VGG: pilas de filtros pequeños

VGG made a bold bet: use only tiny 3x3 convolutions everywhere, and just stack a lot of them.

Two stacked 3x3 layers see the same area as one 5x5, but with fewer parameters and an extra nonlinearity in between.

Instead of fat layers, VGG chose to go deep, reaching 16 or 19 layers and learning richer feature hierarchies.

Every conv keeps the same padding and size, so shapes stay easy to reason about. This uniformity made VGG simple to copy.

After each block of convs, a 2x2 max pool halves the spatial size while doubling the channel count.

The repeating unit is conv, ReLU, conv, ReLU, pool. Chain these blocks and you have most of VGG.

VGG ends with three large dense layers. They hold most of its weights, which is why the model is so big on disk.

That elegance has a price: VGG is slow and memory-hungry. Beauty in design did not mean cheap to run.

Even today, VGGs early layers make excellent reusable features for transfer learning and style transfer.

Because the pattern repeats, you can generate VGG from a config list, looping to add each layer programmatically.

VGGs gift was a clear principle: prefer many small filters over a few large ones. Modern nets still follow it.

Recall the central design choice behind VGG.

VGG proved that depth built from simple 3x3 blocks beats clever big filters. Simple, repeatable, and influential. Well done!