Add truncated gaussian initializer. (#13000)
* Add truncated gaussian initializer. * Fix unitest. * Update API.spec * Fix code style and fix bug. * Fix code style. * Small fix.upload-readme
whs
7 years ago
committed by
GitHub
parent
642cf6ca2f
commit
cf128231c6
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,76 @@
|
||||
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
|
||||
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. */
|
||||
|
||||
#include <thrust/random.h>
|
||||
#include <thrust/transform.h>
|
||||
#include <limits>
|
||||
#include "paddle/fluid/framework/op_registry.h"
|
||||
#include "paddle/fluid/framework/operator.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace operators {
|
||||
|
||||
template <typename T>
|
||||
struct TruncatedNormal {
|
||||
T mean, std;
|
||||
T a_normal_cdf;
|
||||
T b_normal_cdf;
|
||||
unsigned int seed;
|
||||
T numeric_min;
|
||||
|
||||
__host__ __device__ TruncatedNormal(T mean, T std, T numeric_min, int seed)
|
||||
: mean(mean), std(std), seed(seed), numeric_min(numeric_min) {
|
||||
a_normal_cdf = (1.0 + erff(-2.0 / sqrtf(2.0))) / 2.0;
|
||||
b_normal_cdf = (1.0 + erff(2.0 / sqrtf(2.0))) / 2.0;
|
||||
}
|
||||
|
||||
__host__ __device__ T operator()(const unsigned int n) const {
|
||||
thrust::minstd_rand rng;
|
||||
rng.seed(seed);
|
||||
thrust::uniform_real_distribution<T> dist(numeric_min, 1);
|
||||
rng.discard(n);
|
||||
T value = dist(rng);
|
||||
auto p = a_normal_cdf + (b_normal_cdf - a_normal_cdf) * value;
|
||||
return (std::sqrt(2.0) * erfinvf(2 * p - 1) + mean) * std;
|
||||
}
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
class GPUTruncatedGaussianRandomKernel : public framework::OpKernel<T> {
|
||||
public:
|
||||
void Compute(const framework::ExecutionContext& context) const override {
|
||||
auto* tensor = context.Output<framework::Tensor>("Out");
|
||||
T* data = tensor->mutable_data<T>(context.GetPlace());
|
||||
unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
|
||||
if (seed == 0) {
|
||||
std::random_device rd;
|
||||
seed = rd();
|
||||
}
|
||||
T mean = static_cast<T>(context.Attr<float>("mean"));
|
||||
T std = static_cast<T>(context.Attr<float>("std"));
|
||||
thrust::counting_iterator<unsigned int> index_sequence_begin(0);
|
||||
int64_t size = tensor->numel();
|
||||
thrust::transform(
|
||||
index_sequence_begin, index_sequence_begin + size,
|
||||
thrust::device_ptr<T>(data),
|
||||
TruncatedNormal<T>(mean, std, std::numeric_limits<T>::min(), seed));
|
||||
}
|
||||
};
|
||||
|
||||
} // namespace operators
|
||||
} // namespace paddle
|
||||
|
||||
REGISTER_OP_CUDA_KERNEL(
|
||||
truncated_gaussian_random,
|
||||
paddle::operators::GPUTruncatedGaussianRandomKernel<float>);
|
||||
@ -0,0 +1,69 @@
|
||||
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
#
|
||||
# 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.
|
||||
|
||||
from __future__ import print_function
|
||||
|
||||
import unittest
|
||||
import numpy
|
||||
|
||||
import paddle.fluid as fluid
|
||||
import paddle.fluid.core as core
|
||||
from paddle.fluid.op import Operator
|
||||
from paddle.fluid.executor import Executor
|
||||
|
||||
|
||||
class TestTrunctedGaussianRandomOp(unittest.TestCase):
|
||||
def setUp(self):
|
||||
self.op_type = "truncated_gaussian_random"
|
||||
self.inputs = {}
|
||||
self.attrs = {
|
||||
"shape": [10000],
|
||||
"mean": .0,
|
||||
"std": 1.,
|
||||
"seed": 10,
|
||||
}
|
||||
|
||||
self.outputs = ["Out"]
|
||||
|
||||
def test_cpu(self):
|
||||
self.gaussian_random_test(place=fluid.CPUPlace())
|
||||
|
||||
def test_gpu(self):
|
||||
if core.is_compiled_with_cuda():
|
||||
self.gaussian_random_test(place=fluid.CUDAPlace(0))
|
||||
|
||||
def gaussian_random_test(self, place):
|
||||
|
||||
program = fluid.Program()
|
||||
block = program.global_block()
|
||||
vout = block.create_var(name="Out")
|
||||
op = block.append_op(
|
||||
type=self.op_type, outputs={"Out": vout}, attrs=self.attrs)
|
||||
|
||||
op.desc.infer_var_type(block.desc)
|
||||
op.desc.infer_shape(block.desc)
|
||||
|
||||
fetch_list = []
|
||||
for var_name in self.outputs:
|
||||
fetch_list.append(block.var(var_name))
|
||||
|
||||
exe = Executor(place)
|
||||
outs = exe.run(program, fetch_list=fetch_list)
|
||||
tensor = outs[0]
|
||||
self.assertAlmostEqual(numpy.mean(tensor), .0, delta=0.1)
|
||||
self.assertAlmostEqual(numpy.var(tensor), 0.773, delta=0.1)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Loading…
Reference in new issue