Tensorflow1 Cookbook

Contributions

In now, this repo provides general architectures and functions that are useful for the GAN and classification.

I will continue to add useful things to other areas.

Also, your pull requests and issues are always welcome.

And tell me what you want to implement on the issue. I'll implement it.

Functional vs Sequential

Functional API [Template code]

Pros

• More fast than Sequential
• More easy to create a flexible model architecture
  • Easy to use some layer operaions like concatenate, add , ...

Cons

• Define tf.keras.layers.Input
  • You have to know the shape of input tensor
• Define tf.keras.Model
  • You have to create the Model constructor at the end

Sequential API [Template code]

Pros

• Simple to use
  • Similar to Pytorch style

Cons

• Hard to create a flexible model architecture

Example code

• CycleGAN

How to use

1. Import

Funtional API

• ops_functional.py
  • Functional API operations
  • from ops_functional import *

Sequential API

• ops_sequential.py
  • Sequential API operations
  • from ops_sequential import *

Common

• utils.py
  • image processing + something useful functions (e.g. automatic_gpu_usage)
    • automatic_gpu_usage : Automatically manage gpu memory
    • multiple_gpu_usage : You can set gpu memory limit
  • from utils import *

2. Network template

Functional API

from ops_functional import *
from utils import *

automatic_gpu_usage() # for efficient gpu use

input_shape = [img_height, img_width, img_ch]
inputs = tf.keras.layers.Input(input_shape, name='input')

# architecture
x = conv(inputs, channels=64, kernel=3, stride=2, pad=1, pad_type='reflect', use_bias=False, sn=False, name='conv')
x = batch_norm(x, name='batch_norm')
x = relu(x)

x = global_avg_pooling(x)
x = fully_connected(x, units=10, sn=False, name='fc')

model = tf.keras.Model(inputs, s, name='model')
optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)

Sequential API

from ops_sequential import *
from utils import *

automatic_gpu_usage() # for efficient gpu use

model = []

model += [Conv(channels=64, kernel=3, stride=2, pad=1, pad_type='reflect', use_bias=False, sn=False, name='conv')]
model += [BatchNorm(name)]
model += [Relu()]

model += [Global_Avg_Pooling()]
model += [FullyConnected(units=10, sn=False, name='fc')]

model = Sequential(model, name='model')
optimizer = tf.keras.optimizers.Adam(learning_rate=0.01)

3. Data pipeline

img_class = Image_data(img_height, img_width, img_ch, dataset_path, augment_flag)
img_class.preprocess()

img_slice = tf.data.Dataset.from_tensor_slices(img_class.dataset)
gpu_device = '/gpu:0'
img_slice = img_slice. \
                apply(shuffle_and_repeat(dataset_num)). \
                apply(map_and_batch(img_class.image_processing, self.batch_size,
                                    num_parallel_batches=AUTOTUNE,
                                    drop_remainder=True)). \
                apply(prefetch_to_device(gpu_device, AUTOTUNE))

dataset_iter = iter(img_slice)

4. Restore

ckpt = tf.train.Checkpoint(model=model, optimizer=optimizer)
manager = tf.train.CheckpointManager(ckpt, checkpoint_dir, max_to_keep=2)
start_iteration = 0

if manager.latest_checkpoint:
  ckpt.restore(manager.latest_checkpoint)
  start_iteration = int(manager.latest_checkpoint.split('-')[-1])
  print('Latest checkpoint restored!!')
else:
  print('Not restoring from saved checkpoint')

5-1. Train

def train_step(img):
  with tf.GradientTape() as tape:
    logit = model(img)
    
    # calculate loss
    """
    if classification
    your_loss = cross_entropy_loss(logit, label)
    """
    
    loss = your_loss + regularization_loss(model)
  
  train_variable = model.trainable_variables
  gradient = tape.gradient(loss, train_variable)
  optimizer.apply_gradients(zip(gradient, train_variable))
  
  return loss

def train():
  # setup tensorboard
  summary_writer = tf.summary.create_file_writer(log_dir)
  
  for idx in range(start_iteration, total_iteration):
    img = next(dataset_iter)
    
    
    # update network
    loss = train_step(img)
    
    # save to tensorboard
    with summary_writer.as_default():
      tf.summary.scalar('loss', loss, step=idx)
    
    # save ckpt
    manager.save(checkpoint_number=idx + 1)
  
  # save model for final step
  manager.save(checkpoint_number=total_iteration)

5-2. Multi-GPUs train [Template code]

strategy = tf.distribute.MirroredStrategy()
NUM_GPUS = strategy.num_replicas_in_sync

total_iteration = iteration // NUM_GPUS

with strategy.scope():
  # copy & paste
  # 2. Network template
  # 3. Data pipeline
  # 4. Restore

def train_step(img):
  """ SAME """
  
def distribute_train_step(img):
  with strategy.scope():
    loss = strategy.experimental_run_v2(train_step, args=(img))
    
    loss = strategy.reduce(tf.distribute.ReduceOp.MEAN, loss, axis=None)
    
    return loss

def train():
  # setup tensorboard
  summary_writer = tf.summary.create_file_writer(log_dir)
  
  for idx in range(start_iteration, total_iteration):
    img = next(dataset_iter)
    
