Final testing and improvements

5 August, 2021

Introduction

The goal of this week was to ensure that the component works properly, for that I created a program that connect to this distributed component and execute some operations using it. These things appear more or less easy, but I had a lot of problems connecting the program to the proxy of the component.

Testing the component

Using Ice a distributed framework based in RPC we can provide the components an easily interaction with other components for that reason is used in RoboComp and LearnBlock. So I write the following program using Ice to interact with the component:

import Ice
import sys
import time
import numpy as np
import subprocess

import cv2

Ice.loadSlice("-I ./src/ --all ./DetectionComponent.ice")

from RoboCompDetectionComponent import *

def connectComponent(ic, stringProxy, _class, tries=4):

    i = 0
    while True:
        try:
            i += 1
            print(f"Try {i}")
            basePrx = ic.stringToProxy(stringProxy)
            proxy = _class.checkedCast(basePrx)
            print("Connection Successful:", stringProxy)
            break
        except Ice.Exception as e:
            if i is tries:
                print("Cannot connect to the proxy:", stringProxy)
                return None
            else:
                time.sleep(1.5)
    return proxy


if __name__ == "__main__":
    # Initialize the communicator of Ice
    ic = Ice.initialize(sys.argv)

    # Connecting to the component in the port 10010 of localhost
    component_proxy = connectComponent(ic, "detectioncomponent:tcp -h localhost -p 10010", DetectionComponentPrx,10)

    # Test the setter and getter of the threshold of the component
    print(component_proxy.getThreshold())
    component_proxy.setThreshold(0.1)
    print(component_proxy.getThreshold())
    component_proxy.setThreshold(0.3)

    # Test the processing of the image by the component

    # Prepare the image for the component
    img = cv2.imread("../test/test.jpg")
    img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
    img = cv2.resize(img,(448,448))

    # Create the input of the component according to the interface
    frame = TImage()
    frame.width = img.shape[0]
    frame.height = img.shape[1]
    frame.depth = img.shape[2]
    frame.image = bytes(img)

    # Get the predictions of the component
    predictions = component_proxy.processImage(frame)

    # Print the predictions
    for pred in predictions:
        print(f"Box location=({pred.x},{pred.y}),Width={pred.w},Height={pred.h},Label={pred.label}")

    # Destroy the communicator to ensure the program finishes properly
    ic.destroy()

The program basically executes all the operations defined by the interface that I have previously prepared, for the thereshold there is a getter and a setter so we can check if they work. Finally, tests the principal operation of the component that given an image as an input returns a list with the boxes and labels predicted for that image. Reads an image locate in the test directory and returns its predictions, the image has as attributes the width, height, depth and a sequence of bytes that contains the proper image. The greater problem that appear was related to connecting to the proxy of the component, I finally discover that the thread of the component finished with the program so in the end of the program I create a infinite loop that only finished when the user press Control+c in the terminal.

Improvements and conclusion

After the testing, appeared some errors that needed to be solved, but no one of them was important. The component does not stand out for being fast, but as I can only test it an environment with a CPU I expect that the perfomance will dramatically increases using a GPU. After that, I only need to prepare the pull request for the official repository of LearnBlock from my fork. Basically this pull request will englobe adding a directory call Detect with the files needed from the model to work properly, also I needed to add the interfaces in IDSL and Ice format in the interfaces directory.

Alejandro Fernández Camello