Edifice.Vision.FocalNet (Edifice v0.2.0)

Copy Markdown View Source

FocalNet: Focal Modulation Networks for vision (Yang et al., 2022).

Replaces self-attention with focal modulation, which aggregates context at multiple granularity levels using hierarchical depthwise convolutions and gated aggregation. This provides a simple yet effective alternative to attention that captures both local and global context.

Architecture

Image [batch, channels, height, width]
      |
+-----v--------------------+
| Patch Embedding           |  Split into P x P patches, linear project
+---------------------------+
      |
      v
[batch, num_patches, hidden_size]
      |
+-----v--------------------+
| FocalNet Block x N        |
|                           |
| Focal Modulation:         |
|   q = Dense(x)            |
|   For each level l:       |
|     ctx += gelu(conv_l)   |
|   gate = sigmoid(Dense(x))|
|   out = q * gate * ctx    |
|   + Residual              |
|                           |
| FFN:                      |
|   Dense(4*h) -> GELU      |
|   -> Dense(h)             |
|   + Residual              |
+---------------------------+
      |
      v
+---------------------------+
| LayerNorm -> Mean Pool    |
+---------------------------+
      |
      v
[batch, hidden_size]

Usage

model = FocalNet.build(
  image_size: 224,
  patch_size: 16,
  hidden_size: 256,
  num_layers: 4,
  focal_levels: 3,
  num_classes: 1000
)

References

Summary

Types

Options for build/1.

Functions

Build a FocalNet model.

Get the output size of a FocalNet model.

Types

build_opt()

@type build_opt() ::
  {:focal_kernel, pos_integer()}
  | {:focal_levels, pos_integer()}
  | {:hidden_size, pos_integer()}
  | {:image_size, pos_integer()}
  | {:in_channels, pos_integer()}
  | {:num_classes, pos_integer() | nil}
  | {:num_layers, pos_integer()}
  | {:patch_size, pos_integer()}

Options for build/1.

Functions

build(opts \\ [])

@spec build([build_opt()]) :: Axon.t()

Build a FocalNet model.

Options

  • :image_size - Input image size, square (default: 224)
  • :patch_size - Patch size, square (default: 16)
  • :in_channels - Number of input channels (default: 3)
  • :hidden_size - Hidden dimension per patch (default: 256)
  • :num_layers - Number of FocalNet blocks (default: 4)
  • :focal_levels - Number of focal context levels (default: 3)
  • :focal_kernel - Base kernel size for focal convolutions (default: 3)
  • :num_classes - Number of output classes (optional)

Returns

An Axon model. Without :num_classes, outputs [batch, hidden_size]. With :num_classes, outputs [batch, num_classes].

output_size(opts \\ [])

@spec output_size(keyword()) :: pos_integer()

Get the output size of a FocalNet model.

Returns :num_classes if set, otherwise :hidden_size.