def _clip_grad(clip_type, clip_value, grad):
dt = F.dtype(grad)
if clip_type == 0:
new_grad = C.clip_by_value(
grad, F.cast(F.tuple_to_array((-clip_value, )), dt),
F.cast(F.tuple_to_array((clip_value, )), dt))
else:
new_grad = nn.ClipByNorm()(grad,
F.cast(F.tuple_to_array((clip_value, )),
dt))
return new_grad
def _clip_grad(clip_value, grad):
"""
Clip gradients.
Inputs:
clip_value (float): Specifies how much to clip.
grad (tuple[Tensor]): Gradients.
Outputs:
tuple[Tensor], clipped gradients.
"""
dt = ops.dtype(grad)
new_grad = nn.ClipByNorm()(grad, ops.cast(ops.tuple_to_array((clip_value,)), dt))
return new_grad
def _clip_grad(clip_type, clip_value, grad):
"""
Clip gradients.
Inputs:
clip_type (int): The way to clip, 0 for 'value', 1 for 'norm'.
clip_value (float): Specifies how much to clip.
grad (tuple[Tensor]): Gradients.
Outputs:
tuple[Tensor], clipped gradients.
"""
if clip_type not in (0, 1):
return grad
dt = F.dtype(grad)
if clip_type == 0:
new_grad = C.clip_by_value(grad, F.cast(F.tuple_to_array((-clip_value,)), dt),
F.cast(F.tuple_to_array((clip_value,)), dt))
else:
new_grad = nn.ClipByNorm()(grad, F.cast(F.tuple_to_array((clip_value,)), dt))
return new_grad
def __init__(self):
super(_ClipGradients, self).__init__()
self.clip_by_norm = nn.ClipByNorm()
self.dtype = P.DType()
from mindspore.ops import operations as P
from mindspore.ops import functional as F
from mindspore.ops import composite as C
from mindspore.common.tensor import Tensor
from mindspore.common.parameter import Parameter, ParameterTuple
from mindspore.common import dtype as mstype
from mindspore.nn.wrap.grad_reducer import DistributedGradReducer
from mindspore.train.parallel_utils import ParallelMode
from mindspore.communication.management import get_group_size
from mindspore import context
from .bert_model import BertModel
GRADIENT_CLIP_TYPE = 1
GRADIENT_CLIP_VALUE = 1.0
_nn_clip_by_norm = nn.ClipByNorm()
clip_grad = C.MultitypeFuncGraph("clip_grad")
# pylint: disable=consider-using-in
@clip_grad.register("Number", "Number", "Tensor")
def _clip_grad(clip_type, clip_value, grad):
"""
Clip gradients.
Inputs:
clip_type (int): The way to clip, 0 for 'value', 1 for 'norm'.
clip_value (float): Specifies how much to clip.
grad (tuple[Tensor]): Gradients.
Outputs:
def test_clip_by_norm():
clip_by_norm = nn.ClipByNorm()
x = Tensor(np.array([[-2, 0, 0], [0, 3, 4]]).astype(np.float32))
clip_norm = Tensor(np.array([1]).astype(np.float32))
clip_by_norm(x, clip_norm)
def __init__(self):
super(Network, self).__init__()
self.norm_value = Tensor(np.array([1]).astype(np.float32))
self.clip = nn.ClipByNorm()
