Custom Architectures

How to build a complex, non-standard architecture using Recipes.

from torchvision_customizer import Recipe, build_recipe

# Define a recipe for a network with mixed block types and irregular stages/channels
recipe = Recipe(
    name="ExperimentalNet",
    stem="conv(32, k=3, s=2)",
    stages=[
        "depthwise(32) x 1",
        "residual(48) x 3 | downsample",
        "residual+se(96) x 4 | downsample",
        "dense(128, growth_rate=32) | downsample", # Single dense block
    ],
    head="linear(100)"
)

model = build_recipe(recipe)
print(model.explain())