Skip to content

P
Paddle

机器未来 / Paddle
与 Fork 源项目一致

Fork自 PaddlePaddle / Paddle

通知 1
Star 1
Fork 0
P
Paddle
提交 e7e4f084 编写于 作者: D dengkaipeng
浏览文件
操作

ignore pred overlap gt > 0.7. test=develop

上级 bd6deb1a
显示空白变更内容
内联 并排
Showing with 668 addition and 125 deletion
paddle/fluid/operators/yolov3_loss_op.cc
浏览文件 @ e7e4f084
...@@ -35,12 +35,15 @@ class Yolov3LossOp : public framework::OperatorWithKernel { ...@@ -35,12 +35,15 @@ class Yolov3LossOp : public framework::OperatorWithKernel {
auto dim_gtlabel = ctx->GetInputDim("GTLabel"); auto dim_gtlabel = ctx->GetInputDim("GTLabel");
auto anchors = ctx->Attrs().Get<std::vector<int>>("anchors"); auto anchors = ctx->Attrs().Get<std::vector<int>>("anchors");
int anchor_num = anchors.size() / 2; int anchor_num = anchors.size() / 2;
auto anchor_mask = ctx->Attrs().Get<std::vector<int>>("anchor_mask");
int mask_num = anchor_mask.size();
auto class_num = ctx->Attrs().Get<int>("class_num"); auto class_num = ctx->Attrs().Get<int>("class_num");
PADDLE_ENFORCE_EQ(dim_x.size(), 4, "Input(X) should be a 4-D tensor."); PADDLE_ENFORCE_EQ(dim_x.size(), 4, "Input(X) should be a 4-D tensor.");
PADDLE_ENFORCE_EQ(dim_x[2], dim_x[3], PADDLE_ENFORCE_EQ(dim_x[2], dim_x[3],
"Input(X) dim[3] and dim[4] should be euqal."); "Input(X) dim[3] and dim[4] should be euqal.");
PADDLE_ENFORCE_EQ(dim_x[1], anchor_num * (5 + class_num), PADDLE_ENFORCE_EQ(
"Input(X) dim[1] should be equal to (anchor_number * (5 " dim_x[1], mask_num * (5 + class_num),
"Input(X) dim[1] should be equal to (anchor_mask_number * (5 "
"+ class_num))."); "+ class_num)).");
PADDLE_ENFORCE_EQ(dim_gtbox.size(), 3, PADDLE_ENFORCE_EQ(dim_gtbox.size(), 3,
"Input(GTBox) should be a 3-D tensor"); "Input(GTBox) should be a 3-D tensor");
...@@ -55,6 +58,11 @@ class Yolov3LossOp : public framework::OperatorWithKernel { ...@@ -55,6 +58,11 @@ class Yolov3LossOp : public framework::OperatorWithKernel {
"Attr(anchors) length should be greater then 0."); "Attr(anchors) length should be greater then 0.");
PADDLE_ENFORCE_EQ(anchors.size() % 2, 0, PADDLE_ENFORCE_EQ(anchors.size() % 2, 0,
"Attr(anchors) length should be even integer."); "Attr(anchors) length should be even integer.");
for (size_t i = 0; i < anchor_mask.size(); i++) {
PADDLE_ENFORCE_LT(
anchor_mask[i], anchor_num,
"Attr(anchor_mask) should not crossover Attr(anchors).");
}
PADDLE_ENFORCE_GT(class_num, 0, PADDLE_ENFORCE_GT(class_num, 0,
"Attr(class_num) should be an integer greater then 0."); "Attr(class_num) should be an integer greater then 0.");
...@@ -74,7 +82,7 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -74,7 +82,7 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
public: public:
void Make() override { void Make() override {
AddInput("X", AddInput("X",
"The input tensor of YOLO v3 loss operator, " "The input tensor of YOLOv3 loss operator, "
"This is a 4-D tensor with shape of [N, C, H, W]." "This is a 4-D tensor with shape of [N, C, H, W]."
"H and W should be same, and the second dimention(C) stores" "H and W should be same, and the second dimention(C) stores"
"box locations, confidence score and classification one-hot" "box locations, confidence score and classification one-hot"
...@@ -99,13 +107,20 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker { ...@@ -99,13 +107,20 @@ class Yolov3LossOpMaker : public framework::OpProtoAndCheckerMaker {
AddAttr<int>("class_num", "The number of classes to predict."); AddAttr<int>("class_num", "The number of classes to predict.");
AddAttr<std::vector<int>>("anchors", AddAttr<std::vector<int>>("anchors",
"The anchor width and height, " "The anchor width and height, "
"it will be parsed pair by pair."); "it will be parsed pair by pair.")
AddAttr<int>("input_size", .SetDefault(std::vector<int>{});
"The input size of YOLOv3 net, " AddAttr<std::vector<int>>("anchor_mask",
"generally this is set as 320, 416 or 608.") "The mask index of anchors used in "
.SetDefault(406); "current YOLOv3 loss calculation.")
.SetDefault(std::vector<int>{});
AddAttr<int>("downsample",
"The downsample ratio from network input to YOLOv3 loss "
"input, so 32, 16, 8 should be set for the first, second, "
"and thrid YOLOv3 loss operators.")
.SetDefault(32);
AddAttr<float>("ignore_thresh", AddAttr<float>("ignore_thresh",
"The ignore threshold to ignore confidence loss."); "The ignore threshold to ignore confidence loss.")
