Docker for Raspberry Pi + Python Server, not a bug, a suggestion, need Docker improvement.

[BaptisteOlympio]

I am working on Raspberry Pi 5 and OpenFace. I have succed to modify my docker images so that it is working on my Raspberry Pi.
My spec on my Raspberry Pi 5 :

    OS: Debian GNU/Linux 12 (bookworm)
    Kernel: 6.6.74+rpt-rpi-2712
    CPU: Cortex-A76 (4) @ 71% 2400.0000 1500.0000MHz
    Memory: 1474MiB / 8048MiB 

Here is my Dockerfile for my Raspberry Pi. I changed :

• the image from Ubuntu from Ubuntu:18.04 to arm64v8/ubuntu:18.04
• In the section == Building OpenFace == I modify the cmake with adding some flags :

    CXX=g++-8 CC=gcc-8 cmake -D CMAKE_BUILD_TYPE=RELEASE \
    -D OpenBLAS_INCLUDE_DIR=/usr/include \
    -D OpenBLAS_LIB=/usr/lib/aarch64-linux-gnu/libopenblas.so \
    -G Ninja .. \
    -DCMAKE_CXX_FLAGS="-march=armv8-a" \
    -DCMAKE_C_FLAGS="-march=armv8-a" && \

I mostly did this with ChatGPT : here is the prompt for explaination : https://chatgpt.com/share/6787e7eb-3a74-8010-865a-a659b565ea1d

# ==================== Building Model Layer ===========================
# This is a little trick to improve caching and minimize rebuild time
# and bandwidth. Note that RUN commands only cache-miss if the prior layers
# miss, or the dockerfile changes prior to this step.
# To update these patch files, be sure to run build with --no-cache
FROM alpine as model_data
RUN apk --no-cache --update-cache add wget
WORKDIR /data/patch_experts

RUN wget -q https://www.dropbox.com/s/7na5qsjzz8yfoer/cen_patches_0.25_of.dat &&\
    wget -q https://www.dropbox.com/s/k7bj804cyiu474t/cen_patches_0.35_of.dat &&\
    wget -q https://www.dropbox.com/s/ixt4vkbmxgab1iu/cen_patches_0.50_of.dat &&\
    wget -q https://www.dropbox.com/s/2t5t1sdpshzfhpj/cen_patches_1.00_of.dat

## ==================== Install Ubuntu Base libs ===========================
## This will be our base image for OpenFace, and also the base for the compiler
## image. We only need packages which are linked

FROM arm64v8/ubuntu:18.04 as ubuntu_base

LABEL maintainer="Michael McDermott <mikemcdermott23@gmail.com>"

ARG DEBIAN_FRONTEND=noninteractive

# todo: minimize this even more
RUN apt-get update -qq  &&\
    apt-get install -qq curl  &&\
    apt-get install -qq --no-install-recommends \
    	libopenblas-dev liblapack-dev \
    	libavcodec-dev libavformat-dev libswscale-dev \
    	libtbb2 libtbb-dev libjpeg-dev \
	libpng-dev libtiff-dev  &&\
	rm -rfv /var/lib/apt/lists/*

RUN apt-get update && apt-get install -y libopenblas-dev

RUN echo OUI
RUN ldconfig -p | grep openblas

## ==================== Build-time dependency libs ======================
## This will build and install opencv and dlib into an additional dummy
## directory, /root/diff, so we can later copy in these artifacts,
## minimizing docker layer size
## Protip: ninja is faster than `make -j` and less likely to lock up system
FROM ubuntu_base as cv_deps

WORKDIR /root/build-dep
ARG DEBIAN_FRONTEND=noninteractive

RUN apt-get update -qq && apt-get install -qq -y \
        cmake ninja-build pkg-config build-essential checkinstall\
        g++-8 &&\
    rm -rf /var/lib/apt/lists/* &&\
    update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-8 800 --slave /usr/bin/g++ g++ /usr/bin/g++-8

##        llvm clang-3.7 libc++-dev libc++abi-dev  \
## ==================== Building dlib ===========================

RUN curl http://dlib.net/files/dlib-19.13.tar.bz2 -LO &&\
    tar xf dlib-19.13.tar.bz2 && \
    rm dlib-19.13.tar.bz2 &&\
    mv dlib-19.13 dlib &&\
    mkdir -p dlib/build &&\
    cd dlib/build &&\
    cmake -DCMAKE_BUILD_TYPE=Release -G Ninja .. &&\
    ninja && \
    ninja install && \
    DESTDIR=/root/diff ninja install &&\
    ldconfig

## ==================== Building OpenCV ======================
ENV OPENCV_VERSION=4.1.0

RUN curl https://github.com/opencv/opencv/archive/${OPENCV_VERSION}.tar.gz -LO &&\
    tar xf ${OPENCV_VERSION}.tar.gz && \
    rm ${OPENCV_VERSION}.tar.gz &&\
    mv opencv-${OPENCV_VERSION} opencv && \
    mkdir -p opencv/build && \
    cd opencv/build && \
    cmake -D CMAKE_BUILD_TYPE=RELEASE \
        -D CMAKE_INSTALL_PREFIX=/usr/local \
        -D WITH_TBB=ON -D WITH_CUDA=OFF \
        -DWITH_QT=OFF -DWITH_GTK=OFF\
        -G Ninja .. && \
    ninja && \
    ninja install &&\
    DESTDIR=/root/diff ninja install

