clip_grad_value

shrew::prelude

Function clip_grad_value 

Source 
pub fn clip_grad_value<B>(
    grads: &GradStore<B>,
    params: &[Tensor<B>],
    max_value: f64,
) -> Result<GradStore<B>, Error>
where
    B: Backend,
Expand description

Clamp each gradient element to [-max_value, max_value].

This is a simpler but less commonly used strategy than norm clipping.

Returns clipped_grads.