.SetDefault(0.7);
AddComment(R"DOC( AddComment(R"DOC(
This operator generate yolov3 loss by given predict result and ground This operator generate yolov3 loss by given predict result and ground
truth boxes. truth boxes.
......
paddle/fluid/operators/yolov3_loss_op.h
浏览文件 @ e7e4f084
...@@ -321,6 +321,182 @@ static void CalcYolov3LossGrad(T* input_grad_data, const Tensor& loss_grad, ...@@ -321,6 +321,182 @@ static void CalcYolov3LossGrad(T* input_grad_data, const Tensor& loss_grad,
obj_mask_data, n, an_num, grid_num, class_num, class_num); obj_mask_data, n, an_num, grid_num, class_num, class_num);
} }
static int mask_index(std::vector<int> mask, int val) {
for (int i = 0; i < mask.size(); i++) {
if (mask[i] == val) {
return i;
}
}
return -1;
}
template <typename T>
struct Box {
float x, y, w, h;
};
template <typename T>
static inline T sigmoid(T x) {
return 1.0 / (1.0 + std::exp(-x));
}
template <typename T>
static inline void sigmoid_arrray(T* arr, int len) {
for (int i = 0; i < len; i++) {
arr[i] = sigmoid(arr[i]);
}
}
template <typename T>
static inline Box<T> get_yolo_box(const T* x, std::vector<int> anchors, int i,
int j, int an_idx, int grid_size,
int input_size, int index, int stride) {
Box<T> b;
b.x = (i + sigmoid<T>(x[index])) / grid_size;
b.y = (j + sigmoid<T>(x[index + stride])) / grid_size;
b.w = std::exp(x[index + 2 * stride]) * anchors[2 * an_idx] / input_size;
b.h = std::exp(x[index + 3 * stride]) * anchors[2 * an_idx + 1] / input_size;
return b;
}
template <typename T>
static inline Box<T> get_gt_box(const T* gt, int batch, int max_boxes,
int idx) {
Box<T> b;
b.x = gt[(batch * max_boxes + idx) * 4];
b.y = gt[(batch * max_boxes + idx) * 4 + 1];
b.w = gt[(batch * max_boxes + idx) * 4 + 2];
b.h = gt[(batch * max_boxes + idx) * 4 + 3];
return b;
}
template <typename T>
static inline T overlap(T c1, T w1, T c2, T w2) {
T l1 = c1 - w1 / 2.0;
T l2 = c2 - w2 / 2.0;
T left = l1 > l2 ? l1 : l2;
T r1 = c1 + w1 / 2.0;
T r2 = c2 + w2 / 2.0;
T right = r1 < r2 ? r1 : r2;
return right - left;
}
template <typename T>
static inline T box_iou(Box<T> b1, Box<T> b2) {
T w = overlap(b1.x, b1.w, b2.x, b2.w);
T h = overlap(b1.y, b1.h, b2.y, b2.h);
T inter_area = (w < 0 || h < 0) ? 0.0 : w * h;
T union_area = b1.w * b1.h + b2.w * b2.h - inter_area;
return inter_area / union_area;
}
static inline int entry_index(int batch, int an_idx, int hw_idx, int an_num,
int an_stride, int stride, int entry) {
return (batch * an_num + an_idx) * an_stride + entry * stride + hw_idx;
}
template <typename T>
static void CalcBoxLocationLoss(T* loss, const T* input, Box<T> gt,
std::vector<int> anchors, int an_idx,
int box_idx, int gi, int gj, int grid_size,
int input_size, int stride) {
T tx = gt.x * grid_size - gi;
T ty = gt.y * grid_size - gj;
T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);
T scale = 2.0 - gt.w * gt.h;
loss[0] += SCE<T>(input[box_idx], tx) * scale;
loss[0] += SCE<T>(input[box_idx + stride], ty) * scale;
loss[0] += L1Loss<T>(input[box_idx + 2 * stride], tw) * scale;
loss[0] += L1Loss<T>(input[box_idx + 3 * stride], th) * scale;
}
template <typename T>
static void CalcBoxLocationLossGrad(T* input_grad, const T loss, const T* input,
Box<T> gt, std::vector<int> anchors,
int an_idx, int box_idx, int gi, int gj,
int grid_size, int input_size, int stride) {
T tx = gt.x * grid_size - gi;
T ty = gt.y * grid_size - gj;
T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);
T scale = 2.0 - gt.w * gt.h;
input_grad[box_idx] = SCEGrad<T>(input[box_idx], tx) * scale * loss;
input_grad[box_idx + stride] =
SCEGrad<T>(input[box_idx + stride], ty) * scale * loss;
input_grad[box_idx + 2 * stride] =
L1LossGrad<T>(input[box_idx + 2 * stride], tw) * scale * loss;
input_grad[box_idx + 3 * stride] =
L1LossGrad<T>(input[box_idx + 3 * stride], th) * scale * loss;
}
template <typename T>
static inline void CalcLabelLoss(T* loss, const T* input, const int index,
const int label, const int class_num,
const int stride) {
for (int i = 0; i < class_num; i++) {
loss[0] += SCE<T>(input[index + i * stride], (i == label) ? 1.0 : 0.0);
}
}
template <typename T>
static inline void CalcLabelLossGrad(T* input_grad, const T loss,
const T* input, const int index,
const int label, const int class_num,
const int stride) {
for (int i = 0; i < class_num; i++) {
input_grad[index + i * stride] =
SCEGrad<T>(input[index + i * stride], (i == label) ? 1.0 : 0.0) * loss;
}
}
template <typename T>
static inline void CalcObjnessLoss(T* loss, const T* input, const int* objness,
const int n, const int an_num, const int h,
const int w, const int stride,
const int an_stride) {
for (int i = 0; i < n; i++) {
for (int j = 0; j < an_num; j++) {
for (int k = 0; k < h; k++) {
for (int l = 0; l < w; l++) {
int obj = objness[k * w + l];
if (obj >= 0) {
loss[i] += SCE<T>(input[k * w + l], static_cast<T>(obj));
}
}
}
objness += stride;
input += an_stride;
}
}
}
template <typename T>
static inline void CalcObjnessLossGrad(T* input_grad, const T* loss,
const T* input, const int* objness,
const int n, const int an_num,
const int h, const int w,
const int stride, const int an_stride) {
for (int i = 0; i < n; i++) {