## ==================== Building OpenFace ===========================
FROM cv_deps as openface
WORKDIR /root/openface

COPY ./ ./

COPY --from=model_data /data/patch_experts/* \
    /root/openface/lib/local/LandmarkDetector/model/patch_experts/


RUN mkdir -p build && cd build && \
    CXX=g++-8 CC=gcc-8 cmake -D CMAKE_BUILD_TYPE=RELEASE \
    -D OpenBLAS_INCLUDE_DIR=/usr/include \
    -D OpenBLAS_LIB=/usr/lib/aarch64-linux-gnu/libopenblas.so \
    -G Ninja .. \
    -DCMAKE_CXX_FLAGS="-march=armv8-a" \
    -DCMAKE_C_FLAGS="-march=armv8-a" && \
    ninja && \
    DESTDIR=/root/diff ninja install


## ==================== Streamline container ===========================
## Clean up - start fresh and only copy in necessary stuff
## This shrinks the image from ~8 GB to ~1.6 GB
FROM ubuntu_base as final

WORKDIR /root

# Copy in only necessary libraries
COPY --from=openface /root/diff /

# Since we "imported" the build artifacts, we need to reconfigure ld
RUN ldconfig


## ==================== Python ===========================

WORKDIR /root/openface/opengaze

COPY ./opengaze .

# Install Python and FastAPI dependencies
RUN apt-get update && apt-get install -y python3 python3-pip && \
    pip3 install -r requirements.txt
    # pip3 install --no-cache-dir fastapi uvicorn==0.16.0

# # Copy FastAPI application
# COPY main.py /root/main.py

# # Expose FastAPI port
EXPOSE 8000

I also had to modify the CMakeLists.txt :
from

if (${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
    execute_process(COMMAND ${CMAKE_CXX_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
    if (GCC_VERSION VERSION_LESS 8.0)
		MESSAGE(FATAL_ERROR "Need a 8.0 or newer GCC compiler. Current GCC: ${GCC_VERSION}")
    else ()
        set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3")
    endif ()
endif ()

to

if (${CMAKE_CXX_COMPILER_ID} STREQUAL "GNU")
    execute_process(COMMAND ${CMAKE_CXX_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
    if (GCC_VERSION VERSION_LESS 8.0)
		MESSAGE(FATAL_ERROR "Need a 8.0 or newer GCC compiler. Current GCC: ${GCC_VERSION}")
    else ()
	if(NOT CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64") 
        	set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -msse -msse2 -msse3")
	endif ()
    endif ()
endif ()

I also add at the end a python app with fastapi where I can send a photo to my docker with HTTP.

from fastapi import FastAPI, File, UploadFile, HTTPException
import uvicorn
import os
import subprocess
import csv
import json

app = FastAPI()

# Ensure the /tmp/openface directory exists
os.makedirs("/tmp/openface", exist_ok=True)

@app.get("/start-record/")
async def start_record() :
    command = []
    return

@app.post("/upload-photo/")
async def upload_photo(file: UploadFile = File(...)):
    
    # Define the path where the file will be saved
    file_path = os.path.join("/tmp/openface", file.filename.split("/")[-1])
    
    try:
        # Save the file to the specified path
        with open(file_path, "wb") as buffer:
            buffer.write(await file.read())
        
        # Check if the file was saved correctly
        if not(os.path.exists(file_path)):
            # return {"message": f"Photo '{file.filename}' was correctly downloaded and saved to /tmp/openface."}
            raise HTTPException(status_code=500, detail="Failed to save the photo.")
        
        # TODO faire le subprocess, importer pandas et renvoyer la réponse sous forme de json ou csv c'est mpoins lourd
        command = ["FeatureExtraction", "-gaze", "-f", file_path, "-of", "/tmp/openface/output.csv"]
        # command = f"FeatureExtraction -gaze -f {file_path} -of /tmp/openface/output.csv"
        # print(command)
        subprocess.check_call(command)
        # print("ok")
        
        csv_path = "/tmp/openface/output.csv"

        # Read the CSV file
        with open(csv_path, 'r') as csvfile:
            reader = csv.DictReader(
                csvfile, delimiter=',', skipinitialspace=True)

            # Filter out the rows that have no gaze data
            out = [{"face_id": row["face_id"],
                    "confidence": row["confidence"],
                    "gaze_angle_x": row["gaze_angle_x"],
                    "gaze_angle_y": row["gaze_angle_y"]
                    }
                   for row in reader]
        
            # print(out)

            # use gaze_data = list(reader) to get all the rows
            gaze_data = list(out)

        # Convert the gaze data to JSON string
        gaze_data_json = json.loads(json.dumps(gaze_data))
        return gaze_data_json
        
        
        # return f"FeatureExtraction -gaze -f {file_path} -of /tmp/openface/output.csv"
        
    except Exception as e:
        print(e)
        raise HTTPException(status_code=500, detail=str(e))

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)

I am not an expert on Docker so I am looking for improvement to My Dockerfile.