How to initiate parameters of layers

1

Dose PyTorch possess any elegant way to initiate parameters of some layers when necessary?Dose some functions,such as conv2,Linear,has direct interfaces for users to do initialization?
Thank you for your answers!

3 Likes
2

Have a look of this

m = t.nn.Conv2d(16, 33, 3, stride=2)
m.weight.data.normal_(0, 0.01)
m.bias.data.fill_(0)

Also nn.init have implemented many initialization methods.

m = t.nn.Conv2d(16, 33, 3, stride=2)
xavier_uniform(m.weight.data)
5 Likes
3

Thanks for you reply!!!:+1: