MadnisLoss

class MadnisLoss : public madspace::FunctionGenerator

MadNIS training objective for the adaptive samplers.

Combines the per-channel integrand and sampling-probability values into the variance-based loss used to train the Flow and ChannelWeightNetwork [1]. A dedicated study of madspace with MadNIS is left to a future publication, so this class is described from the code only.

batch is the leading batch dimension. c indexes the channels.

Arguments

• chan<c>_integrand – float, shape (batch,) – integrand values.

• chan<c>_sample_prob – float, shape (batch,) – sampling probabilities.

• chan<c>_cwnet_inputs, chan<c>_channel_weight_values / chan<c>_channel_weight_indices – present only with a ChannelWeightNetwork.

Returns

• loss – float, scalar – the training loss.

• abs_means – float, shape (c,) – per-channel mean of |integrand|.

• variances – float, shape (c,) – per-channel integrand variance.

References

• [1] T. Heimel et al., “MadNIS - Neural multi-channel importance sampling”, https://arxiv.org/abs/2212.06172

Public Functions

MadnisLoss(const std::vector<std::shared_ptr<FunctionGenerator>> &functions, const std::optional<ChannelWeightNetwork> &cwnet, double softclip_threshold = 0.0, std::size_t compressed_channel_weight_count = 50)

Parameters:

• functions – The per-channel integrands.

• cwnet – Optional channel-weight network to train jointly.

• softclip_threshold – Soft-clipping threshold on the loss contribution; 0 disables it.

• compressed_channel_weight_count – Number of channel-weight entries kept per event.