提交 51796aa6 编写于 作者: L lirongzhen1

fix sparse feature bug for auto parallel

上级 7fbed0ce
......@@ -200,14 +200,14 @@ def get_bprop_mirror_operator(self):
float_one = F.scalar_cast(1.0, F.dtype(grad))
num = F.scalar_cast(dev_num, F.dtype(grad))
grad = mul(grad, cast(F.scalar_to_array(float_one/num), F.dtype(grad)))
dx = (indices, grad, dout.dense_shape())
dx = IndexedSlices(indices, grad, dout.dense_shape())
else:
if F.issubclass_(F.typeof(dout), mstype.tensor):
dx = all_reduce(dout)
else:
indices = all_gather(dout.indices())
grad = all_gather(dout.values())
dx = (indices, grad, dout.dense_shape())
dx = IndexedSlices(indices, grad, dout.dense_shape())
return (dx,)
return bprop
......
......@@ -21,9 +21,8 @@ from mindspore import context
from mindspore.common.parameter import Parameter
from mindspore.common.tensor import Tensor
from mindspore.ops import composite as C, operations as P
from mindspore.ops.operations.comm_ops import AllReduce, _MirrorOperator
from mindspore.ops.operations.comm_ops import AllReduce
from mindspore.common.api import _executor
from mindspore.communication.management import HCCL_WORLD_COMM_GROUP
from mindspore.nn import TrainOneStepCell, Adam
......@@ -60,30 +59,37 @@ def test_bprop_with_sparse_feature_allreduce():
_executor.compile(net, x)
def test_bprop_with_sparse_feature_mirror():
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="hybrid_parallel")
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
context.set_context(enable_sparse=True)
class Net(nn.Cell):
def __init__(self, axis=0, shape=None):
def __init__(self, shape=None):
super(Net, self).__init__()
if shape is None:
shape = [8, 8]
self.mirror = _MirrorOperator(group=HCCL_WORLD_COMM_GROUP)
self.gatherv2 = P.SparseGatherV2()
weight = Tensor(np.ones([64, 64]), dtype=ms.float32)
self.weight = Parameter(weight, "w")
self.index = Tensor(np.ones(shape), dtype=ms.int32)
self.axis = axis
self.embeddinglookup = nn.EmbeddingLookup()
self.embeddinglookup.embeddinglookup.set_strategy(((1, 1), (8, 1)))
def construct(self, x):
out = self.mirror(x)
out = self.gatherv2(out, self.index, self.axis)
def construct(self, x, b):
out = self.embeddinglookup(self.weight, self.index)
return out
net = GradWrap(Net())
x = Tensor(np.ones([64, 64]), dtype=ms.float32)
_x = Tensor(np.ones([126, 64, 32]), dtype=ms.float32)
_b = Tensor(np.ones([126, 64, 32]), dtype=ms.float32)
_executor.compile(net, x)
def compile_net(net):
optimizer = Adam(net.trainable_params(), learning_rate=0.1, loss_scale=1024.0, weight_decay=0.9)
train_net = TrainOneStepCell(net, optimizer)
_executor.compile(train_net, _x, _b)
net = Net()
compile_net(net)
def test_bprop_with_sparse_feature_dataparallel():
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册