the performance of Stackelberg GAN on fashion-MNIST dataset.
19 | * CIFAR-10: This folder contains code implementing the proposed Stackelberg GAN in CIFAR-10 using TensorFlow (version>=1.4.1). models.py constructed the model of Stackelberg GAN. main.py conducted experiment based on CIFAR-10 dataset. In main.py, model.fit() trains the Stackelberg GAN model based on the given dataset, while model.predict() outputs the generated examples.
20 |
21 | ## Using the code
22 | The command `python xxx.py --help` gives the help information about how to run the code.
23 |
24 | ## Architecture of Stackelberg GAN
25 |
26 | Stackelberg GAN is a general framework which can be built on top of all variants of standard GANs. The key idea is to apply multiple generators which team up to play against the discriminator.
27 |
28 |
29 | 
30 |
31 |
32 |
33 |
34 | ## Experimental Results
35 |
36 | ### Mixture of Gaussians
37 |
38 |
39 | 
40 |
41 |
42 | We test the performance of varying architectures of GANs on a synthetic mixture of Gaussians dataset with 8 modes and 0.01 standard deviation. We observe the following phenomena:
43 |
44 | *Naïvely increasing capacity of one-generator architecture does not alleviate mode collapse*. It shows
45 | that the multi-generator architecture in the Stackelberg GAN effectively alleviates the mode collapse issue.
46 | Though naïvely increasing capacity of one-generator architecture alleviates mode dropping issue, for more
47 | challenging mode collapse issue, the effect is not obvious.
48 |
49 | #### Running Example
50 |
51 | 
52 |
53 |
54 | *Stackelberg GAN outperforms multi-branch models.* We compare performance of multi-branch GAN (i.e., classic GAN with multi-branch architecture for its generator) and Stackelberg GAN. The performance of Stackelberg GAN is also better than multi-branch GAN of much larger capacity.
55 |
56 | #### Running Example
57 |
58 | 
59 |
60 |
61 | *Generators tend to learn balanced number of modes when they have same capacity*. We observe that
62 | for varying number of generators, each generator in the Stackelberg GAN tends to learn equal number of
63 | modes when the modes are symmetric and every generator has same capacity.
64 |
65 | #### Running Example
66 |
67 | 
68 |
69 |
70 | ### MNIST Dataset
71 | The following figure shows the diversity of generated digits by Stackelberg GAN with varying number of generators. *Left Figure:*
72 | Digits generated by the standard GAN. It shows that the standard GAN generates many "1"’s which are not very diverse. *Middle Figure:* Digits generated by the Stackelberg GAN with 5 generators, where every two rows correspond to one generator. *Right Figure:* Digits generated by the Stackelberg GAN with 10 generators, where each row corresponds to one generator. As the number of generators increases, the images tend to be more diverse.
73 |
74 | #### Running Example
75 |
76 | 
77 |
78 |
79 | ### Fashion-MNIST Dataset
80 | The following figure shows the diversity of generated fashions by Stackelberg GAN with varying number of generators. *Left Figure:*
81 | Examples generated by the standard GAN. It shows that the standard GAN fails to generate bags. *Middle Figure:* Examples generated by the Stackelberg GAN with 5 generators, where every two rows correspond to one generator. *Right Figure:* Examples generated by the Stackelberg GAN with 10 generators, where each row corresponds to one generator.
82 |
83 | #### Running Example
84 |
85 | 
86 |
87 |
88 | ### CIFAR-10/Tiny ImageNet Dataset
89 | The following figure shows the examples generated by Stackelberg GAN with 10 generators on CIFAR-10 and Tiny ImageNet. *Left Figure:*
90 | Examples generated by the Stackelberg GAN on CIFAR-10. *Right Figure:* Examples generated by the Stackelberg GAN on ImageNet.
91 |
92 | #### Running Example
93 |
94 | 
95 |
96 |
97 |
98 |
99 | ## Reference
100 | For technical details and full experimental results, see [the paper](https://arxiv.org/abs/1811.08010).
101 | ```
102 | @article{Zhang2018stackelberg,
103 | author = {Hongyang Zhang and Susu Xu and Jiantao Jiao and Pengtao Xie and Ruslan Salakhutdinov and Eric P. Xing},
104 | title = {Stackelberg GAN: Towards Provable Minimax Equilibrium via Multi-Generator Architectures},
105 | journal={arXiv preprint arXiv:1811.08010},
106 | year = {2018}
107 | }
108 | ```
109 |
110 | ## Contact
111 | Please contact hongyanz@cs.cmu.edu if you have any question on the codes.
112 |
--------------------------------------------------------------------------------
/Stackelberg GAN/Gaussian mixture/gan_branch_mG.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import numpy as np
4 | import math
5 |
6 | import torchvision.transforms as transforms
7 | from torchvision.utils import save_image
8 |
9 | from torch.utils.data import DataLoader
10 | from torchvision import datasets
11 | from torch.autograd import Variable
12 |
13 | import torch.nn as nn
14 | import torch.nn.functional as F
15 | import torch
16 | import shutil
17 |
18 | import matplotlib.pyplot as plt
19 | plt.switch_backend('agg')
20 | import seaborn as sns
21 | import matplotlib
22 |
23 | os.makedirs('images_branch_mG', exist_ok=True)
24 | shutil.rmtree('images_branch_mG')
25 | os.makedirs('images_branch_mG', exist_ok=True)
26 |
27 | parser = argparse.ArgumentParser()
28 | parser.add_argument('--n_epochs', type=int, default=200, help='number of epochs of training')
29 | parser.add_argument('--batch_size', type=int, default=64, help='size of the batches')
30 | parser.add_argument('--lr', type=float, default=0.0002, help='adam: learning rate')
31 | parser.add_argument('--b1', type=float, default=0.5, help='adam: decay of first order momentum of gradient')
32 | parser.add_argument('--b2', type=float, default=0.999, help='adam: decay of first order momentum of gradient')
33 | parser.add_argument('--n_cpu', type=int, default=8, help='number of cpu threads to use during batch generation')
34 | parser.add_argument('--latent_dim', type=int, default=2, help='dimensionality of the latent space')
35 | parser.add_argument('--img_size', type=int, default=28, help='size of each image dimension')
36 | parser.add_argument('--channels', type=int, default=1, help='number of image channels')
37 | parser.add_argument('--sample_interval', type=int, default=400, help='interval betwen image samples')
38 | parser.add_argument('--n_paths_D', type=int, default=1, help='number of paths of discriminator')
39 | parser.add_argument('--n_paths_G', type=int, default=1, help='number of paths of generator')
40 | opt = parser.parse_args()
41 | print(opt)
42 |
43 |
44 | cuda = True if torch.cuda.is_available() else False
45 |
46 | class Generator(nn.Module):
47 | def __init__(self):
48 | super(Generator, self).__init__()
49 |
50 | def block(in_feat, out_feat, normalize=True):
51 | layers = [nn.Linear(in_feat, out_feat)]
52 | if normalize:
53 | layers.append(nn.BatchNorm1d(out_feat, 0.8))
54 | layers.append(nn.LeakyReLU(0.2, inplace=True))
55 | return layers
56 |
57 | modules = nn.ModuleList()
58 | for _ in range(opt.n_paths_G):
59 | modules.append(nn.Sequential(
60 | *block(opt.latent_dim, 128, normalize=False),
61 | *block(128, 256),
62 | *block(256, 512),
63 | *block(512, 1024),
64 | nn.Linear(1024, 2),
65 | nn.Tanh()
66 | ))
67 | self.paths = modules
68 |
69 | def forward(self, z):
70 | img = torch.zeros(z.shape[0], 2).cuda()
71 | for path in self.paths:
