Edifice Model Explorer Demo

Mix.install([
  {:edifice, path: Path.join(__DIR__, "..")},
  {:kino, "~> 0.14"}
])

Setup

Register the Smart Cell so it appears in the “Smart” cell menu.

Edifice.SmartCell.ModelExplorer.register()

Usage

Click the + Smart button below and select Edifice Model Explorer.

The cell provides:

• Family dropdown — pick from transformer, recurrent, ssm, vision, etc.
• Architecture dropdown — filtered to the selected family
• Option inputs — common options (embed_dim, num_layers, etc.) plus family-specific ones
• Variable name — the output variable for the built model

After evaluating, a summary line shows layer count, parameter count, and memory estimate.

Manual equivalent

If you prefer code, the Smart Cell generates calls like:

model = Edifice.build(:mamba, embed_dim: 128, num_layers: 4, state_size: 16)

{encoder, decoder} = Edifice.build(:vae, embed_dim: 256, latent_size: 64)

Inspect a model

model = Edifice.build(:decoder_only, embed_dim: 128, num_heads: 4, num_layers: 2, seq_len: 32)
IO.puts(Edifice.Display.as_table(model))

Compare architectures

for arch <- [:mamba, :decoder_only, :lstm, :min_gru, :mlp] do
  m = Edifice.build(arch, embed_dim: 128, num_layers: 2, num_heads: 4, seq_len: 32)
  layers = Axon.reduce_nodes(m, 0, fn _, acc -> acc + 1 end)
  table = Edifice.Display.as_table(m)
  lines = String.split(table, "\n")
  IO.puts("#{arch}: #{layers} layers — #{Enum.at(lines, -2) |> String.trim()}, #{Enum.at(lines, -1) |> String.trim()}")
end

:ok

Wire into a training recipe

model = Edifice.build(:mlp, embed_dim: 64, hidden_sizes: [128, 64])
IO.inspect(Edifice.Recipes.describe(:classify, num_classes: 10), label: "Recipe config")

# To train:
# loop = Edifice.Recipes.classify(model, num_classes: 10, log: false)
# state = Axon.Loop.run(loop, train_data, %{}, epochs: 5)