Skip to content

P
Paddle-Lite

PaddlePaddle / Paddle-Lite

通知 331
Star 4
Fork 1
P
Paddle-Lite
未验证 提交 44b045ae 编写于 作者: R Ruilong Liu 提交者： GitHub
浏览文件
操作

Merge branch 'develop' into develop

上级 434a0179 04ea1b77
隐藏空白更改
内联 并排
Showing with 181 addition and 106 deletion
README.md
浏览文件 @ 44b045ae
......@@ -27,7 +27,7 @@ Paddle-Moible是PaddlePaddle组织下的项目，是一个致力于嵌入式平
- **ARM CPU**
![](http://mms-graph.bj.bcebos.com/paddle-mobile%2F2018_07_18.png)
![](http://mms-graph.bj.bcebos.com/paddle-mobile%2F2018_07_29.png)
arm cpu是paddle-mobile的主要支持方向，cpu的通用性一直是其优势。嵌入式深度学习，需要大量的cpu汇编实现。我们正在紧锣密鼓的编码，为的是能充分硬件的每一点加速能力。
arm cpu的优化工作还在进行中，现在使用了常规的cpu优化。在arm a73上paddle-mobile arm-v7现在单核运行一次mobilenet1.0是110+ms，显然这不是我们的最终目标，我们正在用大量的汇编改写，后续性能仍会有巨大提升空间, 目前只支持armv7, 未来我们也会支持armv8。
......
src/common/types.cpp
浏览文件 @ 44b045ae
......@@ -17,39 +17,39 @@ limitations under the License. */
namespace paddle_mobile {
const std::string G_OP_TYPE_CONV = "conv2d";
const std::string G_OP_TYPE_BATCHNORM = "batch_norm";
const std::string G_OP_TYPE_BOX_CODER = "box_coder";
const std::string G_OP_TYPE_CONCAT = "concat";
const std::string G_OP_TYPE_ELEMENTWISE_ADD = "elementwise_add";
const std::string G_OP_TYPE_FUSION_CONV_ADD_RELU = "fusion_conv_add_relu";
const std::string G_OP_TYPE_FUSION_CONV_ADD_BN_RELU = "fusion_conv_add_bn_relu";
const std::string G_OP_TYPE_FUSION_DWCONV_BN_RELU = "fusion_dwconv_bn_relu";
const std::string G_OP_TYPE_FUSION_CONV_BN_RELU = "fusion_conv_bn_relu";
const std::string G_OP_TYPE_FC = "fusion_fc";
const std::string G_OP_TYPE_FUSION_CONV_ADD = "fusion_conv_add";
const std::string G_OP_TYPE_LRN = "lrn";
const std::string G_OP_TYPE_MUL = "mul";
const std::string G_OP_TYPE_MULTICLASS_NMS = "multiclass_nms";
const std::string G_OP_TYPE_POOL2D = "pool2d";
const std::string G_OP_TYPE_PRIOR_BOX = "prior_box";
const std::string G_OP_TYPE_RELU = "relu";
const std::string G_OP_TYPE_RESHAPE = "reshape";
const std::string G_OP_TYPE_SIGMOID = "sigmoid";
const std::string G_OP_TYPE_SOFTMAX = "softmax";
const std::string G_OP_TYPE_TRANSPOSE = "transpose";
const std::string G_OP_TYPE_SPLIT = "split";
const std::string G_OP_TYPE_FEED = "feed";
const std::string G_OP_TYPE_FETCH = "fetch";
const std::string G_OP_TYPE_DEPTHWISE_CONV = "depthwise_conv2d";
const std::string G_OP_TYPE_IM2SEQUENCE = "im2sequence";
const std::string G_OP_TYPE_DROPOUT = "dropout";
const std::string G_OP_TYPE_FUSION_CONV_ADD_BN = "fusion_conv_add_bn";
const std::string G_OP_TYPE_FUSION_POOL_BN = "fusion_pool_bn";
const std::string G_OP_TYPE_FUSION_ELEMENTWISE_ADD_RELU =
const char *G_OP_TYPE_CONV = "conv2d";
const char *G_OP_TYPE_BATCHNORM = "batch_norm";
const char *G_OP_TYPE_BOX_CODER = "box_coder";
const char *G_OP_TYPE_CONCAT = "concat";
const char *G_OP_TYPE_ELEMENTWISE_ADD = "elementwise_add";
const char *G_OP_TYPE_FUSION_CONV_ADD_RELU = "fusion_conv_add_relu";
const char *G_OP_TYPE_FUSION_CONV_ADD_BN_RELU = "fusion_conv_add_bn_relu";
const char *G_OP_TYPE_FUSION_DWCONV_BN_RELU = "fusion_dwconv_bn_relu";
const char *G_OP_TYPE_FUSION_CONV_BN_RELU = "fusion_conv_bn_relu";