72 | img += path(z)
73 | img = img/opt.n_paths_G
74 | return img
75 |
76 | class Discriminator(nn.Module):
77 | def __init__(self):
78 | super(Discriminator, self).__init__()
79 | modules = nn.ModuleList()
80 | for _ in range(opt.n_paths_D):
81 | modules.append(nn.Sequential(
82 | nn.Linear(2, 512),
83 | nn.LeakyReLU(0.2, inplace=True),
84 | nn.Linear(512, 256),
85 | nn.LeakyReLU(0.2, inplace=True),
86 | nn.Linear(256, 1),
87 | nn.Sigmoid()
88 | ))
89 | self.paths = modules
90 |
91 | def forward(self, img):
92 | img_flat = img
93 | validity = []
94 | for path in self.paths:
95 | validity.append(path(img_flat))
96 | validity = torch.cat(validity, dim=1)
97 | return validity
98 |
99 | # Loss function
100 | adversarial_loss = torch.nn.BCELoss()
101 |
102 | # Initialize generator and discriminator
103 | generator = Generator()
104 | discriminator = Discriminator()
105 |
106 | if cuda:
107 | generator.cuda()
108 | discriminator.cuda()
109 | adversarial_loss.cuda()
110 |
111 | # Configure data loader
112 | n_mixture = 8
113 | radius = 1
114 | std = 0.01
115 | thetas = np.linspace(0, 2 * (1-1/n_mixture) * np.pi, n_mixture)
116 | xs, ys = radius * np.sin(thetas), radius * np.cos(thetas)
117 | data_size = 1000*n_mixture
118 | data = torch.zeros(data_size, 2)
119 | for i in range(data_size):
120 | coin = np.random.randint(0, n_mixture)
121 | data[i, :] = torch.normal(mean=torch.Tensor([xs[coin], ys[coin]]), std=std*torch.ones(1, 2))
122 |
123 |
124 | # Optimizers
125 | optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
126 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
127 |
128 | Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
129 |
130 | # ----------
131 | # Training
132 | # ----------
133 | n_batch = math.ceil(data_size/opt.batch_size)
134 |
135 | for epoch in range(opt.n_epochs):
136 | colors = matplotlib.cm.rainbow(np.linspace(0, 1, 1+opt.n_paths_G))
137 | plt.plot(data[:, 0].cpu().numpy(), data[:, 1].cpu().numpy(), color=colors[0], marker='.', linestyle='None')
138 | z = Variable(Tensor(np.random.normal(0, 1, (8000//opt.n_paths_G, opt.latent_dim))))
139 | for k in range(opt.n_paths_G):
140 | gen_data = generator.paths[k](z).detach()
141 | plt.plot(gen_data[:, 0].cpu().numpy(), gen_data[:, 1].cpu().numpy(), color=colors[1+k], marker='.', linestyle='None')
142 |
143 | for i in range(n_batch):
144 |
145 | imgs = data[i*opt.batch_size:min((i+1)*opt.batch_size, data_size-1), :]
146 |
147 | # Adversarial ground truths
148 | valid = Variable(Tensor(imgs.size(0), opt.n_paths_D).fill_(1.0), requires_grad=False)
149 | fake = Variable(Tensor(imgs.size(0), opt.n_paths_D).fill_(0.0), requires_grad=False)
150 |
151 | # Configure input
152 | real_imgs = Variable(imgs.type(Tensor))
153 |
154 | # -----------------
155 | # Train Generator
156 | # -----------------
157 |
158 | optimizer_G.zero_grad()
159 |
160 | # Sample noise as generator input
161 | z = Variable(Tensor(np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))))
162 |
163 | # Generate a batch of images
164 | gen_imgs = generator(z)
165 |
166 | # Loss measures generator's ability to fool the discriminator
167 | g_loss = adversarial_loss(discriminator(gen_imgs), valid)
168 |
169 | g_loss.backward()
170 | optimizer_G.step()
171 |
172 | # ---------------------
173 | # Train Discriminator
174 | # ---------------------
175 |
176 | optimizer_D.zero_grad()
177 |
178 | d_loss = 0
179 | real_loss = adversarial_loss(discriminator(real_imgs), valid)
180 |
181 | # Generate a batch of images
182 | gen_imgs = generator(z)
183 |
184 | # Measure discriminator's ability to classify real from generated samples
185 | fake_loss = adversarial_loss(discriminator(gen_imgs.detach()), fake)
186 | d_loss = (real_loss + fake_loss) / 2
187 |
188 | d_loss.backward()
189 | optimizer_D.step()
190 |
191 | print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch+1, opt.n_epochs, i, n_batch,
192 | d_loss.item(), g_loss.item()))
193 |
194 | colors = matplotlib.cm.rainbow(np.linspace(0, 1, 1+opt.n_paths_G))
195 | plt.plot(data[:, 0].cpu().numpy(), data[:, 1].cpu().numpy(), color=colors[0], marker='.', linestyle='None')
196 |
197 | z = Variable(Tensor(np.random.normal(0, 1, (8000//opt.n_paths_G, opt.latent_dim))))
198 | temp = []
199 | for k in range(opt.n_paths_G):
200 | # temp.append(generator.paths[k](z).detach())
201 | # gen_data = torch.cat(temp, dim=0)
202 | # sns.jointplot(gen_data[:, 0].cpu().numpy(), gen_data[:, 1].cpu().numpy(), kind='kde', stat_func=None)
203 | gen_data = generator.paths[k](z).detach()
204 | plt.plot(gen_data[:, 0].cpu().numpy(), gen_data[:, 1].cpu().numpy(), color=colors[1+k], marker='.', linestyle='None')
205 |
206 | plt.savefig('images_branch_mG/%d.png' % epoch)
207 | plt.close('all')
208 |
--------------------------------------------------------------------------------
/Stackelberg GAN/fashion_MNIST/gan_mnist_fashion_classifier.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import numpy as np
4 | import math
5 |
6 | import torchvision.transforms as transforms
7 | from torchvision.utils import save_image
8 |
9 | from torch.utils.data import DataLoader
10 | from torchvision import datasets
11 | from torch.autograd import Variable
12 | from tqdm import tqdm
13 |
14 | import torch.nn as nn
15 | import torch.nn.functional as F
16 | import torch
17 | import shutil
18 |
19 | os.makedirs('images_ensemble_fashionmnist10', exist_ok=True)
20 | shutil.rmtree('images_ensemble_fashionmnist10')
21 | os.makedirs('images_ensemble_fashionmnist10', exist_ok=True)
22 |
23 | parser = argparse.ArgumentParser()
24 | parser.add_argument('--n_epochs', type=int, default=500, help='number of epochs of training')
25 | parser.add_argument('--batch_size', type=int, default=100, help='size of the batches')
26 | parser.add_argument('--lr', type=float, default=0.0002, help='adam: learning rate')
27 | parser.add_argument('--b1', type=float, default=0.5, help='adam: decay of first order momentum of gradient')
28 | parser.add_argument('--b2', type=float, default=0.999, help='adam: decay of first order momentum of gradient')
29 | parser.add_argument('--n_cpu', type=int, default=8, help='number of cpu threads to use during batch generation')
30 | parser.add_argument('--latent_dim', type=int, default=2, help='dimensionality of the latent space')
31 | parser.add_argument('--img_size', type=int, default=28, help='size of each image dimension')
32 | parser.add_argument('--channels', type=int, default=1, help='number of image channels')
33 | parser.add_argument('--sample_interval', type=int, default=400, help='interval betwen image samples')
34 | parser.add_argument('--n_paths_G', type=int, default=10, help='number of paths of generator')
35 | opt = parser.parse_args()
36 | print(opt)
37 |
38 | img_shape = (opt.channels, opt.img_size, opt.img_size)
39 |
40 | cuda = True if torch.cuda.is_available() else False
41 |
42 | class Generator(nn.Module):
43 | def __init__(self):
44 | super(Generator, self).__init__()
45 |
46 | def block(in_feat, out_feat, normalize=True):
47 | layers = [nn.Linear(in_feat, out_feat)]
48 | if normalize:
49 | layers.append(nn.BatchNorm1d(out_feat, 0.8))
50 | layers.append(nn.LeakyReLU(0.2, inplace=True))
51 | return layers
52 |
53 | modules = nn.ModuleList()
54 | for _ in range(opt.n_paths_G):
55 | modules.append(nn.Sequential(
56 | *block(opt.latent_dim, 128, normalize=False),
57 | *block(128, 256),
58 | *block(256, 512),
