def get_gradients(self, loss_or_grads, params):
"""
Note
----
The returned gradients may contain None value
"""
# check valid algorithm
if self.algorithm is None or \
not hasattr(self.algorithm, 'compute_gradients') or \
not hasattr(self.algorithm, 'apply_gradients'):
raise RuntimeError("Optimizer is None, or doesn't has attributes: "
"compute_gradients and apply_gradients.")
with tf.variable_scope(self.name):
# get the gradient
grads_var = self.algorithm.compute_gradients(loss_or_grads,
var_list=params)
grads_var = {g: v for g, v in grads_var if g is not None}
grads = list(grads_var.keys())
params = list(grads_var.values())
# ====== clipnorm ====== #
if self.clipnorm is not None:
if self.clip_alg == 'norm':
grads = [tf.clip_by_norm(g, self.clipnorm)
for g in grads]
elif self.clip_alg == 'total_norm':
grads, _ = tf.clip_by_global_norm(grads, self.clipnorm)
elif self.clip_alg == 'avg_norm':
grads = [tf.clip_by_average_norm(g, self.clipnorm)
for g in grads]
# ====== clipvalue ====== #
if self.clipvalue is not None:
grads = [tf.clip_by_value(g, -self.clipvalue, self.clipvalue)
for g in grads]
# ====== get final norm value ====== #
self._norm = add_role(tf.global_norm(grads, name="GradientNorm"),
GradientsNorm)
return [(g, p) for g, p in zip(grads, params)]
评论列表
文章目录
