lichenever
4 years ago
parent
cffe2c94fe
commit
78e131cf15
File diff suppressed because it is too large
Load Diff
@ -1,91 +0,0 @@
|
||||
# Copyright 2019 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
import numpy as np
|
||||
import mindspore as ms
|
||||
import mindspore.nn as nn
|
||||
from mindspore import Tensor
|
||||
from mindspore import context
|
||||
from mindspore.common.api import _executor
|
||||
from mindspore.ops import composite as C
|
||||
from mindspore.ops import operations as P
|
||||
from tests.ut.python.ops.test_math_ops import VirtualLoss
|
||||
|
||||
|
||||
grad_all = C.GradOperation(get_all=True)
|
||||
|
||||
|
||||
class NetWithLoss(nn.Cell):
|
||||
def __init__(self, network):
|
||||
super(NetWithLoss, self).__init__()
|
||||
self.loss = VirtualLoss()
|
||||
self.network = network
|
||||
|
||||
def construct(self, x, y):
|
||||
predict = self.network(x, y)
|
||||
return self.loss(predict)
|
||||
|
||||
|
||||
class GradWrap(nn.Cell):
|
||||
def __init__(self, network):
|
||||
super(GradWrap, self).__init__()
|
||||
self.network = network
|
||||
|
||||
def construct(self, x, y):
|
||||
return grad_all(self.network)(x, y)
|
||||
|
||||
|
||||
class Net(nn.Cell):
|
||||
def __init__(self, axis=0, stage1=0, stage2=0, strategy1=None, strategy2=None, shape=None, target=""):
|
||||
super().__init__()
|
||||
if shape is None:
|
||||
shape = [64, 64]
|
||||
self.gatherv2 = P.GatherV2().shard(strategy1).add_prim_attr("primitive_target", target)
|
||||
self.mul = P.Mul().shard(strategy2)
|
||||
self.index = Tensor(np.ones(shape), dtype=ms.int32)
|
||||
self.gatherv2.set_stage(stage1)
|
||||
self.mul.set_stage(stage2)
|
||||
self.axis = axis
|
||||
|
||||
def construct(self, x, y):
|
||||
out = self.gatherv2(x, self.index, self.axis)
|
||||
out = self.mul(out, y)
|
||||
return out
|
||||
|
||||
|
||||
def test_gatherv2_semi_samestage1():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, \
|
||||
parallel_mode="semi_auto_parallel", pipeline_stages=2)
|
||||
strategy1 = ((1, 2), (1, 1))
|
||||
strategy2 = ((2, 1, 1), (2, 1, 1))
|
||||
net = GradWrap(NetWithLoss(Net(0, 0, 0, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 64]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
net.set_train()
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_samestage2():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=5, \
|
||||
parallel_mode="semi_auto_parallel", pipeline_stages=2)
|
||||
strategy1 = ((1, 2), (1, 1))
|
||||
strategy2 = ((2, 1, 1), (2, 1, 1))
|
||||
net = GradWrap(NetWithLoss(Net(0, 1, 1, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 64]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
net.set_train()
|
||||
_executor.compile(net, x, y)
|
||||
@ -0,0 +1,109 @@
|
||||
# Copyright 2020 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
# ============================================================================
|
||||
import numpy as np
|
||||
import mindspore as ms
|
||||
import mindspore.nn as nn
|
||||
from mindspore import context
|
||||
from mindspore import Tensor
|
||||
from mindspore.ops import operations as P
|
||||
from mindspore.common.parameter import Parameter
|
||||
from mindspore.common.initializer import initializer
|
||||
from mindspore.train.model import Model
|
||||
|
||||
|
||||
class DatasetLenet():
|
||||
def __init__(self, data, label, length=3):
|
||||
self.data = data
|
||||
self.label = label
|
||||
self.index = 1
|
||||
self.length = length
|
||||
|
||||
def __iter__(self):
|
||||
return self
|
||||
|
||||
def __next__(self):
|
||||
if self.index >= self.length:
|
||||
raise StopIteration
|
||||
self.index += 1
|
||||
return self.data, self.label
|
||||
|
||||
def reset(self):
|
||||
self.index = 0
|
||||
|
||||
def get_dataset_size(self):
|
||||
return 32
|
||||
|
||||
def get_repeat_count(self):
|
||||
return 1
|
||||
|
||||
def get_batch_size(self):
|
||||
return 32
|
||||
|
||||
def create_tuple_iterator(self, num_epochs=1):
|
||||
return self
|
||||
|
||||
|
||||
class MatMulCell(nn.Cell):
|
||||
def __init__(self, strategy1, strategy2):
|
||||
super().__init__()
|
||||
self.param = Parameter(initializer("zeros", [64, 64]), name="param")
|
||||
self.param1 = Parameter(initializer("zeros", [64, 64]), name="param1")
|
||||
self.matmul = P.MatMul().shard(strategy1)
|
||||
self.matmul1 = P.MatMul().shard(strategy2)
|
||||
|
||||
def construct(self, x):
|
||||
out = self.matmul(x, self.param)
|
||||
out = self.matmul1(out, self.param1)
|
||||
return out
|
||||
|
||||
|
||||
class Net(nn.Cell):
|
||||
def __init__(self, strategy1, strategy2):
|
||||
super().__init__()
|
||||
self.block = nn.CellList()
|
||||
for i in range(2):
|
||||
cell = MatMulCell(strategy1, strategy2)
|
||||
cell.stage = i
|
||||
self.block.append(cell)
|
||||
|
||||
def construct(self, x):
|
||||
for i in range(2):
|
||||
x = self.block[i](x)
|
||||
return x
|
||||
|
||||
|
||||
class PipelineSplit(nn.Cell):
|
||||
def __init__(self, strategy1, strategy2):
|
||||
super().__init__()
|
||||
self.cell = Net(strategy1, strategy2)
|
||||
|
||||
def construct(self, x, label):
|
||||
x = self.cell(x)
|
||||
return x
|
||||
|
||||
|
||||
def test_pipeline_split():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=4, pipeline_stages=2)
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel")
|
||||
data = Tensor(np.ones([32, 64]), dtype=ms.float32)
|
||||
label = Tensor(np.ones([64, 64]), dtype=ms.float32)
|
||||
strategy1 = ((4, 1), (1, 1))
|
||||
strategy2 = ((2, 1), (1, 1))
|
||||
net = PipelineSplit(strategy1, strategy2)
|
||||
params = net.cell.block[1].trainable_params()
|
||||
dataset = DatasetLenet(data, label, 3)
|
||||
optimizer = nn.Lamb(params, learning_rate=0.01)
|
||||
model = Model(net, optimizer=optimizer)
|
||||
model.train(2, dataset, dataset_sink_mode=False)
|
||||
Loading…
Reference in new issue