def wrapper(tensor, **kwargs):
r"""Manages arguments of `tf.sg_opt`.
Args:
tensor: automatically passed by decorator
kwargs:
in_dim: An integer. The size of input dimension, which is set to the last one by default.
dim: An integer. The size of output dimension. Has the same value as in_dim by default.
ln: Boolean. If True, layer normalization is applied.
bias: Boolean. If True, biases are added. As a default, it is set to True
name: A name for the layer. As a default, the function name is assigned.
reuse: `True` or `None`; if `True`, we go into reuse mode for this `layer` scope
as well as all sub-scopes; if `None`, we just inherit the parent scope reuse.
"""
# kwargs parsing
opt = tf.sg_opt(kwargs) + _context
# set default argument
try:
shape = tensor.get_shape().as_list()
# dropout off
opt += tf.sg_opt(shape=shape,
in_dim=shape[-1],
dim=shape[-1],
dout=0)
# disable bias when normalization on
opt += tf.sg_opt(bias=not opt.ln)
finally:
pass
# automatic layer naming
if opt.name is None:
# layer function name will be used as layer name
opt.name = func.__name__.replace('sg_', 'lyr-')
# find existing layer names
exist_layers = []
for t in tf.global_variables():
scope_name = tf.get_variable_scope().name
prefix = scope_name + '/' if len(scope_name) > 0 else ''
i = t.name.rfind(prefix + opt.name)
if i >= 0:
exist_layers.append(t.name[i:].split('/')[-2])
exist_layers = list(set(exist_layers))
# layer name numbering
if len(exist_layers) == 0:
opt.name += '_1'
else:
opt.name += '_%d' % (
max([int(n.split('_')[-1]) for n in exist_layers]) + 1)
# all layer variables start with 'lyr-' prefix
with tf.variable_scope(opt.name, reuse=opt.reuse) as scope:
# call layer function
out = func(tensor, opt)
# apply dropout
if opt.dout:
out = tf.cond(_phase, lambda: tf.nn.dropout(out, 1 - opt.dout),
lambda: out)
# rename tensor
out = tf.identity(out, 'out')
# add final output summary
if scope.reuse:
tf.sg_summary_activation(out)
# save node info for reuse
out._sugar = tf.sg_opt(func=func,
arg=tf.sg_opt(kwargs) + _context,
prev=tensor,
is_layer=True,
name=opt.name)
# inject reuse function
out.sg_reuse = types.MethodType(sg_reuse, out)
return out
def wrapper(tensor, **kwargs):
r"""Manages arguments of `tf.sg_opt`.
Args:
tensor: A `tensor` (automatically passed by decorator).
kwargs:
shape: A list of integers. The shape of `tensor`. Inferred if not specified.
in_dim: An integer. The size of input dimension, which is set to the last one by default.
dim: An integer. The size of output dimension. Has the same value as in_dim by default.
bn: Boolean. If True, batch normalization is applied.
ln: Boolean. If True, layer normalization is applied.
dout: A float of range [0, 100). A dropout rate. Set to 0 by default.
bias: Boolean. If True, biases are added. As a default, it is set to True
name: A name for the layer. As a default, the function name is assigned.
act: A name of activation function. e.g., `sigmoid`, `tanh`, etc.
reuse: `True` or `None`; if `True`, we go into reuse mode for this `layer` scope
as well as all sub-scopes; if `None`, we just inherit the parent scope reuse.
"""
from . import sg_initializer as init
from . import sg_activation
# kwargs parsing
opt = tf.sg_opt(kwargs) + _context
# set default argument
try:
shape = tensor.get_shape().as_list()
# batch normalization off, layer normalization off, dropout off
opt += tf.sg_opt(shape=shape,
in_dim=shape[-1],
dim=shape[-1],
bn=False,
ln=False,
dout=0)
assert not (
opt.bn and opt.ln
), 'one of batch normalization and layer normalization is available.'