const char *G_OP_TYPE_FC = "fusion_fc";
const char *G_OP_TYPE_FUSION_CONV_ADD = "fusion_conv_add";
const char *G_OP_TYPE_LRN = "lrn";
const char *G_OP_TYPE_MUL = "mul";
const char *G_OP_TYPE_MULTICLASS_NMS = "multiclass_nms";
const char *G_OP_TYPE_POOL2D = "pool2d";
const char *G_OP_TYPE_PRIOR_BOX = "prior_box";
const char *G_OP_TYPE_RELU = "relu";
const char *G_OP_TYPE_RESHAPE = "reshape";
const char *G_OP_TYPE_SIGMOID = "sigmoid";
const char *G_OP_TYPE_SOFTMAX = "softmax";
const char *G_OP_TYPE_TRANSPOSE = "transpose";
const char *G_OP_TYPE_SPLIT = "split";
const char *G_OP_TYPE_FEED = "feed";
const char *G_OP_TYPE_FETCH = "fetch";
const char *G_OP_TYPE_DEPTHWISE_CONV = "depthwise_conv2d";
const char *G_OP_TYPE_IM2SEQUENCE = "im2sequence";
const char *G_OP_TYPE_DROPOUT = "dropout";
const char *G_OP_TYPE_FUSION_CONV_ADD_BN = "fusion_conv_add_bn";
const char *G_OP_TYPE_FUSION_POOL_BN = "fusion_pool_bn";
const char *G_OP_TYPE_FUSION_ELEMENTWISE_ADD_RELU =
"fusion_elementwise_add_relu";
const std::string G_OP_TYPE_FUSION_FC_RELU = "fusion_fc_relu";
const std::string G_OP_TYPE_REGION = "region";
const char *G_OP_TYPE_FUSION_FC_RELU = "fusion_fc_relu";
const char *G_OP_TYPE_REGION = "region";
std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
......
src/common/types.h
浏览文件 @ 44b045ae
......@@ -79,40 +79,40 @@ enum PMStatus {
PMWrongDevice = 0x08 /*!< un-correct device. */
};
extern const std::string G_OP_TYPE_CONV;
extern const std::string G_OP_TYPE_BATCHNORM;
extern const std::string G_OP_TYPE_BOX_CODER;
extern const std::string G_OP_TYPE_CONCAT;
extern const std::string G_OP_TYPE_ELEMENTWISE_ADD;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD_RELU;
extern const std::string G_OP_TYPE_FC;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD_BN_RELU;
extern const std::string G_OP_TYPE_FUSION_DWCONV_BN_RELU;
extern const std::string G_OP_TYPE_FUSION_CONV_BN_RELU;
extern const std::string G_OP_TYPE_LRN;
extern const std::string G_OP_TYPE_MUL;
extern const std::string G_OP_TYPE_MULTICLASS_NMS;
extern const std::string G_OP_TYPE_POOL2D;
extern const std::string G_OP_TYPE_PRIOR_BOX;
extern const std::string G_OP_TYPE_RELU;
extern const std::string G_OP_TYPE_RESHAPE;
extern const std::string G_OP_TYPE_SIGMOID;
extern const std::string G_OP_TYPE_SOFTMAX;
extern const std::string G_OP_TYPE_TRANSPOSE;
extern const std::string G_OP_TYPE_SPLIT;
extern const std::string G_OP_TYPE_FEED;
extern const std::string G_OP_TYPE_FETCH;
extern const std::string G_OP_TYPE_DEPTHWISE_CONV;
extern const std::string G_OP_TYPE_IM2SEQUENCE;
extern const std::string G_OP_TYPE_DROPOUT;
extern const std::string G_OP_TYPE_FUSION_CONV_ADD_BN;
extern const std::string G_OP_TYPE_FUSION_POOL_BN;
extern const std::string G_OP_TYPE_FUSION_ELEMENTWISE_ADD_RELU;
extern const std::string G_OP_TYPE_FUSION_FC_RELU;
extern const std::string G_OP_TYPE_REGION;
extern const char *G_OP_TYPE_CONV;
extern const char *G_OP_TYPE_BATCHNORM;
extern const char *G_OP_TYPE_BOX_CODER;
extern const char *G_OP_TYPE_CONCAT;
extern const char *G_OP_TYPE_ELEMENTWISE_ADD;
extern const char *G_OP_TYPE_FUSION_CONV_ADD_RELU;
