Skip to content
Snippets Groups Projects
Commit 0288f6b5 authored by Yifan Zhao's avatar Yifan Zhao
Browse files

Fixed col major and a few other issues

parent d3717850
No related branches found
No related tags found
No related merge requests found
Hide whitespace changes
Inline Side-by-side
hpvm/test/epoch_dnn/host_code/eigen_utils.h
+ 28
− 28
View file @ 0288f6b5
...@@ -3,10 +3,11 @@ ...@@ -3,10 +3,11 @@
#include <unsupported/Eigen/CXX11/Tensor> #include <unsupported/Eigen/CXX11/Tensor>
using Eigen::Index; using Eigen::Index;
using Eigen::Tensor; template <typename Scalar, int N>
template <int N> using BTensor = Eigen::Tensor<bool, N>; using RTensor = Eigen::Tensor<Scalar, N, Eigen::RowMajor>;
template <int N> using FTensor = Eigen::Tensor<float, N>; template <int N> using BTensor = RTensor<bool, N>;
template <int N> using ITensor = Eigen::Tensor<Index, N>; template <int N> using FTensor = RTensor<float, N>;
template <int N> using ITensor = RTensor<Index, N>;
template <int N> using IArray = std::array<Index, N>; template <int N> using IArray = std::array<Index, N>;
template <typename Nested, int ODim> template <typename Nested, int ODim>
...@@ -17,8 +18,8 @@ using BcastReshape = Eigen::TensorBroadcastingOp<const IArray<ODim>, ...@@ -17,8 +18,8 @@ using BcastReshape = Eigen::TensorBroadcastingOp<const IArray<ODim>,
const Reshape<Nested, ODim>>; const Reshape<Nested, ODim>>;
template <int N, typename Scalar, int Dim> template <int N, typename Scalar, int Dim>
Reshape<Tensor<Scalar, Dim>, Dim + N> Reshape<RTensor<Scalar, Dim>, Dim + N>
unsqueeze(const Tensor<Scalar, Dim> &input, const IArray<N> &dims) { unsqueeze(const RTensor<Scalar, Dim> &input, const IArray<N> &dims) {
IArray<Dim + N> reshapeTo; IArray<Dim + N> reshapeTo;
std::unordered_set<Index> dimSet(dims.begin(), dims.end()); std::unordered_set<Index> dimSet(dims.begin(), dims.end());
size_t fromDim = 0; size_t fromDim = 0;
...@@ -54,14 +55,14 @@ BcastReshape<TensorOp, ODim> broadcast(const TensorOp &iTensor, ...@@ -54,14 +55,14 @@ BcastReshape<TensorOp, ODim> broadcast(const TensorOp &iTensor,
} }
template <size_t ODim, typename Scalar, int IDim> template <size_t ODim, typename Scalar, int IDim>
BcastReshape<Tensor<Scalar, IDim>, ODim> BcastReshape<RTensor<Scalar, IDim>, ODim>
broadcast(const Tensor<Scalar, IDim> &iTensor, const IArray<ODim> &shape) { broadcast(const RTensor<Scalar, IDim> &iTensor, const IArray<ODim> &shape) {
return broadcast(iTensor, iTensor.dimensions(), shape); return broadcast(iTensor, iTensor.dimensions(), shape);
} }
template <size_t ODim, typename Scalar, int IDim> template <size_t ODim, typename Scalar, int IDim>
BcastReshape<Reshape<Tensor<Scalar, IDim>, IDim + 1>, ODim> BcastReshape<Reshape<RTensor<Scalar, IDim>, IDim + 1>, ODim>
broadcast(const Tensor<Scalar, IDim> &iTensor, const IArray<ODim> &shape, broadcast(const RTensor<Scalar, IDim> &iTensor, const IArray<ODim> &shape,
Index unsqueezeDim) { Index unsqueezeDim) {
auto unsqueezed = unsqueeze<1>(iTensor, {unsqueezeDim}); auto unsqueezed = unsqueeze<1>(iTensor, {unsqueezeDim});
return broadcast(unsqueezed, unsqueezed.dimensions(), shape); return broadcast(unsqueezed, unsqueezed.dimensions(), shape);
...@@ -75,9 +76,8 @@ FTensor<1> arange(Index from, Index to) { ...@@ -75,9 +76,8 @@ FTensor<1> arange(Index from, Index to) {
std::tuple<FTensor<1>, FTensor<1>> meshgrid(Index h, Index w) { std::tuple<FTensor<1>, FTensor<1>> meshgrid(Index h, Index w) {
IArray<1> reshapeTo({h * w}); IArray<1> reshapeTo({h * w});
