提交 84a525a3 编写于 作者: C chenweihang

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into squeeze_op

Fixed-point quantization uses lower bits, for example, 2-bit, 3-bit or 8-bit fixed point to represent weights and activations, which usually are in singe-precision float-point with 32 bits. The fixed-point representation has advantages in reducing memory bandwidth, lowering power consumption and computational resources as well as the model storage requirements. It is especially important for the inference in embedded-device deployment.
According to some experiments, the apporach to quantize the model trained in float point directly works effectively on the large models, like the VGG model having many parameters. But the accuracy drops a lot for the small model. In order to improve the tradeoff between accuracy and latency, many quantized training apporaches are proposed.
This document is to design a quantized training framework on Fluid. The first part will introduce how to quantize, The second part will describe the quantized training framework. The last part will illustrate how to calculate the quantization scale.
### How to quantize
There are many ways to quantize the float value to fixed-point value. For example:
$$ r = min(max(x, a), b)$$
$$ s = \frac{b - a}{n - 1} $$
$$ q = \left \lfloor \frac{r - a}{s} \right \rceil $$
where, $x$ is the float value to be quantized, $[a, b]$ is the quantization range, $a$ is the minimum value and $b$ is the maximal value. $\left \lfloor \right \rceil$ denotes rounding to the nearest integer. If the quantization level is $k$, $n$ is $2^k$, for example, $k$ is 8 and $n$ is 256. $q$ is the quantized integer.
The quantization we applied is parameterized by the number of quantization levels and maximum absolute value:
$$ M = max(abs(x)) $$
$$ q = \left \lfloor \frac{x}{M} * (n - 1) \right \rceil $$
where, $x$ is the float value to be quantized, $M$ is maximum absolute value. $\left \lfloor \right \rceil$ denotes rounding to the nearest integer. For 8 bit quantization, $n=2^{8}=256$. $q$ is the quantized integer.
Wether the *min-max* quantization or *max-abs* quantization, they also can be represent:
$q = scale * r + b$
We call *min-max*, *max-abs* as the quantization arguments, also call them quantization scale or quantization range.
How to calculate the quantization scale (or maximum absolute value) for inference will be described in the last part.
### Training Framework
#### Forward pass
The forward pass is simulated quantization, see Figure 1.
The training framework is as following figure.
<p align="center">
<img src="quantization_forward.png" width="300" height="340"><br/>
Figure 1. Forward in training with simulated quantization.
</p>
- Firstly, both input and weight will be quantized to 8-bit integers.
- Second, do the multiplication (or convolution) operation with integers.
- Third, dequantize the multiplication (or convolution) results to 32-bit float point.
- Finally, do bias-addition in float type of 32 bit. Here, the bias is not quantized.
For general matrix multiplication (GEMM), quantize for $X$ and $W$:
$$ X_q = \left \lfloor \frac{X}{X_m} * (n - 1) \right \rceil $$
$$ W_q = \left \lfloor \frac{W}{W_m} * (n - 1) \right \rceil $$
Do GEMM:
$$ Y = X_q * W_q $$
Dequantize $Y$:
$$
\begin{align}
Y_{dq} &=\frac{Y}{(n - 1) * (n - 1)} * X_m * W_m \\\
&=\frac{X_q * W_q}{(n - 1) * (n - 1)} * X_m * W_m \\\
&=(\frac{X_q}{n - 1} * X_m) * (\frac{W_q}{n - 1} * W_m)
\end{align}
$$
From these formulas, dequantization also can be moved before GEMM, do dequantization for $Xq$ and $Wq$ at first, then do GEMM. The forward workflow in training is equivalent to following framework.
<p align="center">
<img src="quantization_equivalent_forward.png" width="300" height="330"><br/>
Figure 2. Equivalent forward in training with simulated quantization.
</p>
We use this equivalent workflow in the training. In our desigin, there is a quantization transpiler to insert the quantization operator and the de-quantization operator in the Fluid `ProgramDesc`. Since the outputs of quantization and de-quantization operator are still in floating point, they are called faked quantization and de-quantization operator. And the training framework is called simulated quantization.
#### Backward pass
See Figure 3. The gradients are calculated by dequantized weights and activations. All inputs and outputs are float point with 32-bit. And in the weight updating process, the gradients will be added to the original weight, not the quantized or dequantized weights.
<p align="center">
<img src="quantization_backward_and_optimization.png"><br/>
Figure 3. Backward and weight updating in training with simulated quantization.
</p>
So the quantization transipler will change some inputs of the corresponding backward operators.
### How to calculate quantization scale
There are two strategies to calculate quantization scale, we call them dynamic and static strategy. The dynamic strategy calculates the quantization scale value each iteration. The static strategy keeps the quantization scale for different inputs.
For weights, we apply the dynamic strategy in the training, that is to say, the quantization scale will be recalculated during each iteration until the traning is finished.
For activations, the quantization scales are estimated during training, then used in inference. There are several different ways to estimate them:
1. Calculate the mean of maximum absolute during a window.
2. Calculate the max of maximum absolute during a window.
3. Calculate the running mean of maximum absolute during a window, as follows:
$$ Vt = (1 - k) * V + k * V_{t-1} $$
where, $V$ is the maximum absolute value of current batch, $Vt$ is the running mean value. $k$ is a factor, such as 0.9.
