Of course, I’m going to think some of my first articles were crap if I now have a better feel for what isn’t crappy. Hat

The last line computes the gradients of the weights of the models, which we are watching with the gradient tape. We could now perform an update, using an optimizer like the Stochastic Gradient Descent.,My name is Dimitris Poulopoulos and I’m a machine learning engineer working for Arrikto. I have worked on designing and implementing AI and software solutions for major clients such as the European Commission, Eurostat, IMF, the European Central Bank, OECD, and IKEA.,But what about operator precedence? Well if you add a TF ND array with a NumPy ND array, the TF one takes precedence. Also, if you pass a TF ND array to a TensorFlow function, the return type will be a TensorFlow tensor. Thus, by design, TensorFlow data types have higher precedence than those of NumPy:,We also saw how to expand what you can do with NumPy, using auto-differentiation, to create and train a Machine Learning model from scratch, on GPU and using TensorFlow-related speed-up tricks, like compilation and auto-vectorization.,Multivariate adaptive regression splines algorithm is best summarized as an improved version of linear regression that can model non-linear relationships between the variables. While I demonstrated examples using 1 and 2 independent variables, remember that you can add as many variables as you like.,One of the greatest writers ever, Franz Kafka, burned over 90 percent of his work. Some days I look at my writing and feel like pulling a Kafka. I want to throw it all out.,Now we are ready to perform a forward pass through this model, calculate the loss and compute the gradients of the model, which, as we said previously, are not TF tensors, but TF ND arrays:,TensorFlow’s GradientTape can be used to enable auto-differentiation for TensorFlow NumPy code. GradientTape APIs can also return ND array outputs. To watch this in action, let’s define and instantiate a simple model:,This model is a simple two-layer neural network, with a ReLU activation between the two linearities. Its weights (w1, w2, bias) are not TF variables, but TF ND arrays. The predict method simulates a forward pass through the layers of the network.,I can now more easily identify what isn’t working in my writing. Then I can either change it or drop it. I only had a small seed of this ability when I first started writing. Now I have a well-watered sapling. Someday hopefully I’ll have a towering oak of creativity, editing prowess, and artistic sensibility. But my job now is to just keep watering that sapling.,Despite the shortcomings, this is a powerful tool to work around cases where Tensorflow doesn’t still support certain NumPy operations. Moreover, we can pass a TF ND array to functions that expect a NumPy ND array, like matplotlib’s plotting functions.,TensorFlow ND arrays can interoperate with NumPy functions and the other way around. However, the conversion of a TF ND array to and from a NumPy ND array may trigger actual data copies, slashing the performance.,If I thought a certain article was good before but now I think it is bad then something changed. I realize now that what changed is my own editorial and artistic standards. They are higher now.,It all looks like complete garbage to me. Perhaps a lot of it is. Many writers say they look back on their first few years of writing and cringe. I’m still in my first few years and luckily I cringe some days.,My name is Dimitris Poulopoulos and I’m a machine learning engineer working for Arrikto. I have worked on designing and implementing AI and software solutions for major clients such as the European Commission, Eurostat, IMF, the European Central Bank, OECD, and IKEA.