// auto b = broadcast<1>(arange(0, h), {h, w}, 1); FTensor<1> linx = broadcast<2>(arange(0, w), {h, w}, 0).reshape(reshapeTo);
FTensor<1> linx = broadcast<2>(arange(0, h), {h, w}, 1).reshape(reshapeTo); FTensor<1> liny = broadcast<2>(arange(0, h), {h, w}, 1).reshape(reshapeTo);
FTensor<1> liny = broadcast<2>(arange(0, w), {h, w}, 0).reshape(reshapeTo);
return std::make_pair(linx, liny); return std::make_pair(linx, liny);
} }
...@@ -91,15 +91,15 @@ template <int Dim> FTensor<4> softmax(const FTensor<4> &input) { ...@@ -91,15 +91,15 @@ template <int Dim> FTensor<4> softmax(const FTensor<4> &input) {
} }
template <int AlongDim, typename Scalar, int IDim> template <int AlongDim, typename Scalar, int IDim>
Tensor<Scalar, IDim> maskSelect(const Tensor<Scalar, IDim> &input, RTensor<Scalar, IDim> maskSelect(const RTensor<Scalar, IDim> &input,
const BTensor<1> &mask) { const BTensor<1> &mask) {
size_t nSelected = 0; size_t nSelected = 0;
for (Index i = 0; i < mask.dimension(0); i++) for (Index i = 0; i < mask.dimension(0); i++)
if (mask[i]) if (mask[i])
nSelected += 1; nSelected += 1;
IArray<IDim> retShape = input.dimensions(); IArray<IDim> retShape = input.dimensions();
retShape[AlongDim] = nSelected; retShape[AlongDim] = nSelected;
Tensor<Scalar, IDim> ret(retShape); RTensor<Scalar, IDim> ret(retShape);
for (Index i = 0, j = 0; i < mask.dimension(0); i++) for (Index i = 0, j = 0; i < mask.dimension(0); i++)
if (mask[i]) { if (mask[i]) {
ret.chip(j, AlongDim) = input.chip(i, AlongDim); ret.chip(j, AlongDim) = input.chip(i, AlongDim);
...@@ -109,9 +109,9 @@ Tensor<Scalar, IDim> maskSelect(const Tensor<Scalar, IDim> &input, ...@@ -109,9 +109,9 @@ Tensor<Scalar, IDim> maskSelect(const Tensor<Scalar, IDim> &input,
} }
template <int AlongDim, typename Scalar, int IDim> template <int AlongDim, typename Scalar, int IDim>
Tensor<Scalar, IDim> &maskAssign(Tensor<Scalar, IDim> &tensor, RTensor<Scalar, IDim> &maskAssign(RTensor<Scalar, IDim> &tensor,
const BTensor<1> &mask, const BTensor<1> &mask,
const Tensor<Scalar, IDim> &values) { const RTensor<Scalar, IDim> &values) {
for (Index i = 0, j = 0; i < mask.dimension(0); i++) for (Index i = 0, j = 0; i < mask.dimension(0); i++)
if (mask[i]) { if (mask[i]) {
tensor.chip(i, AlongDim) = values.chip(j, AlongDim); tensor.chip(i, AlongDim) = values.chip(j, AlongDim);
...@@ -121,8 +121,8 @@ Tensor<Scalar, IDim> &maskAssign(Tensor<Scalar, IDim> &tensor, ...@@ -121,8 +121,8 @@ Tensor<Scalar, IDim> &maskAssign(Tensor<Scalar, IDim> &tensor,
} }
template <int AlongDim, typename Scalar, int IDim> template <int AlongDim, typename Scalar, int IDim>
Tensor<Scalar, IDim> &maskAssign(Tensor<Scalar, IDim> &tensor, RTensor<Scalar, IDim> &maskAssign(RTensor<Scalar, IDim> &tensor,
const BTensor<1> &mask, const Scalar &value) { const BTensor<1> &mask, const Scalar &value) {
for (Index i = 0, j = 0; i < mask.dimension(0); i++) for (Index i = 0, j = 0; i < mask.dimension(0); i++)
if (mask[i]) { if (mask[i]) {
tensor.chip(i, AlongDim).setConstant(value); tensor.chip(i, AlongDim).setConstant(value);
...@@ -132,8 +132,8 @@ Tensor<Scalar, IDim> &maskAssign(Tensor<Scalar, IDim> &tensor, ...@@ -132,8 +132,8 @@ Tensor<Scalar, IDim> &maskAssign(Tensor<Scalar, IDim> &tensor,
} }
template <typename Scalar, int IDim> template <typename Scalar, int IDim>
Tensor<Scalar, IDim> dimSelect(const Tensor<Scalar, IDim> &input, size_t dim, RTensor<Scalar, IDim> dimSelect(const RTensor<Scalar, IDim> &input, size_t dim,
Index from, Index to) { Index from, Index to) {
Index dimSize = input.dimension(dim); Index dimSize = input.dimension(dim);
if (from < 0) if (from < 0)
