A model architecture is the layout of the model. The layout decides whether the model reads pixels, words, sound, or a mixture. It also decides whether the answer is a label, more text, a new image, or something else.
The trained weights are different. They are the learned settings inside that shape. Architecture is the shape. Weights are what that shape learned from examples.
ClothingCNN is one choice
In the forward clothing course, Fashion-MNIST goes into ClothingCNN and a clothing label comes out. ClothingCNN is an image-classifier architecture. That means it is one good shape for this job. It is not the only kind of model.
Different shapes still run on a chip
From the metal to the model starts at the hardware and works back up through the runtime, UI, conversion, training, and pipeline. A language model and an image model can have different shapes. Both still become work on the chip.
Next, follow a language model as it reads a sequence and writes the next token.
Key concepts
- Architecture is the layout: what the model reads, how information moves, what it writes.
- Weights are the learned settings inside that layout.
- ClothingCNN is an image-classifier shape: 28x28 clothing pixels in, a clothing label out.
- Different shapes still become work on the chip when the app runs.
Takeaways
- You can say architecture versus weights in one sentence.
- You can describe ClothingCNN as pixels, then model shape, then label.
- You can point to the forward course for the build and From the metal to the model for the hardware path.