...@@ -27,7 +27,8 @@ anchor_generator_op.cu) ...@@ -27,7 +27,8 @@ anchor_generator_op.cu)
detection_library(target_assign_op SRCS target_assign_op.cc detection_library(target_assign_op SRCS target_assign_op.cc
target_assign_op.cu) target_assign_op.cu)
detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc detection_library(polygon_box_transform_op SRCS polygon_box_transform_op.cc
polygon_box_transform_op.cu) polygon_box_transform_op.cu)
detection_library(rpn_target_assign_op SRCS rpn_target_assign_op.cc)
# Export local libraries to parent # Export local libraries to parent
set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE) set(DETECTION_LIBRARY ${LOCAL_DETECTION_LIBS} PARENT_SCOPE)
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <random>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
using LoDTensor = framework::LoDTensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
class RpnTargetAssignOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("DistMat"),
"Input(DistMat) of RpnTargetAssignOp should not be null");
PADDLE_ENFORCE(
ctx->HasOutput("LocationIndex"),
"Output(LocationIndex) of RpnTargetAssignOp should not be null");
PADDLE_ENFORCE(
ctx->HasOutput("ScoreIndex"),
"Output(ScoreIndex) of RpnTargetAssignOp should not be null");
PADDLE_ENFORCE(
ctx->HasOutput("TargetLabel"),
"Output(TargetLabel) of RpnTargetAssignOp should not be null");
auto in_dims = ctx->GetInputDim("DistMat");
PADDLE_ENFORCE_EQ(in_dims.size(), 2,
"The rank of Input(DistMat) must be 2.");
}
};
template <typename T>
class RpnTargetAssignKernel : public framework::OpKernel<T> {
public:
void ScoreAssign(const T* dist_data, const Tensor& anchor_to_gt_max,
const int row, const int col, const float pos_threshold,
const float neg_threshold, int64_t* target_label_data,
std::vector<int>* fg_inds, std::vector<int>* bg_inds) const {
int fg_offset = fg_inds->size();
int bg_offset = bg_inds->size();
for (int64_t i = 0; i < row; ++i) {
const T* v = dist_data + i * col;
T max_dist = *std::max_element(v, v + col);
for (int64_t j = 0; j < col; ++j) {
T val = dist_data[i * col + j];
if (val == max_dist) target_label_data[j] = 1;
}
}
// Pick the fg/bg and count the number
for (int64_t j = 0; j < col; ++j) {
if (anchor_to_gt_max.data<T>()[j] > pos_threshold) {
target_label_data[j] = 1;
} else if (anchor_to_gt_max.data<T>()[j] < neg_threshold) {
target_label_data[j] = 0;
}
if (target_label_data[j] == 1) {
fg_inds->push_back(fg_offset + j);
} else if (target_label_data[j] == 0) {
bg_inds->push_back(bg_offset + j);
}
}
}
void ReservoirSampling(const int num, const int offset,
std::minstd_rand engine,
std::vector<int>* inds) const {
std::uniform_real_distribution<float> uniform(0, 1);
if (inds->size() > num) {
for (int i = num; i < inds->size(); ++i) {
int rng_ind = std::floor(uniform(engine) * i);
if (rng_ind < num)
std::iter_swap(inds->begin() + rng_ind + offset,
inds->begin() + i + offset);
}
}
}
void RpnTargetAssign(const framework::ExecutionContext& ctx,
const Tensor& dist, const float pos_threshold,
const float neg_threshold, const int rpn_batch_size,
const int fg_num, std::minstd_rand engine,
std::vector<int>* fg_inds, std::vector<int>* bg_inds,
int64_t* target_label_data) const {
auto* dist_data = dist.data<T>();
int64_t row = dist.dims()[0];
int64_t col = dist.dims()[1];
int fg_offset = fg_inds->size();
int bg_offset = bg_inds->size();
// Calculate the max IoU between anchors and gt boxes
Tensor anchor_to_gt_max;
anchor_to_gt_max.mutable_data<T>(
framework::make_ddim({static_cast<int64_t>(col), 1}),
platform::CPUPlace());
auto& place = *ctx.template device_context<platform::CPUDeviceContext>()
.eigen_device();
auto x = EigenMatrix<T>::From(dist);
auto x_col_max = EigenMatrix<T>::From(anchor_to_gt_max);
x_col_max.device(place) =
x.maximum(Eigen::DSizes<int, 1>(0))
.reshape(Eigen::DSizes<int, 2>(static_cast<int64_t>(col), 1));
// Follow the Faster RCNN's implementation
ScoreAssign(dist_data, anchor_to_gt_max, row, col, pos_threshold,
neg_threshold, target_label_data, fg_inds, bg_inds);
// Reservoir Sampling
ReservoirSampling(fg_num, fg_offset, engine, fg_inds);
int bg_num = rpn_batch_size - fg_inds->size();
ReservoirSampling(bg_num, bg_offset, engine, bg_inds);
}
void Compute(const framework::ExecutionContext& context) const override {
auto* dist = context.Input<LoDTensor>("DistMat");
auto* loc_index = context.Output<Tensor>("LocationIndex");
auto* score_index = context.Output<Tensor>("ScoreIndex");
auto* tgt_lbl = context.Output<Tensor>("TargetLabel");
auto col = dist->dims()[1];
int64_t n = dist->lod().size() == 0UL
? 1
: static_cast<int64_t>(dist->lod().back().size() - 1);
if (dist->lod().size()) {
PADDLE_ENFORCE_EQ(dist->lod().size(), 1UL,
"Only support 1 level of LoD.");
}
int rpn_batch_size = context.Attr<int>("rpn_batch_size_per_im");
float pos_threshold = context.Attr<float>("rpn_positive_overlap");
float neg_threshold = context.Attr<float>("rpn_negative_overlap");
float fg_fraction = context.Attr<float>("fg_fraction");
int fg_num = static_cast<int>(rpn_batch_size * fg_fraction);
int64_t* target_label_data =
tgt_lbl->mutable_data<int64_t>({n * col, 1}, context.GetPlace());
auto& dev_ctx = context.device_context<platform::CPUDeviceContext>();
math::SetConstant<platform::CPUDeviceContext, int64_t> iset;
iset(dev_ctx, tgt_lbl, static_cast<int>(-1));
std::vector<int> fg_inds;
std::vector<int> bg_inds;
std::random_device rnd;
std::minstd_rand engine;
int seed =
context.Attr<bool>("fix_seed") ? context.Attr<int>("seed") : rnd();
engine.seed(seed);
if (n == 1) {
RpnTargetAssign(context, *dist, pos_threshold, neg_threshold,
rpn_batch_size, fg_num, engine, &fg_inds, &bg_inds,
target_label_data);
} else {
auto lod = dist->lod().back();
for (size_t i = 0; i < lod.size() - 1; ++i) {
Tensor one_ins = dist->Slice(lod[i], lod[i + 1]);
RpnTargetAssign(context, one_ins, pos_threshold, neg_threshold,
rpn_batch_size, fg_num, engine, &fg_inds, &bg_inds,
target_label_data + i * col);
}
}
int* loc_index_data = loc_index->mutable_data<int>(
{static_cast<int>(fg_inds.size())}, context.GetPlace());
int* score_index_data = score_index->mutable_data<int>(
{static_cast<int>(fg_inds.size() + bg_inds.size())},
context.GetPlace());
memcpy(loc_index_data, reinterpret_cast<int*>(&fg_inds[0]),
fg_inds.size() * sizeof(int));
memcpy(score_index_data, reinterpret_cast<int*>(&fg_inds[0]),
fg_inds.size() * sizeof(int));
memcpy(score_index_data + fg_inds.size(),
reinterpret_cast<int*>(&bg_inds[0]), bg_inds.size() * sizeof(int));
}
};
class RpnTargetAssignOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput(
"DistMat",
"(LoDTensor or Tensor) this input is a 2-D LoDTensor with shape "
"[K, M]. It is pair-wise distance matrix between the entities "
"represented by each row and each column. For example, assumed one "
"entity is A with shape [K], another entity is B with shape [M]. The "
"DistMat[i][j] is the distance between A[i] and B[j]. The bigger "
"the distance is, the better macthing the pairs are. Please note, "
"This tensor can contain LoD information to represent a batch of "
"inputs. One instance of this batch can contain different numbers of "
"entities.");
AddAttr<float>(
"rpn_positive_overlap",
"Minimum overlap required between an anchor and ground-truth "
"box for the (anchor, gt box) pair to be a positive example.")
