deepstream框架下多路源模型部署参考步骤

目录

1.dynamic onnx的生成

2.tensorRT engine模型的生成

3.配置文件的参数指定

---

在deepstream框架下，想测试SCRFD模型的多路分布式运行，当修改配置文件中相应的batch参数后，运行报错如下

0:00:08.390136614  8575 0xaaab0daf9c70 WARN                 nvinfer gstnvinfer.cpp:643:gst_nvinfer_logger: NvDsInferContext[UID 1]: Warning from NvDsInferContextImpl::checkBackendParams()  [UID = 1]: Backend has maxBatchSize 1 whereas 2 has been requested
0:00:08.390549772  8575 0xaaab0daf9c70 WARN                 nvinfer gstnvinfer.cpp:643:gst_nvinfer_logger: NvDsInferContext[UID 1]: Warning from NvDsInferContextImpl::generateBackendContext()  [UID = 1]: deserialized backend context :/home/leedarson/workspace/models/scrfd_500m_bnkps_shape640x640.engine failed to match config params, trying rebuild
0:00:08.396210854  8575 0xaaab0daf9c70 INFO                 nvinfer gstnvinfer.cpp:646:gst_nvinfer_logger: NvDsInferContext[UID 1]: Info from NvDsInferContextImpl::buildModel()  [UID = 1]: Trying to create engine from model files
ERROR: failed to build network since there is no model file matched.
ERROR: failed to build network.
0:00:10.735614046  8575 0xaaab0daf9c70 ERROR                nvinfer gstnvinfer.cpp:640:gst_nvinfer_logger: NvDsInferContext[UID 1]: Error in NvDsInferContextImpl::buildModel()  [UID = 1]: build engine file failed

 查看报错信息，大概是SCRFD模型engine的maxBatchSize=1，而使用时却给的参数值为2，所以engine的创建失败。查看deepstream中给出的多路视频运行的配置文件示例，尽管模型实现的功能相同，但是不同的batch使用的模型却不同，这才想起onnx可以分为模型有静态和动态两种，所谓静态onnx模型的输入和输出都是指定好的，转模型的时候不能随意更改，所以如果batchsize一旦给定就不能更改，我一开始生成的SCRFD模型就是静态的，给定batchsize=1，尝试batchsize>1时运行就会报错；而动态onnx模型，可以支持动态给定输入和输出，在转其他模型或者执行的时候指定相应的参数即可。因此如果需要在deepstream框架上进行多路分布式部署运行模型，大概需要三步：1.生成dynamic onnx；2.将dynamic onnx在模型的运行平台上转tensorRT engine；3.在配置文件中给定相应的参数。

1.dynamic onnx的生成

insightface中提供的有SCRFD模型转onnx的Python文件，刚开始按照提示，指定的命令行参数如下运行生成onnx

python3 tools/scrfd2onnx.py configs/scrfd/scrfd_500m_bnkps.py model/scrfd_0.5gkps.pth  --input-img data/test.jpg --shape 640 640

可以在Netron看到生成的onnx模型输入输入结构如下

有太多的参数未指定，在转相应的tensorRT engine时就遇到了如下的报错

[02/20/2023-11:07:38] [E] Error[4]: [graphShapeAnalyzer.cpp::processCheck::587] Error Code 4: Internal Error (IAssertionLayer (Unnamed Layer* 4) [Assertion]: condition[0] is false: 0. For input: 'input.1' all named dimensions that share the same name must be equal. Note: Named dimensions were present on the following axes: 0 (name: '?'), 2 (name: '?')) [02/20/2023-11:07:38] [E] Error[2]: [builder.cpp::buildSerializedNetwork::636] Error Code 2: Internal Error (Assertion engine != nullptr failed. )

根据trtexec的报错信息，可以理解大概是onnx的输入参数问题。回到insightface的scrfd2onnx.py文件中，查看其对于动态维度的指定代码如下

# Define input and outputs names, which are required to properly define # dynamic axes input_names = ['input.1'] output_names = ['score_8', 'score_16', 'score_32',                 'bbox_8', 'bbox_16', 'bbox_32',                 ] # If model graph contains keypoints strides add keypoints to outputs if 'stride_kps' in str(model):     output_names += ['kps_8', 'kps_16', 'kps_32'] # Define dynamic axes for export dynamic_axes = None if dynamic:     dynamic_axes = {out: {0: '?', 1: '?'} for out in output_names}     dynamic_axes[input_names[0]] = {         0: '?',         2: '?',         3: '?'     }

