log_basis_weights is too big

[hummingg]

Some elements in log_basis_weights are up to 90 running with my own datasets, causing loss NaN. How to solve it? Please!

self.shallow_net = ResidualLayer(hidden_size, n_types * (n_bases + 1))
# ...
log_basis_weights = self.shallow_net(history_emb).view(
            batch_size, T, self.n_types, -1
        )
# ...
cumulants = integrals.unsqueeze(2).mul(log_basis_weights.exp()).sum(-1)