.SetDefault(0.7);
AddAttr<float>(
"rpn_negative_overlap",
"Maximum overlap allowed between an anchor and ground-truth "
"box for the (anchor, gt box) pair to be a negative examples.")
.SetDefault(0.3);
AddAttr<float>(
"fg_fraction",
"Target fraction of RoI minibatch that "
"is labeled foreground (i.e. class > 0), 0-th class is background.")
.SetDefault(0.25);
AddAttr<int>("rpn_batch_size_per_im",
"Total number of RPN examples per image.")
.SetDefault(256);
AddAttr<bool>("fix_seed",
"A flag indicating whether to use a fixed seed to generate "
"random mask. NOTE: DO NOT set this flag to true in "
"training. Setting this flag to true is only useful in "
"unittest.")
.SetDefault(false);
AddAttr<int>("seed", "RpnTargetAssign random seed.").SetDefault(0);
AddOutput(
"LocationIndex",
"(Tensor), The indexes of foreground anchors in all RPN anchors, the "
"shape of the LocationIndex is [F], F depends on the value of input "
"tensor and attributes.");
AddOutput(
"ScoreIndex",
"(Tensor), The indexes of foreground and background anchors in all "
"RPN anchors(The rest anchors are ignored). The shape of the "
"ScoreIndex is [F + B], F and B depend on the value of input "
"tensor and attributes.");
AddOutput("TargetLabel",
"(Tensor<int64_t>), The target labels of each anchor with shape "
"[K * M, 1], "
"K and M is the same as they are in DistMat.");
AddComment(R"DOC(
This operator can be, for given the IoU between the ground truth bboxes and the
anchors, to assign classification and regression targets to each prediction.
The Score index and LocationIndex will be generated according to the DistMat.
The rest anchors would not contibute to the RPN training loss
ScoreIndex is composed of foreground anchor indexes(positive labels) and
background anchor indexes(negative labels). LocationIndex is exactly same
as the foreground anchor indexes since we can not assign regression target to
the background anchors.
The classification targets(TargetLabel) is a binary class label (of being
an object or not). Following the paper of Faster-RCNN, the positive labels
are two kinds of anchors: (i) the anchor/anchors with the highest IoU
overlap with a ground-truth box, or (ii) an anchor that has an IoU overlap
higher than rpn_positive_overlap(0.7) with any ground-truth box. Note that
a single ground-truth box may assign positive labels to multiple anchors.
A non-positive anchor is when its IoU ratio is lower than rpn_negative_overlap
(0.3) for all ground-truth boxes. Anchors that are neither positive nor
negative do not contribute to the training objective.
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(rpn_target_assign, ops::RpnTargetAssignOp,
ops::RpnTargetAssignOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(rpn_target_assign, ops::RpnTargetAssignKernel<float>,
ops::RpnTargetAssignKernel<double>);
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/fake_quantize_op.h"
#include <string>
namespace paddle {
namespace operators {
class FakeQuantizeOp : public framework::OperatorWithKernel {
public:
FakeQuantizeOp(const std::string &type,
const framework::VariableNameMap &inputs,
const framework::VariableNameMap &outputs,
const framework::AttributeMap &attrs)
: OperatorWithKernel(type, inputs, outputs, attrs) {}
void InferShape(framework::InferShapeContext *ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of FakeQuantizeOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("Out"),
"Output(Out) of FakeQuantizeOp should not be null.");
PADDLE_ENFORCE(ctx->HasOutput("OutMovingScale"),
"OutMovingScale(Out) of FakeQuantizeOp should not be null");
// if (ctx->HasInput("InMovingScale")) {
ctx->SetOutputDim("OutMovingScale", ctx->GetInputDim("InMovingScale"));
//}
// if (ctx->HasInput("InScales")) {
PADDLE_ENFORCE(ctx->HasOutput("OutScales"),
"OutScales(Out) of FakeQuantizeOp should not be null");
ctx->SetOutputDim("OutScales", ctx->GetInputDim("InScales"));
// PADDLE_ENFORCE_EQ(ctx->Inputs("InScales")[0],
// ctx->Outputs("OutScales")[0],
// "Mean and MeanOut should share the same memory");
//}
ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
ctx->ShareLoD("X", /*->*/ "Out");
}
};
class FakeQuantizeOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X", "(Tensor) Input tensor of scale operator.");
AddInput("InScales", "(Tensor) scale buffer, used in static quantization.")
