Introducing W&B

Key Takeaways

[Music] here we have a pretty typical training script using a modern machine learning library in this case I'm using pytorch but it doesn't matter this could be using plain python or any other machine learning library it might look familiar to you we have some definition of the configurations of the model or training script that we're trying we have some setup of the data the model and some training details but how the model will be updated during training and then we have the part of the code that actually iterates over the data and updates the model because machine learning is such an experimental field we'll have to actually run the code before we know how this specific configuration will perform and keep track of some metrics that we care about to see this as we run our code we're printing the value of the loss to keep track of it as the model trains and also because we might want to use this model for some Downstream tasks we're saving the model weights locally so once we've set up this training script we might want to run it many times using different configurations and iterate on the various aspects of the script whether we're trying a new model architecture or adding more data or whether we're changing the configurations like learning rate or or the number of epochs that we train for on each time we try one of these changes we want to note down how that affected our model's performance and maybe make some plots to visualize that if you've trained more than a handful of models and you've been through this process you'll likely know that it becomes pretty difficult to keep track of what you've tried and how that affected the metrics that you're trying to improve so that's just what weights and biases was created for it was created to solve many of the issues that arise when running machine learning experiments like this in this series of short videos we'll go through how we refactor this code to use weights and biases to automatically track and organize machine learning projects by just changing a few lines of code we'll log everything to weights and biases and we'll get shareable dashboards that'll automatically visualize what we've tried and how that affected our model performance foreign