Skip to content

Layer — Abstract Base Layer Class

The foundational contract for all MAGNET layer components.


Overview

Layer defines the shared interface for every MAGNET building block: trainable parameters in params, gradient cache in grads, a forward pass method, an optional backward hook, and callable function behavior (layer(x)). It does not perform any computation on its own; subclasses implement forward() and optionally backward().


Key Features

  • Parameter Storage: params dictionary holds trainable Tensor objects
  • Gradient Cache: grads dictionary for manual or autograd-based gradients
  • Callable Interface: Call layer(x, training=True) directly
  • Mode Tracking: set_mode() records training/inference state
  • Parameter Iterator: parameters() yields trainable tensors to optimizers
  • Gradient Sync: sync_grads() copies tensor gradients for inspection

Parameters

Parameter Type Default Description
name str | None None Display name; defaults to class name
verbose bool False Enable debug-level logging

Attributes

| Attribute | Type | Description | | :--- | :--- | :--- | :--- | | params | dict | Trainable Tensor objects | | grads | dict | Gradient arrays | | trainable | bool | Whether parameters are trainable | | _training | bool | Current mode (True = training) |


Methods

forward(x, training=True) (abstract)

Transform input in the forward pass. Subclasses must override.

backward(grad_output) (optional)

Backward pass hook for manual gradient computation.

__call__(x, training=True)

Alias for forward() so layers can be called like functions.

parameters()

Yield all trainable tensors owned by the layer.

sync_grads()

Copy tensor gradients into grads for inspection.

get_params_and_grads()

Yield (parameter, gradient) pairs for optimizers.

set_mode(training)

Record whether the layer is in training or inference mode.


Example Usage

from machinegnostics.magnet import Layer, Tensor
import numpy as np

class Scale(Layer):
    def __init__(self):
        super().__init__()
        self.params["w"] = Tensor(np.array([2.0]), requires_grad=True)
    def forward(self, x, training=True):
        x = x if isinstance(x, Tensor) else Tensor(x)
        return x * self.params["w"]
    def backward(self, grad_output):
        raise NotImplementedError

layer = Scale()
print(layer(np.array([3.0])).data)  # [6.0]

Author: Nirmal Parmar