input → pre-trained model → output — The Essence of AI Inference in One Line

The Core: The Neural Network Does Exactly One Line

Look at the source separation pipeline again:

song.mp3
  ↓ torchaudio.load()        ← file I/O (math)
  ↓ torch.stft()             ← Fourier transform (math)
  ↓ model(magnitude)         ← ★ ONLY neural network ★
  ↓ magnitude × mask         ← matrix multiply (math)
  ↓ torch.istft()            ← inverse Fourier (math)
vocals.wav

model(magnitude) — this one line is the only part using the neural network.

Everything else is deterministic math.

Pattern: input → model → output

All AI inference follows this structure:

output = model(input)

Example 1: Image Classification (Simplest)

output = model(image)  # [28×28 pixels] → [10 probabilities]
predicted = output.argmax().item()  # → 7

Example 2: Sound Classification

result = classifier("dog_bark.wav")
# → {'label': 'Dog', 'score': 0.98}

Inside: wav → mel spectrogram → model() → 527 class probabilities

Example 3: Speech Recognition

result = model.transcribe("speech.mp3")
# → "Hello, how are you?"

Inside: audio → mel spectrogram → model() → token sequence → text

Example 4: Source Separation

sources = apply_model(model, wav)
vocals = sources[3]  # 4 stems, index 3 = vocals

Inside: wav → chunks → STFT → model() → masks → apply → ISTFT

Example 5: Text-to-Speech

tts.tts_to_file(text="Hello", language="en", file_path="hello.wav")

Inside: text → phonemes → model() → audio tokens → vocoder → wav

Common Pattern

Pre/post-processing differs per task, but the core is always model(input) — one line.