!8323 [Auto parallel] Supporting for-loop in strategy-searching
From: @xiaoda_zh Reviewed-by: Signed-off-by:pull/8323/MERGE
mindspore-ci-bot
4 years ago
committed by
Gitee
commit
4e07f43dff
@ -0,0 +1,129 @@
|
||||
# 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
|
||||
from mindspore import context, Tensor, Parameter
|
||||
from mindspore.nn import Cell
|
||||
import mindspore.nn as nn
|
||||
from mindspore.ops import operations as P, functional as F
|
||||
from mindspore.common.initializer import initializer
|
||||
import mindspore.common.dtype as mstype
|
||||
from mindspore.common.api import _executor
|
||||
from tests.dataset_mock import MindData
|
||||
|
||||
|
||||
class Dataset(MindData):
|
||||
def __init__(self, predict, label, length=3):
|
||||
super(Dataset, self).__init__(size=length)
|
||||
self.predict = predict
|
||||
self.label = label
|
||||
self.index = 0
|
||||
self.length = length
|
||||
|
||||
def __iter__(self):
|
||||
return self
|
||||
|
||||
def __next__(self):
|
||||
if self.index >= self.length:
|
||||
raise StopIteration
|
||||
self.index += 1
|
||||
return self.predict, self.label
|
||||
|
||||
def reset(self):
|
||||
self.index = 0
|
||||
|
||||
|
||||
class LayerNorm(nn.Cell):
|
||||
def __init__(self, normalized_shape, eps=1e-5):
|
||||
super(LayerNorm, self).__init__()
|
||||
self.gamma = Parameter(initializer('ones', normalized_shape), name="gamma")
|
||||
self.beta = Parameter(initializer('zeros', normalized_shape), name="beta")
|
||||
self.mean = P.ReduceMean(keep_dims=True)
|
||||
self.eps = eps
|
||||
self.sub = P.Sub()
|
||||
self.add = P.TensorAdd()
|
||||
self.mul = P.Mul()
|
||||
self.div = P.RealDiv()
|
||||
|
||||
def construct(self, x):
|
||||
mean = self.mean(x, -1)
|
||||
variance = self.mean(F.square(self.sub(x, mean)))
|
||||
output = self.div(self.sub(x, mean), F.sqrt(self.add(variance, self.eps)))
|
||||
rescaled_output = self.add(self.mul(output, self.gamma), self.beta)
|
||||
return rescaled_output
|
||||
|
||||
|
||||
class SubNet(Cell):
|
||||
def __init__(self, index):
|
||||
super().__init__()
|
||||
self.matmul = P.MatMul()
|
||||
self.relu = P.ReLU()
|
||||
self.weight = Parameter(Tensor(np.ones([128, 128]), dtype=ms.float32), "matmul_w"+str(index))
|
||||
self.layernorm1 = LayerNorm((128,)).to_float(mstype.float32)
|
||||
|
||||
def construct(self, x):
|
||||
x = self.layernorm1(x)
|
||||
out = self.matmul(x, self.weight)
|
||||
out = self.relu(out)
|
||||
return out
|
||||
|
||||
|
||||
class Net(Cell):
|
||||
def __init__(self, mul_weight, num_layers, strategy1=None, strategy2=None):
|
||||
super().__init__()
|
||||
self.mul = P.Mul().shard(strategy1)
|
||||
self.neg = P.Neg().shard(strategy2)
|
||||
self.mul_weight = Parameter(mul_weight, "w1")
|
||||
self.num_layers = num_layers
|
||||
self.layers = nn.CellList()
|
||||
for i in range(num_layers):
|
||||
self.layers.append(SubNet(i))
|
||||
|
||||
def construct(self, x):
|
||||
for i in range(self.num_layers):
|
||||
x = self.layers[i](x)
|
||||
out = self.mul(x, self.mul_weight)
|
||||
out = self.neg(out)
|
||||
return out
|
||||
|
||||
|
||||
class Full(Cell):
|
||||
def __init__(self, mul_weight, num_layers, strategy1=None, strategy2=None):
|
||||
super().__init__()
|
||||
self.network = Net(mul_weight, num_layers, strategy1, strategy2)
|
||||
self.relu = P.ReLU()
|
||||
|
||||
def construct(self, x):
|
||||
out = self.network(x)
|
||||
out = self.relu(out)
|
||||
return out
|
||||
|
||||
|
||||
_x = Tensor(np.ones([512, 128]), dtype=ms.float32)
|
||||
_b = Tensor(np.ones([32]), dtype=ms.int32)
|
||||
_w1 = Tensor(np.ones([512, 128]), dtype=ms.float32)
|
||||
|
||||
|
||||
def test_auto_parallel():
|
||||
context.set_context(save_graphs=True)
|
||||
context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=16, global_rank=0)