.AsDispensable();
AddInput("InMovingScale", "Last scale, used in static quantization.")
.AsDispensable();
AddInput("InCurrentIter",
"Last iteration number, used in static quantization.")
.AsDispensable();
AddOutput("Out", "(Tensor) Output of quantized low level tensor.");
AddOutput("OutScales",
"(Tensor) scale buffer, used in static quantization.")
.AsDispensable();
AddOutput("OutMovingScale", " Current scale");
AddOutput("OutCurrentIter", "Current iteration number.").AsDispensable();
AddAttr<std::string>("quantize_type",
"(string, default abs_max)"
"The scaling tpe of the quantize operator.")
.SetDefault("abs_max");
AddAttr<int>("window_size", "(int, default 10000)").SetDefault(10000);
AddAttr<int>("bit_length", "(int, default 8)")
.SetDefault(8)
.AddCustomChecker([](const int &bit_length) {
PADDLE_ENFORCE(bit_length >= 1 && bit_length <= 16,
"'bit_length' should be between 1 and 16.");
});
AddAttr<bool>("is_test", "").SetDefault(false);
AddComment(R"DOC(
FakeQuantize operator
quantize_type = abs_max:
$$scale = max(abs(x))$$
quantize_type = range_abs_max:
$$scale = max(max(abs(x)), history_abs_max)$$
quantize_type = moving_average_abs_max:
$$scale = 0.1*scale+0.9*new_abs_max)$$
$$Out = scale*X$$
)DOC");
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(fake_quantize, ops::FakeQuantizeOp, ops::FakeQuantizeOpMaker,
paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(
fake_quantize,
ops::FakeQuantizeKernel<paddle::platform::CPUDeviceContext, float>,
ops::FakeQuantizeKernel<paddle::platform::CPUDeviceContext, double>);
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include <string>
#include "paddle/fluid/operators/fake_quantize_op.h"
#include "paddle/fluid/platform/cuda_primitives.h"
namespace paddle {
namespace operators {
template <typename T>
__global__ void FindAbsMaxKernel(const int n, const T* in, T* out) {
int bid = threadIdx.x + blockIdx.x * blockDim.x;
int tid = threadIdx.x;
extern __shared__ T shared_max_data[];
if (gridDim.x > 1) {
shared_max_data[tid] = T(0);
for (int i = bid; i < n; i += blockDim.x * gridDim.x) {
T tmp = fabs(in[i]);
if (tmp > shared_max_data[tid]) {
shared_max_data[tid] = tmp;
}
}
} else {
if (bid < n) {
shared_max_data[tid] = fabs(in[bid]);
} else {
shared_max_data[tid] = T(0);
}
}
__syncthreads();
for (int i = blockDim.x / 2; i > 0; i >>= 1) {
if (tid < i && shared_max_data[tid] < shared_max_data[tid + i]) {
shared_max_data[tid] = shared_max_data[tid + i];
}
__syncthreads();
}
if (tid == 0) {
out[blockIdx.x] = shared_max_data[0];
}
}
float FindAbsMaxGpu(const platform::CUDADeviceContext& ctx, const float* array,
int length) {
float host_max;
int kNumTheads = 1024;
int gridDimx = (kNumTheads - 1 + length) / kNumTheads;
gridDimx = (gridDimx > kNumTheads) ? kNumTheads : gridDimx;
framework::Tensor t;
float* device_max = t.mutable_data<float>(framework::make_ddim({gridDimx}),
platform::CUDAPlace());
FindAbsMaxKernel<float><<<gridDimx, kNumTheads, kNumTheads * sizeof(float),
ctx.stream()>>>(length, array, device_max);
FindAbsMaxKernel<
float><<<1, kNumTheads, kNumTheads * sizeof(float), ctx.stream()>>>(
gridDimx, device_max, device_max);
PADDLE_ENFORCE_EQ(
cudaMemcpy(&host_max, device_max, sizeof(float), cudaMemcpyDeviceToHost),
cudaSuccess, "cudaMemcpy failed");
return host_max;
}
template <typename T>
__global__ void ApplySaturateKernel(const int n, const T* in, T* out,
int* num_saturate, const T min,
const T max) {
int bid = threadIdx.x + blockIdx.x * blockDim.x;
int tid = threadIdx.x;
extern __shared__ int shared_count[];
shared_count[tid] = 0;
for (int i = bid; i < n; i += blockDim.x * gridDim.x) {
if (in[i] > max) {
out[i] = max;
shared_count[tid] += 1;
} else if (in[i] < min) {
out[i] = min;
shared_count[tid] += 1;
} else {
out[i] = in[i];
}
}
__syncthreads();
for (int i = blockDim.x / 2; i > 0; i >>= 1) {
if (tid < i) {
shared_count[tid] += shared_count[tid + i];
}
__syncthreads();
}
if (tid == 0) {
num_saturate[blockIdx.x] = shared_count[0];
}
}
template <typename T>
__global__ void ReduceKernel(const int n, const T* in, T* out) {
int tid = threadIdx.x;
extern __shared__ T shared_sum[];
if (tid < n) {
shared_sum[tid] = in[tid];
} else {
shared_sum[tid] = T(0);
}
__syncthreads();
// blockDim.x must >= n
for (int i = (n + 1) / 2; i > 0; i >>= 1) {
if (tid < i) {
shared_sum[tid] += shared_sum[tid + i];
}
__syncthreads();
}
if (tid == 0) {
out[0] = shared_sum[0];
}
}
template <typename T>
int ApplySaturateGpu(const platform::CUDADeviceContext& ctx, const int n,
const T* in, T* out, const T min, const T max) {
int host_num_saturate;
int kNumTheads = 1024;
int gridDimx = (n + kNumTheads - 1) / kNumTheads;
gridDimx = (gridDimx > kNumTheads) ? kNumTheads : gridDimx;
framework::Tensor t;
int* device_num_saturate = t.mutable_data<int>(
framework::make_ddim({gridDimx}), platform::CUDAPlace());
ApplySaturateKernel<
T><<<gridDimx, kNumTheads, kNumTheads * sizeof(T), ctx.stream()>>>(
n, in, out, device_num_saturate, min, max);
ReduceKernel<int><<<1, kNumTheads, kNumTheads * sizeof(T), ctx.stream()>>>(
gridDimx, device_num_saturate, device_num_saturate);
PADDLE_ENFORCE_EQ(cudaSuccess,
cudaMemcpy(&host_num_saturate, device_num_saturate,
sizeof(int), cudaMemcpyDeviceToHost),
"cudaMemcpy failed");
return host_num_saturate;
}
template <typename DeviceContext, typename T>