extern const char *G_OP_TYPE_FC;
extern const char *G_OP_TYPE_FUSION_CONV_ADD;
extern const char *G_OP_TYPE_FUSION_CONV_ADD_BN_RELU;
extern const char *G_OP_TYPE_FUSION_DWCONV_BN_RELU;
extern const char *G_OP_TYPE_FUSION_CONV_BN_RELU;
extern const char *G_OP_TYPE_LRN;
extern const char *G_OP_TYPE_MUL;
extern const char *G_OP_TYPE_MULTICLASS_NMS;
extern const char *G_OP_TYPE_POOL2D;
extern const char *G_OP_TYPE_PRIOR_BOX;
extern const char *G_OP_TYPE_RELU;
extern const char *G_OP_TYPE_RESHAPE;
extern const char *G_OP_TYPE_SIGMOID;
extern const char *G_OP_TYPE_SOFTMAX;
extern const char *G_OP_TYPE_TRANSPOSE;
extern const char *G_OP_TYPE_SPLIT;
extern const char *G_OP_TYPE_FEED;
extern const char *G_OP_TYPE_FETCH;
extern const char *G_OP_TYPE_DEPTHWISE_CONV;
extern const char *G_OP_TYPE_IM2SEQUENCE;
extern const char *G_OP_TYPE_DROPOUT;
extern const char *G_OP_TYPE_FUSION_CONV_ADD_BN;
extern const char *G_OP_TYPE_FUSION_POOL_BN;
extern const char *G_OP_TYPE_FUSION_ELEMENTWISE_ADD_RELU;
extern const char *G_OP_TYPE_FUSION_FC_RELU;
extern const char *G_OP_TYPE_REGION;
extern std::unordered_map<
std::string, std::pair<std::vector<std::string>, std::vector<std::string>>>
......
test/CMakeLists.txt
浏览文件 @ 44b045ae
......@@ -145,6 +145,10 @@ else ()
ADD_EXECUTABLE(test-conv-add-relu-op operators/test_conv_add_relu_op.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-conv-add-relu-op paddle-mobile)
# gen test
ADD_EXECUTABLE(test-conv-add-bn-relu-op operators/test_fusion_conv_add_bn_relu_op.cpp test_helper.h test_include.h executor_for_test.h)
target_link_libraries(test-conv-add-bn-relu-op paddle-mobile)
#add_library(test-lib-size SHARED common/test_lib_size.h common/test_lib_size.cpp)
endif()
test/executor_for_test.h
浏览文件 @ 44b045ae
......@@ -43,7 +43,7 @@ template <typename DeviceType, typename OpType>
class Executor4Test : public Executor<DeviceType> {
public:
Executor4Test(Program<DeviceType> p, string op_type,
bool use_optimize = false)
bool use_optimize = false, int predict_op_count = 1)
: Executor<DeviceType>() {
this->use_optimize_ = use_optimize;
this->program_ = p;
......@@ -57,12 +57,14 @@ class Executor4Test : public Executor<DeviceType> {
LOG(paddle_mobile::LogLevel::kLOG_ERROR)
<< "to_predict_program_ == nullptr";
}
const std::vector<std::shared_ptr<BlockDesc>> blocks =
this->to_predict_program_->Blocks();
for (std::shared_ptr<BlockDesc> block_desc : blocks) {
std::vector<std::shared_ptr<OpDesc>> ops = block_desc->Ops();
for (std::shared_ptr<OpDesc> op : ops) {
if (op->Type() == op_type) {
for (int i = 0; i < ops.size(); ++i) {
auto op = ops[i];
if (op->Type() == op_type && i < predict_op_count) {
DLOG << "匹配到: " << op->Type();
/// test first meeting op in program
......@@ -72,11 +74,17 @@ class Executor4Test : public Executor<DeviceType> {
op->Type(), op->GetInputs(), op->GetOutputs(),
op->GetAttrMap(), this->program_.scope);
this->ops_of_block_[*block_desc.get()].push_back(op_ptr);
break;
}
}
}
this->InitMemory();
std::shared_ptr<paddle_mobile::framework::BlockDesc> to_predict_block =
this->to_predict_program_->Block(0);
