今天為大家推薦一個實用的GitHub項目：TensorFlow-Cookbook。 這是一個易用的TensorFlow代碼集，包含了對GAN有用的一些通用架構(gòu)和函數(shù)。

今天為大家推薦一個實用的GitHub項目：TensorFlow-Cookbook。

這是一個易用的TensorFlow代碼集，作者是來自韓國的AI研究科學家Junho Kim，內(nèi)容涵蓋了譜歸一化卷積、部分卷積、pixel shuffle、幾種歸一化函數(shù)、 tf-datasetAPI，等等。

作者表示，這個repo包含了對GAN有用的一些通用架構(gòu)和函數(shù)。

項目正在進行中，作者將持續(xù)為其他領(lǐng)域添加有用的代碼，目前正在添加的是 tf-Eager mode的代碼。歡迎提交pull requests和issues。

Github地址 :

https://github.com/taki0112/Tensorflow-Cookbook

如何使用

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

使用DatasetAPI向網(wǎng)絡插入數(shù)據(jù)

Image_Data_Class = ImageData(img_size, img_ch, augment_flag) 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)

了解更多，請閱讀：

https://github.com/taki0112/Tensorflow-DatasetAPI

Option

padding='SAME'

pad = ceil[ (kernel - stride) / 2 ]

pad_type

'zero' or 'reflect'

sn

usespectral_normalizationor not

Ra

userelativistic ganor not

loss_func

gan

lsgan

hinge

wgan

wgan-gp

dragan

注意

如果你不想共享變量，請以不同的方式設置所有作用域名稱。

權(quán)重(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)

初始化(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

正則化(Regularization)

l2_decay: tf.contrib.layers.l2_regularizer(0.0001)

orthogonal_regularizer: orthogonal_regularizer(0.0001) & orthogonal_regularizer_fully(0.0001)

卷積(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 conv (NVIDIAPartial Convolution)

x = 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='SAME', sn=True, scope='dilate_conv')

Deconvolution

basic deconv

x = deconv(x, channels=64, kernel=3, stride=2, padding='SAME', use_bias=True, sn=True, scope='deconv')

Fully-connected

x = fully_conneted(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]

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

Normalization

x = batch_norm(x, is_training=is_training, scope='batch_norm') x = instance_norm(x, scope='instance_norm') x = layer_norm(x, scope='layer_norm') x = group_norm(x, groups=32, scope='group_norm') x = pixel_norm(x) x = batch_instance_norm(x, scope='batch_instance_norm') x = condition_batch_norm(x, z, is_training=is_training, scope='condition_batch_norm'): x = adaptive_instance_norm(x, gamma, beta):

如何使用condition_batch_norm，請參考：

https://github.com/taki0112/BigGAN-Tensorflow

如何使用adaptive_instance_norm，請參考：

https://github.com/taki0112/MUNIT-Tensorflow

Activation

x = relu(x) x = lrelu(x, alpha=0.2) x = tanh(x) x = sigmoid(x) x = swish(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)

histogram_loss表示圖像像素值在顏色分布上的差異。

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)

如何使用gradient_penalty，請參考：

https://github.com/taki0112/BigGAN-Tensorflow/blob/master/BigGAN_512.py#L180

kl-divergence (z ~ N(0, 1))

loss = kl_loss(mean, logvar)

Author

Junho Kim

Github地址 :

https://github.com/taki0112/Tensorflow-Cookbook