Image Classification - ResNet18

This example demonstrates how to deploy the ResNet18 model for image classification inference using the Python interface of hbm_runtime. It is intended for RDK S100 devices equipped with a BPU chip. The sample code is located in /app/pydev_demo/01_classification_sample/01_resnet18/.

This example demonstrates how to deploy the ResNet18 model for image classification inference using the Python interface of hbm_runtime. It is intended for RDK S600 devices equipped with a BPU chip. The sample code is located in /app/pydev_demo/classification_sample/resnet18/.

Model Description

• Overview:

  ResNet (Residual Network) is a deep convolutional neural network architecture proposed by Microsoft Research. Its core idea is to introduce "residual connections," which use skip connections across layers to mitigate the vanishing gradient problem in deep networks, enabling effective training of networks with dozens or even hundreds of layers. The ResNet18 variant used in this example is a lightweight version with 18 layers, widely used for image classification, feature extraction, and similar tasks.

• HBM model name: resnet18_224x224_nv12.hbm

• Input format: NV12, size 224x224

• Output: Softmax probability distribution over 1000 classes

• Model download URL (automatically downloaded by the program):

  https://archive.d-robotics.cc/downloads/rdk_model_zoo/rdk_s100/ResNet/resnet18_224x224_nv12.hbm

  https://archive.d-robotics.cc/downloads/rdk_model_zoo/rdk_s600/ResNet/resnet18_224x224_nv12.hbm

Features

• Model loading

  Loads the specified model using hbm_runtime, parses input and output names and shapes, and prepares them for subsequent inference.

• Input preprocessing

  Resizes the BGR image to 224x224 and converts it to NV12 format (Y and UV planes separated).

• Inference execution

  Performs model forward inference through the .run() method, with optional scheduling parameters (priority and core binding).

• Post-processing

  Reads the output tensor, parses top-K classification results (Top-5), and displays class labels and probability values.

Environment Dependencies

This sample has no special environment requirements. Just ensure that the dependencies in pydev are installed.

pip install -r ../../requirements.txt

pip install -r ../../requirements.txt --break-system-packages

Directory Structure

.
├── resnet18.py                 # Main program; uses `hbm_runtime` to run ResNet18 classification
└── README.md                   # Usage instructions

Parameters

Parameter	Description	Default Value
--model-path	Model file path (.hbm format)	/opt/hobot/model/s100/basic/resnet18_224x224_nv12.hbm
--test-img	Test image path	/app/res/assets/zebra_cls.jpg
--label-file	Class label mapping file path	/app/res/labels/imagenet1000_clsidx_to_labels.txt
--priority	Model priority (0~255; higher values mean higher priority)	0
--bpu-cores	List of BPU core IDs used for inference (e.g. --bpu-cores 0 1)	[0]

Parameter	Description	Default Value
--model-path	Model file path (.hbm format)	/opt/hobot/model/s600/basic/resnet18_224x224_nv12.hbm
--test-img	Test image path	/app/res/assets/zebra_cls.jpg
--label-file	Class label mapping file path	/app/res/labels/imagenet1000_clsidx_to_labels.txt
--priority	Model priority (0~255; higher values mean higher priority)	0
--bpu-cores	List of BPU core IDs used for inference (e.g. --bpu-cores 0 1)	[0]

Quick Start

• Run the model

  • With default parameters

    python resnet18.py

  • Run with specified parameters

    python resnet18.py \
    --model-path /opt/hobot/model/s100/basic/resnet18_224x224_nv12.hbm \
    --test-img /app/res/assets/zebra_cls.jpg \
    --label-file /app/res/labels/imagenet1000_clsidx_to_labels.txt

    python resnet18.py \
    --model-path /opt/hobot/model/s600/basic/resnet18_224x224_nv12.hbm \
    --test-img /app/res/assets/zebra_cls.jpg \
    --label-file /app/res/labels/imagenet1000_clsidx_to_labels.txt

• View results

  Top-5 Predictions:
  zebra: 0.9979
  impala, Aepyceros melampus: 0.0005
  cheetah, chetah, Acinonyx jubatus: 0.0005
  gazelle: 0.0004
  prairie chicken, prairie grouse, prairie fowl: 0.0002

Notes

• If the specified model path does not exist, the program will attempt to download the model automatically.