Why isn't data type an arguement while creating nn modules?

1

I was looking at code of Layer norm and the tensors are implicitly being assigned float32 data type.
Why has this been done?

2

Hi,

All module’s parameters are created with the default tensor type torch.Tensor.
If you want to change this, you can either change the default tensor type with torch.set_default_tensor_type() or by changing the type of the layer afterwards with your_layer.double() to convert to double or your_layer.to(new_type).

1 Like
3

Thanks, didn’t knew about that.