99 lines
3.5 KiB
C++
Executable File
99 lines
3.5 KiB
C++
Executable File
// 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/convolution.hpp"
|
|
#include "fdeep/filter.hpp"
|
|
#include "fdeep/shape2.hpp"
|
|
#include "fdeep/shape5.hpp"
|
|
#include "fdeep/layers/layer.hpp"
|
|
|
|
#include <fplus/fplus.hpp>
|
|
|
|
#include <algorithm>
|
|
#include <cstddef>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
namespace fdeep { namespace internal
|
|
{
|
|
|
|
// Convolve depth slices separately.
|
|
class depthwise_conv_2d_layer : public layer
|
|
{
|
|
public:
|
|
explicit depthwise_conv_2d_layer(
|
|
const std::string& name, std::size_t input_depth,
|
|
const shape5& filter_shape,
|
|
std::size_t k, const shape2& strides, padding p,
|
|
bool padding_valid_offset_depth_1,
|
|
bool padding_same_offset_depth_1,
|
|
bool padding_valid_offset_depth_2,
|
|
bool padding_same_offset_depth_2,
|
|
const shape2& dilation_rate,
|
|
const float_vec& depthwise_weights,
|
|
const float_vec& bias)
|
|
: layer(name),
|
|
filters_depthwise_(fplus::transform(generate_im2col_single_filter_matrix,
|
|
generate_filters(dilation_rate, filter_shape,
|
|
input_depth, depthwise_weights, bias))),
|
|
strides_(strides),
|
|
padding_(p),
|
|
padding_valid_offset_depth_1_(padding_valid_offset_depth_1),
|
|
padding_same_offset_depth_1_(padding_same_offset_depth_1),
|
|
padding_valid_offset_depth_2_(padding_valid_offset_depth_2),
|
|
padding_same_offset_depth_2_(padding_same_offset_depth_2)
|
|
{
|
|
assertion(k > 0, "needs at least one filter");
|
|
assertion(filter_shape.volume() > 0, "filter must have volume");
|
|
assertion(strides.area() > 0, "invalid strides");
|
|
assertion(filters_depthwise_.size() == input_depth,
|
|
"invalid number of filters");
|
|
}
|
|
protected:
|
|
tensor5s apply_impl(const tensor5s& inputs) const override
|
|
{
|
|
assertion(inputs.size() == 1, "only one input tensor allowed");
|
|
|
|
const auto input_slices = tensor5_to_depth_slices(inputs.front());
|
|
|
|
assertion(input_slices.size() == filters_depthwise_.size(),
|
|
"invalid input depth");
|
|
|
|
const bool use_offset = input_slices.size() == 1 ?
|
|
((padding_ == padding::valid && padding_valid_offset_depth_1_) ||
|
|
(padding_ == padding::same && padding_same_offset_depth_1_)) :
|
|
((padding_ == padding::valid && padding_valid_offset_depth_2_) ||
|
|
(padding_ == padding::same && padding_same_offset_depth_2_));
|
|
|
|
const auto convolve_slice =
|
|
[&](const tensor5& slice, const im2col_filter_matrix& f) -> tensor5
|
|
{
|
|
assertion(f.filter_shape_.depth_ == 1, "invalid filter depth");
|
|
const auto result = convolve(strides_, padding_,
|
|
use_offset, f, slice);
|
|
assertion(result.shape().depth_ == 1, "invalid conv output");
|
|
return result;
|
|
};
|
|
|
|
assertion(input_slices.size() == filters_depthwise_.size(),
|
|
"invalid depthwise filter count");
|
|
return {concatenate_tensor5s_depth(fplus::zip_with(
|
|
convolve_slice, input_slices, filters_depthwise_))};
|
|
}
|
|
|
|
std::vector<im2col_filter_matrix> filters_depthwise_;
|
|
shape2 strides_;
|
|
padding padding_;
|
|
bool padding_valid_offset_depth_1_;
|
|
bool padding_same_offset_depth_1_;
|
|
bool padding_valid_offset_depth_2_;
|
|
bool padding_same_offset_depth_2_;
|
|
};
|
|
|
|
} } // namespace fdeep, namespace internal
|