# disable bias when normalization on
opt += tf.sg_opt(bias=not (opt.bn or opt.ln))
finally:
pass
# automatic layer naming
if opt.name is None:
# layer function name will be used as layer name
opt.name = func.__name__.replace('sg_', '')
# find existing layer names
exist_layers = []
for t in tf.global_variables():
scope_name = tf.get_variable_scope().name
prefix = scope_name + '/' if len(scope_name) > 0 else ''
i = t.name.rfind(prefix + opt.name)
if i >= 0:
exist_layers.append(t.name[i:].split('/')[-2])
exist_layers = list(set(exist_layers))
# layer name numbering
if len(exist_layers) == 0:
opt.name += '_1'
else:
opt.name += '_%d' % (
max([int(n.split('_')[-1]) for n in exist_layers]) + 1)
# all layer variables start with 'lyr-' prefix
with tf.variable_scope(opt.name, reuse=opt.reuse) as scope:
# call layer function
out = func(tensor, opt)
# apply batch normalization
if opt.bn:
# offset, scale parameter
beta = init.constant('beta', opt.dim, summary=False)
gamma = init.constant('gamma', opt.dim, value=1, summary=False)
# offset, scale parameter
mean_running = init.constant('mean', opt.dim, summary=False)
variance_running = init.constant('variance',
opt.dim,
value=1,
summary=False)
# calc batch mean, variance
mean, variance = tf.nn.moments(
out, axes=range(len(out.get_shape()) - 1))
# update running mean, variance
def update_running_stat():
decay = 0.99
update_op = [
mean_running.assign(mean_running * decay + mean *
(1 - decay)),
variance_running.assign(variance_running * decay +
variance * (1 - decay))
]
with tf.control_dependencies(update_op):
return tf.identity(mean), tf.identity(variance)
# select mean, variance by training phase
m, v = tf.cond(
_phase,
update_running_stat, # updated running stat and batch mean, variance
lambda:
(mean_running, variance_running)) # saved mean, variance
# apply batch normalization
out = tf.nn.batch_normalization(out, m, v, beta, gamma,
tf.sg_eps)
# apply normalization parameters
if opt.ln:
# offset, scale parameter
beta = init.constant('beta', opt.dim, summary=False)
gamma = init.constant('gamma', opt.dim, value=1, summary=False)
# calc layer mean, variance for final axis
mean, variance = tf.nn.moments(out,
axes=[len(out.get_shape()) - 1],
keep_dims=True)
# apply normalization
out = (out - mean) / tf.sqrt(variance + tf.sg_eps)
# apply parameter
out = gamma * out + beta
# apply activation
if opt.act:
out = getattr(sg_activation, 'sg_' + opt.act.lower())(out)
# apply dropout
if opt.dout:
out = tf.cond(_phase, lambda: tf.nn.dropout(out, 1 - opt.dout),
lambda: out)
# rename tensor
out = tf.identity(out, 'out')
# add final output summary
if not scope.reuse:
tf.sg_summary_activation(out)
# save node info for reuse
out._sugar = tf.sg_opt(func=func,
arg=tf.sg_opt(kwargs) + _context,
prev=tensor,
is_layer=True,
name=opt.name)
# inject reuse function
out.sg_reuse = types.MethodType(sg_reuse, out)
return out
def wrapper(tensor, **kwargs):
import sg_initializer as init
import sg_activation
# kwargs parsing
opt = tf.sg_opt(kwargs) + _context
# set default argument
try:
shape = tensor.get_shape().as_list()
# batch normalization off, layer normalization off, dropout off
opt += tf.sg_opt(shape=shape,
in_dim=shape[-1],
dim=shape[-1],
bn=False,
ln=False,
dout=0)
assert not (
opt.bn and opt.ln
), 'one of batch normalization and layer normalization is available.'
