MNN/source/geometry/GeometryImageOp.cpp

204 lines
8.5 KiB
C++
Raw Normal View History

2020-11-05 16:41:56 +08:00
//
// GeometryImageOp.cpp
// MNN
//
// Created by MNN on 2020/05/07.
// Copyright © 2018, Alibaba Group Holding Limited
//
#include "ConvertUtils.hpp"
#include "geometry/GeometryComputer.hpp"
#include "geometry/GeometryComputerUtils.hpp"
#include "shape/SizeComputer.hpp"
namespace MNN {
/**
if coordinate_transformation_mode is "half_pixel",
x_original = (x_resized + 0.5) / scale - 0.5,
if coordinate_transformation_mode is "pytorch_half_pixel",
x_original = length_resized > 1 ? (x_resized + 0.5) / scale - 0.5 : 0,
if coordinate_transformation_mode is "align_corners",
x_original = x_resized * (length_original - 1) / (length_resized - 1),
if coordinate_transformation_mode is "asymmetric",
x_original = x_resized / scale,
if coordinate_transformation_mode is "tf_half_pixel_for_nn",
x_original = (x_resized + 0.5) / scale,
if coordinate_transformation_mode is "tf_crop_and_resize",
x_original = length_resized > 1 ? start_x * (length_original - 1) + x_resized * (end_x - start_x) * (length_original - 1) / (length_resized - 1) : 0.5 * (start_x + end_x) * (length_original - 1).
*/
static void _ConverterInterp(const Interp* resize, InterpT* dstInfo, int inW, int inH, int outW, int outH) {
switch (resize->ctm()) {
case CoordinateTransformationMode_NotSet:
{
// For compability, old model's nearest don't support halfpixels
if (resize->halfPixelCenters() && resize->resizeType() != 1) {
2020-11-05 16:41:56 +08:00
dstInfo->heightScale = (float)(inH) / (float)(outH);
dstInfo->widthScale = (float)(inW) / (float)(outW);
dstInfo->widthOffset = 0.5f * dstInfo->widthScale - 0.5f;
dstInfo->heightOffset = 0.5f * dstInfo->heightScale - 0.5f;
} else if (resize->alignCorners()) {
if (outH == 1) {
dstInfo->heightScale = 0.0f;
} else {
dstInfo->heightScale = (float)(inH - 1) / (float)(outH - 1);
}
if (outW == 1) {
dstInfo->widthScale = 0.0f;
} else {
dstInfo->widthScale = (float)(inW - 1) / (float)(outW - 1);
}
} else {
dstInfo->heightScale = (float)(inH) / (float)(outH);
dstInfo->widthScale = (float)(inW) / (float)(outW);
}
break;
}
case CoordinateTransformationMode_AlignCorners:
{
if (outH == 1) {
dstInfo->heightScale = 0.0f;
} else {
dstInfo->heightScale = (float)(inH - 1) / (float)(outH - 1);
}
if (outW == 1) {
dstInfo->widthScale = 0.0f;
} else {
dstInfo->widthScale = (float)(inW - 1) / (float)(outW - 1);
}
break;
}
case CoordinateTransformationMode_HalfPixels:
{
dstInfo->heightScale = (float)(inH) / (float)(outH);
dstInfo->widthScale = (float)(inW) / (float)(outW);
dstInfo->widthOffset = 0.5f * dstInfo->widthScale - 0.5f;
dstInfo->heightOffset = 0.5f * dstInfo->heightScale - 0.5f;
break;
}
case CoordinateTransformationMode_PytorchHalfPixels:
{
if (outH > 1) {
dstInfo->heightScale = (float)inH / (float)outH;
dstInfo->heightOffset = 0.5f * dstInfo->heightScale - 0.5f;
} else {
dstInfo->heightScale = 0.0f;
}
if (outW > 1) {
dstInfo->widthScale = (float)inW / (float)outW;
dstInfo->widthOffset = 0.5f * dstInfo->widthScale - 0.5f;
} else {
dstInfo->widthScale = 0.0f;
}
break;
}
case CoordinateTransformationMode_Asymmetric:
{
dstInfo->heightScale = (float)(inH) / (float)(outH);
dstInfo->widthScale = (float)(inW) / (float)(outW);
break;
}
case CoordinateTransformationMode_TensorflowHalfPixels:
{
dstInfo->heightScale = (float)(inH) / (float)(outH);
