Output "-nan" for some layer

[dbsxdbsx]

I use tch-rs "0.7.2" with libtorch-win-shared-with-deps-1.11.0+cu113 on win10. Sometimes, within the output of this layer:

let shared_feature_layer = nn::seq()
        .add(nn::linear(
            p / "actor_feature",
            input_len,
            128,
            Default::default(),
        ))
        .add_fn(|xs| xs.relu());

It outputs nan like this:

Columns 1 to 26-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind)
-nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind) -nan(ind
...

The input is like this:

 0.0190  0.0000  0.2186  0.0000  0.7207  0.7616  0.7431  0.0000  0.0000  0.0000
[ CPUFloatType{1,10} ]

I don't know what trigges this error.