Skip to main content

InceptionV3

This tutorial provides guidance on running the InceptionV3 model on RevyOS using either the CPU or NPU. InceptionV3 is a multi-branch convolutional neural network architecture, capable of efficiently extracting multi-scale features for image classification tasks.

Initial Environment Setup

Before proceeding, please ensure you have completed the environment setup section.

Obtaining Example Code​

The example code for this tutorial is available on Github. Clone it locally using the following command:

$ git clone https://github.com/zhangwm-pt/lpi4a-example.git

The relevant code for this tutorial is located in the classification/inceptionv3 directory.

Obtaining the Model​

The model used in this tutorial is from the TensorFlow model repository. You can download the InceptionV3 model using the following commands:

Environment Preparation​

Before downloading the model, please ensure that the TensorFlow environment is properly set up. It is recommended to use a virtual environment.

$ python3 -m venv tf
$ source tf/bin/activate
$ pip install tensorflow==1.15 tf_slim pytest

Downloading Model Weights​

Follow these steps to obtain the TensorFlow model weights:

$ git clone https://github.com/tensorflow/models.git
$ cd models/research/slim
$ wget http://download.tensorflow.org/models/inception_v3_2016_08_28.tar.gz
$ tar xzvf inception_v3_2016_08_28.tar.gz

Export the TensorFlow model graph using:

$ python export_inference_graph.py \
    --alsologtostderr \
    --model_name=inception_v3 \
    --output_file=./inception_v3_inf_graph.pb

Freeze the model graph and weights into a single model file:

$ python -m tensorflow.python.tools.freeze_graph \
    --input_graph=./inception_v3_inf_graph.pb \
    --input_checkpoint=./inception_v3.ckpt \
    --input_binary=true \
    --output_graph=./inception_v3.pb \
    --output_node_names=InceptionV3/Predictions/Reshape_1
About File Relationships and Conversion
  • inception_v3.ckpt: TensorFlow 1.x checkpoint file containing training weights; must be used with the model structure and cannot be used for inference alone.
  • inception_v3_info_graph.pt: Exported model computation graph structure, does not contain weights, cannot be run directly, only used as input for freezing the model.
  • inception_v3.pb: The frozen inference model, containing both structure and parameters, suitable for deployment and HHB model compilation.
About Github Network Proxy

If you encounter network issues accessing GitHub from mainland China, consider using a network proxy tool to accelerate access.

Model Information​

GFLOPsparamsaccuracyinput nameoutput nameshapelayoutchannel orderscale valuemean values
11.524Mtop1 77.9%, top5 93.8%inputInceptionV3/Logits/Predictions/Reshape_11, 299, 299, 3NHWCRGB0.0039127.5 127.5 127.5

Model Conversion and Compilation​

On an x86 machine, use the HHB tool to convert the .pb model into a computation graph and glue code suitable for RevyOS. Before proceeding, ensure you have started the HHB container and cloned the example repository as described in the environment setup section.

Model Conversion with HHB​

In this step, the .pb model is converted into a format compatible with the HHB platform.

Navigate to the classification/inceptionv3 directory and execute the following commands:

$ hhb -D --model-file ./inception_v3.pb --data-scale 0.0039 --data-mean "127.5 127.5 127.5"  \
    --board th1520  --postprocess save_and_top5 --input-name "input" --output-name "InceptionV3/Predictions/Reshape_1" \
    --input-shape "1 299 299 3" --calibrate-dataset persian_cat.jpg  --quantization-scheme "int8_asym"
About Parameters
  • -D: Specifies the HHB process to stop at the executable generation stage
  • --model-file: Specifies the input model file
  • --data-mean: Specifies the mean values
  • --data-scale: Specifies the scale value
  • --board: Target platform, C920 (CPU) or TH1520 (NPU)
  • --input-name: Model input tensor name
  • --output-name: Model output tensor name
  • --input-shape: Model input tensor shape
  • --postprocess: Specifies the post-processing behavior for the generated glue code. save_and_top5 saves the output and prints the top 5 results
  • --quantization-scheme: Specifies the quantization type

You can run hhb --help to view all available parameters and options.

About HHB Generated Files

After execution, an hhb_out subdirectory will be generated in the current directory, containing files such as hhb_runtime, model.c, and others:

  • hhb.bm: HHB model file, including quantized weights and related data
  • hhb_runtime: Executable for the development board, compiled from the C files in the directory
  • main.c: Reference entry for the generated example program
  • model.c: Model structure representation file
  • model.params: Model weights file
  • io.c: Example program with file I/O helper functions
  • io.h: Declarations for I/O helper functions
  • process.c: Example program with image preprocessing functions
  • process.h: Declarations for preprocessing functions

Compiling the Application​

The glue code generated by HHB only tests the model's functionality. For complete image preprocessing and postprocessing, an application using OpenCV is provided to load the model and perform inference.