dstInfo->widthScale = (float)(inW) / (float)(outW);
dstInfo->widthOffset = 0.5f * dstInfo->widthScale;
dstInfo->heightOffset = 0.5f * dstInfo->heightScale;
break;
}
case CoordinateTransformationMode_TensorflowCropAndResize:
{
//FIXME: Not support now
MNN_ERROR("Don't support CoordinateTransformationMode_TensorflowCropAndResize currently\n");
break;
}
default:
break;
}
}
class GeometryImageOp : public GeometryComputer {
public:
virtual bool onCompute(const Op* op, const std::vector<Tensor*>& inputs, const std::vector<Tensor*>& outputs,
Context& context, CommandBuffer& res) const override {
auto newOutputs = outputs;
auto newInputs = inputs;
auto originOutput = outputs[0];
auto output = originOutput;
auto inputDes = TensorUtils::getDescribe(newInputs[0]);
auto format = inputDes->dimensionFormat;
if (MNN_DATA_FORMAT_NC4HW4 != format) {
std::shared_ptr<Tensor> newInput(new Tensor(newInputs[0], Tensor::CAFFE_C4, false));
ConvertUtils::compute(newInputs[0], newInput.get(), res);
newInputs[0] = newInput.get();
res.extras.emplace_back(std::move(newInput));
std::shared_ptr<Tensor> newOutput(new Tensor(originOutput, Tensor::CAFFE_C4, false));
output = newOutput.get();
newOutputs[0] = output;
res.extras.emplace_back(newOutput);
}
if (OpType_Resize == op->type()) {
// Turn resize to interp
std::unique_ptr<OpT> interp(new OpT);
interp->type = OpType_Interp;
interp->main.type = OpParameter_Interp;
interp->main.value = new InterpT;
interp->main.AsInterp()->widthScale = (float)inputs[0]->width() / (float)outputs[0]->width();
interp->main.AsInterp()->heightScale = (float)inputs[0]->height() / (float)outputs[0]->height();
interp->main.AsInterp()->resizeType = 2; // bilinear
res.command.emplace_back(GeometryComputerUtils::makeCommand(interp.get(), newInputs, newOutputs));
}
else if (OpType_Interp == op->type()) {
// Compute cord transform for interp
std::unique_ptr<OpT> interp(new OpT);
interp->type = OpType_Interp;
auto resize = op->main_as_Interp();
interp->main.type = OpParameter_Interp;
interp->main.value = new InterpT;
interp->main.AsInterp()->resizeType = resize->resizeType();
auto inW = inputs[0]->width();
auto inH = inputs[0]->height();
auto outW = outputs[0]->width();
auto outH = outputs[0]->height();
auto dstInfo = interp->main.AsInterp();
_ConverterInterp(resize, dstInfo, inW, inH, outW, outH);
res.command.emplace_back(GeometryComputerUtils::makeCommand(interp.get(), newInputs, newOutputs));
} else {
Command cmd;
cmd.op = op;
cmd.inputs = std::move(newInputs);
cmd.outputs = std::move(newOutputs);
res.command.emplace_back(std::move(cmd));
}
if (originOutput != output) {
ConvertUtils::compute(output, originOutput, res);
}
return true;
}
virtual std::vector<bool> onGetOutputVirtual(const Op* op, const std::vector<Tensor*>& inputs,
const std::vector<Tensor*>& outputs) const override {
std::vector<bool> res(outputs.size(), false);
auto outputDes = TensorUtils::getDescribe(outputs[0]);
if (MNN_DATA_FORMAT_NC4HW4 != outputDes->dimensionFormat) {
res[0] = true;
}
return res;
}
};
static void _create() {
std::shared_ptr<GeometryComputer> comp(new GeometryImageOp);
GeometryComputer::registerGeometryComputer(
comp, {OpType_ConvInt8, OpType_ConvolutionDepthwise, OpType_DeconvolutionDepthwise,
OpType_Pooling, OpType_Interp, OpType_Resize, OpType_Int8ToFloat, OpType_FloatToInt8});
}
REGISTER_GEOMETRY(GeometryImageOp, _create);
} // namespace MNN