The Longest YouTube Video I’ve EVER Seen…
There are 24 hours in a day. 23 hours, 56 minutes, and 4 seconds, to be specific. I spent the past week watching 25 hours of YouTube. Why? One of my main tasks this week while working with the Oasis Lab was to learn machine learning through PyTorch. While it wasn’t easy, having an understanding of how to code in Python beforehand made the job much easier.
Rather than bore you with a monotonous summary of everything that I learned from the video, I will highlight the most important parts. Let’s start with the basics. The typical PyTorch workflow and the workflow I plan to follow while working with this program start with gathering the data, typically from databases like Kaggle. The next step is to build a model or choose a pre-trained model. Seems simple, right? In reality, it’s not so simple. You must select a loss function and optimizer, and you must build a training loop before this can be achieved. Next, you have to fit the model to the training data, so that you can make predictions and evaluate a model based on your main data. Once saved, your work is done and ready to be used elsewhere.
The neural network for this model and any model through PyTorch can work through 3 different types of classification: binary, multi-class, or multi-label classification, each dependent on what you need the model to do. More specifically, binary refers to when a target can be defined as one of two options. Meanwhile, multi-class is similar but there are more than 2 options. Multi-label differentiates itself from the other two because it can be defined as more than one option.
Another important consideration when working with machine learning is tracking your experiments. Rather than simply printing out results, more complicated models will need more elaborate methods which come in 3 different forms: TensorBoard, Weights and biases experiment tracking, or MLFlow. While setting up TensorBoard may be one of the easier methods to set up of the three, the latter two provide a generally more agreeable user experience with better capabilities and resources. MLFlow is fully open source and has a wide range of applications, but it is slightly harder to set up a remote tracking server compared to other services. If time permits, I plan on utilizing the latter two methods given their benefits, especially prioritizing the use of MLFlow.
After watching the seemingly endless video and coding the examples shown in the video, I believe I have a firm enough grasp of the concept of machine learning through PyTorch to begin working on a model myself. But before I can get started on that, I will need to find the appropriate datasets on car crashes and determine the criteria for the model to use from the dataset. Make sure to stay tuned for next week as I continue my work with machine learning in relation to car crashes.
Comments:
All viewpoints are welcome but profane, threatening, disrespectful, or harassing comments will not be tolerated and are subject to moderation up to, and including, full deletion.