torch.js has not been released yet.

torch.nn.ForwardHook

export type ForwardHook = (
  module: Module,
  input: Tensor[],
  output: Tensor | Tensor[]
) => Tensor | Tensor[] | void;

Hook called after a module's forward pass.

ForwardHook is executed after the module's forward() method completes. It receives the module, its inputs, and the computed outputs. Can modify or validate the output before it propagates further.

Signature:

(module: Module, input: Tensor[], output: Tensor | Tensor[]) => Tensor | Tensor[] | void

Parameters:

module: The module whose forward just completed
input: Original input tensors (may be modified from pre-hooks)
output: The computed output from forward (single tensor or array)

Return Value:

Return modified output to change downstream computation
Return void/undefined to use original output unchanged
Useful for output normalization, validation, or debugging

Use Cases:

Log outputs for inspection/debugging
Monitor activations during training
Modify outputs (e.g., clip values, apply constraints)
Implement custom regularization on outputs
Validate output properties (shape, range, NaN/Inf)
Extract intermediate representations

Examples

const hook: ForwardHook = (module, input, output) => {
  // Inspect output activation statistics
  if (Array.isArray(output)) {
    output.forEach(t => console.log('Output min/max:', t.min().item(), t.max().item()));
  } else {
    console.log('Output shape:', output.shape);
  }
  // Modify output: clip to [-1, 1]
  return torch.clamp(output, -1, 1);
};

layer.register_forward_hook(hook);