DeepFaceLab/core/leras/optimizers/AdaBelief.py

import numpy as np
from core.leras import nn
from tensorflow.python.ops import control_flow_ops, state_ops

tf = nn.tf

class AdaBelief(nn.OptimizerBase):
    def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999, lr_dropout=1.0, lr_cos=0, clipnorm=0.0, name=None, **kwargs):
        super().__init__(name=name)

        if name is None:
            raise ValueError('name must be defined.')

        self.lr = lr
        self.beta_1 = beta_1
        self.beta_2 = beta_2
        self.lr_dropout = lr_dropout
        self.lr_cos = lr_cos
        self.clipnorm = clipnorm

        with tf.device('/CPU:0') :
            with tf.variable_scope(self.name):
                self.iterations = tf.Variable(0, dtype=tf.int64, name='iters')

        self.ms_dict = {}
        self.vs_dict = {}
        self.lr_rnds_dict = {}

    def get_weights(self):
        return [self.iterations] + list(self.ms_dict.values()) + list(self.vs_dict.values())

    def initialize_variables(self, trainable_weights, vars_on_cpu=True, lr_dropout_on_cpu=False):
        # 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):
            ms = { v.name : tf.get_variable ( f'ms_{v.name}'.replace(':','_'), v.shape, dtype=v.dtype, initializer=tf.initializers.constant(0.0), trainable=False) for v in trainable_weights }
            vs = { v.name : tf.get_variable ( f'vs_{v.name}'.replace(':','_'), v.shape, dtype=v.dtype, initializer=tf.initializers.constant(0.0), trainable=False) for v in trainable_weights }
            self.ms_dict.update (ms)
            self.vs_dict.update (vs)
            
            if self.lr_dropout != 1.0:
                e = tf.device('/CPU:0') if lr_dropout_on_cpu else None
                if e: e.__enter__()                    
                lr_rnds = [ nn.random_binomial( v.shape, p=self.lr_dropout, dtype=v.dtype) for v in trainable_weights ]
                if e: e.__exit__(None, None, None)                
                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(tf.cast(g, tf.float32))) 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, tf.cast(norm, g.dtype) )

            ms = self.ms_dict[ v.name ]
            vs = self.vs_dict[ v.name ]
            
            m_t = self.beta_1*ms + (1.0-self.beta_1) * g
            v_t = self.beta_2*vs + (1.0-self.beta_2) * tf.square(g-m_t)

            lr = tf.constant(self.lr, g.dtype)
            if self.lr_cos != 0:
                lr *= (tf.cos(  tf.cast(self.iterations, g.dtype) * (2*3.1415926535/ float(self.lr_cos) )  ) + 1.0) / 2.0

            v_diff = - lr * m_t / (tf.sqrt(v_t) + np.finfo( g.dtype.as_numpy_dtype ).resolution )
            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(ms, m_t))
            updates.append (state_ops.assign(vs, v_t))
            updates.append (state_ops.assign(v, new_v))

        return control_flow_ops.group ( *updates, name=self.name+'_updates')
nn.AdaBelief = AdaBelief