提交 e8240167 编写于 作者: H He, Kai

add reduce to mean_normalize

上级 aa88beaf
......@@ -381,6 +381,20 @@ public:
FixedTensor::calc_precision_recall(in, &out_);
}
void div(const Tensor *lhs, const Tensor *rhs, Tensor *out) override {
auto lhs_tuple = from_tensor(lhs);
auto rhs_tuple = from_tensor(rhs);
auto out_tuple = from_tensor(out);
auto lhs_ = std::get<0>(lhs_tuple).get();
auto rhs_ = std::get<0>(rhs_tuple).get();
auto out_ = std::get<0>(out_tuple).get();
lhs_->long_div(rhs_, out_);
}
private:
template <typename T>
std::tuple<
......
......@@ -93,6 +93,8 @@ public:
Tensor* out) = 0;
virtual void calc_precision_recall(const Tensor* tp_fp_fn, Tensor* out) = 0;
virtual void div(const Tensor *lhs, const Tensor *rhs, Tensor *out) = 0;
};
} // mpc
......
......@@ -39,6 +39,9 @@ class MpcMeanNormalizationOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(
ctx->HasInput("SampleNum"), true,
platform::errors::InvalidArgument("Input(Sample) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("TotalNum"), true,
platform::errors::InvalidArgument("Input(TotalNum) should not be null."));
PADDLE_ENFORCE_EQ(ctx->HasOutput("Range"), true,
platform::errors::InvalidArgument(
"Output(Range) should not be null."));
......@@ -46,13 +49,11 @@ class MpcMeanNormalizationOp : public framework::OperatorWithKernel {
platform::errors::InvalidArgument(
"Output(Meanor) should not be null."));
int64_t total_sample_num =
static_cast<int64_t>(ctx->Attrs().Get<int>("total_sample_num"));
auto min_dims = ctx->GetInputDim("Min");
auto max_dims = ctx->GetInputDim("Max");
auto mean_dims = ctx->GetInputDim("Mean");
auto sample_num_dims = ctx->GetInputDim("SampleNum");
auto total_num_dims = ctx->GetInputDim("TotalNum");
if (ctx->IsRuntime()) {
PADDLE_ENFORCE_EQ(min_dims, max_dims,
......@@ -77,7 +78,7 @@ class MpcMeanNormalizationOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(
sample_num_dims.size(), 2,
platform::errors::InvalidArgument(
"The dimension of Input(SampleNum) should be equal to 3 "
"The dimension of Input(SampleNum) should be equal to 2 "
"(share_num, party_num). But received (%d)",
sample_num_dims.size()));
......@@ -87,6 +88,27 @@ class MpcMeanNormalizationOp : public framework::OperatorWithKernel {
"The party num of Input(SampleNum) and Input(Min) "
"should be equal But received (%d) != (%d)",
sample_num_dims[1], min_dims[1]));
PADDLE_ENFORCE_EQ(
total_num_dims.size(), 2,
platform::errors::InvalidArgument(
"The dimension of Input(TotalNum) "
"should be 2, But received (%d) != (%d)",
total_num_dims.size(), 2));
PADDLE_ENFORCE_EQ(
sample_num_dims[0], total_num_dims[0],
platform::errors::InvalidArgument(
"The share num of Input(SampleNum) and Input(TotalNum) "
"should be equal But received (%d) != (%d)",
sample_num_dims[0], total_num_dims[0]));
PADDLE_ENFORCE_EQ(
total_num_dims[1], 1,
platform::errors::InvalidArgument(
"The shape of Input(TotalNum) "
"should be [share_num, 1] But dims[1] received (%d) != (%d)",
total_num_dims[1], 1));
}
ctx->SetOutputDim("Range", {mean_dims[0], mean_dims[2]});
......@@ -121,6 +143,9 @@ class MpcMeanNormalizationOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor, default Tensor<int64_t>) A 1-D tensor with shape [P], "
"where P is the party num. Each element contains "
"sample num of party_i.");
AddInput("TotalNum",
"(Tensor, default Tensor<int64_t>) A 1-D tensor with shape [1], "
"Element contains sum of sample num of party_i.");
AddOutput("Range",
"(Tensor, default Tensor<int64_t>) A 1-D tensor with shape [N], "
"where N is the feature num. Each element contains "
......@@ -129,10 +154,9 @@ class MpcMeanNormalizationOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor, default Tensor<int64_t>) A 1-D tensor with shape [N], "
"where N is the feature num. Each element contains "
"global mean of feature_i.");
AddAttr<int>("total_sample_num", "(int) Sum of sample nums from all party.");
AddComment(R"DOC(
Mean normalization Operator.
