haotian/duan8
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Arcface trt8 (#1076)

Browse Source 
* gen_weights for latest pretrained_models

* tensorrt-8 migrations

* updated docs
This commit is contained in:
Subtain Malik 2022-08-19 14:55:08 +05:00 committed by GitHub
parent 46dccf11ea
commit 3accb0d8e2
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 89 additions and 62 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

25
arcface/README.md
View File

@ -1,8 +1,15 @@
# arcface
### TensortRT 8
The mxnet implementation is from [deepinsight/insightface.](https://github.com/deepinsight/insightface)
The pretrained models are from [LResNet50E-IR,ArcFace@ms1m-refine-v1](https://github.com/deepinsight/insightface/wiki/Model-Zoo#32-lresnet50e-irarcfacems1m-refine-v1), [LResNet100E-IR,ArcFace@ms1m-refine-v2](https://github.com/deepinsight/insightface/wiki/Model-Zoo#31-lresnet100e-irarcfacems1m-refine-v2) and [MobileFaceNet,ArcFace@ms1m-refine-v1](https://github.com/deepinsight/insightface/wiki/Model-Zoo#34-mobilefacenetarcfacems1m-refine-v1)
**Updated Pretrained Weights:** ArcFace-R100 [Insight Face Google Drive](https://drive.google.com/file/d/1Hc5zUfBATaXUgcU2haUNa7dcaZSw95h2/view)
---
**Previous Pre-trained models:** The pretrained models are from [LResNet50E-IR,ArcFace@ms1m-refine-v1](https://github.com/deepinsight/insightface/wiki/Model-Zoo#32-lresnet50e-irarcfacems1m-refine-v1), [LResNet100E-IR,ArcFace@ms1m-refine-v2](https://github.com/deepinsight/insightface/wiki/Model-Zoo#31-lresnet100e-irarcfacems1m-refine-v2) and [MobileFaceNet,ArcFace@ms1m-refine-v1](https://github.com/deepinsight/insightface/wiki/Model-Zoo#34-mobilefacenetarcfacems1m-refine-v1)
---
The two input images used in this project are joey0.ppm and joey1.ppm, download them from [Google Drive.](https://drive.google.com/drive/folders/1ctqpkRCRKyBZRCNwo9Uq4eUoMRLtFq1e). The input image is 112x112, and generated from `get_input()` in `insightface/deploy/face_model.py`, which is cropped and aligned face image.
@ -17,18 +24,18 @@ The two input images used in this project are joey0.ppm and joey1.ppm, download
## Run
1.Generate .wts file from mxnet implementation of pretrained model. The following example described how to generate arcface-r50.wts from mxnet implementation of LResNet50E-IR,ArcFace@ms1m-refine-v1.
1.Generate .wts file from mxnet implementation of pretrained model. The following example described how to generate arcface-r100.wts from mxnet implementation of LResNet100E-IR,ArcFace@ms1m-refine-v1.
```
git clone https://github.com/deepinsight/insightface
cd insightface
git checkout 3866cd77a6896c934b51ed39e9651b791d78bb57
cd deploy
// copy tensorrtx/arcface/gen_wts.py to here(insightface/deploy)
// download model-r50-am-lfw.zip and unzip here(insightface/deploy)
// download model-r100-ii.zip and unzip here(insightface/deploy)
python gen_wts.py
// a file 'arcface-r50.wts' will be generated.
// a file 'arcface-r100.wts' will be generated.
// the master branch of insightface should work, if not, you can checkout 94ad870abb3203d6f31b049b70dd080dc8f33fca
// arcface-r100.wts/arcface-mobilefacenet.wts can be generated in similar way from mxnet implementation of LResNet100E-IR,ArcFace@ms1m-refine-v1/MobileFaceNet,ArcFace@ms1m-refine-v1 pretrained model.
// arcface-r50.wts/arcface-mobilefacenet.wts can be generated in similar way from mxnet implementation of LResNet50E-IR,ArcFace@ms1m-refine-v1/MobileFaceNet,ArcFace@ms1m-refine-v1 pretrained model.
```
2.Put .wts file into tensorrtx/arcface, build and run
@ -40,13 +47,13 @@ mkdir build
cd build
cmake ..
make
sudo ./arcface-r50 -s // serialize model to plan file i.e. 'arcface-r50.engine'
sudo ./arcface-r50 -d // deserialize plan file and run inference
sudo ./arcface-r100 -s // serialize model to plan file i.e. 'arcface-r100.engine'
sudo ./arcface-r100 -d // deserialize plan file and run inference
or
sudo ./arcface-r100 -s // serialize model to plan file i.e. 'arcface-r100.engine'
sudo ./arcface-r100 -d // deserialize plan file and run inference
sudo ./arcface-r50 -s // serialize model to plan file i.e. 'arcface-r50.engine'
sudo ./arcface-r50 -d // deserialize plan file and run inference
or

