import tensorflow as tf
from tensorflow.python.training import optimizer
class DenseLossScaleOptimizer:
def __init__(self, opt: optimizer.Optimizer, loss_scale: int) -> None:
if not isinstance(opt, optimizer.Optimizer):
raise ValueError('"opt" must be an instance of Optimizer, but got: %s' % type(opt))
self._optimizer = opt
self._loss_scale = tf.convert_to_tensor(loss_scale, tf.float32)
_update_lr_loss_scale(self._optimizer, loss_scale)
def compute_gradients(self, loss, var_list=None):
return self._optimizer.compute_gradients(loss * self._loss_scale, var_list=var_list)
def apply_gradients(self, avg_grads):
return self._optimizer.apply_gradients(avg_grads)
class SparseLossScaleOptimizer:
def __init__(self, opt: optimizer.Optimizer, loss_scale: int) -> None:
if not isinstance(opt, optimizer.Optimizer):
raise ValueError('"opt" must be an instance of Optimizer, but got: %s' % type(opt))
self._optimizer = opt
self._loss_scale = tf.convert_to_tensor(loss_scale, tf.float32)
_update_lr_loss_scale(self._optimizer, loss_scale)
def compute_gradients(self, loss, var_list=None):
return tf.gradients(loss * self._loss_scale, var_list)
def apply_gradients(self, grads_and_vars):
return self._optimizer.apply_gradients(grads_and_vars)
def _update_lr_loss_scale(opt, loss_scale):
if loss_scale <= 0:
raise RuntimeError("the loss_scale must be greater than zero.")
loss_scale = tf.convert_to_tensor(loss_scale, tf.float32)
if hasattr(opt, "_lr"):
opt._lr = opt._lr / loss_scale
elif hasattr(opt, "_learning_rate"):
opt._learning_rate = opt._learning_rate / loss_scale
else:
raise RuntimeError("`opt` should have a `_learning_rate` or `_lr` named field.")