When given Input(Min), Input(Max), Input(Mean) and Input(SampleNum),
When given Input(Min), Input(Max), Input(Mean), Input(SampleNum) and Input(TotalNum)
this operator can be used to compute global range and mean for further feature
scaling.
Output(Range) is the global range of all features.
......
......@@ -12,6 +12,8 @@ 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. */
#include <algorithm>
#include "paddle/fluid/framework/op_registry.h"
#include "mpc_op.h"
......@@ -28,6 +30,7 @@ class MpcMeanNormalizationKernel : public MpcOpKernel<T> {
const Tensor* max = context.Input<Tensor>("Max");
const Tensor* mean = context.Input<Tensor>("Mean");
const Tensor* sample_num = context.Input<Tensor>("SampleNum");
const Tensor* total_num = context.Input<Tensor>("TotalNum");
Tensor* range = context.Output<Tensor>("Range");
Tensor* mean_out = context.Output<Tensor>("MeanOut");
......@@ -65,9 +68,6 @@ class MpcMeanNormalizationKernel : public MpcOpKernel<T> {
range->mutable_data<T>(
framework::make_ddim({share_num, feat_num}), context.GetPlace(), 0);
// TODO: get total_sample_num by reduing size
int total_sample_num = context.Attr<int>("total_sample_num");
Tensor sample_num_;
sample_num_.ShareDataWith(*sample_num);
......@@ -84,8 +84,20 @@ class MpcMeanNormalizationKernel : public MpcOpKernel<T> {
mean_out->mutable_data<T>(
framework::make_ddim({share_num, feat_num}), context.GetPlace(), 0);
Tensor total_num_;
total_num_.mutable_data<T>(
framework::make_ddim({share_num, feat_num}), context.GetPlace(), 0);
// broadcasting total_num to shape [share_num, feat_num]
for (int i = 0; i < share_num; ++i) {
std::fill(total_num_.data<T>() + i * feat_num,
total_num_.data<T>() + (i + 1) * feat_num,
total_num->data<T>()[i]);
}
mpc::MpcInstance::mpc_instance()->mpc_protocol()
->mpc_operators()->scale(mean_out, 1.0 / total_sample_num, mean_out);
->mpc_operators()->div(mean_out, &total_num_, mean_out);
}
};
......
......@@ -17,10 +17,11 @@ mpc data preprocessing op layers.
from paddle.fluid.data_feeder import check_type, check_dtype
from ..framework import check_mpc_variable_and_dtype
from ..mpc_layer_helper import MpcLayerHelper
from .math import reduce_sum
__all__ = ['mean_normalize']
def mean_normalize(f_min, f_max, f_mean, sample_num, total_sample_num):
def mean_normalize(f_min, f_max, f_mean, sample_num):
'''
Mean normalization is a method used to normalize the range of independent
variables or features of data.
......@@ -40,7 +41,6 @@ def mean_normalize(f_min, f_max, f_mean, sample_num, total_sample_num):
sample_num (Variable): A 1-D tensor with shape [P], where P is the
party num. Each element contains sample num
of party_i.
total_sample_num (int): Sum of sample nums from all party.