auto &ops = this->ops_of_block_[*to_predict_block.get()];
for (const auto &op : ops) {
op->Init();
}
}
template <typename T = LoDTensor>
......@@ -130,9 +138,6 @@ class Executor4Test : public Executor<DeviceType> {
auto *output_tensor = con_output->GetMutable<LoDTensor>();
output_tensor->mutable_data<float>(dDim);
std::shared_ptr<Tensor> out_tensor = std::make_shared<LoDTensor>();
out_tensor.reset(output_tensor);
std::shared_ptr<paddle_mobile::framework::BlockDesc> to_predict_block =
this->to_predict_program_->Block(0);
for (int j = 0; j < this->ops_of_block_[*to_predict_block.get()].size();
......@@ -141,6 +146,7 @@ class Executor4Test : public Executor<DeviceType> {
op->Run();
}
return out_tensor;
return std::make_shared<paddle_mobile::framework::Tensor>(
paddle_mobile::framework::Tensor(*output_tensor));
}
};
test/net/test_mobilenet+ssd.cpp
浏览文件 @ 44b045ae
......@@ -20,22 +20,20 @@ int main() {
paddle_mobile::PaddleMobile<paddle_mobile::CPU> paddle_mobile;
paddle_mobile.SetThreadNum(4);
auto time1 = time();
auto isok = paddle_mobile.Load(g_mobilenet_ssd_gesture + "/model",
g_mobilenet_ssd_gesture + "/params", true);
auto isok = paddle_mobile.Load(
std::string(g_mobilenet_ssd_gesture) + "/model",
std::string(g_mobilenet_ssd_gesture) + "/params", true);
// auto isok = paddle_mobile.Load(g_mobilenet_ssd, false);
if (isok) {
auto time2 = time();
std::cout << "load cost :" << time_diff(time1, time2) << "ms" << std::endl;
std::vector<float> input;
std::vector<int64_t> dims{1, 3, 300, 300};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 300, 300}, static_cast<float>(0),
static_cast<float>(1));
GetInput<float>(g_hand, &input, dims);
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
paddle_mobile.Predict(input, dims);
auto output = paddle_mobile.Predict(input, dims);
auto time4 = time();
std::cout << "predict cost :" << time_diff(time3, time4) << "ms"
<< std::endl;
......
test/net/test_mobilenet.cpp
浏览文件 @ 44b045ae
......@@ -24,19 +24,21 @@ int main() {
auto time2 = time();
std::cout << "load cost :" << time_diff(time1, time1) << "ms" << std::endl;
std::vector<float> input;
std::vector<int64_t> dims{1, 3, 224, 224};
Tensor input_tensor;
SetupTensor<float>(&input_tensor, {1, 3, 224, 224}, static_cast<float>(0),
static_cast<float>(1));
std::vector<float> input(input_tensor.data<float>(),
input_tensor.data<float>() + input_tensor.numel());
auto time3 = time();
auto vec_result = paddle_mobile.Predict(input, dims);
auto time4 = time();
std::cout << "predict cost :" << time_diff(time3, time4) << "ms"
<< std::endl;
GetInput<float>(g_test_image_1x3x224x224, &input, dims);
for (int i = 0; i < 10; ++i) {
auto time3 = time();
auto vec_result = paddle_mobile.Predict(input, dims);
auto time4 = time();
std::vector<float>::iterator biggest =
std::max_element(std::begin(vec_result), std::end(vec_result));
std::cout << " Max element is " << *biggest << " at position "
<< std::distance(std::begin(vec_result), biggest) << std::endl;
std::cout << "predict cost :" << time_diff(time3, time4) << "ms"
<< std::endl;
}
}
return 0;
......