59 | *block(512, 1024),
60 | nn.Linear(1024, int(np.prod(img_shape))),
61 | nn.Tanh()
62 | ))
63 | self.paths = modules
64 |
65 | def forward(self, z):
66 | img = []
67 | for path in self.paths:
68 | img.append(path(z).view(img.size(0), *img_shape))
69 | img = torch.cat(img, dim=0)
70 | return img
71 |
72 | class Discriminator(nn.Module):
73 | def __init__(self):
74 | super(Discriminator, self).__init__()
75 | self.fc1 = nn.Linear(int(np.prod(img_shape)), 512)
76 | self.lr1 = nn.LeakyReLU(0.2, inplace=True)
77 | self.fc2 = nn.Linear(512, 256)
78 | self.lr2 = nn.LeakyReLU(0.2, inplace=True)
79 | modules = nn.ModuleList()
80 | modules.append(nn.Sequential(
81 | nn.Linear(256, 1),
82 | nn.Sigmoid(),
83 | ))
84 | modules.append(nn.Sequential(
85 | nn.Linear(256, 10),
86 | ))
87 | self.paths = modules
88 |
89 | def forward(self, img):
90 | img_flat = img.view(img.size(0), -1)
91 | img_flat = self.lr2(self.fc2(self.lr1(self.fc1(img_flat))))
92 | validity = self.paths[0](img_flat)
93 | classifier = F.log_softmax(self.paths[1](img_flat), dim=1)
94 | return validity, classifier
95 |
96 | # Loss function
97 | adversarial_loss = torch.nn.BCELoss()
98 |
99 | # Initialize generator and discriminator
100 | generator = Generator()
101 | discriminator = Discriminator()
102 |
103 | if cuda:
104 | generator.cuda()
105 | discriminator.cuda()
106 | adversarial_loss.cuda()
107 |
108 | # Configure data loader
109 | os.makedirs('../data/fashionmnist', exist_ok=True)
110 | dataloader = torch.utils.data.DataLoader(
111 | datasets.FashionMNIST('../data/fashionmnist', train=True, download=True,
112 | transform=transforms.Compose([
113 | transforms.ToTensor(),
114 | transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
115 | ])),
116 | batch_size=opt.batch_size, shuffle=True)
117 |
118 |
119 | # Optimizers
120 | optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
121 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
122 |
123 | Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
124 |
125 | # ----------
126 | # Training
127 | # ----------
128 |
129 | for epoch in tqdm(range(opt.n_epochs)):
130 | for i, (imgs, _) in enumerate(dataloader):
131 |
132 | # Adversarial ground truths
133 | valid = Variable(Tensor(imgs.size(0), 1).fill_(1.0), requires_grad=False)
134 | fake = Variable(Tensor(imgs.size(0), 1).fill_(0.0), requires_grad=False)
135 |
136 | # Configure input
137 | real_imgs = Variable(imgs.type(Tensor))
138 |
139 | # -----------------
140 | # Train Generator
141 | # -----------------
142 |
143 | optimizer_G.zero_grad()
144 |
145 | # Sample noise as generator input
146 | z = Variable(Tensor(np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))))
147 |
148 | g_loss = 0
149 | for k in range(opt.n_paths_G):
150 |
151 | # Generate a batch of images
152 | gen_imgs = generator.paths[k](z)
153 |
154 | # Loss measures generator's ability to fool the discriminator
155 | validity, classifier = discriminator(gen_imgs)
156 | g_loss += adversarial_loss(validity, valid)
157 |
158 | # Loss measures classifier's ability to classify various generators
159 | target = Variable(Tensor(imgs.size(0)).fill_(k), requires_grad=False)
160 | target = target.type(torch.cuda.LongTensor)
161 | g_loss += F.nll_loss(classifier, target)*0.1
162 |
163 | g_loss.backward()
164 | optimizer_G.step()
165 |
166 | # ------------------------------------
167 | # Train Discriminator and Classifier
168 | # ------------------------------------
169 |
170 | optimizer_D.zero_grad()
171 |
172 | d_loss = 0
173 | validity, classifier = discriminator(real_imgs)
174 | real_loss = adversarial_loss(validity, valid)
175 | temp = []
176 | for k in range(opt.n_paths_G):
177 |
178 | # Generate a batch of images
179 | gen_imgs = generator.paths[k](z).view(imgs.shape[0], *img_shape)
180 | temp.append(gen_imgs[0:(100//opt.n_paths_G), :])
181 |
182 | # Loss measures discriminator's ability to classify real from generated samples
183 | validity, classifier = discriminator(gen_imgs.detach())
184 | fake_loss = adversarial_loss(validity, fake)
185 | d_loss += (real_loss + fake_loss) / 2
186 |
187 | # Loss measures classifier's ability to classify various generators
188 | target = Variable(Tensor(imgs.size(0)).fill_(k), requires_grad=False)
189 | target = target.type(torch.cuda.LongTensor)
190 | d_loss += F.nll_loss(classifier, target)*0.1
191 |
192 | plot_imgs = torch.cat(temp, dim=0)
193 | plot_imgs.detach()
194 |
195 | d_loss.backward()
196 | optimizer_D.step()
197 |
198 | #print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, opt.n_epochs, i, len(dataloader),
199 | #d_loss.item(), g_loss.item()))
200 |
201 | batches_done = epoch * len(dataloader) + i
202 | if batches_done % opt.sample_interval == 0:
203 | save_image(plot_imgs[:100], 'images_ensemble_fashionmnist10/%d.png' % batches_done, nrow=10, normalize=True)
204 |
205 |
--------------------------------------------------------------------------------
/Stackelberg GAN/Gaussian mixture/gan_stackelberg_mG.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import numpy as np
4 | import math
5 |
6 | import torchvision.transforms as transforms
7 | from torchvision.utils import save_image
8 |
9 | from torch.utils.data import DataLoader
10 | from torchvision import datasets
11 | from torch.autograd import Variable
12 |
13 | import torch.nn as nn
14 | import torch.nn.functional as F
15 | import torch
16 | import shutil
17 |
18 | import matplotlib.pyplot as plt
19 | plt.switch_backend('agg')
20 | import seaborn as sns
21 | import matplotlib
22 |
23 | #os.environ["CUDA_VISIBLE_DEVICES"] = "1"
24 |
25 | os.makedirs('images_ensemble_mG', exist_ok=True)
26 | shutil.rmtree('images_ensemble_mG')
27 | os.makedirs('images_ensemble_mG', exist_ok=True)
28 |
29 | parser = argparse.ArgumentParser()
30 | parser.add_argument('--n_epochs', type=int, default=200, help='number of epochs of training')
31 | parser.add_argument('--batch_size', type=int, default=64, help='size of the batches')
32 | parser.add_argument('--lr', type=float, default=0.0002, help='adam: learning rate')
33 | parser.add_argument('--b1', type=float, default=0.5, help='adam: decay of first order momentum of gradient')
34 | parser.add_argument('--b2', type=float, default=0.999, help='adam: decay of first order momentum of gradient')
35 | parser.add_argument('--n_cpu', type=int, default=8, help='number of cpu threads to use during batch generation')
36 | parser.add_argument('--latent_dim', type=int, default=2, help='dimensionality of the latent space')
37 | parser.add_argument('--img_size', type=int, default=28, help='size of each image dimension')
38 | parser.add_argument('--channels', type=int, default=1, help='number of image channels')
39 | parser.add_argument('--sample_interval', type=int, default=400, help='interval betwen image samples')
40 | parser.add_argument('--n_paths_D', type=int, default=1, help='number of paths of discriminator')
41 | parser.add_argument('--n_paths_G', type=int, default=8, help='number of paths of generator')
42 | opt = parser.parse_args()
43 | print(opt)
44 |
45 | #np.random.seed(1)
46 | #torch.manual_seed(1)
47 | #if torch.cuda.is_available():
48 | # torch.cuda.manual_seed(1)
49 |
50 | cuda = True if torch.cuda.is_available() else False
51 |
52 | class Generator(nn.Module):
53 | def __init__(self):
54 | super(Generator, self).__init__()
55 |
56 | def block(in_feat, out_feat, normalize=True):