In the classification/inceptionv3 directory, compile the application with:

$ export OPENCV_DIR=../../modules/opencv/ # Set the path to OpenCV
$ riscv64-unknown-linux-gnu-g++ main.cpp -I${OPENCV_DIR}/include/opencv4 -L${OPENCV_DIR}/lib  \
    -lopencv_imgproc   -lopencv_imgcodecs -L${OPENCV_DIR}/lib/opencv4/3rdparty/ -llibjpeg-turbo 
    -llibwebp -llibpng -llibtiff -llibopenjp2    -lopencv_core -ldl  -lpthread -lrt -lzlib \
    -lcsi_cv -latomic -static -o inceptionv3_example
About OpenCV

The example code uses OpenCV for model input preprocessing. Please ensure OpenCV is installed as described in the environment setup section.

Compilation Parameter Description
  • -I../prebuilt_opencv/include/opencv4: Header file search path, pointing to the OpenCV headers
  • -L../prebuilt_opencv/lib: Library search path, pointing to the precompiled OpenCV binaries
  • -lopencv_imgproc -lopencv_imgcodecs -lopencv_core: OpenCV libraries
  • -llibjpeg-turbo -llibwebp -llibpng -llibtiff -llibopenjp2 -lcsi_cv: OpenCV dependencies
  • -static: Static linking
  • -o inceptionv3_example: Output executable name

After successful compilation, the inceptionv3_example file will be generated in the example directory.

Uploading and Running the Application​

Upload to the Development Board​

Package all files in this directory and upload them to the development board. For example, use the scp command to upload to /home/debian/inceptionv3:

$ scp -r ../inceptionv3/ debian@<board_ip>:/home/debian/inceptionv3/

Alternatively, you may use other methods such as USB storage devices or network sharing.

Running the Program​

On the development board, navigate to /home/debian/inceptionv3. Ensure the SHL library is installed and LD_LIBRARY_PATH is configured. Then run:

$ ./inceptionv3_example
hhb_runtime Error

If you encounter the following error:

hhb_out/hhb_runtime: error while loading shared libraries: libshl_th1520.so.2: cannot open shared object file: No such file or directory

Ensure LD_LIBRARY_PATH is correctly set. If the issue persists, run pip show shl-python to check the version.

If the version is 3.x.x, it is too high. The program requires shl-python version 2.x. Downgrade with:

$ pip install shl-python==2.6.17
About NPU Device Permissions

If you encounter the following error:

FATAL: could not open driver '/dev/vha0': Permission denied

Check if the current user has read/write permissions for /dev/vha0. Set permissions with:

$ sudo chmod 0666 /dev/vha0

It is recommended to configure udev rules for automatic permission setting. Consult AI or documentation for udev configuration.

About Long NPU Inference Time

In theory, the program should run quickly. However, the first run may take over 5 minutes due to JIT compilation when loading the model on the NPU. Due to HHB runtime design, JIT compilation occurs on every run, resulting in long execution times.

For more details, refer to Common Issues and Solutions.

Sample output:​

In this tutorial, the input is a picture of a Persian cat. The expected result for ResNet50 is that the largest value is at index 283, corresponding to Persian cat. Persian cat

$ ./inceptionv3_example
 ********** preprocess image **********
 ********** run model **********
INFO: NNA clock:363733 [kHz]
INFO: Heap :anonymous (0x2)
INFO: Heap :dmabuf (0x2)
INFO: Heap :unified (0x5)
WARNING: Mapping to the on chip ram failed (128 > 0), continuing...
FATAL: Importing 268203 bytes of CPU memory has failed (wrong memory alignment)
Run graph execution time: 19.44006ms, FPS=51.44

=== tensor info ===
shape: 1 3 299 299
data pointer: 0x1ab83af0

=== tensor info ===
shape: 1 1001
data pointer: 0x3f7f78f000
The max_value of output: 0.899233
The min_value of output: 0.000000
The mean_value of output: 0.000906
The std_value of output: 0.000807
 ============ top5: ===========
284: 0.899233
282: 0.003670
288: 0.003670
  0: 0.000000
  1: 0.000000
 ********** postprocess result **********
 ********** probability top5: **********
Persian cat
tabby, tabby cat
lynx, catamount
background
tench, Tinca tinca