将其更改为如下后

# Define input and outputs names, which are required to properly define # dynamic axes input_names = ['input.1'] output_names = ['score_8', 'score_16', 'score_32',                 'bbox_8', 'bbox_16', 'bbox_32',                 ] # If model graph contains keypoints strides add keypoints to outputs if 'stride_kps' in str(model):     output_names += ['kps_8', 'kps_16', 'kps_32'] # Define dynamic axes for export dynamic_axes = None if dynamic:     dynamic_axes = {out: {0: 'batch_size'} for out in output_names}     dynamic_axes[input_names[0]] = {         0: 'batch_size'     }

执行如下命令，生成相应的dynamic onnx模型。

python3 tools/scrfd2onnx.py configs/scrfd/scrfd_500m_bnkps.py model/scrfd_0.5gkps.pth  --input-img data/test.jpg --shape 640 640

生成的dynamic onnx模型输入输出参数如下图。

2.tensorRT engine模型的生成

deepstream框架下模型的推理运行使用的是tensorRT组件，所以如果要在deepstream框架上运行模型，需要将待运行的模型在将要运行的硬件平台上转为tensorRT能进行推理的engine文件。对于已经生成的dynamic onnx模型文件，执行如下命令进行engine模型文件的生成。

/usr/src/tensorrt/bin/trtexec --onnx=/home/user/workspace/models/scrfd_500m_bnkps_dynamic.onnx --minShapes=input.1:1x3x640x640 --optShapes=input.1:8x3x640x640 --maxShapes=input.1:8x3x640x640 --saveEngine=/home/leedarson/workspace/models/scrfd_500m_bnkps_dynamic.engine --fp16

参考相应的static onnx转engine命令如下。

/usr/src/tensorrt/bin/trtexec --onnx=./scrfd_500m_bnkps.onnx --fp16 --saveEngine=./scrfd_500m_bnkps.engine --device=0

可以看到对于dynamic onnx需要指定minShapes、optShapes和maxShapes，通常情况下optShapes和maxShapes可以给成相同的值。

注：生成tensorRT engine文件的硬件平台必须和模型推理部署的硬件平台保持一致，即转模型和推理部署都必须要在同一个硬件设备上执行，否则会导致推理部署失败。

3.配置文件的参数指定

在生成tensorRT engine文件时，已指定maxShapes=input.1:8x3x640x640，因此该engine最多可以指定8路同时运行，配置文件中batchsize<=8都可以，注意保持[source0]、[streammux]、[primary-gie]、[tiled-display]等组件的batchsize相同即可。模型SCRFD的参考配置参数如下。