57 | layers = [nn.Linear(in_feat, out_feat)]
58 | if normalize:
59 | layers.append(nn.BatchNorm1d(out_feat, 0.8))
60 | layers.append(nn.LeakyReLU(0.2, inplace=True))
61 | return layers
62 |
63 | modules = nn.ModuleList()
64 | for _ in range(opt.n_paths_G):
65 | modules.append(nn.Sequential(
66 | *block(opt.latent_dim, 32, normalize=False),
67 | nn.Linear(32, 2),
68 | nn.Tanh()
69 | ))
70 | self.paths = modules
71 |
72 | def forward(self, z):
73 | img = []
74 | for path in self.paths:
75 | img.append(path(z))
76 | img = torch.cat(img, dim=1)
77 | return img
78 |
79 | class Discriminator(nn.Module):
80 | def __init__(self):
81 | super(Discriminator, self).__init__()
82 | modules = nn.ModuleList()
83 | for _ in range(opt.n_paths_D):
84 | modules.append(nn.Sequential(
85 | nn.Linear(2, 512),
86 | nn.LeakyReLU(0.2, inplace=True),
87 | nn.Linear(512, 256),
88 | nn.LeakyReLU(0.2, inplace=True),
89 | nn.Linear(256, 1),
90 | nn.Sigmoid()
91 | ))
92 | self.paths = modules
93 |
94 | def forward(self, img):
95 | img_flat = img
96 | validity = []
97 | for path in self.paths:
98 | validity.append(path(img_flat))
99 | validity = torch.cat(validity, dim=1)
100 | return validity
101 |
102 | # Loss function
103 | adversarial_loss = torch.nn.BCELoss()
104 |
105 | # Initialize generator and discriminator
106 | generator = Generator()
107 | discriminator = Discriminator()
108 |
109 | if cuda:
110 | generator.cuda()
111 | discriminator.cuda()
112 | adversarial_loss.cuda()
113 |
114 | # Configure data loader
115 | n_mixture = 8
116 | radius = 1
117 | std = 0.01
118 | thetas = np.linspace(0, 2 * (1-1/n_mixture) * np.pi, n_mixture)
119 | xs, ys = radius * np.sin(thetas), radius * np.cos(thetas)
120 | data_size = 1000*n_mixture
121 | data = torch.zeros(data_size, 2)
122 | for i in range(data_size):
123 | coin = np.random.randint(0, n_mixture)
124 | data[i, :] = torch.normal(mean=torch.Tensor([xs[coin], ys[coin]]), std=std*torch.ones(1, 2))
125 |
126 |
127 | # Optimizers
128 | optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
129 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
130 |
131 | Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
132 |
133 | # ----------
134 | # Training
135 | # ----------
136 | n_batch = math.ceil(data_size/opt.batch_size)
137 |
138 | for epoch in range(opt.n_epochs):
139 | colors = matplotlib.cm.rainbow(np.linspace(0, 1, 1+opt.n_paths_G))
140 | plt.plot(data[:, 0].cpu().numpy(), data[:, 1].cpu().numpy(), color=colors[0], marker='.', linestyle='None')
141 | z = Variable(Tensor(np.random.normal(0, 1, (8000//opt.n_paths_G, opt.latent_dim))))
142 | for k in range(opt.n_paths_G):
143 | gen_data = generator.paths[k](z).detach()
144 | plt.plot(gen_data[:, 0].cpu().numpy(), gen_data[:, 1].cpu().numpy(), color=colors[1+k], marker='.', linestyle='None')
145 |
146 | for i in range(n_batch):
147 |
148 | imgs = data[i*opt.batch_size:min((i+1)*opt.batch_size, data_size-1), :]
149 |
150 | # Adversarial ground truths
151 | valid = Variable(Tensor(imgs.size(0), opt.n_paths_D).fill_(1.0), requires_grad=False)
152 | fake = Variable(Tensor(imgs.size(0), opt.n_paths_D).fill_(0.0), requires_grad=False)
153 |
154 | # Configure input
155 | real_imgs = Variable(imgs.type(Tensor))
156 |
157 | # -----------------
158 | # Train Generator
159 | # -----------------
160 |
161 | optimizer_G.zero_grad()
162 |
163 | # Sample noise as generator input
164 | z = Variable(Tensor(np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))))
165 |
166 | g_loss = 0
167 | for k in range(opt.n_paths_G):
168 |
169 | # Generate a batch of images
170 | gen_imgs = generator.paths[k](z)
171 |
172 | # Loss measures generator's ability to fool the discriminator
173 | g_loss += adversarial_loss(discriminator(gen_imgs), valid)
174 |
175 | g_loss.backward()
176 | optimizer_G.step()
177 |
178 | # ---------------------
179 | # Train Discriminator
180 | # ---------------------
181 |
182 | optimizer_D.zero_grad()
183 |
184 | d_loss = 0
185 | real_loss = adversarial_loss(discriminator(real_imgs), valid)
186 | for k in range(opt.n_paths_G):
187 |
188 | # Generate a batch of images
189 | gen_imgs = generator.paths[k](z)
190 |
191 | # Measure discriminator's ability to classify real from generated samples
192 | fake_loss = adversarial_loss(discriminator(gen_imgs.detach()), fake)
193 | d_loss += (real_loss + fake_loss) / 2
194 |
195 | d_loss.backward()
196 | optimizer_D.step()
197 |
198 | print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f] [D value: %f]" % (epoch+1, opt.n_epochs, i, n_batch,
199 | d_loss.item(), g_loss.item(), discriminator(real_imgs).mean(0).mean().item()))
200 |
201 | colors = matplotlib.cm.rainbow(np.linspace(0, 1, 1+opt.n_paths_G))
202 | plt.plot(data[:, 0].cpu().numpy(), data[:, 1].cpu().numpy(), color=colors[0], marker='.', linestyle='None')
203 |
204 | z = Variable(Tensor(np.random.normal(0, 1, (8000//opt.n_paths_G, opt.latent_dim))))
205 | temp = []
206 | for k in range(opt.n_paths_G):
207 | # temp.append(generator.paths[k](z).detach())
208 | # gen_data = torch.cat(temp, dim=0)
209 | # sns.jointplot(gen_data[:, 0].cpu().numpy(), gen_data[:, 1].cpu().numpy(), kind='kde', stat_func=None)
210 | gen_data = generator.paths[k](z).detach()
211 | plt.plot(gen_data[:, 0].cpu().numpy(), gen_data[:, 1].cpu().numpy(), color=colors[1+k], marker='.', linestyle='None')
212 |
213 | plt.savefig('images_ensemble_mG/%d.png' % epoch)
214 | plt.close('all')
215 |
--------------------------------------------------------------------------------
/Stackelberg GAN/MNIST/gan_mnist_classifier.py:
--------------------------------------------------------------------------------
1 | import argparse
2 | import os
3 | import numpy as np
4 | import math
5 |
6 | import torchvision.transforms as transforms
7 | from torchvision.utils import save_image
8 |
9 | from torch.utils.data import DataLoader
10 | from torchvision import datasets
11 | from torch.autograd import Variable
12 | from tqdm import tqdm
13 |
14 | import torch.nn as nn
15 | import torch.nn.functional as F
16 | import torch
17 | import shutil
18 |
19 | os.makedirs('images_ensemble_mnist10_classifier_small', exist_ok=True)
20 | shutil.rmtree('images_ensemble_mnist10_classifier_small')
21 | os.makedirs('images_ensemble_mnist10_classifier_small', exist_ok=True)
22 |
23 | parser = argparse.ArgumentParser()
24 | parser.add_argument('--n_epochs', type=int, default=500, help='number of epochs of training')
25 | parser.add_argument('--batch_size', type=int, default=100, help='size of the batches')
26 | parser.add_argument('--lr', type=float, default=0.0002, help='adam: learning rate')
27 | parser.add_argument('--b1', type=float, default=0.5, help='adam: decay of first order momentum of gradient')
28 | parser.add_argument('--b2', type=float, default=0.999, help='adam: decay of first order momentum of gradient')
29 | parser.add_argument('--n_cpu', type=int, default=8, help='number of cpu threads to use during batch generation')
30 | parser.add_argument('--latent_dim', type=int, default=100, help='dimensionality of the latent space')
31 | parser.add_argument('--img_size', type=int, default=28, help='size of each image dimension')
32 | parser.add_argument('--channels', type=int, default=1, help='number of image channels')
33 | parser.add_argument('--sample_interval', type=int, default=400, help='interval betwen image samples')
34 | parser.add_argument('--n_paths_G', type=int, default=10, help='number of paths of generator')
