What’s the difference between Model methods predict() and __call__()?

Let’s answer with an extract from Deep Learning with Python, Second Edition:

Both y = model.predict(x) and y = model(x) (where x is an array of input data) mean “run the model on x and retrieve the output y.” Yet they aren’t exactly the same thing.

predict() loops over the data in batches (in fact, you can specify the batch size via predict(x, batch_size=64)), and it extracts the NumPy value of the outputs. It’s schematically equivalent to this:

def predict(x):
    y_batches = []
    for x_batch in get_batches(x):
        y_batch = model(x_batch).numpy()
        y_batches.append(y_batch)
    return np.concatenate(y_batches)

This means that predict() calls can scale to very large arrays. Meanwhile, model(x) happens in-memory and doesn’t scale. On the other hand, predict() is not differentiable: you cannot retrieve its gradient if you call it in a GradientTape scope.

You should use model(x) when you need to retrieve the gradients of the model call, and you should use predict() if you just need the output value. In other words, always use predict() unless you’re in the middle of writing a low-level gradient descent loop (as we are now).