from += dimSize; from += dimSize;
...@@ -167,8 +167,8 @@ FTensor<IDim> concat(const std::array<FTensor<IDim>, N> &tensors) { ...@@ -167,8 +167,8 @@ FTensor<IDim> concat(const std::array<FTensor<IDim>, N> &tensors) {
} }
template <typename OutScalar, typename InScalar, typename FuncT> template <typename OutScalar, typename InScalar, typename FuncT>
Tensor<OutScalar, 2> outerAction(const Tensor<InScalar, 1> &xs, RTensor<OutScalar, 2> outerAction(const RTensor<InScalar, 1> &xs,
const FuncT &func) { const FuncT &func) {
Index nElem = xs.size(); Index nElem = xs.size();
auto lhs = broadcast<2>(xs, {nElem, nElem}, 0), auto lhs = broadcast<2>(xs, {nElem, nElem}, 0),
rhs = broadcast<2>(xs, {nElem, nElem}, 1); rhs = broadcast<2>(xs, {nElem, nElem}, 1);
...@@ -176,10 +176,10 @@ Tensor<OutScalar, 2> outerAction(const Tensor<InScalar, 1> &xs, ...@@ -176,10 +176,10 @@ Tensor<OutScalar, 2> outerAction(const Tensor<InScalar, 1> &xs,
} }
template <typename Scalar> template <typename Scalar>
Tensor<Scalar, 1> scatter1D(const Tensor<Index, 1> &indices, RTensor<Scalar, 1> scatter1D(const RTensor<Index, 1> &indices,
const Tensor<Scalar, 1> &values) { const RTensor<Scalar, 1> &values) {
assert(indices.size() == values.size()); assert(indices.size() == values.size());
Tensor<Scalar, 1> ret(indices.size()); RTensor<Scalar, 1> ret(indices.size());
for (size_t i = 0; i < indices.size(); i++) for (size_t i = 0; i < indices.size(); i++)
ret[indices[i]] = values[i]; ret[indices[i]] = values[i];
return ret; return ret;
......
hpvm/test/epoch_dnn/host_code/host_bin
+ 0
− 0
View file @ 0288f6b5
No preview for this file type
hpvm/test/epoch_dnn/host_code/yolo.cpp
+ 8
− 14
View file @ 0288f6b5
...@@ -141,7 +141,7 @@ FTensor<2> NMS::forward(const FTensor<2> &boxes, size_t nImages) { ...@@ -141,7 +141,7 @@ FTensor<2> NMS::forward(const FTensor<2> &boxes, size_t nImages) {
FTensor<2> sortByCol(const FTensor<2> &input, Index col, bool descending, FTensor<2> sortByCol(const FTensor<2> &input, Index col, bool descending,
ITensor<1> &order) { ITensor<1> &order) {
using TupleTy = Eigen::Tuple<Index, float>; using TupleTy = Eigen::Tuple<Index, float>;
Tensor<TupleTy, 1> sortCol = input.chip<1>(col).index_tuples(); RTensor<TupleTy, 1> sortCol = input.chip<1>(col).index_tuples();
TupleTy *sortValPtr = sortCol.data(); TupleTy *sortValPtr = sortCol.data();
auto cmp = [descending](const TupleTy &lhs, const TupleTy &rhs) { auto cmp = [descending](const TupleTy &lhs, const TupleTy &rhs) {
Index lhs_ = lhs.first, rhs_ = rhs.first; Index lhs_ = lhs.first, rhs_ = rhs.first;
...@@ -167,7 +167,6 @@ BTensor<2> triu1(BTensor<2> &xs) { ...@@ -167,7 +167,6 @@ BTensor<2> triu1(BTensor<2> &xs) {
BTensor<1> NMS::nms(const FTensor<2> &boxes) { BTensor<1> NMS::nms(const FTensor<2> &boxes) {
// Sort by descending score // Sort by descending score
Index nBoxes = boxes.dimension(0); Index nBoxes = boxes.dimension(0);
std::cout << nBoxes << " boxes\n";
ITensor<1> order(nBoxes); ITensor<1> order(nBoxes);
FTensor<2> sortedBoxes = sortByCol(boxes, 5, true, order); FTensor<2> sortedBoxes = sortByCol(boxes, 5, true, order);
FTensor<1> x1 = sortedBoxes.chip<1>(1), y1 = sortedBoxes.chip<1>(2), FTensor<1> x1 = sortedBoxes.chip<1>(1), y1 = sortedBoxes.chip<1>(2),
...@@ -177,8 +176,6 @@ BTensor<1> NMS::nms(const FTensor<2> &boxes) { ...@@ -177,8 +176,6 @@ BTensor<1> NMS::nms(const FTensor<2> &boxes) {
// Compute dx and dy between each pair of boxes (these mat contain // Compute dx and dy between each pair of boxes (these mat contain
// every pair twice...) // every pair twice...)
