ResNet50 Example Code

The NPU ResNet50 example directory (npu_resnet50_sample) mainly includes example code for calling NPU Runtime interfaces, as well as model directories, binary programs: src and models, bin.

Installing ResNet50 Example

Please follow the ENNP SDK Download to download ai-release.

cd ai-release/EIC7x_Release_20241230/softwares/packages
sudo apt install ./es-sdk-sample-npu-resnet50.deb

After installation, the sample is located in the /opt/eswin/sample-code/npu_sample/npu_resnet50_sample directory.

Directory Structure

The detailed directory structure is as follows:

debian@rockos-eswin:/opt/eswin/sample-code/npu_sample/npu_resnet50_sample$ tree
.
├── bin
│   └── sample_npu
├── models
│   └── resnet50
│       ├── es_resnet50_classes.txt
│       ├── es_resnet50_post_process.json
│       ├── es_resnet50_pre_process.json
│       ├── git_resnet50_mix_4x3x224x224_dyn_latency.model
│       ├── git_resnet50_mix_4x3x224x224_dyn_latency.ofmap_order.txt
│       ├── git_resnet50_mix_4x3x224x224_dyn.model
│       ├── git_resnet50_mix_4x3x224x224_dyn_throughput.model
│       ├── git_resnet50_mix_4x3x224x224_dyn_throughput.ofmap_order.txt
│       ├── input
│       │   └── pictures
│       │       └── input0
│       │           ├── ILSVRC2012_val_00003966.JPEG
│       │           ├── ILSVRC2012_val_00009032.JPEG
│       │           ├── ILSVRC2012_val_00011712.JPEG
│       │           ├── ILSVRC2012_val_00016098.JPEG
│       │           ├── ILSVRC2012_val_00017212.JPEG
│       │           ├── ILSVRC2012_val_00017773.JPEG
│       │           ├── ILSVRC2012_val_00021709.JPEG
│       │           ├── ILSVRC2012_val_00022662.JPEG
│       │           ├── ILSVRC2012_val_00025785.JPEG
│       │           ├── ILSVRC2012_val_00026363.JPEG
│       │           ├── ILSVRC2012_val_00027830.JPEG
│       │           ├── ILSVRC2012_val_00030734.JPEG
│       │           ├── ILSVRC2012_val_00031637.JPEG
│       │           ├── ILSVRC2012_val_00033024.JPEG
│       │           ├── ILSVRC2012_val_00033154.JPEG
│       │           ├── ILSVRC2012_val_00033435.JPEG
│       │           ├── ILSVRC2012_val_00034978.JPEG
│       │           ├── ILSVRC2012_val_00040463.JPEG
│       │           └── ILSVRC2012_val_00049043.JPEG
│       └── model.json
└── src
    ├── build.sh
    ├── CMakeLists.txt
    ├── README.md
    ├── sample_npu_comm.cpp
    ├── sample_npu_comm.h
    ├── sample_npu.cpp
    └── utils
        ├── json
        │   ├── cJSON.cpp
        │   └── cJSON.h
        ├── postprocess
        │   ├── EsPostProcess.cpp
        │   └── EsPostProcess.h
        ├── preprocess
        │   ├── EsHwPreProcess.cpp
        │   ├── EsHwPreProcess.h
        │   └── IPreprocess.h
        ├── sample_npu_utils.cpp
        ├── sample_npu_utils.h
        └── vdec
            ├── common
            │   ├── sample_comm.h
            │   ├── sample_comm_sys.cpp
            │   ├── sample_comm_vdec.cpp
            │   └── sample_comm_vps.cpp
            ├── EsVdec.cpp
            ├── EsVdec.h
            ├── IVdec.h
            └── MemoryPool.h

src Directory

The entry point for the NPU ResNet50 example code is the es_ai_inference.cpp file. The README.md document includes the compilation method for the ResNet50 sample code and the command to run it.

The utils directory mainly contains auxiliary utility libraries. For example, it includes pre-processing for images and post-processing for inference results.

models Directory

• Configuration files:

  • es_resnet50_classes.txt contains class label information.
  • es_resnet50_pre_process.json contains the pre-processing strategy and parameters.
  • es_resnet50_post_process.json contains parameters for the post-processing operator.

• input directory: This directory stores the raw image files for inference.

• Binary directory (bin): This directory contains the binary executable programs.

Usage

For more details, please refer to /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/src/README.md.

debian@rockos-eswin:/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/bin$ ./sample_npu --help
[E][ES_MEM] open_mmz_vb_dev:  348 open /dev/mmz_vb failed!
Usage:
-h, --help                display help info
-s, --sample=type         set sample type (
                          1.SAMPLE_SYNC
                          2.SAMPLE_ASYNC
                          3.SAMPLE_MULTI_CONTEXT
                          4.SAMPLE_MULTI_STREAM
                          5.SAMPLE_MULTI_MODEL
                          6.SAMPLE_COMPOSITE_MODEL
                          7.SAMPLE_D2D
                          8.SAMPLE_PEPELINE)
-m, --model=dir           set model dirs (Multiple model dirs are separated by commas)
-i, --input=dir           set input dirs (Multiple input dirs are separated by commas)
-o, --output=dir          set output dirs (Multiple input dirs are separated by commas)
-p, --pre_process=path    set pre-process config file (Multiple config files are separated by commas)
-q, --post_process=path   set post-process config file (Multiple config files are separated by commas)
-t, --classify=path       set classify file (Multiple classify files are separated by commas)
-n, --numbers=n           set numbers of context/stream.

Example Usage

cd /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/bin

sudo ./sample_npu -s 1 -m /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/ -i /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/ -p /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/es_resnet50_pre_process.json -q /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/es_resnet50_post_process.json -t /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/es_resnet50_classes.txt

debian@rockos-eswin:/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/bin$ sudo ./sample_npu -s 1 -m /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/ -i /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/ -p /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/es_resnet50_pre_process.json -q /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/es_resnet50_post_process.json -t /opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/es_resnet50_classes.txt
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00009032.JPEG: label(742: ('printer'), confidence(0.999691)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00040463.JPEG: label(259: ('Pomeranian'), confidence(0.999973)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00003966.JPEG: label(734: ('police van'), confidence(0.999887)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00021709.JPEG: label(144: ('pelican'), confidence(0.999998)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00031637.JPEG: label(50: ('American alligator'), confidence(0.997608)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00034978.JPEG: label(563: ('fountain pen'), confidence(0.999710)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00016098.JPEG: label(717: ('pickup'), confidence(0.999992)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00049043.JPEG: label(349: ('bighorn'), confidence(0.593028)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00033024.JPEG: label(763: ('power drill'), confidence(0.992695)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00026363.JPEG: label(94: ('cheetah'), confidence(0.999983)
/opt/eswin/sample-code/npu_sample/npu_resnet50_sample/models/resnet50/input/pictures/input0/ILSVRC2012_val_00025785.JPEG: label(831: ('sandbar'), confidence(0.999724)