35 | parser.add_argument('--classifier_para', type=float, default=1.0, help='regularization parameter for classifier')
36 | opt = parser.parse_args()
37 | print(opt)
38 |
39 | img_shape = (opt.channels, opt.img_size, opt.img_size)
40 |
41 | cuda = True if torch.cuda.is_available() else False
42 |
43 | class Generator(nn.Module):
44 | def __init__(self):
45 | super(Generator, self).__init__()
46 |
47 | def block(in_feat, out_feat, normalize=True):
48 | layers = [nn.Linear(in_feat, out_feat)]
49 | if normalize:
50 | layers.append(nn.BatchNorm1d(out_feat, 0.8))
51 | layers.append(nn.LeakyReLU(0.2, inplace=True))
52 | return layers
53 |
54 | modules = nn.ModuleList()
55 | for _ in range(opt.n_paths_G):
56 | modules.append(nn.Sequential(
57 | *block(opt.latent_dim, 128),
58 | *block(128, 512),
59 | #*block(256, 512),
60 | #*block(512, 512),
61 | #*block(512, 1024),
62 | nn.Linear(512, int(np.prod(img_shape))),
63 | nn.Tanh()
64 | ))
65 | self.paths = modules
66 |
67 | def forward(self, z):
68 | img = []
69 | for path in self.paths:
70 | img.append(path(z).view(img.size(0), *img_shape))
71 | img = torch.cat(img, dim=0)
72 | return img
73 |
74 | class Discriminator(nn.Module):
75 | def __init__(self):
76 | super(Discriminator, self).__init__()
77 | self.fc1 = nn.Linear(int(np.prod(img_shape)), 512)
78 | self.lr1 = nn.LeakyReLU(0.2, inplace=True)
79 | self.fc2 = nn.Linear(512, 256)
80 | self.lr2 = nn.LeakyReLU(0.2, inplace=True)
81 | modules = nn.ModuleList()
82 | modules.append(nn.Sequential(
83 | nn.Linear(256, 1),
84 | nn.Sigmoid(),
85 | ))
86 | modules.append(nn.Sequential(
87 | nn.Linear(256, 10),
88 | ))
89 | self.paths = modules
90 |
91 | def forward(self, img):
92 | img_flat = img.view(img.size(0), -1)
93 | img_flat = self.lr2(self.fc2(self.lr1(self.fc1(img_flat))))
94 | validity = self.paths[0](img_flat)
95 | classifier = F.log_softmax(self.paths[1](img_flat), dim=1)
96 | return validity, classifier
97 |
98 | # Loss function
99 | adversarial_loss = torch.nn.BCELoss()
100 |
101 | # Initialize generator and discriminator
102 | generator = Generator()
103 | discriminator = Discriminator()
104 |
105 | if cuda:
106 | generator.cuda()
107 | discriminator.cuda()
108 | adversarial_loss.cuda()
109 |
110 | # Configure data loader
111 | os.makedirs('../data/mnist', exist_ok=True)
112 | dataloader = torch.utils.data.DataLoader(
113 | datasets.MNIST('../data/mnist', train=True, download=True,
114 | transform=transforms.Compose([
115 | transforms.ToTensor(),
116 | transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
117 | ])),
118 | batch_size=opt.batch_size, shuffle=True)
119 |
120 |
121 | # Optimizers
122 | optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
123 | optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
124 |
125 | Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor
126 |
127 | # ----------
128 | # Training
129 | # ----------
130 |
131 | # print("--------Loading Model--------")
132 | # checkpoint = torch.load('checkpoint_images_ensemble_fashionmnist10_classifier.tar')
133 | # generator.load_state_dict(checkpoint['g_state_dict'])
134 | # discriminator.load_state_dict(checkpoint['d_state_dict'])
135 |
136 | for epoch in tqdm(range(opt.n_epochs)):
137 | for i, (imgs, _) in enumerate(dataloader):
138 |
139 | # Adversarial ground truths
140 | valid = Variable(Tensor(imgs.size(0), 1).fill_(1.0), requires_grad=False)
141 | fake = Variable(Tensor(imgs.size(0), 1).fill_(0.0), requires_grad=False)
142 |
143 | # Configure input
144 | real_imgs = Variable(imgs.type(Tensor))
145 |
146 | # -----------------
147 | # Train Generator
148 | # -----------------
149 |
150 | optimizer_G.zero_grad()
151 |
152 | # Sample noise as generator input
153 | z = Variable(Tensor(np.random.normal(0, 1, (imgs.shape[0], opt.latent_dim))))
154 |
155 | g_loss = 0
156 | for k in range(opt.n_paths_G):
157 |
158 | # Generate a batch of images
159 | gen_imgs = generator.paths[k](z)
160 |
161 | # Loss measures generator's ability to fool the discriminator
162 | validity, classifier = discriminator(gen_imgs)
163 | g_loss += adversarial_loss(validity, valid)
164 |
165 | # Loss measures classifier's ability to classify various generators
166 | target = Variable(Tensor(imgs.size(0)).fill_(k), requires_grad=False)
167 | target = target.type(torch.cuda.LongTensor)
168 | g_loss += F.nll_loss(classifier, target)*opt.classifier_para
169 |
170 | g_loss.backward()
171 | optimizer_G.step()
172 |
173 | # ------------------------------------
174 | # Train Discriminator and Classifier
175 | # ------------------------------------
176 |
177 | optimizer_D.zero_grad()
178 |
179 | d_loss = 0
180 | validity, classifier = discriminator(real_imgs)
181 | real_loss = adversarial_loss(validity, valid)
182 | temp = []
183 | for k in range(opt.n_paths_G):
184 |
185 | # Generate a batch of images
186 | gen_imgs = generator.paths[k](z).view(imgs.shape[0], *img_shape)
187 | temp.append(gen_imgs[0:(100//opt.n_paths_G), :])
188 |
189 | # Loss measures discriminator's ability to classify real from generated samples
190 | validity, classifier = discriminator(gen_imgs.detach())
191 | fake_loss = adversarial_loss(validity, fake)
192 | d_loss += (real_loss + fake_loss) / 2
193 |
194 | # Loss measures classifier's ability to classify various generators
195 | target = Variable(Tensor(imgs.size(0)).fill_(k), requires_grad=False)
196 | target = target.type(torch.cuda.LongTensor)
197 | d_loss += F.nll_loss(classifier, target)*opt.classifier_para
198 |
199 | plot_imgs = torch.cat(temp, dim=0)
200 | plot_imgs.detach()
201 |
202 | d_loss.backward()
203 | optimizer_D.step()
204 |
205 | #print ("[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f]" % (epoch, opt.n_epochs, i, len(dataloader),
206 | #d_loss.item(), g_loss.item()))
207 |
208 | batches_done = epoch * len(dataloader) + i
209 | if batches_done % opt.sample_interval == 0:
210 | save_image(plot_imgs[:100], 'images_ensemble_mnist10_classifier_small/%d.png' % batches_done, nrow=10, normalize=True)
211 |
212 | #if epoch % 10 == 0:
213 | #torch.save({
214 | #'epoch': epoch + 1,
215 | #'g_state_dict': generator.state_dict(),
216 | #'d_state_dict': discriminator.state_dict(),
217 | #}, 'checkpoint_images_ensemble_fashionmnist10_classifier.tar')
218 |
--------------------------------------------------------------------------------
/LICENSE:
--------------------------------------------------------------------------------
1 | Apache License
2 | Version 2.0, January 2004
3 | http://www.apache.org/licenses/
4 |
5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6 |
7 | 1. Definitions.
8 |
9 | "License" shall mean the terms and conditions for use, reproduction,
10 | and distribution as defined by Sections 1 through 9 of this document.
11 |
12 | "Licensor" shall mean the copyright owner or entity authorized by
13 | the copyright owner that is granting the License.
14 |
15 | "Legal Entity" shall mean the union of the acting entity and all
16 | other entities that control, are controlled by, or are under common
17 | control with that entity. For the purposes of this definition,
18 | "control" means (i) the power, direct or indirect, to cause the
19 | direction or management of such entity, whether by contract or
20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the
21 | outstanding shares, or (iii) beneficial ownership of such entity.