config_face.txt

[application] enable-perf-measurement=1 perf-measurement-interval-sec=5 [tiled-display] enable=1 rows=2 columns=4 width=1080 height=720 gpu-id=0 #(0): nvbuf-mem-default - Default memory allocated, specific to particular platform #(1): nvbuf-mem-cuda-pinned - Allocate Pinned/Host cuda memory applicable for Tesla #(2): nvbuf-mem-cuda-device - Allocate Device cuda memory applicable for Tesla #(3): nvbuf-mem-cuda-unified - Allocate Unified cuda memory applicable for Tesla #(4): nvbuf-mem-surface-array - Allocate Surface Array memory, applicable for Jetson nvbuf-memory-type=0 [source0] enable=1 #Type - 1=CameraV4L2 2=URI 3=MultiURI type=3 uri=file:/home/leedarson/workspace/testData/face5.mp4 num-sources=8 gpu-id=0 # (0): memtype_device   - Memory type Device # (1): memtype_pinned   - Memory type Host Pinned # (2): memtype_unified  - Memory type Unified cudadec-memtype=0 [sink0] enable=1 #Type - 1=FakeSink 2=EglSink 3=File type=2 sync=1 source-id=0 gpu-id=0 nvbuf-memory-type=0 [sink1] enable=1 type=3 #1=mp4 2=mkv container=1 #1=h264 2=h265 codec=1 #encoder type 0=Hardware 1=Software enc-type=0 sync=0 #iframeinterval=10 bitrate=2000000 #H264 Profile - 0=Baseline 2=Main 4=High #H265 Profile - 0=Main 1=Main10 profile=0 output-file=out.mp4 source-id=0 [sink2] enable=0 #Type - 1=FakeSink 2=EglSink 3=File 4=RTSPStreaming type=4 #1=h264 2=h265 codec=1 #encoder type 0=Hardware 1=Software enc-type=0 sync=0 bitrate=4000000 #H264 Profile - 0=Baseline 2=Main 4=High #H265 Profile - 0=Main 1=Main10 profile=0 # set below properties in case of RTSPStreaming rtsp-port=8554 udp-port=5400 [osd] enable=1 gpu-id=0 border-width=1 text-size=15 text-color=1;1;1;1; text-bg-color=0.3;0.3;0.3;1 font=Serif show-clock=0 clock-x-offset=800 clock-y-offset=820 clock-text-size=12 clock-color=1;0;0;0 nvbuf-memory-type=0 [streammux] gpu-id=0 ##Boolean property to inform muxer that sources are live live-source=0 buffer-pool-size=1 batch-size=8 ##time out in usec, to wait after the first buffer is available ##to push the batch even if the complete batch is not formed batched-push-timeout=40000 ## Set muxer output width and height width=1920 height=1080 ##Enable to maintain aspect ratio wrt source, and allow black borders, works ##along with width, height properties enable-padding=0 nvbuf-memory-type=0 [primary-gie] enable=1 gpu-id=0 model-engine-file=/home/user/workspace/models/scrfd_500m_bnkps_dynamic.engine batch-size=8 #Required by the app for OSD, not a plugin property bbox-border-color0=1;0;0;1 bbox-border-color1=0;1;1;1 bbox-border-color2=0;0;1;1 bbox-border-color3=0;1;0;1 interval=0 gie-unique-id=1 nvbuf-memory-type=0 config-file=config_detection.txt [tracker] enable=1 # For NvDCF and DeepSORT tracker, tracker-width and tracker-height must be a multiple of 32, respectively tracker-width=640 tracker-height=384 ll-lib-file=/opt/nvidia/deepstream/deepstream/lib/libnvds_nvmultiobjecttracker.so # ll-config-file required to set different tracker types # ll-config-file=config_tracker_IOU.yml ll-config-file=/opt/nvidia/deepstream/deepstream-6.1/samples/configs/deepstream-app/config_tracker_NvDCF_perf.yml # ll-config-file=config_tracker_NvDCF_accuracy.yml # ll-config-file=config_tracker_DeepSORT.yml gpu-id=0 enable-batch-process=1 enable-past-frame=1 display-tracking-id=1 [tests] file-loop=0

config_detection.txt内容如下。

[property] gpu-id=0 model-engine-file=/home/user/workspace/models/scrfd_500m_bnkps_dynamic.engine batch-size=4 net-scale-factor=0.0078125 offsets=127.5;127.5;127.5 force-implicit-batch-dim=1 model-color-format=1 # 0=FP32, 1=INT8, 2=FP16 mode network-mode=2 process-mode=1 num-detected-classes=6 labelfile-path=labels.txt # number of consecutive batches to skip for inference interval=0 # custom detection parser parse-bbox-func-name=NvDsInferParseCustomSCRFD custom-lib-path=nvdsinfer_custom_impl_face/libnvdsinfer_custom_impl_face.so output-blob-names=score_8;bbox_8;kps_8;score_16;bbox_16;kps_16;score_32;bbox_32;kps_32 force-implicit-batch-dim=1 [class-attrs-all] # bbox threshold pre-cluster-threshold=0.5 # nms threshold post-cluster-threshold=0.45

执行如下命令

deepstream-app -c config_face.txt

对比deepstream推理SCRFD时解析的网络参数如下。

dynamic engine推理时解析的网络层参数

static engine推理时解析的网络层参数

运行结果如下

8路640x480大小的视频同时在NVIDIA NX平台上运行，帧率大概16fps；

4路640x480大小的视频同时在NVIDIA NX平台上运行，帧率大概30fps。