for (int j = 0; j < an_num; j++) {
for (int k = 0; k < h; k++) {
for (int l = 0; l < w; l++) {
int obj = objness[k * w + l];
if (obj >= 0) {
input_grad[k * w + l] =
SCEGrad<T>(input[k * w + l], static_cast<T>(obj)) * loss[i];
}
}
}
objness += stride;
input += an_stride;
input_grad += an_stride;
}
}
}
template <typename T> template <typename T>
class Yolov3LossKernel : public framework::OpKernel<T> { class Yolov3LossKernel : public framework::OpKernel<T> {
public: public:
...@@ -330,55 +506,158 @@ class Yolov3LossKernel : public framework::OpKernel<T> { ...@@ -330,55 +506,158 @@ class Yolov3LossKernel : public framework::OpKernel<T> {
auto* gt_label = ctx.Input<Tensor>("GTLabel"); auto* gt_label = ctx.Input<Tensor>("GTLabel");
auto* loss = ctx.Output<Tensor>("Loss"); auto* loss = ctx.Output<Tensor>("Loss");
auto anchors = ctx.Attr<std::vector<int>>("anchors"); auto anchors = ctx.Attr<std::vector<int>>("anchors");
auto anchor_mask = ctx.Attr<std::vector<int>>("anchor_mask");
int class_num = ctx.Attr<int>("class_num"); int class_num = ctx.Attr<int>("class_num");
int input_size = ctx.Attr<int>("input_size");
float ignore_thresh = ctx.Attr<float>("ignore_thresh"); float ignore_thresh = ctx.Attr<float>("ignore_thresh");
int downsample = ctx.Attr<int>("downsample");
const int n = input->dims()[0]; const int n = input->dims()[0];
const int h = input->dims()[2]; const int h = input->dims()[2];
const int w = input->dims()[3]; const int w = input->dims()[3];
const int an_num = anchors.size() / 2; const int an_num = anchors.size() / 2;
const int mask_num = anchor_mask.size();
const int b = gt_box->dims()[1];
int input_size = downsample * h;
Tensor conf_mask, obj_mask; const T* input_data = input->data<T>();
Tensor tx, ty, tw, th, tweight, tconf, tclass; const T* gt_box_data = gt_box->data<T>();
conf_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace()); const int* gt_label_data = gt_label->data<int>();
obj_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tx.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
ty.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tw.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
th.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tweight.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tconf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tclass.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
math::SetConstant<platform::CPUDeviceContext, T> constant;
constant(ctx.template device_context<platform::CPUDeviceContext>(),
&conf_mask, static_cast<T>(1.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(),
&obj_mask, static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tx,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &ty,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tw,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &th,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(),
&tweight, static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tconf,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tclass,
static_cast<T>(0.0));
PreProcessGTBox<T>(*gt_box, *gt_label, ignore_thresh, anchors, input_size,
h, &conf_mask, &obj_mask, &tx, &ty, &tw, &th, &tweight,
&tconf, &tclass);
T* loss_data = loss->mutable_data<T>({n}, ctx.GetPlace()); T* loss_data = loss->mutable_data<T>({n}, ctx.GetPlace());
memset(loss_data, 0, n * sizeof(T)); memset(loss_data, 0, n * sizeof(int));
CalcYolov3Loss<T>(loss_data, *input, tx, ty, tw, th, tweight, tconf, tclass,
conf_mask, obj_mask); Tensor objness;
int* objness_data =
objness.mutable_data<int>({n, mask_num, h, w}, ctx.GetPlace());
memset(objness_data, 0, objness.numel() * sizeof(int));
const int stride = h * w;
const int an_stride = (class_num + 5) * stride;
for (int i = 0; i < n; i++) {
for (int j = 0; j < mask_num; j++) {
for (int k = 0; k < h; k++) {
for (int l = 0; l < w; l++) {
int box_idx =
entry_index(i, j, k * w + l, mask_num, an_stride, stride, 0);
Box<T> pred =
get_yolo_box(input_data, anchors, l, k, anchor_mask[j], h,
input_size, box_idx, stride);
T best_iou = 0;
// int best_t = 0;
for (int t = 0; t < b; t++) {
if (isZero<T>(gt_box_data[i * b * 4 + t * 4]) &&
isZero<T>(gt_box_data[i * b * 4 + t * 4 + 1])) {
continue;
}
Box<T> gt = get_gt_box(gt_box_data, i, b, t);
T iou = box_iou(pred, gt);
if (iou > best_iou) {
best_iou = iou;
// best_t = t;
}
}
if (best_iou > ignore_thresh) {
int obj_idx = (i * mask_num + j) * stride + k * w + l;
objness_data[obj_idx] = -1;
}
}
}
}
for (int t = 0; t < b; t++) {
if (isZero<T>(gt_box_data[i * b * 4 + t * 4]) &&
isZero<T>(gt_box_data[i * b * 4 + t * 4 + 1])) {
continue;
}
Box<T> gt = get_gt_box(gt_box_data, i, b, t);
int gi = static_cast<int>(gt.x * w);
int gj = static_cast<int>(gt.y * h);
Box<T> gt_shift = gt;
gt_shift.x = 0.0;
gt_shift.y = 0.0;
T best_iou = 0.0;
int best_n = 0;
for (int an_idx = 0; an_idx < an_num; an_idx++) {
Box<T> an_box;
an_box.x = 0.0;
an_box.y = 0.0;
an_box.w = anchors[2 * an_idx] / static_cast<T>(input_size);
an_box.h = anchors[2 * an_idx + 1] / static_cast<T>(input_size);
float iou = box_iou<T>(an_box, gt_shift);
// TO DO: iou > 0.5 ?