22 |
23 | "You" (or "Your") shall mean an individual or Legal Entity
24 | exercising permissions granted by this License.
25 |
26 | "Source" form shall mean the preferred form for making modifications,
27 | including but not limited to software source code, documentation
28 | source, and configuration files.
29 |
30 | "Object" form shall mean any form resulting from mechanical
31 | transformation or translation of a Source form, including but
32 | not limited to compiled object code, generated documentation,
33 | and conversions to other media types.
34 |
35 | "Work" shall mean the work of authorship, whether in Source or
36 | Object form, made available under the License, as indicated by a
37 | copyright notice that is included in or attached to the work
38 | (an example is provided in the Appendix below).
39 |
40 | "Derivative Works" shall mean any work, whether in Source or Object
41 | form, that is based on (or derived from) the Work and for which the
42 | editorial revisions, annotations, elaborations, or other modifications
43 | represent, as a whole, an original work of authorship. For the purposes
44 | of this License, Derivative Works shall not include works that remain
45 | separable from, or merely link (or bind by name) to the interfaces of,
46 | the Work and Derivative Works thereof.
47 |
48 | "Contribution" shall mean any work of authorship, including
49 | the original version of the Work and any modifications or additions
50 | to that Work or Derivative Works thereof, that is intentionally
51 | submitted to Licensor for inclusion in the Work by the copyright owner
52 | or by an individual or Legal Entity authorized to submit on behalf of
53 | the copyright owner. For the purposes of this definition, "submitted"
54 | means any form of electronic, verbal, or written communication sent
55 | to the Licensor or its representatives, including but not limited to
56 | communication on electronic mailing lists, source code control systems,
57 | and issue tracking systems that are managed by, or on behalf of, the
58 | Licensor for the purpose of discussing and improving the Work, but
59 | excluding communication that is conspicuously marked or otherwise
60 | designated in writing by the copyright owner as "Not a Contribution."
61 |
62 | "Contributor" shall mean Licensor and any individual or Legal Entity
63 | on behalf of whom a Contribution has been received by Licensor and
64 | subsequently incorporated within the Work.
65 |
66 | 2. Grant of Copyright License. Subject to the terms and conditions of
67 | this License, each Contributor hereby grants to You a perpetual,
68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69 | copyright license to reproduce, prepare Derivative Works of,
70 | publicly display, publicly perform, sublicense, and distribute the
71 | Work and such Derivative Works in Source or Object form.
72 |
73 | 3. Grant of Patent License. Subject to the terms and conditions of
74 | this License, each Contributor hereby grants to You a perpetual,
75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76 | (except as stated in this section) patent license to make, have made,
77 | use, offer to sell, sell, import, and otherwise transfer the Work,
78 | where such license applies only to those patent claims licensable
79 | by such Contributor that are necessarily infringed by their
80 | Contribution(s) alone or by combination of their Contribution(s)
81 | with the Work to which such Contribution(s) was submitted. If You
82 | institute patent litigation against any entity (including a
83 | cross-claim or counterclaim in a lawsuit) alleging that the Work
84 | or a Contribution incorporated within the Work constitutes direct
85 | or contributory patent infringement, then any patent licenses
86 | granted to You under this License for that Work shall terminate
87 | as of the date such litigation is filed.
88 |
89 | 4. Redistribution. You may reproduce and distribute copies of the
90 | Work or Derivative Works thereof in any medium, with or without
91 | modifications, and in Source or Object form, provided that You
92 | meet the following conditions:
93 |
94 | (a) You must give any other recipients of the Work or
95 | Derivative Works a copy of this License; and
96 |
97 | (b) You must cause any modified files to carry prominent notices
98 | stating that You changed the files; and
99 |
100 | (c) You must retain, in the Source form of any Derivative Works
101 | that You distribute, all copyright, patent, trademark, and
102 | attribution notices from the Source form of the Work,
103 | excluding those notices that do not pertain to any part of
104 | the Derivative Works; and
105 |
106 | (d) If the Work includes a "NOTICE" text file as part of its
107 | distribution, then any Derivative Works that You distribute must
108 | include a readable copy of the attribution notices contained
109 | within such NOTICE file, excluding those notices that do not
110 | pertain to any part of the Derivative Works, in at least one
111 | of the following places: within a NOTICE text file distributed
112 | as part of the Derivative Works; within the Source form or
113 | documentation, if provided along with the Derivative Works; or,
114 | within a display generated by the Derivative Works, if and
115 | wherever such third-party notices normally appear. The contents
116 | of the NOTICE file are for informational purposes only and
117 | do not modify the License. You may add Your own attribution
118 | notices within Derivative Works that You distribute, alongside
119 | or as an addendum to the NOTICE text from the Work, provided
120 | that such additional attribution notices cannot be construed
121 | as modifying the License.
122 |
123 | You may add Your own copyright statement to Your modifications and
124 | may provide additional or different license terms and conditions
125 | for use, reproduction, or distribution of Your modifications, or
126 | for any such Derivative Works as a whole, provided Your use,
127 | reproduction, and distribution of the Work otherwise complies with
128 | the conditions stated in this License.
129 |
130 | 5. Submission of Contributions. Unless You explicitly state otherwise,
131 | any Contribution intentionally submitted for inclusion in the Work
132 | by You to the Licensor shall be under the terms and conditions of
133 | this License, without any additional terms or conditions.
134 | Notwithstanding the above, nothing herein shall supersede or modify
135 | the terms of any separate license agreement you may have executed
136 | with Licensor regarding such Contributions.
137 |
138 | 6. Trademarks. This License does not grant permission to use the trade
139 | names, trademarks, service marks, or product names of the Licensor,
140 | except as required for reasonable and customary use in describing the
141 | origin of the Work and reproducing the content of the NOTICE file.
142 |
143 | 7. Disclaimer of Warranty. Unless required by applicable law or
144 | agreed to in writing, Licensor provides the Work (and each
145 | Contributor provides its Contributions) on an "AS IS" BASIS,
146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147 | implied, including, without limitation, any warranties or conditions
148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149 | PARTICULAR PURPOSE. You are solely responsible for determining the
150 | appropriateness of using or redistributing the Work and assume any
151 | risks associated with Your exercise of permissions under this License.
152 |
153 | 8. Limitation of Liability. In no event and under no legal theory,
154 | whether in tort (including negligence), contract, or otherwise,
155 | unless required by applicable law (such as deliberate and grossly
156 | negligent acts) or agreed to in writing, shall any Contributor be
157 | liable to You for damages, including any direct, indirect, special,
158 | incidental, or consequential damages of any character arising as a
159 | result of this License or out of the use or inability to use the
160 | Work (including but not limited to damages for loss of goodwill,
161 | work stoppage, computer failure or malfunction, or any and all
162 | other commercial damages or losses), even if such Contributor
163 | has been advised of the possibility of such damages.
164 |
165 | 9. Accepting Warranty or Additional Liability. While redistributing
166 | the Work or Derivative Works thereof, You may choose to offer,
167 | and charge a fee for, acceptance of support, warranty, indemnity,
168 | or other liability obligations and/or rights consistent with this
169 | License. However, in accepting such obligations, You may act only
170 | on Your own behalf and on Your sole responsibility, not on behalf
171 | of any other Contributor, and only if You agree to indemnify,
172 | defend, and hold each Contributor harmless for any liability
173 | incurred by, or claims asserted against, such Contributor by reason
174 | of your accepting any such warranty or additional liability.
175 |
176 | END OF TERMS AND CONDITIONS
177 |
178 | APPENDIX: How to apply the Apache License to your work.
179 |
180 | To apply the Apache License to your work, attach the following
181 | boilerplate notice, with the fields enclosed by brackets "{}"
182 | replaced with your own identifying information. (Don't include
183 | the brackets!) The text should be enclosed in the appropriate
184 | comment syntax for the file format. We also recommend that a
185 | file or class name and description of purpose be included on the
186 | same "printed page" as the copyright notice for easier
187 | identification within third-party archives.
188 |
189 | Copyright {yyyy} {name of copyright owner}
190 |
191 | Licensed under the Apache License, Version 2.0 (the "License");
192 | you may not use this file except in compliance with the License.
