Skip to content

Functional API

The clorch.nn.functional namespace (aliased as F) contains functional versions of standard layers and utilities. These functions are stateless and expect weights and biases to be passed explicitly where applicable.

Core Operations

Linear & Convolution

(require '[clorch.nn.functional :as F])

;; Linear
(F/linear input weight bias)

;; Convolution (1D, 2D, 3D)
(F/conv2d input weight :bias bias :stride 1 :padding 0)

Pooling

(F/max-pool2d input 2)
(F/avg-pool2d input 2)
(F/adaptive-avg-pool2d input [7 7])

Normalization

(F/layer-norm input normalized-shape :weight weight :bias bias :eps eps)
(F/group-norm input num-groups :weight weight :bias bias :eps eps)

Regularization & Utilities

Dropout

;; p is the probability of an element to be zeroed
(F/dropout x 0.5 :training? true)

;; Spatial Dropout
(F/dropout2d x 0.5 :training? true)

Interpolation & Padding

;; Resize image-like tensors
(F/interpolate input :size [224 224] :mode :nearest)

;; Advanced Padding
;; pad is a vector of [left right top bottom ...]
(F/pad input [1 1 1 1] :mode :reflect)

Loss Functions

See the Loss Functions guide for a full list.

(F/mse-loss input target)
(F/cross-entropy logits targets)

When to use Functional vs Modules?

  • Use Modules (nn/): For standard layers with learnable parameters. Modules handle parameter registration and state management (training/eval mode) automatically.
  • Use Functional (F/): For stateless operations (activations, pooling) or when you need total manual control over weights (e.g., custom meta-learning loops or weights sharing).