test/operators/test_fusion_conv_add_bn_relu_op.cpp 0 → 100644
浏览文件 @ 44b045ae
/* 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 "../test_include.h"
#include "operators/fusion_conv_add_bn_relu_op.h"
int main() {
paddle_mobile::Loader<paddle_mobile::CPU> loader;
// ../models/image_classification_resnet.inference.model
auto program = loader.Load(g_mobilenet, true);
PADDLE_MOBILE_ENFORCE(program.originProgram != nullptr,
"program file read fail");
Executor4Test<paddle_mobile::CPU,
paddle_mobile::operators::FusionConvAddBNReluOp<
paddle_mobile::CPU, float>>
executor(program, "fusion_conv_add_bn_relu", true);
std::cout << "executor 4 test: " << std::endl;
paddle_mobile::framework::Tensor input;
GetInput<float>(g_test_image_1x3x224x224_banana, &input, {1, 3, 224, 224});
// // use SetupTensor if not has local input image .
// SetupTensor<float>(&input, {1, 3, 224, 224}, static_cast<float>(0),
// static_cast<float>(1));
DLOG << " fuck: " << input;
auto out_ddim = paddle_mobile::framework::make_ddim({1, 32, 112, 112});
std::cout << "before predict: " << std::endl;
auto output =
executor.Predict(input, "data", "conv2_1_dw_bn.tmp_2", out_ddim);
std::cout << "after predict " << std::endl;
auto output_ptr = output->data<float>();
int stride = output->numel() / 100;
for (int i = 0; i < 100; i++) {
DLOG << " index:" << i * stride << " value: " << output_ptr[i * stride];
}
// for (int i = 0; i < 100; i++) {
// DLOG << " index:" << i << " value: "<< output_ptr[i];
// }
// for (int j = 0; j < output->numel(); ++j) {
// std::cout << " (index: " << j << " value: " << output_ptr[j] << ") ";
// }
std::cout << std::endl;
return 0;
}
test/test_helper.h
浏览文件 @ 44b045ae
......@@ -24,18 +24,21 @@ limitations under the License. */
#include "framework/ddim.h"
#include "framework/tensor.h"
static const std::string g_mobilenet_ssd = "../models/mobilenet+ssd";
static const std::string g_mobilenet_ssd_gesture =
"../models/mobilenet+ssd_gesture";
static const std::string g_squeezenet = "../models/squeezenet";
static const std::string g_googlenet = "../models/googlenet";
static const std::string g_mobilenet = "../models/mobilenet";
static const std::string g_resnet_50 = "../models/resnet_50";
static const std::string g_resnet = "../models/resnet";
static const std::string g_googlenet_combine = "../models/googlenet_combine";
static const std::string g_yolo = "../models/yolo";
static const std::string g_test_image_1x3x224x224 =
static const char *g_mobilenet_ssd = "../models/mobilenet+ssd";
static const char *g_mobilenet_ssd_gesture = "../models/mobilenet+ssd_gesture";
static const char *g_squeezenet = "../models/squeezenet";
static const char *g_googlenet = "../models/googlenet";
static const char *g_mobilenet = "../models/mobilenet";
static const char *g_resnet_50 = "../models/resnet_50";
static const char *g_resnet = "../models/resnet";
static const char *g_googlenet_combine = "../models/googlenet_combine";
static const char *g_yolo = "../models/yolo";
static const char *g_test_image_1x3x224x224 =
"../images/test_image_1x3x224x224_float";
static const char *g_test_image_1x3x224x224_banana =
"../images/input_3x224x224_banana";
static const char *g_hand = "../images/hand_image";
using paddle_mobile::framework::DDim;
using paddle_mobile::framework::Tensor;
......
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