简体中文 | English

RKNN Model Zoo is developed based on the RKNPU SDK toolchain and provides deployment examples for current mainstream algorithms. Include the process of exporting the RKNN model and using Python API and CAPI to infer the RKNN model.

• Support RK3562, RK3566, RK3568, RK3588 , RK3576 platforms.
• Limited support RV1103, RV1106
• Support RK1808, RV1109, RV1126 platforms.

RKNN Model Zoo relies on RKNN-Toolkit2 for model conversion. The Android compilation tool chain is required when compiling the Android demo, and the Linux compilation tool chain is required when compiling the Linux demo. For the installation of these dependencies, please refer to the Quick Start documentation at https://github.com/airockchip/rknn-toolkit2/tree/master/doc.

• Please note that the Android compilation tool chain recommends using version r18 or r19. Using other versions may encounter the problem of Cdemo compilation failure.

In addition to exporting the model from the corresponding respository, the models file are available on https://console.zbox.filez.com/l/8ufwtG (key: rknn).

• This performance data are collected based on the maximum NPU frequency of each platform.
• This performance data calculate the time-consuming of model inference. Does not include the time-consuming of pre-processing and post-processing.
• RK3576 with sparse_weight referring to the performance when enabling the sparse weight for models
• Note: Models with sparse weight (via Kernel) should have improved performance, but may have accuracy drops depending on models.

For Linux develop board:

./build-linux.sh -t <target> -a <arch> -d <build_demo_name> [-b <build_type>] [-m]
    -t : target (rk356x/rk3588/rk3576/rv1106/rk1808/rv1126)
    -a : arch (aarch64/armhf)
    -d : demo name
    -b : build_type(Debug/Release)
    -m : enable address sanitizer, build_type need set to Debug
Note: 'rk356x' represents rk3562/rk3566/rk3568, 'rv1106' represents rv1103/rv1106, 'rv1126' represents rv1109/rv1126

# Here is an example for compiling yolov5 demo for 64-bit Linux RK3566.
./build-linux.sh -t rk356x -a aarch64 -d yolov5

For Android development board:

# For Android develop boards, it's require to set path for Android NDK compilation tool path according to the user environment
export ANDROID_NDK_PATH=~/opts/ndk/android-ndk-r18b
./build-android.sh -t <target> -a <arch> -d <build_demo_name> [-b <build_type>] [-m]
    -t : target (rk356x/rk3588/rk3576)
    -a : arch (arm64-v8a/armeabi-v7a)
    -d : demo name
    -b : build_type (Debug/Release)
    -m : enable address sanitizer, build_type need set to Debug

# Here is an example for compiling yolov5 demo for 64-bit Android RK3566.
./build-android.sh -t rk356x -a arm64-v8a -d yolov5

All demos in RKNN Model Zoo are verified based on the latest RKNPU SDK. If using a lower version for verification, the inference performance and inference results may be wrong.

• RKNPU2 SDK: https://github.com/airockchip/rknn-toolkit2
• RKNPU1 SDK: https://github.com/airockchip/rknn-toolkit

Apache License 2.0