193 | You may obtain a copy of the License at
194 |
195 | http://www.apache.org/licenses/LICENSE-2.0
196 |
197 | Unless required by applicable law or agreed to in writing, software
198 | distributed under the License is distributed on an "AS IS" BASIS,
199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200 | See the License for the specific language governing permissions and
201 | limitations under the License.
202 |
--------------------------------------------------------------------------------
/Stackelberg GAN/CIFAR-10/models.py:
--------------------------------------------------------------------------------
1 | from __future__ import division
2 | from __future__ import print_function
3 | from __future__ import absolute_import
4 | from functools import partial
5 |
6 | import os
7 | import numpy as np
8 | import tensorflow as tf
9 | from ops import lrelu, linear, conv2d, deconv2d
10 | from utils import make_batches, Prior, conv_out_size_same, create_image_grid, merge
11 |
12 | batch_norm = partial(tf.contrib.layers.batch_norm,
13 | decay=0.9,
14 | updates_collections=None,
15 | epsilon=1e-5,
16 | scale=True)
17 |
18 |
19 | class SGAN(object):#Stackelberg GAN
20 |
21 | def __init__(self,
22 | model_name='SGAN',
23 | beta=1.0,
24 | num_z=128,
25 | num_gens=4,
26 | d_batch_size=64,
27 | g_batch_size=32,
28 | z_prior="uniform",
29 | same_input=True,
30 | learning_rate1=0.0002,
31 | learning_rate2=0.0002,
32 | img_size=(32, 32, 3), # (height, width, channels)
33 | g_num_conv_layers=3,
34 | d_num_conv_layers=3,
35 | num_gen_feature_maps=128, # number of feature maps of generator
36 | num_dis_feature_maps=128, # number of feature maps of discriminator
37 | sample_fp=None,
38 | sample_by_gen_fp=None,
39 | num_epochs=25000,
40 | random_seed=6789,
41 | checkpoint_dir=None):
42 | self.beta = beta
43 | self.num_z = num_z
44 | self.num_gens = num_gens
45 | self.d_batch_size = d_batch_size
46 | self.g_batch_size = g_batch_size
47 | self.z_prior = Prior(z_prior)
48 | self.same_input = same_input
49 | self.learning_rate1 = learning_rate1
50 | self.learning_rate2 = learning_rate2
51 | self.num_epochs = num_epochs
52 | self.img_size = img_size
53 | self.g_num_conv_layers = g_num_conv_layers
54 | self.d_num_conv_layers = d_num_conv_layers
55 | self.num_gen_feature_maps = num_gen_feature_maps
56 | self.num_dis_feature_maps = num_dis_feature_maps
57 | self.sample_fp = sample_fp
58 | self.sample_by_gen_fp = sample_by_gen_fp
59 | self.random_seed = random_seed
60 | self.checkpoint_dir = checkpoint_dir
61 |
62 | def _init(self):
63 | self.epoch = 0
64 |
65 | # TensorFlow's initialization
66 | self.tf_graph = tf.Graph()
67 | self.tf_config = tf.ConfigProto()
68 | self.tf_config.gpu_options.allow_growth = True
69 | self.tf_config.log_device_placement = False
70 | self.tf_config.allow_soft_placement = True
71 | self.tf_session = tf.Session(config=self.tf_config, graph=self.tf_graph)
72 |
73 | np.random.seed(self.random_seed)
74 | with self.tf_graph.as_default():
75 | tf.set_random_seed(self.random_seed)
76 |
77 | def _build_model(self):
78 | arr = np.array([i // self.g_batch_size for i in range(self.g_batch_size * self.num_gens)])
79 | d_mul_labels = tf.constant(arr, dtype=tf.int32)
80 |
81 | self.x = tf.placeholder(tf.float32, [None,
82 | self.img_size[0], self.img_size[1], self.img_size[2]],
83 | name="real_data")
84 | self.z = tf.placeholder(tf.float32, [self.g_batch_size * self.num_gens, self.num_z], name='noise')
85 |
86 | # create generator G
87 | self.g = self._create_generator(self.z)
88 |
89 | # create sampler to generate samples
90 | self.sampler = self._create_generator(self.z, train=False, reuse=True)
91 |
92 | # create discriminator D
93 |
94 | d_bin_x_logits, d_mul_x_logits = self._create_discriminator(self.x)
95 | d_bin_g_logits, d_mul_g_logits = self._create_discriminator(self.g, reuse=True)
96 |
97 | # define loss functions
98 | self.d_bin_x_loss = tf.reduce_mean(
99 | tf.nn.sigmoid_cross_entropy_with_logits(
100 | logits=d_bin_x_logits, labels=tf.ones_like(d_bin_x_logits)),
101 | name='d_bin_x_loss')
102 | self.d_bin_g_loss = tf.reduce_mean(
103 | tf.nn.sigmoid_cross_entropy_with_logits(
104 | logits=d_bin_g_logits, labels=tf.zeros_like(d_bin_g_logits)),
105 | name='d_bin_g_loss')
106 | self.d_bin_loss = tf.add(self.d_bin_x_loss, self.d_bin_g_loss, name='d_bin_loss')
107 | self.d_mul_loss = tf.reduce_mean(
108 | tf.nn.sparse_softmax_cross_entropy_with_logits(
109 | logits=d_mul_g_logits, labels=d_mul_labels),
110 | name="d_mul_loss")
111 | self.d_loss = tf.add(self.d_bin_loss, tf.multiply(self.beta, self.d_mul_loss), name="d_loss")
112 |
113 | self.g_bin_loss = tf.reduce_mean(
114 | tf.nn.sigmoid_cross_entropy_with_logits(
115 | logits=d_bin_g_logits, labels=tf.ones_like(d_bin_g_logits)),
116 | name="g_bin_loss")
117 | self.g_mul_loss = tf.multiply(self.beta, self.d_mul_loss, name='g_mul_loss')
118 | self.g_loss = tf.add(self.g_bin_loss, self.g_mul_loss, name="g_loss")
119 |
120 | # create optimizers
121 | self.d_opt = self._create_optimizer(self.d_loss, scope='discriminator',
122 | lr=self.learning_rate1)
123 | self.g_opt = self._create_optimizer(self.g_loss, scope='generator',
124 | lr=self.learning_rate2)
125 | self.saver = tf.train.Saver(max_to_keep=10)
126 | def _create_generator(self, z, train=True, reuse=False, name="generator"):
127 | out_size = [(conv_out_size_same(self.img_size[0], 2),
128 | conv_out_size_same(self.img_size[1], 2),
129 | self.num_gen_feature_maps)]
130 | for i in range(self.g_num_conv_layers - 1):
131 | out_size = [(conv_out_size_same(out_size[0][0], 2),
132 | conv_out_size_same(out_size[0][1], 2),
133 | out_size[0][2] * 2)] + out_size
134 |
135 | print(out_size)
136 | with tf.variable_scope(name) as scope:
137 | if reuse:
138 | scope.reuse_variables()
139 |
140 | z_split = tf.split(z, self.num_gens, axis=0)
141 | h0 = []
142 | for i, var in enumerate(z_split):
143 | h0.append(tf.nn.relu(linear(var, out_size[0][0] * out_size[0][1] * out_size[0][2],
144 | scope='g_h0_linear{}'.format(i), stddev=0.02),name="g_h0_relu{}".format(i)))
145 |
146 | g_out = []
147 | for k, var in enumerate(h0):
148 | var = tf.reshape(var, [self.g_batch_size, out_size[0][0], out_size[0][1], out_size[0][2]])
149 | for i in range(1, self.g_num_conv_layers):
150 | var = tf.nn.relu(
151 | deconv2d(var,
152 | [self.g_batch_size, out_size[i][0], out_size[i][1], out_size[i][2]],
153 | stddev=0.02, name="g{}_h{}_deconv".format(k,i)),
154 | name="g{}_h{}_relu".format(k,i))
155 |
156 | g_out.append(tf.nn.tanh(
157 | deconv2d(var,
158 | [self.g_batch_size, self.img_size[0], self.img_size[1], self.img_size[2]],
159 | stddev=0.02, name="g{}_out_deconv".format(k,i)),
160 | name="g{}_out_tanh".format(k,i)))
161 |
162 | g_out = tf.concat(g_out, axis=0, name="g_out")
163 |
164 |
165 | return g_out
166 |
167 | def _create_discriminator(self, x, train=True, reuse=False, name="discriminator"):
168 | with tf.variable_scope(name) as scope:
169 | if reuse:
170 | scope.reuse_variables()
171 |
172 | h = x
173 | for i in range(self.d_num_conv_layers):
174 | h = lrelu(batch_norm(conv2d(h, self.num_dis_feature_maps * (2 ** i),
175 | stddev=0.02, name="d_h{}_conv".format(i)),
176 | is_training=train,
177 | scope="d_bn{}".format(i)))
178 |
179 | dim = h.get_shape()[1:].num_elements()
180 | h = tf.reshape(h, [-1, dim])
181 | d_bin_logits = linear(h, 1, scope='d_bin_logits')
182 | d_mul_logits = linear(h, self.num_gens, scope='d_mul_logits')
183 | return d_bin_logits, d_mul_logits
184 |
185 | def _create_optimizer(self, loss, scope, lr):
186 | params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope)
187 | opt = tf.train.AdamOptimizer(lr, beta1=0.5)
188 | grads = opt.compute_gradients(loss, var_list=params)
189 | train_op = opt.apply_gradients(grads)
190 | return train_op
191 |
192 | def fit(self, x):
193 | if (not hasattr(self, 'epoch')) or self.epoch == 0:
194 | self._init()
195 | with self.tf_graph.as_default():
196 | self._build_model()
197 | self.tf_session.run(tf.global_variables_initializer())
198 | if self.load():
199 | print('load the checkpoint!')