if (iou > best_iou) {
best_iou = iou;
best_n = an_idx;
}
}
int mask_idx = mask_index(anchor_mask, best_n);
if (mask_idx >= 0) {
int box_idx = entry_index(i, mask_idx, gj * w + gi, mask_num,
an_stride, stride, 0);
CalcBoxLocationLoss<T>(loss_data + i, input_data, gt, anchors, best_n,
box_idx, gi, gj, h, input_size, stride);
int obj_idx = (i * mask_num + mask_idx) * stride + gj * w + gi;
objness_data[obj_idx] = 1;
int label = gt_label_data[i * b + t];
int label_idx = entry_index(i, mask_idx, gj * w + gi, mask_num,
an_stride, stride, 5);
CalcLabelLoss<T>(loss_data + i, input_data, label_idx, label,
class_num, stride);
}
}
}
CalcObjnessLoss<T>(loss_data, input_data + 4 * stride, objness_data, n,
mask_num, h, w, stride, an_stride);
// Tensor conf_mask, obj_mask;
// Tensor tx, ty, tw, th, tweight, tconf, tclass;
// conf_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// obj_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tx.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// ty.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tw.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// th.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tweight.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tconf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tclass.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
//
// math::SetConstant<platform::CPUDeviceContext, T> constant;
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &conf_mask, static_cast<T>(1.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &obj_mask, static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &tx,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &ty,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &tw,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &th,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &tweight, static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &tconf,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &tclass,
// static_cast<T>(0.0));
//
// PreProcessGTBox<T>(*gt_box, *gt_label, ignore_thresh, anchors,
// input_size,
// h, &conf_mask, &obj_mask, &tx, &ty, &tw, &th,
// &tweight,
// &tconf, &tclass);
//
// T* loss_data = loss->mutable_data<T>({n}, ctx.GetPlace());
// memset(loss_data, 0, n * sizeof(T));
// CalcYolov3Loss<T>(loss_data, *input, tx, ty, tw, th, tweight, tconf,
// tclass,
// conf_mask, obj_mask);
} }
}; };
...@@ -389,59 +668,172 @@ class Yolov3LossGradKernel : public framework::OpKernel<T> { ...@@ -389,59 +668,172 @@ class Yolov3LossGradKernel : public framework::OpKernel<T> {
auto* input = ctx.Input<Tensor>("X"); auto* input = ctx.Input<Tensor>("X");
auto* gt_box = ctx.Input<Tensor>("GTBox"); auto* gt_box = ctx.Input<Tensor>("GTBox");
auto* gt_label = ctx.Input<Tensor>("GTLabel"); auto* gt_label = ctx.Input<Tensor>("GTLabel");
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
auto anchors = ctx.Attr<std::vector<int>>("anchors"); auto anchors = ctx.Attr<std::vector<int>>("anchors");
auto anchor_mask = ctx.Attr<std::vector<int>>("anchor_mask");
int class_num = ctx.Attr<int>("class_num"); int class_num = ctx.Attr<int>("class_num");
float ignore_thresh = ctx.Attr<float>("ignore_thresh"); float ignore_thresh = ctx.Attr<float>("ignore_thresh");
auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X")); int downsample = ctx.Attr<int>("downsample");
auto* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
int input_size = ctx.Attr<int>("input_size");
const int n = input->dims()[0]; const int n = input->dims()[0];
const int c = input->dims()[1]; const int c = input->dims()[1];
const int h = input->dims()[2]; const int h = input->dims()[2];
const int w = input->dims()[3]; const int w = input->dims()[3];
const int an_num = anchors.size() / 2; const int an_num = anchors.size() / 2;
const int mask_num = anchor_mask.size();
Tensor conf_mask, obj_mask; const int b = gt_box->dims()[1];
Tensor tx, ty, tw, th, tweight, tconf, tclass; int input_size = downsample * h;
conf_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
obj_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace()); const T* input_data = input->data<T>();
tx.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace()); const T* gt_box_data = gt_box->data<T>();
ty.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace()); const int* gt_label_data = gt_label->data<int>();
tw.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace()); const T* loss_grad_data = loss_grad->data<T>();
th.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tweight.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tconf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
tclass.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
math::SetConstant<platform::CPUDeviceContext, T> constant;
constant(ctx.template device_context<platform::CPUDeviceContext>(),
&conf_mask, static_cast<T>(1.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(),
&obj_mask, static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tx,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &ty,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tw,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &th,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(),