class FakeQuantizeCUDAKernel : public framework::OpKernel<T> {
public:
T FindRangeAbsMax(const platform::CUDADeviceContext& ctx,
framework::Tensor* scale_list, framework::Tensor* out_scale,
const T& cur_scale, int window_size,
int current_iter) const {
T* sl = scale_list->mutable_data<T>(platform::CPUPlace());
T remove_tmp = sl[current_iter];
sl[current_iter] = cur_scale;
T& max_scale = out_scale->mutable_data<T>(platform::CPUPlace())[0];
if (max_scale < cur_scale) {
max_scale = cur_scale;
} else if (fabs(remove_tmp - max_scale) < 1e-6) {
int size = (current_iter > window_size) ? window_size : current_iter;
max_scale = T(FindAbsMaxGpu(ctx, scale_list->data<float>(), size));
}
return max_scale;
}
T FindMovingAverageAbsMmax(framework::Tensor* in_scale,
framework::Tensor* out_scale,
const T& cur_scale) const {
T* ins = in_scale->mutable_data<T>(platform::CPUPlace());
T* outs = out_scale->mutable_data<T>(platform::CPUPlace());
outs[0] = 0.9 * cur_scale + 0.1 * ins[0];
return T(outs[0]);
}
virtual void Compute(const framework::ExecutionContext& context) const {
PADDLE_ENFORCE(platform::is_gpu_place(context.GetPlace()),
"This kernel only runs on GPU device.");
auto& device_ctx = context.cuda_device_context();
auto* tensor = context.Output<framework::Tensor>("Out");
auto* in = context.Input<framework::Tensor>("X");
const bool is_test = context.Attr<bool>("is_test");
tensor->mutable_data<T>(in->place());
context.Output<framework::Tensor>("OutMovingScale")
->mutable_data<T>(
context.Input<framework::Tensor>("InMovingScale")->place());
auto quantize_type =
static_cast<std::string>(context.Attr<std::string>("quantize_type"));
if (quantize_type == std::string("range_abs_max")) {
context.Output<framework::Tensor>("OutScales")
->mutable_data<T>(
context.Input<framework::Tensor>("InScales")->place());
context.Output<framework::Tensor>("OutCurrentIter")
->mutable_data<T>(
context.Input<framework::Tensor>("InCurrentIter")->place());
}
T scale = T(1);
int window_size = context.Attr<int>("window_size");
T bin_cnt = (T)((1 << (context.Attr<int>("bit_length") - 1)) - 1);
if (quantize_type == std::string("abs_max")) {
auto* saving_scale = context.Output<framework::Tensor>("OutMovingScale");
scale = (T)FindAbsMaxGpu(device_ctx, in->data<float>(), in->numel());
saving_scale->mutable_data<T>(platform::CPUPlace())[0] = scale;
auto& device_ctx = context.template device_context<DeviceContext>();
auto* scale_list = context.Output<framework::Tensor>("OutScales");
math::SetConstant<DeviceContext, T> scalar;
scale_list->mutable_data<T>(context.GetPlace());
scalar(device_ctx, scale_list, static_cast<T>(0));
auto* iter = context.Output<framework::Tensor>("OutCurrentIter");
iter->mutable_data<T>(context.GetPlace());
scalar(device_ctx, iter, static_cast<T>(0));
} else if (quantize_type == std::string("range_abs_max")) {
auto* moving_scale = const_cast<framework::Tensor*>(
context.Input<framework::Tensor>("InMovingScale"));
if (is_test) {
scale = moving_scale->mutable_data<T>(platform::CPUPlace())[0];
} else {
auto* it = const_cast<framework::Tensor*>(
context.Input<framework::Tensor>("InCurrentIter"));
auto* iter = context.Output<framework::Tensor>("OutCurrentIter");
int* last_iter = it->mutable_data<int>(platform::CPUPlace());
int* current_iter = iter->mutable_data<int>(platform::CPUPlace());
auto* scale_list = context.Output<framework::Tensor>("OutScales");
auto* saving_scale =
context.Output<framework::Tensor>("OutMovingScale");
scale = (T)FindAbsMaxGpu(device_ctx, in->data<float>(), in->numel());
scale = FindRangeAbsMax(device_ctx, scale_list, saving_scale, scale,
window_size, current_iter[0]);
(*current_iter) = (*last_iter) + 1;
}
} else if (quantize_type == std::string("moving_average_abs_max")) {
auto* moving_scale = const_cast<framework::Tensor*>(
context.Input<framework::Tensor>("InMovingScale"));
if (is_test) {
scale = moving_scale->mutable_data<T>(platform::CPUPlace())[0];
} else {
scale = (T)FindAbsMaxGpu(device_ctx, in->data<float>(), in->numel());
auto* saving_scale =
context.Output<framework::Tensor>("OutMovingScale");
scale = FindMovingAverageAbsMmax(
const_cast<framework::Tensor*>(moving_scale), saving_scale, scale);
}
}
ApplySaturateGpu<T>(device_ctx, in->numel(), in->data<T>(),
tensor->mutable_data<T>(in->place()), -scale, scale);
scale = bin_cnt / scale;
auto& dev =
*context.template device_context<DeviceContext>().eigen_device();
auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
auto eigen_in = framework::EigenVector<T>::Flatten(*tensor);
eigen_out.device(dev) = (scale * eigen_in).round();
}
};
} // namespace operators
} // namespace paddle
REGISTER_OP_CUDA_KERNEL(fake_quantize,
paddle::operators::FakeQuantizeCUDAKernel<
paddle::platform::CUDADeviceContext, float>,
paddle::operators::FakeQuantizeCUDAKernel<
paddle::platform::CUDADeviceContext, double>);
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/clip_op.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/transform.h"
namespace paddle {
namespace operators {
using platform::Transform;
template <typename DeviceContext, typename T>
class FakeQuantizeKernel : public framework::OpKernel<T> {
public:
T FindAbsMax(framework::Tensor* in, int n) const {
T* p = in->mutable_data<T>(platform::CPUPlace());
