DeepFaceLab/core/leras/optimizers.py
2020-03-04 15:08:10 +04:00

110 lines
No EOL
4.6 KiB
Python

import copy
def initialize_optimizers(nn):
tf = nn.tf
from tensorflow.python.ops import state_ops, control_flow_ops
class TFBaseOptimizer(nn.Saveable):
def __init__(self, name=None):
super().__init__(name=name)
def tf_clip_norm(self, g, c, n):
"""Clip the gradient `g` if the L2 norm `n` exceeds `c`.
# Arguments
g: Tensor, the gradient tensor
c: float >= 0. Gradients will be clipped
when their L2 norm exceeds this value.
n: Tensor, actual norm of `g`.
# Returns
Tensor, the gradient clipped if required.
"""
if c <= 0: # if clipnorm == 0 no need to add ops to the graph
return g
condition = n >= c
then_expression = tf.scalar_mul(c / n, g)
else_expression = g
# saving the shape to avoid converting sparse tensor to dense
if isinstance(then_expression, tf.Tensor):
g_shape = copy.copy(then_expression.get_shape())
elif isinstance(then_expression, tf.IndexedSlices):
g_shape = copy.copy(then_expression.dense_shape)
if condition.dtype != tf.bool:
condition = tf.cast(condition, 'bool')
g = tf.cond(condition,
lambda: then_expression,
lambda: else_expression)
if isinstance(then_expression, tf.Tensor):
g.set_shape(g_shape)
elif isinstance(then_expression, tf.IndexedSlices):
g._dense_shape = g_shape
return g
nn.TFBaseOptimizer = TFBaseOptimizer
class TFRMSpropOptimizer(TFBaseOptimizer):
def __init__(self, lr=0.001, rho=0.9, lr_dropout=1.0, epsilon=1e-7, clipnorm=0.0, name=None):
super().__init__(name=name)
if name is None:
raise ValueError('name must be defined.')
self.lr_dropout = lr_dropout
self.clipnorm = clipnorm
with tf.device('/CPU:0') :
with tf.variable_scope(self.name):
self.lr = tf.Variable (lr, name="lr")
self.rho = tf.Variable (rho, name="rho")
self.epsilon = tf.Variable (epsilon, name="epsilon")
self.iterations = tf.Variable(0, dtype=tf.int64, name='iters')
self.accumulators_dict = {}
self.lr_rnds_dict = {}
def get_weights(self):
return [self.lr, self.rho, self.epsilon, self.iterations] + list(self.accumulators_dict.values())
def initialize_variables(self, trainable_weights, vars_on_cpu=True):
# Initialize here all trainable variables used in training
e = tf.device('/CPU:0') if vars_on_cpu else None
if e: e.__enter__()
with tf.variable_scope(self.name):
accumulators = { v.name : tf.get_variable ( f'acc_{v.name}'.replace(':','_'), v.shape, dtype=v.dtype, initializer=tf.initializers.constant(0.0), trainable=False) for v in trainable_weights }
self.accumulators_dict.update ( accumulators)
if self.lr_dropout != 1.0:
lr_rnds = [ nn.tf_random_binomial( v.shape, p=self.lr_dropout, dtype=v.dtype) for v in trainable_weights ]
self.lr_rnds_dict.update ( { v.name : rnd for v,rnd in zip(trainable_weights,lr_rnds) } )
if e: e.__exit__(None, None, None)
def get_update_op(self, grads_vars):
updates = []
if self.clipnorm > 0.0:
norm = tf.sqrt( sum([tf.reduce_sum(tf.square(g)) for g,v in grads_vars]))
updates += [ state_ops.assign_add( self.iterations, 1) ]
for i, (g,v) in enumerate(grads_vars):
if self.clipnorm > 0.0:
g = self.tf_clip_norm(g, self.clipnorm, norm)
a = self.accumulators_dict[ v.name ]
rho = tf.cast(self.rho, a.dtype)
new_a = rho * a + (1. - rho) * tf.square(g)
lr = tf.cast(self.lr, a.dtype)
epsilon = tf.cast(self.epsilon, a.dtype)
v_diff = - lr * g / (tf.sqrt(new_a) + epsilon)
if self.lr_dropout != 1.0:
lr_rnd = self.lr_rnds_dict[v.name]
v_diff *= lr_rnd
new_v = v + v_diff
updates.append (state_ops.assign(a, new_a))
updates.append (state_ops.assign(v, new_v))
return control_flow_ops.group ( *updates, name=self.name+'_updates')
nn.TFRMSpropOptimizer = TFRMSpropOptimizer