4
arcface/gen_wts.py
View File

@ -8,7 +8,7 @@ import numpy as np
parser = argparse.ArgumentParser(description='face model test')
# general
parser.add_argument('--image-size', default='112,112', help='')
parser.add_argument('--model', default='model-r50-am-lfw/model,1', help='path to load model.')
parser.add_argument('--model', default='model-r100-ii/model,0', help='path to load model.')
parser.add_argument('--ga-model', default='', help='path to load model.')
parser.add_argument('--gpu', default=0, type=int, help='gpu id')
parser.add_argument('--det', default=0, type=int, help='mtcnn option, 1 means using R+O, 0 means detect from begining')
@ -18,7 +18,7 @@ args = parser.parse_args()
model = face_model.FaceModel(args)
f = open('arcface-r50.wts', 'w')
f = open('arcface-r100.wts', 'w')
f.write('{}\n'.format(len(model.model.get_params()[0].keys()) + len(model.model.get_params()[1].keys())))
for k, v in model.model.get_params()[0].items():
vr = v.reshape(-1).asnumpy()

9
arcface/logging.h
View File

@ -25,6 +25,13 @@
#include <ostream>
#include <sstream>
#include <string>
#include "macros.h"
#if NV_TENSORRT_MAJOR >= 8
#define TRT_NOEXCEPT noexcept
#else
#define TRT_NOEXCEPT
#endif
using Severity = nvinfer1::ILogger::Severity;
@ -236,7 +243,7 @@ public:
//! Note samples should not be calling this function directly; it will eventually go away once we eliminate the
//! inheritance from nvinfer1::ILogger
//!
void log(Severity severity, const char* msg) override
void log(Severity severity, const char* msg) TRT_NOEXCEPT override
{
LogStreamConsumer(mReportableSeverity, severity) << "[TRT] " << std::string(msg) << std::endl;
}

12
arcface/macros.h Normal file
View File

@ -0,0 +1,12 @@
#ifndef __MACROS_H
#define __MACROS_H
#if NV_TENSORRT_MAJOR >= 8
#define TRT_NOEXCEPT noexcept
#define TRT_CONST_ENQUEUE const
#else
#define TRT_NOEXCEPT
#define TRT_CONST_ENQUEUE
#endif
#endif // __MACROS_H