T abs_max = (T)0.00000001;
for (int i = 0; i < n; i++) {
T tmp = fabs(p[i]);
if (tmp > abs_max) abs_max = tmp;
}
return T(abs_max);
}
T FindRangeAbsMax(framework::Tensor* scale_list, framework::Tensor* out_scale,
const T& cur_scale, int window_size,
int current_iter) const {
T* sl = scale_list->mutable_data<T>(platform::CPUPlace());
T remove_tmp = sl[current_iter];
sl[current_iter] = cur_scale;
T& max_scale = out_scale->mutable_data<T>(platform::CPUPlace())[0];
if (max_scale < cur_scale) {
max_scale = cur_scale;
} else if (fabs(remove_tmp - max_scale) < 1e-6) {
int size = (current_iter > window_size) ? window_size : current_iter;
max_scale = T(FindAbsMax(scale_list, size));
}
return max_scale;
}
T FindMovingAverageAbsMmax(framework::Tensor* in_scale,
framework::Tensor* out_scale,
const T& cur_scale) const {
T* ins = in_scale->mutable_data<T>(platform::CPUPlace());
T* outs = out_scale->mutable_data<T>(platform::CPUPlace());
outs[0] = 0.9 * cur_scale + 0.1 * ins[0];
return T(outs[0]);
}
virtual void Compute(const framework::ExecutionContext& context) const {
auto* tensor = context.Output<framework::Tensor>("Out");
auto* in = context.Input<framework::Tensor>("X");
const bool is_test = context.Attr<bool>("is_test");
tensor->mutable_data<T>(in->place());
auto* oms_tensor = context.Output<framework::Tensor>("OutMovingScale");
oms_tensor->mutable_data<T>(in->place());
auto quantize_type =
static_cast<std::string>(context.Attr<std::string>("quantize_type"));
if (quantize_type == std::string("range_abs_max")) {
auto* oss_tensor = context.Output<framework::Tensor>("OutScales");
oss_tensor->mutable_data<T>(
context.Input<framework::Tensor>("InScales")->place());
auto* oci_tensor = context.Output<framework::Tensor>("OutCurrentIter");
oci_tensor->mutable_data<T>(
context.Input<framework::Tensor>("InCurrentIter")->place());
}
T scale = static_cast<T>(1);
int window_size = context.Attr<int>("window_size");
int bit_length = context.Attr<int>("bit_length");
int bin_cnt = std::pow(2, bit_length - 1) - 1;
auto& dev =
*context.template device_context<DeviceContext>().eigen_device();
auto raw_in = framework::EigenVector<T>::Flatten(*in);
if (quantize_type == std::string("abs_max")) {
auto* saving_scale = context.Output<framework::Tensor>("OutMovingScale");
auto scale_out = framework::EigenVector<T>::Flatten(*saving_scale);
scale_out.device(dev) = raw_in.abs().maximum();
scale = scale_out(0);
auto& device_ctx = context.template device_context<DeviceContext>();
auto* scale_list = context.Output<framework::Tensor>("OutScales");
math::SetConstant<DeviceContext, T> scalar;
scale_list->mutable_data<T>(context.GetPlace());
scalar(device_ctx, scale_list, static_cast<T>(0));
auto* iter = context.Output<framework::Tensor>("OutCurrentIter");
iter->mutable_data<T>(context.GetPlace());
scalar(device_ctx, iter, static_cast<T>(0));
} else if (quantize_type == std::string("range_abs_max")) {
auto* moving_scale = context.Input<framework::Tensor>("InMovingScale");
if (is_test) {
scale = moving_scale->data<T>()[0];
} else {
auto* it = context.Input<framework::Tensor>("InCurrentIter");
auto* iter = context.Output<framework::Tensor>("OutCurrentIter");
const int* last_iter = it->data<int>();
int* current_iter = iter->mutable_data<int>(platform::CPUPlace());
auto* scale_list = context.Output<framework::Tensor>("OutScales");
auto* saving_scale =
context.Output<framework::Tensor>("OutMovingScale");
auto scale_out = framework::EigenVector<T>::Flatten(*saving_scale);
scale_out.device(dev) = raw_in.abs().maximum();
scale = saving_scale->mutable_data<T>(platform::CPUPlace())[0];
scale = FindRangeAbsMax(scale_list, saving_scale, scale, window_size,
current_iter[0]);
saving_scale->mutable_data<T>(platform::CPUPlace())[0] = scale;
(*current_iter) = (*last_iter) + 1;
}
} else if (quantize_type == std::string("moving_average_abs_max")) {
auto* moving_scale = context.Input<framework::Tensor>("InMovingScale");
if (is_test) {
scale = moving_scale->data<T>()[0];
} else {
auto* saving_scale =
context.Output<framework::Tensor>("OutMovingScale");
auto scale_out = framework::EigenVector<T>::Flatten(*saving_scale);
scale_out.device(dev) = raw_in.abs().maximum();
scale = saving_scale->mutable_data<T>(platform::CPUPlace())[0];
scale = FindMovingAverageAbsMmax(
const_cast<framework::Tensor*>(moving_scale), saving_scale, scale);
saving_scale->mutable_data<T>(platform::CPUPlace())[0] = scale;
}
}
Transform<DeviceContext> trans;
trans(context.template device_context<DeviceContext>(), in->data<T>(),
in->data<T>() + in->numel(), tensor->mutable_data<T>(in->place()),
ClipFunctor<T>(-scale, scale));
auto eigen_out = framework::EigenVector<T>::Flatten(*tensor);
auto eigen_in = framework::EigenVector<T>::Flatten(*tensor);
eigen_out.device(dev) = (bin_cnt / scale * eigen_in).round();
}
};
} // namespace operators
} // namespace paddle
...@@ -91,3 +91,9 @@ endif() ...@@ -91,3 +91,9 @@ endif()
install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR} install(DIRECTORY ${PADDLE_PYTHON_PACKAGE_DIR}
DESTINATION opt/paddle/share/wheels DESTINATION opt/paddle/share/wheels
) )
find_program(PATCHELF_EXECUTABLE patchelf)
if(NOT PATCHELF_EXECUTABLE)
message(FATAL_ERROR "patchelf not found, please install it.\n"
"For Ubuntu, the command is: apt-get install -y patchelf.")