    # update network
    loss = distribute_train_step(img)
    
    """
    SAME
    """

Weight

weight_initializer = tf.initializers.RandomNormal(mean=0.0, stddev=0.02)
weight_regularizer = tf.keras.regularizers.l2(0.0001)
weight_regularizer_fully = tf.keras.regularizers.l2(0.0001)

Initialization

• Xavier : tf.initializers.GlorotUniform() or tf.initializers.GlorotNormal()
• He : tf.initializers.VarianceScaling()
• Normal : tf.initializers.RandomNormal(mean=0.0, stddev=0.02)
• Truncated normal : tf.initializers.TruncatedNormal(mean=0.0, stddev=0.02)
• Orthogonal : tf.initializers.Orthogonal0.02)

Regularization

• l2_decay : tf.keras.regularizers.l2(0.0001)
• orthogonal_regularizer : orthogonal_regularizer(0.0001) # orthogonal_regularizer_fully(0.0001)

Option

• padding='SAME'
  • pad = ceil[ (kernel - stride) / 2 ]
• pad_type
  • 'zero' or 'reflect'
• sn
  • use spectral normalization of not

Examples of Functional API

Recurrent

x = various_rnn(x, n_hidden=128, n_layer=2, dropout_rate=0.5, training=True, bidirectional=True, rnn_type='lstm', name='rnn')

LSTM

GRU

Bidirectional

Deep (n_layer > 1)

Convolution

basic conv

x = conv(x, channels=64, kernel=3, stride=2, pad=1, pad_type='reflect', use_bias=True, sn=True, name='conv')

partial conv (NVIDIA Partial Convolution)

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

dilated conv

x = dilate_conv(x, channels=64, kernel=3, rate=2, use_bias=True, padding='VALID', sn=True, name='dilate_conv')

Deconvolution

basic deconv

x = deconv(x, channels=64, kernel=3, stride=1, padding='SAME', use_bias=True, sn=True, name='deconv')

Fully-connected

x = fully_connected(x, units=64, use_bias=True, sn=True, snamecope='fully_connected')

Pixel shuffle

x = conv_pixel_shuffle_down(x, scale_factor=2, use_bias=True, sn=True, name='pixel_shuffle_down')
x = conv_pixel_shuffle_up(x, scale_factor=2, use_bias=True, sn=True, name='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, name='residual_block')
x = resblock_down(x, channels=64, is_training=is_training, use_bias=True, sn=True, name='residual_block_down')
x = resblock_up(x, channels=64, is_training=is_training, use_bias=True, sn=True, name='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, name='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, name='res_denseblock')

• n_rdb ===> The number of RDB
• n_rdb_conv ===> per RDB conv layer

attention block

x = self_attention(x, use_bias=True, sn=True, name='self_attention')
x = self_attention_with_pooling(x, use_bias=True, sn=True, name='self_attention_version_2')

x = squeeze_excitation(x, ratio=16, use_bias=True, sn=True, name='squeeze_excitation')

x = convolution_block_attention(x, ratio=16, use_bias=True, sn=True, name='convolution_block_attention')

x = global_context_block(x, use_bias=True, sn=True, name='gc_block')

x = srm_block(x, use_bias=False, is_training=is_training, name='srm_block')

Normalization

x = batch_norm(x, training=training, name='batch_norm')
x = layer_norm(x, name='layer_norm')
x = instance_norm(x, name='instance_norm')
x = group_norm(x, groups=32, name='group_norm')

x = pixel_norm(x)

x = batch_instance_norm(x, name='batch_instance_norm')
x = layer_instance_norm(x, name='layer_instance_norm')
x = switch_norm(x, scope='switch_norm')

x = condition_batch_norm(x, z, training=training, name='condition_batch_norm'):

x = adaptive_instance_norm(x, gamma, beta)
x = adaptive_layer_instance_norm(x, gamma, beta, smoothing=True, name='adaLIN')

• See this for how to use condition_batch_norm
• See this for how to use adaptive_instance_norm
• See this for how to use adaptive_layer_instance_norm & layer_instance_norm

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 = nearest_up_sample(x, scale_factor=2)
x = bilinear_up_sample(x, scale_factor=2)
x = nearest_down_sample(x, scale_factor=2)
x = bilinear_down_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)

loss = dice_loss(n_classes=10, logit, label)

regularization loss

model_reg_loss = regularization_loss(model)

• If you want to use regularizer, then you should write it

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)

loss = color_consistency_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.
• color_consistency_loss means the color difference between the generated image and the input image.

gan loss

d_loss = discriminator_loss(Ra=True, gan_type='wgan-gp', real_logit=real_logit, fake_logit=fake_logit)
g_loss = generator_loss(Ra=True, gan_type='wgan-gp', real_logit=real_logit, fake_logit=fake_logit)

• Ra
  • use relativistic gan or not
• loss_func
  • gan
  • lsgan
  • hinge
  • wgan-gp
  • dragan
  • realness
  • sphere

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