# disable bias when normalization on
opt += tf.sg_opt(bias=not (opt.bn or opt.ln))
finally:
pass
# automatic layer naming
if opt.name is None:
# layer function name will be used as layer name
opt.name = func.__name__.replace('sg_', '')
# find existing layer names
exist_layers = []
for t in tf.get_collection(tf.GraphKeys.VARIABLES):
scope_name = tf.get_variable_scope().name
prefix = scope_name + '/' if len(scope_name) > 0 else ''
i = t.name.rfind(prefix + 'layers/' + opt.name)
if i >= 0:
exist_layers.append(t.name[i:].split('/')[-2])
exist_layers = list(set(exist_layers))
# layer name numbering
if len(exist_layers) == 0:
opt.name += '_1'
else:
opt.name += '_%d' % (
max([int(n.split('_')[-1]) for n in exist_layers]) + 1)
# all layer variables start with 'layers/' prefix
with tf.variable_scope('layers', reuse=opt.reuse):
with tf.variable_scope(opt.name):
# call layer function
out = func(tensor, opt)
# apply batch normalization
if opt.bn:
# offset, scale parameter
beta = init.constant('beta', opt.dim)
gamma = init.constant('gamma', opt.dim, value=1)
# offset, scale parameter
mean_running = init.constant('mean', opt.dim)
variance_running = init.constant('variance',
opt.dim,
value=1)
# calc batch mean, variance
mean, variance = tf.nn.moments(
out, axes=range(len(out.get_shape()) - 1))
# update running mean, variance
def update_running_stat():
decay = 0.99
update_op = [
mean_running.assign(mean_running * decay + mean *
(1 - decay)),
variance_running.assign(variance_running * decay +
variance * (1 - decay))
]
with tf.control_dependencies(update_op):
return tf.identity(mean), tf.identity(variance)
# select mean, variance by training phase
m, v = tf.cond(
_phase,
update_running_stat, # updated running stat and batch mean, variance
lambda: (mean_running, variance_running)
) # saved mean, variance
# apply batch normalization
out = tf.nn.batch_normalization(out, m, v, beta, gamma,
tf.sg_eps)
# apply normalization parameters
if opt.ln:
# offset, scale parameter
beta = init.constant('beta', opt.dim)
gamma = init.constant('gamma', opt.dim, value=1)
# calc layer mean, variance for final axis
mean, variance = tf.nn.moments(
out, axes=[len(out.get_shape()) - 1], keep_dims=True)
# apply normalization
out = (out - mean) / tf.sqrt(variance + tf.sg_eps)
# apply parameter
out = gamma * out + beta
# apply activation
if opt.act:
out = getattr(sg_activation, 'sg_' + opt.act.lower())(out)
# apply dropout
if opt.dout:
out = tf.cond(_phase,
lambda: tf.nn.dropout(out, 1 - opt.dout),
lambda: out)
# rename tensor
out = tf.identity(out, 'out')
# add final output summary
if opt.reuse is None or not opt.reuse:
tf.sg_summary_activation(out)
# save node info for reuse
out._sugar = tf.sg_opt(func=func,
arg=tf.sg_opt(kwargs) + _context,
prev=tensor,
is_layer=True,
name=opt.name)
# inject reuse function
out.sg_reuse = types.MethodType(sg_reuse, out)
return out
def wrapper(tensor, **kwargs):
# kwargs parsing
opt = tf.sg_opt(kwargs) + _context
# set default argument
try:
shape = tensor.get_shape().as_list()
# dropout off
opt += tf.sg_opt(shape=shape,
in_dim=shape[-1],
dim=shape[-1],
dout=0)
# disable bias when normalization on
opt += tf.sg_opt(bias=not opt.ln)
finally:
pass