Returns:
f_range (Variable): A 1-D tensor with shape [N], where N is the
......@@ -51,121 +51,26 @@ def mean_normalize(f_min, f_max, f_mean, sample_num, total_sample_num):
range of feature_i.
Examples:
.. code-block:: python
from multiprocessing import Manager
from multiprocessing import Process
import numpy as np
import paddle.fluid as fluid
import paddle_fl.mpc as pfl_mpc
import mpc_data_utils as mdu
import paddle_fl.mpc.data_utils.aby3 as aby3
redis_server = "127.0.0.1"
redis_port = 9937
test_f_num = 100
# party i owns 2 + 2*i rows of data
test_row_split = range(2, 10, 2)
def mean_norm_naive(f_mat):
ma = np.amax(f_mat, axis=0)
mi = np.amin(f_mat, axis=0)
return ma - mi, np.mean(f_mat, axis=0)
def gen_data(f_num, sample_nums):
f_mat = np.random.rand(np.sum(sample_nums), f_num)
f_min, f_max, f_mean = [], [], []
prev_idx = 0
for n in sample_nums:
i = prev_idx
j = i + n
ma = np.amax(f_mat[i:j], axis=0)
mi = np.amin(f_mat[i:j], axis=0)
me = np.mean(f_mat[i:j], axis=0)
f_min.append(mi)
f_max.append(ma)
f_mean.append(me)
prev_idx += n
f_min = np.array(f_min).reshape(sample_nums.size, f_num)
f_max = np.array(f_max).reshape(sample_nums.size, f_num)
f_mean = np.array(f_mean).reshape(sample_nums.size, f_num)
return f_mat, f_min, f_max, f_mean
class MeanNormDemo:
def mean_normalize(self, **kwargs):
"""
mean_normalize op ut
:param kwargs:
:return:
"""
role = kwargs['role']
pfl_mpc.init("aby3", role, "localhost", redis_server, redis_port)
mi = pfl_mpc.data(name='mi', shape=self.input_size, dtype='int64')
ma = pfl_mpc.data(name='ma', shape=self.input_size, dtype='int64')
me = pfl_mpc.data(name='me', shape=self.input_size, dtype='int64')
sn = pfl_mpc.data(name='sn', shape=self.input_size, dtype='int64')
# 2 for share, 4 for 4 party, 100 for feat_num
input_size = [2, 4, 100]
mi = pfl_mpc.data(name='mi', shape=input_size, dtype='int64')
ma = pfl_mpc.data(name='ma', shape=input_size, dtype='int64')
me = pfl_mpc.data(name='me', shape=input_size, dtype='int64')
sn = pfl_mpc.data(name='sn', shape=input_size[:-1], dtype='int64')
out0, out1 = pfl_mpc.layers.mean_normalize(f_min=mi, f_max=ma,
f_mean=me, sample_num=sn, total_sample_num=self.total_num)
f_mean=me, sample_num=sn)
exe = fluid.Executor(place=fluid.CPUPlace())
f_range, f_mean = exe.run(feed={'mi': kwargs['min'],
'ma': kwargs['max'], 'me': kwargs['mean'],
'sn': kwargs['sample_num']},fetch_list=[out0, out1])
self.f_range_list.append(f_range)
self.f_mean_list.append(f_mean)
def run(self):
f_nums = test_f_num
sample_nums = np.array(test_row_split)
mat, mi, ma, me = gen_data(f_nums, sample_nums)
self.input_size = [len(sample_nums), f_nums]
self.total_num = mat.shape[0]
# simulating encrypting data
share = lambda x: np.array([x * mdu.mpc_one_share] * 2).astype('int64').reshape(
[2] + list(x.shape))
self.f_range_list = Manager().list()
self.f_mean_list = Manager().list()
proc = list()
for role in range(3):
args = {'role': role, 'min': share(mi), 'max': share(ma),
'mean': share(me), 'sample_num': share(sample_nums)}
p = Process(target=self.mean_normalize, kwargs=args)