44
arcface/prelu.cu
View File

@ -23,7 +23,7 @@ namespace nvinfer1
gamma_.assign((float*)p, (float*)p + (length - sizeof(int)) / sizeof(float));
}
void PReluPlugin::serialize(void* buffer) const
void PReluPlugin::serialize(void* buffer) const TRT_NOEXCEPT
{
*reinterpret_cast<int*>(buffer) = input_size_;
char *p = reinterpret_cast<char*>(buffer);
@ -31,17 +31,17 @@ namespace nvinfer1
memcpy(p, gamma_.data(), gamma_.size() * sizeof(float));
}
size_t PReluPlugin::getSerializationSize() const
size_t PReluPlugin::getSerializationSize() const TRT_NOEXCEPT
{
return sizeof(input_size_) + gamma_.size() * sizeof(float);
}
int PReluPlugin::initialize()
int PReluPlugin::initialize() TRT_NOEXCEPT
{
return 0;
}
Dims PReluPlugin::getOutputDimensions(int index, const Dims* inputs, int nbInputDims)
Dims PReluPlugin::getOutputDimensions(int index, const Dims* inputs, int nbInputDims) TRT_NOEXCEPT
{
assert(nbInputDims == 1);
assert(index == 0);
@ -51,63 +51,63 @@ namespace nvinfer1
}
// Set plugin namespace
void PReluPlugin::setPluginNamespace(const char* pluginNamespace)
void PReluPlugin::setPluginNamespace(const char* pluginNamespace) TRT_NOEXCEPT
{
mPluginNamespace = pluginNamespace;
}
const char* PReluPlugin::getPluginNamespace() const
const char* PReluPlugin::getPluginNamespace() const TRT_NOEXCEPT
{
return mPluginNamespace;
}
// Return the DataType of the plugin output at the requested index
DataType PReluPlugin::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const
DataType PReluPlugin::getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const TRT_NOEXCEPT
{
return DataType::kFLOAT;
}
// Return true if output tensor is broadcast across a batch.
bool PReluPlugin::isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const
bool PReluPlugin::isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const TRT_NOEXCEPT
{
return false;
}
// Return true if plugin can use input that is broadcast across batch without replication.
bool PReluPlugin::canBroadcastInputAcrossBatch(int inputIndex) const
bool PReluPlugin::canBroadcastInputAcrossBatch(int inputIndex) const TRT_NOEXCEPT
{
return false;
}
void PReluPlugin::configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput)
void PReluPlugin::configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) TRT_NOEXCEPT
{
}
// Attach the plugin object to an execution context and grant the plugin the access to some context resource.
void PReluPlugin::attachToContext(cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator)
void PReluPlugin::attachToContext(cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) TRT_NOEXCEPT
{
}
// Detach the plugin object from its execution context.
void PReluPlugin::detachFromContext() {}
void PReluPlugin::detachFromContext() TRT_NOEXCEPT {}
const char* PReluPlugin::getPluginType() const
const char* PReluPlugin::getPluginType() const TRT_NOEXCEPT
{
return "PRelu_TRT";
}
const char* PReluPlugin::getPluginVersion() const
const char* PReluPlugin::getPluginVersion() const TRT_NOEXCEPT
{
return "1";
}
void PReluPlugin::destroy()
void PReluPlugin::destroy() TRT_NOEXCEPT
{
delete this;
}
// Clone the plugin
IPluginV2IOExt* PReluPlugin::clone() const
IPluginV2IOExt* PReluPlugin::clone() const TRT_NOEXCEPT
{
PReluPlugin *p = new PReluPlugin(gamma_);
p->input_size_ = input_size_;
@ -138,7 +138,7 @@ namespace nvinfer1
assert(cudaFree(dev_gamma) == cudaSuccess);
}
int PReluPlugin::enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream)
int PReluPlugin::enqueue(int batchSize, const void*const * inputs, void* TRT_CONST_ENQUEUE* outputs, void* workspace, cudaStream_t stream) TRT_NOEXCEPT
{
//assert(batchSize == 1);
//GPU
@ -158,22 +158,22 @@ namespace nvinfer1
mFC.fields = mPluginAttributes.data();
}
const char* PReluPluginCreator::getPluginName() const
const char* PReluPluginCreator::getPluginName() const TRT_NOEXCEPT
{
return "PRelu_TRT";
}
const char* PReluPluginCreator::getPluginVersion() const
const char* PReluPluginCreator::getPluginVersion() const TRT_NOEXCEPT
{
return "1";
}
const PluginFieldCollection* PReluPluginCreator::getFieldNames()
const PluginFieldCollection* PReluPluginCreator::getFieldNames() TRT_NOEXCEPT
{
return &mFC;
}
IPluginV2IOExt* PReluPluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc)
IPluginV2IOExt* PReluPluginCreator::createPlugin(const char* name, const PluginFieldCollection* fc) TRT_NOEXCEPT
{
std::vector<float> gamma;
const PluginField* fields = fc->fields;
@ -197,7 +197,7 @@ namespace nvinfer1
return obj;
}
IPluginV2IOExt* PReluPluginCreator::deserializePlugin(const char* name, const void* serialData, size_t serialLength)
IPluginV2IOExt* PReluPluginCreator::deserializePlugin(const char* name, const void* serialData, size_t serialLength) TRT_NOEXCEPT
{
// This object will be deleted when the network is destroyed, which will
// call PReluPlugin::destroy()