|
||||
net = Full(_w1, 3)
|
||||
net.set_auto_parallel()
|
||||
net.set_train()
|
||||
_executor.compile(net, _x, phase='train')
|
||||
num_ops = _executor._get_num_parallel_ops(net)
|
||||
expected_num = 16
|
||||
assert num_ops == expected_num
|
||||
@ -0,0 +1,136 @@
|
||||
# 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 Tensor, Parameter, ParameterTuple
|
||||
from mindspore import context
|
||||
from mindspore.common.api import _executor
|
||||
from mindspore.nn.optim import Adam, FTRL
|
||||
from mindspore.ops import composite as C
|
||||
from mindspore.ops import functional as F
|
||||
from mindspore.ops import operations as P
|
||||
from mindspore.parallel._cost_model_context import _set_multi_subgraphs
|
||||
from mindspore.parallel._utils import _reset_op_id as reset_op_id
|
||||
|
||||
|
||||
class SubNet(nn.Cell):
|
||||
def __init__(self, index):
|
||||
super().__init__()
|
||||
self.matmul = P.BatchMatMul()
|
||||
self.relu = P.ReLU()
|
||||
self.weight = Parameter(Tensor(np.ones([8, 8, 8, 8]), dtype=ms.float32), "matmul_w"+str(index))
|
||||
|
||||
def construct(self, x):
|
||||
out = self.matmul(x, self.weight)
|
||||
out = self.relu(out)
|
||||
return out
|
||||
|
||||
|
||||
class Net(nn.Cell):
|
||||
def __init__(self):
|
||||
super(Net, self).__init__()
|
||||
self.mul = P.Mul()
|
||||
self.relu = P.ReLU()
|
||||
self.wd = Parameter(Tensor(np.ones([8, 8, 8, 8]).astype(np.float32)), name="wide")
|
||||
self.wt = Parameter(Tensor(np.ones([8, 8, 8, 8]).astype(np.float32)), name="l")
|
||||
self.layers = nn.CellList()
|
||||
for i in range(3):
|
||||
self.layers.append(SubNet(i))
|
||||
|
||||
def construct(self, x):
|
||||
for i in range(3):
|
||||
x = self.layers[i](x)
|
||||
out = self.mul(x, self.wd)
|
||||
out = self.mul(out, self.wt)
|
||||
out = self.relu(out)
|
||||
return out
|
||||
|
||||
|
||||
class NetWithLoss(nn.Cell):
|
||||
def __init__(self, network):
|
||||
super(NetWithLoss, self).__init__()
|
||||
self.sum = P.ReduceSum()
|
||||
self.mean = P.ReduceMean()
|
||||
self.net = network
|
||||
|
||||
def construct(self, x):
|
||||
predict = self.net(x)
|
||||
loss1 = self.sum(predict, -1)
|
||||
loss2 = self.mean(predict, -1)
|
||||
return loss1, loss2
|
||||
|
||||
|
||||
class IthOutputCell(nn.Cell):
|
||||
def __init__(self, network, output_index):
|
||||
super(IthOutputCell, self).__init__()
|
||||
self.network = network
|
||||
self.output_index = output_index
|
||||
|
||||
def construct(self, x):
|
||||
predict = self.network(x)[self.output_index]
|
||||
return predict
|
||||
|
||||
|
||||
class TrainStepWarp(nn.Cell):
|
||||
def __init__(self, network, sens=1000.0):
|
||||
super(TrainStepWarp, self).__init__()
|
||||
self.network = network
|
||||
self.network.set_train()
|
||||
self.trainable_params = network.trainable_params()
|
||||
weights_w = []
|
||||
weights_d = []
|
||||
for params in self.trainable_params:
|
||||
weights_w.append(params)
|
||||
weights_d.append(params)
|
||||
self.weights_w = ParameterTuple(weights_w)
|
||||
self.weights_d = ParameterTuple(weights_d)
|
||||
self.optimizer_w = FTRL(learning_rate=1e-2, params=self.weights_w, l1=1e-8,
|
||||
l2=1e-8, initial_accum=1.0)
|
||||
self.optimizer_d = Adam(self.weights_d, learning_rate=3.5e-4, eps=1e-8,
|
||||
loss_scale=sens)
|
||||
self.hyper_map = C.HyperMap()
|
||||
self.grad_w = C.GradOperation(get_by_list=True, sens_param=True)
|
||||
self.grad_d = C.GradOperation(get_by_list=True, sens_param=True)
|
||||
self.sens = sens
|
||||
self.loss_net_w = IthOutputCell(network, output_index=0)
|
||||
self.loss_net_d = IthOutputCell(network, output_index=1)
|
||||
|
||||
def construct(self, x):
|
||||
weights_w = self.weights_w
|
||||
weights_d = self.weights_d
|
||||
loss_w, loss_d = self.network(x)