proc.append(p)
p.start()
for p in proc:
p.join()
f_r = aby3.reconstruct(np.array(self.f_range_list))
f_m = aby3.reconstruct(np.array(self.f_mean_list))
plain_r, plain_m = mean_norm_naive(mat)
print("max error in featrue range:", np.max(np.abs(f_r - plain_r)))
print("max error in featrue mean:", np.max(np.abs(f_m - plain_m)))
MeanNormDemo().run()
# feed encrypted data
f_range, f_mean = exe.run(feed={'mi': f_min, 'ma': f_max,
'me': f_mean, 'sn': sample_num}, fetch_list=[out0, out1])
'''
helper = MpcLayerHelper("mean_normalize", **locals())
......@@ -180,6 +85,8 @@ def mean_normalize(f_min, f_max, f_mean, sample_num, total_sample_num):
f_range = helper.create_mpc_variable_for_type_inference(dtype=f_min.dtype)
f_mean_out= helper.create_mpc_variable_for_type_inference(dtype=f_min.dtype)
total_num = reduce_sum(sample_num)
op_type = 'mean_normalize'
helper.append_op(
......@@ -189,14 +96,12 @@ def mean_normalize(f_min, f_max, f_mean, sample_num, total_sample_num):
"Max": f_max,
"Mean": f_mean,
"SampleNum": sample_num,
"TotalNum": total_num,
},
outputs={
"Range": f_range,
"MeanOut": f_mean_out,
},
attrs={
# TODO: remove attr total_sample_num, reducing sample_num instead
"total_sample_num": total_sample_num,
})
)
return f_range, f_mean_out
......@@ -18,6 +18,7 @@ mpc math op layers.
from ..framework import MpcVariable
from ..framework import check_mpc_variable_and_dtype
from ..mpc_layer_helper import MpcLayerHelper
from .ml import reshape
__all__ = [
'mean',
......@@ -194,6 +195,8 @@ def reduce_sum(input, dim=None, keep_dim=False, name=None):
inputs={'X': input},
outputs={'Out': out},
attrs=attrs)
if out.shape == (2,):
out = reshape(out, list(out.shape) + [1])
return out
......@@ -78,10 +78,10 @@ class TestOpMeanNormalize(test_op_base.TestOpBase):
mi = pfl_mpc.data(name='mi', shape=self.input_size, dtype='int64')
ma = pfl_mpc.data(name='ma', shape=self.input_size, dtype='int64')
me = pfl_mpc.data(name='me', shape=self.input_size, dtype='int64')
sn = pfl_mpc.data(name='sn', shape=self.input_size, dtype='int64')
out0, out1 = pfl_mpc.layers.mean_normalize(f_min=mi, f_max=ma, f_mean=me, sample_num=sn, total_sample_num=self.total_num)
sn = pfl_mpc.data(name='sn', shape=self.input_size[:-1], dtype='int64')
out0, out1 = pfl_mpc.layers.mean_normalize(f_min=mi,
f_max=ma, f_mean=me, sample_num=sn)
exe = fluid.Executor(place=fluid.CPUPlace())
......@@ -98,7 +98,6 @@ class TestOpMeanNormalize(test_op_base.TestOpBase):
mat, mi, ma, me = gen_data(f_nums, sample_nums)
self.input_size = [len(sample_nums), f_nums]
self.total_num = mat.shape[0]
share = lambda x: np.array([x * mdu.mpc_one_share] * 2).astype('int64').reshape(
[2] + list(x.shape))
......@@ -116,7 +115,7 @@ class TestOpMeanNormalize(test_op_base.TestOpBase):
plain_r, plain_m = mean_norm_naive(mat)
self.assertTrue(np.allclose(f_r, plain_r, atol=1e-4))
self.assertTrue(np.allclose(f_m, plain_m, atol=1e-3))
self.assertTrue(np.allclose(f_m, plain_m, atol=1e-4))
if __name__ == '__main__':
......
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