&tweight, static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tconf,
static_cast<T>(0.0));
constant(ctx.template device_context<platform::CPUDeviceContext>(), &tclass,
static_cast<T>(0.0));
PreProcessGTBox<T>(*gt_box, *gt_label, ignore_thresh, anchors, input_size,
h, &conf_mask, &obj_mask, &tx, &ty, &tw, &th, &tweight,
&tconf, &tclass);
T* input_grad_data = T* input_grad_data =
input_grad->mutable_data<T>({n, c, h, w}, ctx.GetPlace()); input_grad->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
CalcYolov3LossGrad<T>(input_grad_data, *loss_grad, *input, tx, ty, tw, th, memset(input_grad_data, 0, input_grad->numel() * sizeof(T));
tweight, tconf, tclass, conf_mask, obj_mask);
Tensor objness;
int* objness_data =
objness.mutable_data<int>({n, mask_num, h, w}, ctx.GetPlace());
memset(objness_data, 0, objness.numel() * sizeof(int));
const int stride = h * w;
const int an_stride = (class_num + 5) * stride;
for (int i = 0; i < n; i++) {
for (int j = 0; j < mask_num; j++) {
for (int k = 0; k < h; k++) {
for (int l = 0; l < w; l++) {
int box_idx =
entry_index(i, j, k * w + l, mask_num, an_stride, stride, 0);
Box<T> pred =
get_yolo_box(input_data, anchors, l, k, anchor_mask[j], h,
input_size, box_idx, stride);
T best_iou = 0;
// int best_t = 0;
for (int t = 0; t < b; t++) {
if (isZero<T>(gt_box_data[i * b * 4 + t * 4]) &&
isZero<T>(gt_box_data[i * b * 4 + t * 4 + 1])) {
continue;
}
Box<T> gt = get_gt_box(gt_box_data, i, b, t);
T iou = box_iou(pred, gt);
if (iou > best_iou) {
best_iou = iou;
// best_t = t;
}
}
if (best_iou > ignore_thresh) {
int obj_idx = (i * mask_num + j) * stride + k * w + l;
objness_data[obj_idx] = -1;
}
}
}
}
for (int t = 0; t < b; t++) {
if (isZero<T>(gt_box_data[i * b * 4 + t * 4]) &&
isZero<T>(gt_box_data[i * b * 4 + t * 4 + 1])) {
continue;
}
Box<T> gt = get_gt_box(gt_box_data, i, b, t);
int gi = static_cast<int>(gt.x * w);
int gj = static_cast<int>(gt.y * h);
Box<T> gt_shift = gt;
gt_shift.x = 0.0;
gt_shift.y = 0.0;
T best_iou = 0.0;
int best_n = 0;
for (int an_idx = 0; an_idx < an_num; an_idx++) {
Box<T> an_box;
an_box.x = 0.0;
an_box.y = 0.0;
an_box.w = anchors[2 * an_idx] / static_cast<T>(input_size);
an_box.h = anchors[2 * an_idx + 1] / static_cast<T>(input_size);
float iou = box_iou<T>(an_box, gt_shift);
// TO DO: iou > 0.5 ?
if (iou > best_iou) {
best_iou = iou;
best_n = an_idx;
}
}
int mask_idx = mask_index(anchor_mask, best_n);
if (mask_idx >= 0) {
int box_idx = entry_index(i, mask_idx, gj * w + gi, mask_num,
an_stride, stride, 0);
CalcBoxLocationLossGrad<T>(input_grad_data, loss_grad_data[i],
input_data, gt, anchors, best_n, box_idx,
gi, gj, h, input_size, stride);
int obj_idx = (i * mask_num + mask_idx) * stride + gj * w + gi;
objness_data[obj_idx] = 1;
int label = gt_label_data[i * b + t];
int label_idx = entry_index(i, mask_idx, gj * w + gi, mask_num,
an_stride, stride, 5);
CalcLabelLossGrad<T>(input_grad_data, loss_grad_data[i], input_data,
label_idx, label, class_num, stride);
}
}
}
CalcObjnessLossGrad<T>(input_grad_data + 4 * stride, loss_grad_data,
input_data + 4 * stride, objness_data, n, mask_num,
h, w, stride, an_stride);
// const int n = input->dims()[0];
// const int c = input->dims()[1];
// const int h = input->dims()[2];
// const int w = input->dims()[3];
// const int an_num = anchors.size() / 2;
//
// Tensor conf_mask, obj_mask;
// Tensor tx, ty, tw, th, tweight, tconf, tclass;
// conf_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// obj_mask.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tx.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// ty.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tw.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// th.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tweight.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tconf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
// tclass.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
//
// math::SetConstant<platform::CPUDeviceContext, T> constant;
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &conf_mask, static_cast<T>(1.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &obj_mask, static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &tx,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &ty,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &tw,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(), &th,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &tweight, static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &tconf,
// static_cast<T>(0.0));
// constant(ctx.template device_context<platform::CPUDeviceContext>(),
// &tclass,
// static_cast<T>(0.0));
//
// PreProcessGTBox<T>(*gt_box, *gt_label, ignore_thresh, anchors,
// input_size,
// h, &conf_mask, &obj_mask, &tx, &ty, &tw, &th,
// &tweight,
// &tconf, &tclass);
//
// T* input_grad_data =
// input_grad->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
// CalcYolov3LossGrad<T>(input_grad_data, *loss_grad, *input, tx, ty, tw,
// th,
// tweight, tconf, tclass, conf_mask, obj_mask);
} }
}; };
......
