Custom `nn.Module` with a Non-standard Initialization

medium (<10 mins) initialization custom module weights bias

this month by E

Create a custom nn.Module for a simple feed-forward layer. Instead of the default PyTorch initialization, you'll apply a specific, non-standard initialization scheme. For example, you could initialize weights from a uniform distribution between $[- 1, 1]$ and biases to a fixed value, e.g., $0.5$ . This exercise focuses on a key part of model design often taken for granted. You'll need to define a __init__ method and a reset_parameters method where you implement your custom logic.

Verification: Print the weights and biases of your custom module after initialization. Confirm that the values are within the range you specified and that the biases are set to your chosen constant.