std::cout << "len(x1) = " << x1.dimension(0)
<< ", len(x2) = " << x2.dimension(0) << "\n";
using F2Ref = const FTensor<2> &; using F2Ref = const FTensor<2> &;
auto cwiseMin = [](F2Ref lhs, F2Ref rhs) { return lhs.cwiseMin(rhs); }; auto cwiseMin = [](F2Ref lhs, F2Ref rhs) { return lhs.cwiseMin(rhs); };
auto cwiseMax = [](F2Ref lhs, F2Ref rhs) { return lhs.cwiseMax(rhs); }; auto cwiseMax = [](F2Ref lhs, F2Ref rhs) { return lhs.cwiseMax(rhs); };
...@@ -187,7 +184,6 @@ BTensor<1> NMS::nms(const FTensor<2> &boxes) { ...@@ -187,7 +184,6 @@ BTensor<1> NMS::nms(const FTensor<2> &boxes) {
dy2 = outerAction<float>(y2, cwiseMin), dy2 = outerAction<float>(y2, cwiseMin),
dy1 = outerAction<float>(y1, cwiseMax); dy1 = outerAction<float>(y1, cwiseMax);
FTensor<2> dx = (dx2 - dx1).cwiseMax(0.0f), dy = (dy2 - dy1).cwiseMax(0.0f); FTensor<2> dx = (dx2 - dx1).cwiseMax(0.0f), dy = (dy2 - dy1).cwiseMax(0.0f);
std::cout << "Done with dx and dy\n";
// Compute iou // Compute iou
auto sum = [](F2Ref lhs, F2Ref rhs) { return lhs + rhs; }; auto sum = [](F2Ref lhs, F2Ref rhs) { return lhs + rhs; };
...@@ -216,7 +212,7 @@ BTensor<1> NMS::nms(const FTensor<2> &boxes) { ...@@ -216,7 +212,7 @@ BTensor<1> NMS::nms(const FTensor<2> &boxes) {
return scatter1D(order, keep); return scatter1D(order, keep);
} }
size_t ImageOutputShape[3] = {85, 19, 19}; size_t ImageOutputShape[3] = {85, 15, 20};
FTensor<4> loadTensor(const std::string &filename) { FTensor<4> loadTensor(const std::string &filename) {
ifstream fin(filename); ifstream fin(filename);
...@@ -241,15 +237,11 @@ FTensor<4> loadTensor(const std::string &filename) { ...@@ -241,15 +237,11 @@ FTensor<4> loadTensor(const std::string &filename) {
} }
int main(int argc, char *argv[]) { int main(int argc, char *argv[]) {
if (argc != 3) { if (argc != 2) {
std::cerr << "Usage: " << argv[0] << " IMAGE_PATH SCALE\n"; std::cerr << "Usage: " << argv[0] << " SCALE\n";
return 1; return 1;
} }
// std::string cmd = std::string("./nvdla_runtime --image ") + argv[1] + FTensor<4> inputs = loadTensor("output.dimg") * std::stof(argv[1]);
// " --loadable hpvm-mod.nvdla --rawdump";
// std::cout << "Running NVDLA model as \"" << cmd << "\"\n";
// std::system(cmd.c_str());
FTensor<4> inputs = loadTensor("output.dimg") * std::stof(argv[2]);
std::cout << "Loaded input of shape " << inputs.dimensions() << "\n"; std::cout << "Loaded input of shape " << inputs.dimensions() << "\n";
auto tinyYoloV2Anchors = FTensor<2>(5, 2); auto tinyYoloV2Anchors = FTensor<2>(5, 2);
// This should match that defined in torch_dnn/yolo/model.py // This should match that defined in torch_dnn/yolo/model.py
...@@ -259,9 +251,11 @@ int main(int argc, char *argv[]) { ...@@ -259,9 +251,11 @@ int main(int argc, char *argv[]) {
{9.42, 5.11}, {9.42, 5.11},
{16.62, 10.52}}); {16.62, 10.52}});
GetDarknetBoxes getBoxes(0.5, 32, tinyYoloV2Anchors); GetDarknetBoxes getBoxes(0.5, 32, tinyYoloV2Anchors);
NMS nms(0.5);
FTensor<2> ret1 = getBoxes.forward(inputs); FTensor<2> ret1 = getBoxes.forward(inputs);
NMS nms(0.5);
std::cout << ret1.dimension(0) << " boxes before NMS\n";
FTensor<2> ret2 = nms.forward(ret1, inputs.dimension(0)); FTensor<2> ret2 = nms.forward(ret1, inputs.dimension(0));
std::cout << ret2.dimension(0) << " boxes after NMS\n";
cout << ret2 << "\n"; cout << ret2 << "\n";
return 0; return 0;
} }
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment