Web page
Contributions
In now, this repo contains general architectures and functions that are useful for the GAN and classificstion.
I will continue to add useful things to other areas.
Also, your pull requests and issues are always welcome.
And write what you want to implement on the issue. I'll implement it.
How to use
Import
ops.py
- operations
- from ops import *
utils.py
- image processing
- from utils import *
Network template
def network(x, is_training=True, reuse=False, scope="network"):
with tf.variable_scope(scope, reuse=reuse):
x = conv(...)
...
return logit
Insert data to network using DatasetAPI
Image_Data_Class = ImageData(img_size, img_ch, augment_flag)
trainA_dataset = ['./dataset/cat/trainA/a.jpg',
'./dataset/cat/trainA/b.png',
'./dataset/cat/trainA/c.jpeg',
...]
trainA = tf.data.Dataset.from_tensor_slices(trainA_dataset)
trainA = trainA.map(Image_Data_Class.image_processing, num_parallel_calls=16)
trainA = trainA.shuffle(buffer_size=10000).prefetch(buffer_size=batch_size).batch(batch_size).repeat()
trainA_iterator = trainA.make_one_shot_iterator()
data_A = trainA_iterator.get_next()
logit = network(data_A)
- See this for more information.
Option
padding='SAME'
- pad = ceil[ (kernel - stride) / 2 ]
pad_type
- 'zero' or 'reflect'
sn
- use spectral_normalization or not
Caution
- If you don't want to share variable, set all scope names differently.
Weight
weight_init = tf.truncated_normal_initializer(mean=0.0, stddev=0.02)
weight_regularizer = tf.contrib.layers.l2_regularizer(0.0001)
weight_regularizer_fully = tf.contrib.layers.l2_regularizer(0.0001)
Initialization
Xavier
: tf.contrib.layers.xavier_initializer()He
: tf.contrib.layers.variance_scaling_initializer()Normal
: tf.random_normal_initializer(mean=0.0, stddev=0.02)Truncated_normal
: tf.truncated_normal_initializer(mean=0.0, stddev=0.02)Orthogonal
: tf.orthogonal_initializer(1.0) / # if relu = sqrt(2), the others = 1.0
Regularization
l2_decay
: tf.contrib.layers.l2_regularizer(0.0001)orthogonal_regularizer
: orthogonal_regularizer(0.0001) & orthogonal_regularizer_fully(0.0001)
Convolution
basic conv
x = conv(x, channels=64, kernel=3, stride=2, pad=1, pad_type='reflect', use_bias=True, sn=True, scope='conv')
Partial Convolution)
partial conv (NVIDIAx = partial_conv(x, channels=64, kernel=3, stride=2, use_bias=True, padding='SAME', sn=True, scope='partial_conv')
dilated conv
x = dilate_conv(x, channels=64, kernel=3, rate=2, use_bias=True, padding='VALID', sn=True, scope='dilate_conv')
Deconvolution
basic deconv
x = deconv(x, channels=64, kernel=3, stride=1, padding='SAME', use_bias=True, sn=True, scope='deconv')
Fully-connected
x = fully_connected(x, units=64, use_bias=True, sn=True, scope='fully_connected')
Pixel shuffle
x = conv_pixel_shuffle_down(x, scale_factor=2, use_bias=True, sn=True, scope='pixel_shuffle_down')
x = conv_pixel_shuffle_up(x, scale_factor=2, use_bias=True, sn=True, scope='pixel_shuffle_up')
down
===> [height, width] -> [height // scale_factor, width // scale_factor]up
===> [height, width] -> [height * scale_factor, width * scale_factor]
Block
residual block
x = resblock(x, channels=64, is_training=is_training, use_bias=True, sn=True, scope='residual_block')
x = resblock_down(x, channels=64, is_training=is_training, use_bias=True, sn=True, scope='residual_block_down')
x = resblock_up(x, channels=64, is_training=is_training, use_bias=True, sn=True, scope='residual_block_up')
down
===> [height, width] -> [height // 2, width // 2]up
===> [height, width] -> [height * 2, width * 2]
dense block
x = denseblock(x, channels=64, n_db=6, is_training=is_training, use_bias=True, sn=True, scope='denseblock')
n_db
===> The number of dense-block
residual-dense block
x = res_denseblock(x, channels=64, n_rdb=20, n_rdb_conv=6, is_training=is_training, use_bias=True, sn=True, scope='res_denseblock')
n_rdb
===> The number of RDBn_rdb_conv
===> per RDB conv layer
attention block
x = self_attention(x, channels=64, use_bias=True, sn=True, scope='self_attention')
x = self_attention_with_pooling(x, channels=64, use_bias=True, sn=True, scope='self_attention_version_2')
x = squeeze_excitation(x, channels=64, ratio=16, use_bias=True, sn=True, scope='squeeze_excitation')
x = convolution_block_attention(x, channels=64, ratio=16, use_bias=True, sn=True, scope='convolution_block_attention')
Normalization
x = batch_norm(x, is_training=is_training, scope='batch_norm')
x = layer_norm(x, scope='layer_norm')
x = instance_norm(x, scope='instance_norm')
x = group_norm(x, groups=32, scope='group_norm')
x = pixel_norm(x)
x = batch_instance_norm(x, scope='batch_instance_norm')
x = switch_norm(x, scope='switch_norm')
x = condition_batch_norm(x, z, is_training=is_training, scope='condition_batch_norm'):
x = adaptive_instance_norm(x, gamma, beta)
Activation
x = relu(x)
x = lrelu(x, alpha=0.2)
x = tanh(x)
x = sigmoid(x)
x = swish(x)
x = elu(x)
Pooling & Resize
x = up_sample(x, scale_factor=2)
x = max_pooling(x, pool_size=2)
x = avg_pooling(x, pool_size=2)
x = global_max_pooling(x)
x = global_avg_pooling(x)
x = flatten(x)
x = hw_flatten(x)
Loss
classification loss
loss, accuracy = classification_loss(logit, label)
pixel loss
loss = L1_loss(x, y)
loss = L2_loss(x, y)
loss = huber_loss(x, y)
loss = histogram_loss(x, y)
loss = gram_style_loss(x, y)
histogram_loss
means the difference in the color distribution of the image pixel values.gram_style_loss
means the difference between the styles using gram matrix.
gan loss
d_loss = discriminator_loss(Ra=True, loss_func='wgan-gp', real=real_logit, fake=fake_logit)
g_loss = generator_loss(Ra=True, loss_func='wgan-gp', real=real_logit, fake=fake_logit)
Ra
- use relativistic gan or not
loss_func
- gan
- lsgan
- hinge
- wgan-gp
- dragan
- See this for how to use
gradient_penalty
vdb loss
d_bottleneck_loss = vdb_loss(real_mu, real_logvar, i_c) + vdb_loss(fake_mu, fake_logvar, i_c)
kl-divergence (z ~ N(0, 1))
loss = kl_loss(mean, logvar)
Author
Junho Kim