python/paddle/fluid/layers/detection.py
浏览文件 @ e7e4f084
...@@ -413,9 +413,10 @@ def yolov3_loss(x, ...@@ -413,9 +413,10 @@ def yolov3_loss(x,
gtbox, gtbox,
gtlabel, gtlabel,
anchors, anchors,
anchor_mask,
class_num, class_num,
ignore_thresh, ignore_thresh,
input_size, downsample,
name=None): name=None):
""" """
${comment} ${comment}
...@@ -430,9 +431,10 @@ def yolov3_loss(x, ...@@ -430,9 +431,10 @@ def yolov3_loss(x,
gtlabel (Variable): class id of ground truth boxes, shoud be ins shape gtlabel (Variable): class id of ground truth boxes, shoud be ins shape
of [N, B]. of [N, B].
anchors (list|tuple): ${anchors_comment} anchors (list|tuple): ${anchors_comment}
anchor_mask (list|tuple): ${anchor_mask_comment}
class_num (int): ${class_num_comment} class_num (int): ${class_num_comment}
ignore_thresh (float): ${ignore_thresh_comment} ignore_thresh (float): ${ignore_thresh_comment}
input_size (int): ${input_size_comment} downsample (int): ${downsample_comment}
name (string): the name of yolov3 loss name (string): the name of yolov3 loss
Returns: Returns:
...@@ -452,7 +454,8 @@ def yolov3_loss(x, ...@@ -452,7 +454,8 @@ def yolov3_loss(x,
x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32') x = fluid.layers.data(name='x', shape=[255, 13, 13], dtype='float32')
gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32') gtbox = fluid.layers.data(name='gtbox', shape=[6, 5], dtype='float32')
gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32') gtlabel = fluid.layers.data(name='gtlabel', shape=[6, 1], dtype='int32')
anchors = [10, 13, 16, 30, 33, 23] anchors = [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326]
anchors = [0, 1, 2]
loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80 loss = fluid.layers.yolov3_loss(x=x, gtbox=gtbox, class_num=80
anchors=anchors, ignore_thresh=0.5) anchors=anchors, ignore_thresh=0.5)
""" """
...@@ -466,6 +469,8 @@ def yolov3_loss(x, ...@@ -466,6 +469,8 @@ def yolov3_loss(x,
raise TypeError("Input gtlabel of yolov3_loss must be Variable") raise TypeError("Input gtlabel of yolov3_loss must be Variable")
if not isinstance(anchors, list) and not isinstance(anchors, tuple): if not isinstance(anchors, list) and not isinstance(anchors, tuple):
raise TypeError("Attr anchors of yolov3_loss must be list or tuple") raise TypeError("Attr anchors of yolov3_loss must be list or tuple")
if not isinstance(anchor_mask, list) and not isinstance(anchor_mask, tuple):
raise TypeError("Attr anchor_mask of yolov3_loss must be list or tuple")
if not isinstance(class_num, int): if not isinstance(class_num, int):
raise TypeError("Attr class_num of yolov3_loss must be an integer") raise TypeError("Attr class_num of yolov3_loss must be an integer")
if not isinstance(ignore_thresh, float): if not isinstance(ignore_thresh, float):
...@@ -480,9 +485,10 @@ def yolov3_loss(x, ...@@ -480,9 +485,10 @@ def yolov3_loss(x,
attrs = { attrs = {
"anchors": anchors, "anchors": anchors,
"anchor_mask": anchor_mask,
"class_num": class_num, "class_num": class_num,
"ignore_thresh": ignore_thresh, "ignore_thresh": ignore_thresh,
"input_size": input_size, "downsample": downsample,
} }
helper.append_op( helper.append_op(
......
python/paddle/fluid/tests/test_detection.py
浏览文件 @ e7e4f084
...@@ -463,8 +463,8 @@ class TestYoloDetection(unittest.TestCase): ...@@ -463,8 +463,8 @@ class TestYoloDetection(unittest.TestCase):
x = layers.data(name='x', shape=[30, 7, 7], dtype='float32') x = layers.data(name='x', shape=[30, 7, 7], dtype='float32')
gtbox = layers.data(name='gtbox', shape=[10, 4], dtype='float32') gtbox = layers.data(name='gtbox', shape=[10, 4], dtype='float32')
gtlabel = layers.data(name='gtlabel', shape=[10], dtype='int32') gtlabel = layers.data(name='gtlabel', shape=[10], dtype='int32')
loss = layers.yolov3_loss(x, gtbox, gtlabel, [10, 13, 30, 13], 10, loss = layers.yolov3_loss(x, gtbox, gtlabel, [10, 13, 30, 13],
0.7, 416) [0, 1], 10, 0.7, 32)
self.assertIsNotNone(loss) self.assertIsNotNone(loss)
......