endif()
...@@ -30,6 +30,7 @@ __all__ = [ ...@@ -30,6 +30,7 @@ __all__ = [
'detection_output', 'detection_output',
'ssd_loss', 'ssd_loss',
'detection_map', 'detection_map',
'rpn_target_assign',
'anchor_generator', 'anchor_generator',
] ]
...@@ -44,6 +45,135 @@ for _OP in set(__auto__): ...@@ -44,6 +45,135 @@ for _OP in set(__auto__):
globals()[_OP] = generate_layer_fn(_OP) globals()[_OP] = generate_layer_fn(_OP)
def rpn_target_assign(loc,
scores,
anchor_box,
gt_box,
rpn_batch_size_per_im=256,
fg_fraction=0.25,
rpn_positive_overlap=0.7,
rpn_negative_overlap=0.3):
"""
** Target Assign Layer for region proposal network (RPN) in Faster-RCNN detection. **
This layer can be, for given the Intersection-over-Union (IoU) overlap
between anchors and ground truth boxes, to assign classification and
regression targets to each each anchor, these target labels are used for
train RPN. The classification targets is a binary class label (of being
an object or not). Following the paper of Faster-RCNN, the positive labels
are two kinds of anchors: (i) the anchor/anchors with the highest IoU
overlap with a ground-truth box, or (ii) an anchor that has an IoU overlap
higher than rpn_positive_overlap(0.7) with any ground-truth box. Note
that a single ground-truth box may assign positive labels to multiple
anchors. A non-positive anchor is when its IoU ratio is lower than
rpn_negative_overlap (0.3) for all ground-truth boxes. Anchors that are
neither positive nor negative do not contribute to the training objective.
The regression targets are the encoded ground-truth boxes associated with
the positive anchors.
Args:
loc(Variable): A 3-D Tensor with shape [N, M, 4] represents the
predicted locations of M bounding bboxes. N is the batch size,
and each bounding box has four coordinate values and the layout
is [xmin, ymin, xmax, ymax].
scores(Variable): A 3-D Tensor with shape [N, M, C] represents the
predicted confidence predictions. N is the batch size, C is the
class number, M is number of bounding boxes. For each category
there are total M scores which corresponding M bounding boxes.
anchor_box(Variable): A 2-D Tensor with shape [M, 4] holds M boxes,
each box is represented as [xmin, ymin, xmax, ymax],
[xmin, ymin] is the left top coordinate of the anchor box,
if the input is image feature map, they are close to the origin
of the coordinate system. [xmax, ymax] is the right bottom
coordinate of the anchor box.
gt_box (Variable): The ground-truth boudding boxes (bboxes) are a 2D
LoDTensor with shape [Ng, 4], Ng is the total number of ground-truth
bboxes of mini-batch input.
rpn_batch_size_per_im(int): Total number of RPN examples per image.
fg_fraction(float): Target fraction of RoI minibatch that is labeled
foreground (i.e. class > 0), 0-th class is background.
rpn_positive_overlap(float): Minimum overlap required between an anchor
and ground-truth box for the (anchor, gt box) pair to be a positive
example.
rpn_negative_overlap(float): Maximum overlap allowed between an anchor
and ground-truth box for the (anchor, gt box) pair to be a negative
examples.
Returns:
tuple:
A tuple(predicted_scores, predicted_location, target_label,
target_bbox) is returned. The predicted_scores and
predicted_location is the predicted result of the RPN.
The target_label and target_bbox is the ground truth,
respectively. The predicted_location is a 2D Tensor with shape
[F, 4], and the shape of target_bbox is same as the shape of
the predicted_location, F is the number of the foreground
anchors. The predicted_scores is a 2D Tensor with shape
[F + B, 1], and the shape of target_label is same as the shape
of the predicted_scores, B is the number of the background
anchors, the F and B is depends on the input of this operator.