|
||||
sens_w = P.Fill()(P.DType()(loss_w), P.Shape()(loss_w), self.sens)
|
||||
sens_d = P.Fill()(P.DType()(loss_d), P.Shape()(loss_d), self.sens)
|
||||
grads_w = self.grad_w(self.loss_net_w, weights_w)(x, sens_w)
|
||||
grads_d = self.grad_d(self.loss_net_d, weights_d)(x, sens_d)
|
||||
return F.depend(loss_w, self.optimizer_w(grads_w)), F.depend(loss_d, self.optimizer_d(grads_d))
|
||||
|
||||
|
||||
def test_double_subgraphs():
|
||||
context.set_context(save_graphs=True)
|
||||
context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=8, global_rank=0)
|
||||
net = TrainStepWarp(NetWithLoss(Net()))
|
||||
_set_multi_subgraphs()
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([8, 8, 8, 8]), dtype=ms.float32)
|
||||
reset_op_id()
|
||||
net.set_train()
|
||||
_executor.compile(net, x, phase='train')
|
||||
num_ops = _executor._get_num_parallel_ops(net)
|
||||
expected_num = 7
|
||||
assert expected_num == num_ops
|
||||
@ -0,0 +1,101 @@
|
||||
# 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
|
||||
from mindspore import context, Tensor, Parameter
|
||||
from mindspore.nn import Cell, Momentum
|
||||
from mindspore.nn.loss import SoftmaxCrossEntropyWithLogits
|
||||
import mindspore.nn as nn
|
||||
from mindspore.ops import operations as P
|
||||
from mindspore.train import Model
|
||||
from tests.dataset_mock import MindData
|
||||
|
||||
|
||||
class Dataset(MindData):
|
||||
def __init__(self, predict, label, length=3):
|
||||
super(Dataset, self).__init__(size=length)
|
||||
self.predict = predict
|
||||
self.label = label
|
||||
self.index = 0
|
||||
self.length = length
|
||||
|
||||
def __iter__(self):
|
||||
return self
|
||||
|
||||
def __next__(self):
|
||||
if self.index >= self.length:
|
||||
raise StopIteration
|
||||
self.index += 1
|
||||
return self.predict, self.label
|
||||
|
||||
def reset(self):
|
||||
self.index = 0
|
||||
|
||||
|
||||
class SubNet(Cell):
|
||||
def __init__(self, index):
|
||||
super().__init__()
|
||||
self.matmul = P.MatMul()
|
||||
self.relu = P.ReLU()
|
||||
self.weight = Parameter(Tensor(np.ones([128, 128]), dtype=ms.float32), "matmul_w"+str(index))
|
||||
|
||||
def construct(self, x):
|
||||
out = self.matmul(x, self.weight)
|
||||
out = self.relu(out)
|
||||
return out
|
||||
|
||||
|
||||
class Net(Cell):
|
||||
def __init__(self, mul_weight, num_layers, strategy1=None, strategy2=None):
|
||||
super().__init__()
|
||||
self.mul = P.Mul().shard(strategy1)
|
||||
self.neg = P.Neg().shard(strategy2)
|
||||
self.mul_weight = Parameter(mul_weight, "w1")
|
||||
self.num_layers = num_layers
|
||||
self.layers = nn.CellList()
|
||||
for i in range(num_layers):
|
||||
self.layers.append(SubNet(i))
|
||||
|
||||
def construct(self, x):
|
||||
for i in range(self.num_layers):
|
||||
x = self.layers[i](x)
|
||||
out = self.mul(x, self.mul_weight)
|
||||
out = self.neg(out)
|
||||
return out
|
||||
|
||||
|
||||
_x = Tensor(np.ones([32, 128]), dtype=ms.float32)
|
||||
_b = Tensor(np.ones([32]), dtype=ms.int32)
|
||||
_w1 = Tensor(np.ones([512, 128]), dtype=ms.float32)
|
||||
|
||||
|
||||
def compile_net(net):
|
||||
context.set_context(save_graphs=True)
|
||||
learning_rate = 0.1
|
||||
momentum = 0.9
|
||||
epoch_size = 2
|
||||
dataset = Dataset(_x, _b)
|
||||
loss = SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
|
||||
opt = Momentum(net.trainable_params(), learning_rate, momentum)
|
||||
model = Model(net, loss, optimizer=opt)
|
||||
model.train(epoch_size, dataset, dataset_sink_mode=False)
|
||||
context.reset_auto_parallel_context()
|
||||
|
||||
|
||||
def test_auto_parallel():
|
||||
context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=16, global_rank=0)
|
||||
net = Net(_w1, 3)
|
||||
compile_net(net)
|
||||
Loading…
Reference in new issue