# automatic layer naming
if opt.name is None:
# layer function name will be used as layer name
opt.name = func.__name__.replace('sg_', '')
# find existing layer names
exist_layers = []
for t in tf.get_collection(tf.GraphKeys.VARIABLES):
scope_name = tf.get_variable_scope().name
prefix = scope_name + '/' if len(scope_name) > 0 else ''
i = t.name.rfind(prefix + 'layers/' + opt.name)
if i >= 0:
exist_layers.append(t.name[i:].split('/')[-2])
exist_layers = list(set(exist_layers))
# layer name numbering
if len(exist_layers) == 0:
opt.name += '_1'
else:
opt.name += '_%d' % (
max([int(n.split('_')[-1]) for n in exist_layers]) + 1)
# all layer variables start with 'layers/' prefix
with tf.variable_scope('layers', reuse=opt.reuse):
with tf.variable_scope(opt.name):
# call layer function
out = func(tensor, opt)
# apply dropout
if opt.dout:
out = tf.cond(_phase,
lambda: tf.nn.dropout(out, 1 - opt.dout),
lambda: out)
# rename tensor
out = tf.identity(out, 'out')
# add final output summary
if opt.reuse is None or not opt.reuse:
tf.sg_summary_activation(out)
# save node info for reuse
out._sugar = tf.sg_opt(func=func,
arg=tf.sg_opt(kwargs) + _context,
prev=tensor,
is_layer=True,
name=opt.name)
# inject reuse function
out.sg_reuse = types.MethodType(sg_reuse, out)
return out
def wrapper(tensor, **kwargs):
r"""Manages arguments of `tf.sg_opt`.
Args:
tensor: A `tensor` (automatically passed by decorator).
kwargs:
shape: A list of integers. The shape of `tensor`. Inferred if not specified.
in_dim: An integer. The size of input dimension, which is set to the last one by default.
dim: An integer. The size of output dimension. Has the same value as in_dim by default.
bn: Boolean. If True, batch normalization is applied.
ln: Boolean. If True, layer normalization is applied.
scale: If true, multiple by a trainable gamma variable. When the activation is
linear (relu included), this can be disabled because it can be implicitly
learned by the next layer. The default is True.
dout: A float of range [0, 100). A dropout rate. Set to 0 by default.
bias: Boolean. If True, biases are added. As a default, it is set to True
name: A name for the layer. As a default, the function name is assigned.
act: A name of activation function. e.g., `sigmoid`, `tanh`, etc.
reuse: `True` or `None`; if `True`, we go into reuse mode for this `layer` scope
as well as all sub-scopes; if `None`, we just inherit the parent scope reuse.
regularizer: A string. None, 'l1' or 'l2'. The default is None
summary: If True, summaries are added. The default is True.
"""
from . import sg_initializer as init
from . import sg_activation
# kwargs parsing
opt = tf.sg_opt(kwargs) + sg_get_context()
# set default argument
try:
shape = tensor.get_shape().as_list()
# batch normalization off, layer normalization off, dropout off
opt += tf.sg_opt(shape=shape,
in_dim=shape[-1],
dim=shape[-1],
bn=False,
ln=False,
dout=0,
summary=True,
scale=True)
if opt.regularizer == 'l1':
opt.regularizer = lambda x: tf.reduce_mean(tf.abs(x))
elif opt.regularizer == 'l2':
opt.regularizer = lambda x: tf.square(
tf.reduce_mean(tf.square(x)))
else:
opt.regularizer = None
assert not (
opt.bn and opt.ln
), 'one of batch normalization and layer normalization is available.'
