Don't Panic — TensorFlow Fundamentals

TensorFlow is the machinery that takes numerical data and turns it into a machine-learning model that can run somewhere other than the notebook where it was assembled. That is a larger job than persuading a few layers to produce numbers. The layers are the polite part. Data, evaluation, and the destination runtime are where the paperwork begins.

The basic cargo is a tensor, a multidimensional array with a shape and a data type. Images, labels, batches, and model outputs all travel in this container. Tensors do not change in place, so trainable weights live in variables, the mutable state that training is allowed to alter. This separation is less mystical than it sounds. One thing carries values; the other is the thing the optimizer keeps adjusting until the loss stops complaining quite so loudly.

The training loop has an admirably direct plot. Inputs and labels enter a model. The model predicts. A loss measures error. Gradients describe how that error changes with respect to the weights, and an optimizer changes the weights. Keras, TensorFlow's high-level API, packages the routine as compile, fit, evaluate, and predict. Start there. A custom loop is not a graduation ceremony; it is for when the packaged loop no longer describes the work.

The surprise is that training success is not the finish line. A model can fit training data and still fail on new examples. Evaluation uses data that did not update the weights, which is why it is a check rather than a victory parade. Features, labels, shapes, and data types must also agree. A model cannot rescue a dataset that arrives in the wrong arrangement, however earnestly it applies gradients.

TensorFlow normally runs operations eagerly, so Python executes them as it reaches them. When a computation needs graph execution or export, tf.function captures TensorFlow operations into a graph. That graph and its variables can become a SavedModel, the portable artifact for a server, browser, mobile, or other TensorFlow runtime. Export is therefore part of the model's job description, not a decorative zip file at the end.

Continue with the Intro for the full map of tensors, variables, data pipelines, and deployment choices. Use Slides when the relationships need a compact diagram. Keep the Cheatsheet nearby when selecting Keras, Core, tf.data, tf.function, or SavedModel. Then use the practice session to make a small classifier train, evaluate, export, and reload. The machine is not offended by a careful sequence. It rather depends on one.