200 | else:
201 | print('cannot load the checkpoint and init all the varibale')
202 |
203 | num_data = x.shape[0] - x.shape[0] % self.d_batch_size
204 | batches = make_batches(num_data, self.d_batch_size)
205 | best_is = 0.0
206 | while (self.epoch < self.num_epochs):
207 | for batch_idx, (batch_start, batch_end) in enumerate(batches):
208 | batch_size = batch_end - batch_start
209 |
210 | x_batch = x[batch_start:batch_end]
211 | if self.same_input:
212 | z_batch = self.z_prior.sample([self.g_batch_size, self.num_z]).astype(np.float32)
213 | z_batch = np.vstack([z_batch] * self.num_gens)
214 | else:
215 | z_batch = self.z_prior.sample([self.g_batch_size * self.num_gens, self.num_z]).astype(np.float32)
216 |
217 | # update discriminator D
218 | d_bin_loss, d_mul_loss, d_loss, _ = self.tf_session.run(
219 | [self.d_bin_loss, self.d_mul_loss, self.d_loss, self.d_opt],
220 | feed_dict={self.x: x_batch, self.z: z_batch})
221 |
222 | # update generator G
223 | g_bin_loss, g_mul_loss, g_loss, _ = self.tf_session.run(
224 | [self.g_bin_loss, self.g_mul_loss, self.g_loss, self.g_opt],
225 | feed_dict={self.z: z_batch})
226 |
227 | self.epoch += 1
228 | print("Epoch: [%4d/%4d] d_bin_loss: %.5f, d_mul_loss: %.5f, d_loss: %.5f,"
229 | " g_bin_loss: %.5f, g_mul_loss: %.5f, g_loss: %.5f" % (self.epoch, self.num_epochs,
230 | d_bin_loss, d_mul_loss, d_loss, g_bin_loss, g_mul_loss, g_loss))
231 | # print("Epoch: [%4d/%4d] d_bin_loss: %.5f,g_bin_loss: %.5f" % (self.epoch, self.num_epochs,
232 | # d_bin_loss, g_bin_loss))
233 | if self.epoch%10 == 0:
234 | self._samples(self.sample_fp.format(epoch=self.epoch+1))
235 |
236 | if not os.path.exists(self.checkpoint_dir):
237 | os.makedirs(self.checkpoint_dir)
238 | self.saver.save(self.tf_session, os.path.join(self.checkpoint_dir, "classifier_mode_checkpoint"))#+str(self.num_gens)+"epoch_"+str(self.num_epochs)+"num_g_maps_"+str(self.num_gen_feature_maps)))
239 | self._samples_by_gen(self.sample_by_gen_fp)
240 |
241 | def predict(self):
242 | if (not hasattr(self, 'epoch')) or self.epoch == 0:
243 | self._init()
244 | with self.tf_graph.as_default():
245 | self._build_model()
246 | self.tf_session.run(tf.global_variables_initializer())
247 | if self.load():
248 | print('load the checkpoint!')
249 | else:
250 | print('cannot load the checkpoint and init all the varibale')
251 | self._samples_by_gen(self.sample_by_gen_fp)
252 |
253 |
254 | def _generate(self, num_samples=100):
255 | sess = self.tf_session
256 | batch_size = self.g_batch_size * self.num_gens
257 | num = ((num_samples - 1) // batch_size + 1) * batch_size
258 | z = self.z_prior.sample([num, self.num_z]).astype(np.float32)
259 | x = np.zeros([num, self.img_size[0], self.img_size[1], self.img_size[2]],
260 | dtype=np.float32)
261 | batches = make_batches(num, batch_size)
262 | for batch_idx, (batch_start, batch_end) in enumerate(batches):
263 | z_batch = z[batch_start:batch_end]
264 | x[batch_start:batch_end] = sess.run(self.sampler,
265 | feed_dict={self.z: z_batch})
266 | f_x = np.reshape(x, [self.num_gens,-1, self.img_size[0], self.img_size[1], self.img_size[2]])
267 | f_x = f_x[:,0::7,:,:,:]
268 | f_x = np.reshape(f_x,[num_samples,self.img_size[0], self.img_size[1], self.img_size[2]])
269 | # idx = np.random.permutation(num)[:num_samples]
270 | # x = (x[idx] + 1.0) / 2.0
271 | # x = x[idx]
272 | x = (f_x+1)/2
273 | return x
274 |
275 | def _samples(self, filepath, tile_shape=(10, 10)):
276 | if not os.path.exists(os.path.dirname(filepath)):
277 | os.makedirs(os.path.dirname(filepath))
278 |
279 | num_samples = tile_shape[0] * tile_shape[1]
280 | x = self._generate(num_samples)
281 | imgs = create_image_grid(x, img_size=self.img_size, tile_shape=tile_shape)
282 | import scipy.misc
283 | scipy.misc.imsave(filepath, imgs)
284 |
285 | def _samples_by_gen(self, filepath):
286 | if not os.path.exists(filepath):
287 | os.makedirs(filepath)
288 |
289 | num_samples = self.num_gens * 4000
290 | tile_shape = (1,1)
291 |
292 | sess = self.tf_session
293 | img_per_gen = num_samples // self.num_gens
294 | # x = np.zeros([num_samples, self.img_size[0], self.img_size[1], self.img_size[2]],
295 | # dtype=np.float32)
296 | counter = 0
297 | for i in range(0, img_per_gen, self.g_batch_size):
298 | z_batch = self.z_prior.sample([self.g_batch_size * self.num_gens, self.num_z]).astype(np.float32)
299 | samples = sess.run(self.sampler, feed_dict={self.z: z_batch})
300 |
301 | for gen in range(self.num_gens):
302 |
303 | tmp_ = samples[gen * self.g_batch_size:gen * self.g_batch_size + min(self.g_batch_size, img_per_gen)]
304 | for x in tmp_:
305 | counter = counter+1
306 | x = x.reshape(1, 32, 32, 3)
307 | x = (x + 1.0) / 2.0
308 | imgs = merge(x, [1,1])
309 | import scipy.misc
310 | scipy.misc.imsave(os.path.join(filepath, "samples_"+str(counter)+".png"), imgs)
311 | def load(self):
312 | print('Being to load the checkpoint')
313 | ckpt = tf.train.get_checkpoint_state(self.checkpoint_dir)
314 | if ckpt and ckpt.model_checkpoint_path:
315 | self.saver.restore(self.tf_session, ckpt.model_checkpoint_path)
316 | return True
317 | else:
318 | return False
319 |
--------------------------------------------------------------------------------