python/paddle/fluid/tests/unittests/test_yolov3_loss_op.py
浏览文件 @ e7e4f084
...@@ -22,32 +22,42 @@ from op_test import OpTest ...@@ -22,32 +22,42 @@ from op_test import OpTest
from paddle.fluid import core from paddle.fluid import core
# def l1loss(x, y, weight):
def l1loss(x, y, weight): # n = x.shape[0]
n = x.shape[0] # x = x.reshape((n, -1))
x = x.reshape((n, -1)) # y = y.reshape((n, -1))
y = y.reshape((n, -1)) # weight = weight.reshape((n, -1))
weight = weight.reshape((n, -1)) # return (np.abs(y - x) * weight).sum(axis=1)
return (np.abs(y - x) * weight).sum(axis=1) #
#
# def mse(x, y, weight):
# n = x.shape[0]
# x = x.reshape((n, -1))
# y = y.reshape((n, -1))
# weight = weight.reshape((n, -1))
# return ((y - x)**2 * weight).sum(axis=1)
#
#
# def sce(x, label, weight):
# n = x.shape[0]
# x = x.reshape((n, -1))
# label = label.reshape((n, -1))
# weight = weight.reshape((n, -1))
# sigmoid_x = expit(x)
# term1 = label * np.log(sigmoid_x)
# term2 = (1.0 - label) * np.log(1.0 - sigmoid_x)
# return ((-term1 - term2) * weight).sum(axis=1)
def mse(x, y, weight): def l1loss(x, y):
n = x.shape[0] return abs(x - y)
x = x.reshape((n, -1))
y = y.reshape((n, -1))
weight = weight.reshape((n, -1))
return ((y - x)**2 * weight).sum(axis=1)
def sce(x, label, weight): def sce(x, label):
n = x.shape[0]
x = x.reshape((n, -1))
label = label.reshape((n, -1))
weight = weight.reshape((n, -1))
sigmoid_x = expit(x) sigmoid_x = expit(x)
term1 = label * np.log(sigmoid_x) term1 = label * np.log(sigmoid_x)
term2 = (1.0 - label) * np.log(1.0 - sigmoid_x) term2 = (1.0 - label) * np.log(1.0 - sigmoid_x)
return ((-term1 - term2) * weight).sum(axis=1) return -term1 - term2
def box_iou(box1, box2): def box_iou(box1, box2):
...@@ -160,6 +170,121 @@ def YoloV3Loss(x, gtbox, gtlabel, attrs): ...@@ -160,6 +170,121 @@ def YoloV3Loss(x, gtbox, gtlabel, attrs):
return loss_x + loss_y + loss_w + loss_h + loss_obj + loss_class return loss_x + loss_y + loss_w + loss_h + loss_obj + loss_class
def sigmoid(x):
return 1.0 / (1.0 + np.exp(-1.0 * x))
def batch_xywh_box_iou(box1, box2):
b1_left = box1[:, :, 0] - box1[:, :, 2] / 2
b1_right = box1[:, :, 0] + box1[:, :, 2] / 2
b1_top = box1[:, :, 1] - box1[:, :, 3] / 2
b1_bottom = box1[:, :, 1] + box1[:, :, 3] / 2
b2_left = box2[:, :, 0] - box2[:, :, 2] / 2
b2_right = box2[:, :, 0] + box2[:, :, 2] / 2
b2_top = box2[:, :, 1] - box2[:, :, 3] / 2
b2_bottom = box2[:, :, 1] + box2[:, :, 3] / 2
left = np.maximum(b1_left[:, :, np.newaxis], b2_left[:, np.newaxis, :])
right = np.minimum(b1_right[:, :, np.newaxis], b2_right[:, np.newaxis, :])
top = np.maximum(b1_top[:, :, np.newaxis], b2_top[:, np.newaxis, :])
bottom = np.minimum(b1_bottom[:, :, np.newaxis],
b2_bottom[:, np.newaxis, :])
inter_w = np.clip(right - left, 0., 1.)
inter_h = np.clip(bottom - top, 0., 1.)
