Author(s: Michelangiolo Mazzeschi Computer Vision A detailed guide to how to make a chess model. My repo contains training data. The way ML engineers tackle the most challenging tasks in Machine Learning technology is changing with no-code interfaces. Coding the structures necessary to solve computer vision issues, such as object detection or classification, is one of the most difficult coding tasks that I’ve ever faced. Datature is the interface that I will use throughout this article to conduct an object detection experiment. (See this link). Datature Banner I’ll use an image set pre-labeled of a chessboard in this article to build my object detection model. The appropriate label has been applied to each chess piece. Computer vision can perform object detection. A computer vision object detection model can identify objects within an image, and assign the appropriate label to them. Regardless of what libraries we use, this is still one of the most difficult solutions that can be built using code. This solution can be hard coded. The problem with hard-coding is the difficulty of the code, particularly the preprocessing for images. The images must be properly imported, preprocessed, augmented and then fed to a neural network model. Convolutional neural networks, which are used in computer vision tasks, can take hours to create. Worst of all, there is so much that could go wrong when using advanced models for image recognition. My first experience with ResNet object detection was when I couldn’t install an older version of the buggy library due to conflicting TensorFlow. Computer vision code is now more difficult than ever because of the demands on computing power. No matter what version of your computer you have, or how powerful it is (even if it has a GPU), there are not enough resources to run a complex object detection model with training sets of 200 pictures in a day. It is a nightmare. You can only use a virtual machine instead. Nexus I came across Nexus in my search for alternatives to running my own object detection algorithm. It is a platform without code that makes it easy to set up your computer vision algorithms. This platform can integrate with your MLOps process. Once your model is trained it can then be exported to Kubernetes as an artifact. The platform provides the computing power you need to handle intensive tasks, as well as a lot of customization, such as preprocessing options. You get 500 minutes free computing power. This is sufficient to run multiple models and even experiment with object detection. The 230 image model I used took 14 minutes for me to build. MLOps: How to Make or Break Your ML Team MLOps are a process that is used by industry professionals for creating and developing Machine Learning models. This is because we are able to easily build a process that can be integrated with Cloud tools, and that allows us to make changes quickly. While this may seem important if you are handling your model by yourself, imagine working with several models from multiple teams with different “best practices”. This is where things can get complicated. CD4ML, which is a common adaptation to the MLOps process, refers to continuous delivery for machine learning. This framework has the goal of maximising MLOps’ efficiency. MLOps following CD4ML. This is the equivalent to the CD4ML process using the Nexus Platform. I’ll now show you how to create your own object detector model with the Nexus Platform. You can create your first Nexus project once you are logged in. You can have several projects on the platform, which is useful if you have many datasets to manage and specific people working on them. It’s also an excellent way to test different parameters while training your model. Screenshot from the Nexus platform. Files in my repo will contain all of the files that I used for the experiment. Training images, annotations and uploading images. For my example, I’ll select the 231 files in the training folder. This entire process can take a while. I can split my image set between training and testing sets using the platform. The default for test sets is.3. Uploading images Uploading labels I’ve been able upload images to the platform so far but haven’t labeled them. It is difficult to train a supervisory learning algorithm if the images are not labeled. This requires both features and labels. Upload annotation interface The labels for this dataset will be saved in a CSV file that contains the name of each image. It is common for image processing. The platform will let you import multiple object mask labels each one representing a piece of chess. You can also export these files if necessary. After the labels are imported, we can perform an Exploratory Data Analys (a value counting operation) to gain insight into how the data is distributed. It is clear that 2 of the images don’t have labels so we will label them. EDA for the labeling process The annotator is built-in. This section lets us examine each image detail and check the labels. Although they have many tools, I will only be using a boundingbox as my object detection model is being trained. Screenshot of platform’s annotation GIF of manual platform editor Training the model. This is the heart of this platform. The training phase is very easy to setup. You will be able to save up to five hours, especially if you are new at coding complex CNNs. It streamlines model building by making it easy to remove complex code in just a few clicks. You also have the option of choosing preprocessing or augmentation methods. You have a wide range of pre-processing options to choose from. The majority of the time, augmenting images are required when the data set is small or has very few details.
