Training vs Inference

The Idea

Training is the loop that adjusts parameters until predictions match labels. It is expensive and runs rarely. Inference is a single pass over frozen parameters to answer one input. Same model, two completely different cost profiles.

Real-World Example

A pianist preparing a concert. Months of rehearsal are slow, repeated and full of corrections. The performance itself runs once, in real time, at full speed, with no chance to redo a bar.

The Code

data = [(1.0, "cold"), (2.0, "cold"), (8.0, "hot"), (9.0, "hot")]

# TRAINING: fit one parameter from the labelled examples
cold = [x for x, y in data if y == "cold"]
hot = [x for x, y in data if y == "hot"]
threshold = (sum(cold) / len(cold) + sum(hot) / len(hot)) / 2

# INFERENCE: apply the frozen parameter to a new input
def predict(x):
    return "hot" if x > threshold else "cold"

print(threshold, predict(6.5))   # 5.0 hot

data = [(2.0, "low"), (4.0, "low"), (10.0, "high"), (12.0, "high")]
low = [x for x, y in data if y == "low"]
high = [x for x, y in data if y == "high"]
threshold = (sum(low) / len(low) + sum(high) / len(high)) / 2

def predict(x):
  return "high" if x > threshold else "low"

print(threshold, predict(7.0))