52 lines
1.1 KiB
C++
52 lines
1.1 KiB
C++
|
// Copyright 2016, Tobias Hermann.
|
||
|
// https://github.com/Dobiasd/frugally-deep
|
||
|
// Distributed under the MIT License.
|
||
|
// (See accompanying LICENSE file or at
|
||
|
// https://opensource.org/licenses/MIT)
|
||
|
|
||
|
#pragma once
|
||
|
|
||
|
#include "fdeep/layers/layer.hpp"
|
||
|
|
||
|
#include <fplus/fplus.hpp>
|
||
|
|
||
|
#include <algorithm>
|
||
|
#include <cassert>
|
||
|
#include <cstddef>
|
||
|
#include <vector>
|
||
|
#include <string>
|
||
|
|
||
|
namespace fdeep { namespace internal
|
||
|
{
|
||
|
|
||
|
// Abstract base class for actication layers
|
||
|
// https://en.wikipedia.org/wiki/Activation_function
|
||
|
class activation_layer : public layer
|
||
|
{
|
||
|
public:
|
||
|
explicit activation_layer(const std::string& name) :
|
||
|
layer(name)
|
||
|
{
|
||
|
}
|
||
|
tensor5s apply_impl(const tensor5s& inputs) const override
|
||
|
{
|
||
|
const auto f = [this](const tensor5& t) -> tensor5
|
||
|
{
|
||
|
return transform_input(t);
|
||
|
};
|
||
|
return fplus::transform(f, inputs);
|
||
|
}
|
||
|
|
||
|
protected:
|
||
|
virtual tensor5 transform_input(const tensor5& input) const = 0;
|
||
|
};
|
||
|
|
||
|
inline tensor5s apply_activation_layer(
|
||
|
const activation_layer_ptr& ptr,
|
||
|
const tensor5s& input)
|
||
|
{
|
||
|
return ptr == nullptr ? input : ptr->apply(input);
|
||
|
}
|
||
|
|
||
|
} } // namespace fdeep, namespace internal
|