Examples:
.. code-block:: python
loc = layers.data(name='location', shape=[2, 80],
append_batch_size=False, dtype='float32')
scores = layers.data(name='scores', shape=[2, 40],
append_batch_size=False, dtype='float32')
anchor_box = layers.data(name='anchor_box', shape=[20, 4],
append_batch_size=False, dtype='float32')
gt_box = layers.data(name='gt_box', shape=[10, 4],
append_batch_size=False, dtype='float32')
loc_pred, score_pred, loc_target, score_target =
fluid.layers.detection_output(loc=location,
scores=scores,
anchor_box=anchor_box,
gt_box=gt_box)
"""
helper = LayerHelper('rpn_target_assign', **locals())
# 1. Compute the regression target bboxes
target_bbox = box_coder(
prior_box=anchor_box,
target_box=gt_box,
code_type='encode_center_size',
box_normalized=False)
# 2. Compute overlaps between the prior boxes and the gt boxes overlaps
iou = iou_similarity(x=gt_box, y=anchor_box)
# 3. Assign target label to anchors
loc_index = helper.create_tmp_variable(dtype=anchor_box.dtype)
score_index = helper.create_tmp_variable(dtype=anchor_box.dtype)
target_label = helper.create_tmp_variable(dtype=anchor_box.dtype)
helper.append_op(
type="rpn_target_assign",
inputs={'Overlap': iou, },
outputs={
'LocationIndex': loc_index,
'ScoreIndex': score_index,
'TargetLabel': target_label,
},
attrs={
'rpn_batch_size_per_im': rpn_batch_size_per_im,
'rpn_positive_overlap': rpn_positive_overlap,
'rpn_negative_overlap': rpn_negative_overlap,
'fg_fraction': fg_fraction,
})
# 4. Reshape and gather the target entry
scores = nn.reshape(x=scores, shape=(-1, 1))
loc = nn.reshape(x=loc, shape=(-1, 4))
target_label = nn.reshape(x=target_label, shape=(-1, 1))
target_bbox = nn.reshape(x=target_bbox, shape=(-1, 4))
predicted_scores = nn.gather(scores, score_index)
predicted_location = nn.gather(loc, loc_index)
target_label = nn.gather(target_label, score_index)
target_bbox = nn.gather(target_bbox, loc_index)
return predicted_scores, predicted_loc, target_label, target_bbox
def detection_output(loc, def detection_output(loc,
scores, scores,
prior_box, prior_box,
...@@ -388,7 +518,6 @@ def target_assign(input, ...@@ -388,7 +518,6 @@ def target_assign(input,
Returns: Returns:
tuple: tuple:
A tuple(out, out_weight) is returned. out is a 3D Tensor with A tuple(out, out_weight) is returned. out is a 3D Tensor with
shape [N, P, K], N and P is the same as they are in shape [N, P, K], N and P is the same as they are in
`neg_indices`, K is the same as it in input of X. If `neg_indices`, K is the same as it in input of X. If
......
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
from op_test import OpTest
class TestFakeQuantizeOp(OpTest):
def setUp(self):
self.op_type = "fake_quantize"
self.attrs = {
'bit_length': 8,
'quantize_type': 'abs_max',
'window_size': 10000
}
self.inputs = {
'X': np.random.random((10, 10)).astype("float32"),
'InScales': np.zeros(self.attrs['window_size']).astype("float32"),
'InCurrentIter': np.zeros(1).astype("float32"),
'InMovingScale': np.zeros(1).astype("float32")
}
self.scale = {
'abs_max': np.max(np.abs(self.inputs['X'])).astype("float32")
}
self.outputs = {
'Out': np.round(self.inputs['X'] / self.scale['abs_max'] * (
(1 << (self.attrs['bit_length'] - 1)) - 1)),
'OutScales': np.zeros(self.attrs['window_size']).astype("float32"),
'OutMovingScale':
np.array([self.scale['abs_max']]).astype("float32"),
'OutCurrentIter': np.zeros(1).astype("float32")
}
def test_check_output(self):
self.check_output()
if __name__ == "__main__":
unittest.main()
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import unittest
import numpy as np
import paddle.fluid.core as core
from op_test import OpTest
def rpn_target_assign(iou, rpn_batch_size_per_im, rpn_positive_overlap,
rpn_negative_overlap, fg_fraction):
iou = np.transpose(iou)
anchor_to_gt_max = iou.max(axis=1)
gt_to_anchor_argmax = iou.argmax(axis=0)
gt_to_anchor_max = iou[gt_to_anchor_argmax, np.arange(iou.shape[1])]
anchors_with_max_overlap = np.where(iou == gt_to_anchor_max)[0]
tgt_lbl = np.ones((iou.shape[0], ), dtype=np.int32) * -1
tgt_lbl[anchors_with_max_overlap] = 1
tgt_lbl[anchor_to_gt_max >= rpn_positive_overlap] = 1
num_fg = int(fg_fraction * rpn_batch_size_per_im)
fg_inds = np.where(tgt_lbl == 1)[0]
if len(fg_inds) > num_fg:
disable_inds = np.random.choice(
fg_inds, size=(len(fg_inds) - num_fg), replace=False)
tgt_lbl[disable_inds] = -1
fg_inds = np.where(tgt_lbl == 1)[0]
num_bg = rpn_batch_size_per_im - np.sum(tgt_lbl == 1)
bg_inds = np.where(anchor_to_gt_max < rpn_negative_overlap)[0]
if len(bg_inds) > num_bg:
enable_inds = bg_inds[np.random.randint(len(bg_inds), size=num_bg)]
tgt_lbl[enable_inds] = 0
bg_inds = np.where(tgt_lbl == 0)[0]
loc_index = fg_inds
score_index = np.hstack((fg_inds, bg_inds))
tgt_lbl = np.expand_dims(tgt_lbl, axis=1)
return loc_index, score_index, tgt_lbl
class TestRpnTargetAssignOp(OpTest):
def setUp(self):
iou = np.random.random((10, 8)).astype("float32")
self.op_type = "rpn_target_assign"
self.inputs = {'DistMat': iou}
self.attrs = {
'rpn_batch_size_per_im': 256,
'rpn_positive_overlap': 0.95,
'rpn_negative_overlap': 0.3,
'fg_fraction': 0.25,
'fix_seed': True
}
loc_index, score_index, tgt_lbl = rpn_target_assign(iou, 256, 0.95, 0.3,
0.25)
self.outputs = {
'LocationIndex': loc_index,
'ScoreIndex': score_index,
'TargetLabel': tgt_lbl,
}
def test_check_output(self):
self.check_output()
class TestRpnTargetAssignOp2(OpTest):
def setUp(self):
iou = np.random.random((10, 20)).astype("float32")
self.op_type = "rpn_target_assign"
self.inputs = {'DistMat': iou}
self.attrs = {
'rpn_batch_size_per_im': 128,
'rpn_positive_overlap': 0.5,
'rpn_negative_overlap': 0.5,
'fg_fraction': 0.5,
'fix_seed': True
}
loc_index, score_index, tgt_lbl = rpn_target_assign(iou, 128, 0.5, 0.5,
0.5)
self.outputs = {
'LocationIndex': loc_index,
'ScoreIndex': score_index,
'TargetLabel': tgt_lbl,
}
def test_check_output(self):
self.check_output()
if __name__ == '__main__':
unittest.main()
Markdown is supported
0% .
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