inter_area = inter_w * inter_h
b1_area = (b1_right - b1_left) * (b1_bottom - b1_top)
b2_area = (b2_right - b2_left) * (b2_bottom - b2_top)
union = b1_area[:, :, np.newaxis] + b2_area[:, np.newaxis, :] - inter_area
return inter_area / union
def YOLOv3Loss(x, gtbox, gtlabel, attrs):
n, c, h, w = x.shape
b = gtbox.shape[1]
anchors = attrs['anchors']
an_num = len(anchors) // 2
anchor_mask = attrs['anchor_mask']
mask_num = len(anchor_mask)
class_num = attrs["class_num"]
ignore_thresh = attrs['ignore_thresh']
downsample = attrs['downsample']
input_size = downsample * h
x = x.reshape((n, mask_num, 5 + class_num, h, w)).transpose((0, 1, 3, 4, 2))
loss = np.zeros((n)).astype('float32')
pred_box = x[:, :, :, :, :4].copy()
grid_x = np.tile(np.arange(w).reshape((1, w)), (h, 1))
grid_y = np.tile(np.arange(h).reshape((h, 1)), (1, w))
pred_box[:, :, :, :, 0] = (grid_x + sigmoid(pred_box[:, :, :, :, 0])) / w
pred_box[:, :, :, :, 1] = (grid_y + sigmoid(pred_box[:, :, :, :, 1])) / h
mask_anchors = []
for m in anchor_mask:
mask_anchors.append((anchors[2 * m], anchors[2 * m + 1]))
anchors_s = np.array(
[(an_w / input_size, an_h / input_size) for an_w, an_h in mask_anchors])
anchor_w = anchors_s[:, 0:1].reshape((1, mask_num, 1, 1))
anchor_h = anchors_s[:, 1:2].reshape((1, mask_num, 1, 1))
pred_box[:, :, :, :, 2] = np.exp(pred_box[:, :, :, :, 2]) * anchor_w
pred_box[:, :, :, :, 3] = np.exp(pred_box[:, :, :, :, 3]) * anchor_h
pred_box = pred_box.reshape((n, -1, 4))
pred_obj = x[:, :, :, :, 4].reshape((n, -1))
objness = np.zeros(pred_box.shape[:2])
ious = batch_xywh_box_iou(pred_box, gtbox)
ious_max = np.max(ious, axis=-1)
objness = np.where(ious_max > ignore_thresh, -np.ones_like(objness),
objness)
gtbox_shift = gtbox.copy()
gtbox_shift[:, :, 0] = 0
gtbox_shift[:, :, 1] = 0
anchors = [(anchors[2 * i], anchors[2 * i + 1]) for i in range(0, an_num)]
anchors_s = np.array(
[(an_w / input_size, an_h / input_size) for an_w, an_h in anchors])
anchor_boxes = np.concatenate(
[np.zeros_like(anchors_s), anchors_s], axis=-1)
anchor_boxes = np.tile(anchor_boxes[np.newaxis, :, :], (n, 1, 1))
ious = batch_xywh_box_iou(gtbox_shift, anchor_boxes)
iou_matches = np.argmax(ious, axis=-1)
for i in range(n):
for j in range(b):
if gtbox[i, j, 2:].sum() == 0:
continue
if iou_matches[i, j] not in anchor_mask:
continue
an_idx = anchor_mask.index(iou_matches[i, j])
gi = int(gtbox[i, j, 0] * w)
gj = int(gtbox[i, j, 1] * h)
tx = gtbox[i, j, 0] * w - gi
ty = gtbox[i, j, 1] * w - gj
tw = np.log(gtbox[i, j, 2] * input_size / mask_anchors[an_idx][0])
th = np.log(gtbox[i, j, 3] * input_size / mask_anchors[an_idx][1])
scale = 2.0 - gtbox[i, j, 2] * gtbox[i, j, 3]
loss[i] += sce(x[i, an_idx, gj, gi, 0], tx) * scale
loss[i] += sce(x[i, an_idx, gj, gi, 1], ty) * scale
loss[i] += l1loss(x[i, an_idx, gj, gi, 2], tw) * scale
loss[i] += l1loss(x[i, an_idx, gj, gi, 3], th) * scale
objness[i, an_idx * h * w + gj * w + gi] = 1
for label_idx in range(class_num):
loss[i] += sce(x[i, an_idx, gj, gi, 5 + label_idx],
int(label_idx == gtlabel[i, j]))
for j in range(mask_num * h * w):
if objness[i, j] >= 0:
loss[i] += sce(pred_obj[i, j], objness[i, j])
return loss
class TestYolov3LossOp(OpTest): class TestYolov3LossOp(OpTest):
def setUp(self): def setUp(self):
self.initTestCase() self.initTestCase()
...@@ -171,13 +296,14 @@ class TestYolov3LossOp(OpTest): ...@@ -171,13 +296,14 @@ class TestYolov3LossOp(OpTest):
self.attrs = { self.attrs = {
"anchors": self.anchors, "anchors": self.anchors,
"anchor_mask": self.anchor_mask,
"class_num": self.class_num, "class_num": self.class_num,
"ignore_thresh": self.ignore_thresh, "ignore_thresh": self.ignore_thresh,
"input_size": self.input_size, "downsample": self.downsample,
} }
self.inputs = {'X': x, 'GTBox': gtbox, 'GTLabel': gtlabel} self.inputs = {'X': x, 'GTBox': gtbox, 'GTLabel': gtlabel}
self.outputs = {'Loss': YoloV3Loss(x, gtbox, gtlabel, self.attrs)} self.outputs = {'Loss': YOLOv3Loss(x, gtbox, gtlabel, self.attrs)}
def test_check_output(self): def test_check_output(self):
place = core.CPUPlace() place = core.CPUPlace()
...@@ -189,15 +315,19 @@ class TestYolov3LossOp(OpTest): ...@@ -189,15 +315,19 @@ class TestYolov3LossOp(OpTest):
place, ['X'], place, ['X'],
'Loss', 'Loss',
no_grad_set=set(["GTBox", "GTLabel"]), no_grad_set=set(["GTBox", "GTLabel"]),
max_relative_error=0.31) max_relative_error=0.15)
def initTestCase(self): def initTestCase(self):
self.anchors = [12, 12] self.anchors = [
10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198,
373, 326
]
self.anchor_mask = [0, 1, 2]
self.class_num = 5 self.class_num = 5
self.ignore_thresh = 0.5 self.ignore_thresh = 0.7
self.input_size = 416 self.downsample = 32
self.x_shape = (1, len(self.anchors) // 2 * (5 + self.class_num), 3, 3) self.x_shape = (3, len(self.anchor_mask) * (5 + self.class_num), 5, 5)
self.gtbox_shape = (1, 5, 4) self.gtbox_shape = (3, 10, 4)
if __name__ == "__main__": if __name__ == "__main__":
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或