# disable bias when normalization on
opt += tf.sg_opt(bias=not (opt.bn or opt.ln))
finally:
pass
# automatic layer naming
if opt.name is None:
# layer function name will be used as layer name
opt.name = func.__name__.replace('sg_', '')
# find existing layer names
exist_layers = []
for t in tf.global_variables():
scope_name = tf.get_variable_scope().name
prefix = scope_name + '/' if len(scope_name) > 0 else ''
i = t.name.rfind(prefix + opt.name)
if i >= 0:
exist_layers.append(t.name[i:].split('/')[-2])
exist_layers = list(set(exist_layers))
# layer name numbering
if len(exist_layers) == 0:
opt.name += '_1'
else:
opt.name += '_%d' % (
max([int(n.split('_')[-1]) for n in exist_layers]) + 1)
with tf.variable_scope(opt.name, reuse=opt.reuse) as scope:
# call layer function
out = func(tensor, opt)
out_shape = out.get_shape()
# apply batch normalization
if opt.bn:
beta = init.constant('beta', opt.dim, summary=opt.summary)
gamma = init.constant('gamma',
opt.dim,
value=1,
summary=opt.summary,
trainable=opt.scale)
# offset, scale parameter ( for inference )
mean_running = init.constant('mean',
opt.dim,
trainable=False,
summary=opt.summary)
variance_running = init.constant('variance',
opt.dim,
value=1,
trainable=False,
summary=opt.summary)
# use fused batch norm if ndims in [2, 3, 4]
if out_shape.ndims in [2, 3, 4]:
# add HW dims if necessary, fused_batch_norm requires shape to be NHWC
if out_shape.ndims == 2:
out = tf.expand_dims(out, axis=1)
out = tf.expand_dims(out, axis=2)
elif out_shape.ndims == 3:
out = tf.expand_dims(out, axis=2)
fused_eps = tf.sg_eps if tf.sg_eps > 1e-5 else 1e-5
out, mean, variance = tf.cond(
_phase,
lambda: tf.nn.fused_batch_norm(
out, gamma, beta, epsilon=fused_eps),
lambda: tf.nn.fused_batch_norm(out,
gamma,
beta,
mean=mean_running,
variance=
variance_running,
epsilon=fused_eps,
is_training=False),
)
# restore original shape if HW dims was added
if out_shape.ndims == 2:
out = tf.squeeze(out, axis=[1, 2])
elif out_shape.ndims == 3:
out = tf.squeeze(out, axis=2)
# fallback to naive batch norm
else:
mean, variance = tf.nn.moments(
out, axes=list(range(len(out.get_shape()) - 1)))
out = tf.cond(
_phase, lambda: tf.nn.batch_normalization(
out, mean, variance, beta, gamma, tf.sg_eps),
lambda: tf.nn.batch_normalization(
out, mean_running, variance_running, beta, gamma,
tf.sg_eps))
decay = 0.99
tf.add_to_collection(
tf.GraphKeys.UPDATE_OPS,
mean_running.assign(mean_running * decay + mean *
(1 - decay)))
tf.add_to_collection(
tf.GraphKeys.UPDATE_OPS,
variance_running.assign(variance_running * decay +
variance * (1 - decay)))
# apply layer normalization
if opt.ln:
# offset, scale parameter
beta = init.constant('beta', opt.dim, summary=opt.summary)
if opt.scale:
gamma = init.constant('gamma',
opt.dim,
value=1,
summary=opt.summary)
# calc layer mean, variance for final axis
mean, variance = tf.nn.moments(out,
axes=[len(out.get_shape()) - 1],
keep_dims=True)
# apply normalization
out = (out - mean) / tf.sqrt(variance + tf.sg_eps)
# apply parameter
if opt.scale:
out = gamma * out + beta
else:
out = out + beta
# apply activation
if opt.act:
out = getattr(sg_activation, 'sg_' + opt.act.lower())(out)
# apply dropout
if opt.dout:
out = tf.cond(_phase, lambda: tf.nn.dropout(out, 1 - opt.dout),
lambda: out)
# rename tensor
out = tf.identity(out, 'out')
# add final output summary
if opt.summary:
tf.sg_summary_activation(out)
# save node info for reuse
out._sugar = tf.sg_opt(func=func,
arg=tf.sg_opt(kwargs) + sg_get_context(),
prev=tensor,
is_layer=True,
name=opt.name)
# inject reuse function
out.sg_reuse = types.MethodType(sg_reuse, out)
return out