57
arcface/prelu.h
View File

@ -4,6 +4,7 @@
#include <string>
#include <vector>
#include "NvInfer.h"
#include "macros.h"
namespace nvinfer1
{
@ -15,53 +16,53 @@ namespace nvinfer1
~PReluPlugin();
int getNbOutputs() const override
int getNbOutputs() const TRT_NOEXCEPT override
{
return 1;
}
Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) override;
Dims getOutputDimensions(int index, const Dims* inputs, int nbInputDims) TRT_NOEXCEPT override;
int initialize() override;
int initialize() TRT_NOEXCEPT override;
virtual void terminate() override {};
virtual void terminate() TRT_NOEXCEPT override {};
virtual size_t getWorkspaceSize(int maxBatchSize) const override { return 0;}
virtual size_t getWorkspaceSize(int maxBatchSize) const TRT_NOEXCEPT override { return 0;}
virtual int enqueue(int batchSize, const void*const * inputs, void** outputs, void* workspace, cudaStream_t stream) override;
virtual int enqueue(int batchSize, const void*const * inputs, void*TRT_CONST_ENQUEUE* outputs, void* workspace, cudaStream_t stream) TRT_NOEXCEPT override;
virtual size_t getSerializationSize() const override;
virtual size_t getSerializationSize() const TRT_NOEXCEPT override;
virtual void serialize(void* buffer) const override;
virtual void serialize(void* buffer) const TRT_NOEXCEPT override;
bool supportsFormatCombination(int pos, const PluginTensorDesc* inOut, int nbInputs, int nbOutputs) const override {
bool supportsFormatCombination(int pos, const PluginTensorDesc* inOut, int nbInputs, int nbOutputs) const TRT_NOEXCEPT override {
return inOut[pos].format == TensorFormat::kLINEAR && inOut[pos].type == DataType::kFLOAT;
}
const char* getPluginType() const override;
const char* getPluginType() const TRT_NOEXCEPT override;
const char* getPluginVersion() const override;
const char* getPluginVersion() const TRT_NOEXCEPT override;
void destroy() override;
void destroy() TRT_NOEXCEPT override;
IPluginV2IOExt* clone() const override;
IPluginV2IOExt* clone() const TRT_NOEXCEPT override;
void setPluginNamespace(const char* pluginNamespace) override;
void setPluginNamespace(const char* pluginNamespace) TRT_NOEXCEPT override;
const char* getPluginNamespace() const override;
const char* getPluginNamespace() const TRT_NOEXCEPT override;
DataType getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const override;
DataType getOutputDataType(int index, const nvinfer1::DataType* inputTypes, int nbInputs) const TRT_NOEXCEPT override;
bool isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const override;
bool isOutputBroadcastAcrossBatch(int outputIndex, const bool* inputIsBroadcasted, int nbInputs) const TRT_NOEXCEPT override;
bool canBroadcastInputAcrossBatch(int inputIndex) const override;
bool canBroadcastInputAcrossBatch(int inputIndex) const TRT_NOEXCEPT override;
void attachToContext(
cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) override;
cudnnContext* cudnnContext, cublasContext* cublasContext, IGpuAllocator* gpuAllocator) TRT_NOEXCEPT override;
void configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) override;
void configurePlugin(const PluginTensorDesc* in, int nbInput, const PluginTensorDesc* out, int nbOutput) TRT_NOEXCEPT override;
void detachFromContext() override;
void detachFromContext() TRT_NOEXCEPT override;
int input_size_;
private:
@ -78,22 +79,22 @@ namespace nvinfer1
~PReluPluginCreator() override = default;
const char* getPluginName() const override;
const char* getPluginName() const TRT_NOEXCEPT override;
const char* getPluginVersion() const override;
const char* getPluginVersion() const TRT_NOEXCEPT override;
const PluginFieldCollection* getFieldNames() override;
const PluginFieldCollection* getFieldNames() TRT_NOEXCEPT override;
IPluginV2IOExt* createPlugin(const char* name, const PluginFieldCollection* fc) override;
IPluginV2IOExt* createPlugin(const char* name, const PluginFieldCollection* fc) TRT_NOEXCEPT override;
IPluginV2IOExt* deserializePlugin(const char* name, const void* serialData, size_t serialLength) override;
IPluginV2IOExt* deserializePlugin(const char* name, const void* serialData, size_t serialLength) TRT_NOEXCEPT override;
void setPluginNamespace(const char* libNamespace) override
void setPluginNamespace(const char* libNamespace) TRT_NOEXCEPT override
{
mNamespace = libNamespace;
}
const char* getPluginNamespace() const override
const char* getPluginNamespace() const TRT_NOEXCEPT override
{
return mNamespace.c_str();
}