MadnisConfig¶

class madspace.MadnisConfig(self: madspace._madspace_py.MadnisConfig)¶

Bases: pybind11_object

Tunable parameters of a MadnisTraining run.

property adam_beta1¶: Adam first-moment decay rate.

property adam_beta2¶: Adam second-moment decay rate.

property adam_eps¶: Adam numerical-stability constant.

property adam_weight_decay¶: Adam L2 weight-decay coefficient.

property batch_size_offset¶: Constant added to every channel’s requested batch size.

property batch_size_per_channel¶: Minimum training-batch events requested per active channel.

property batches¶: Number of training batches.

property buffer_capacity¶: Maximum number of samples kept in the replay buffer per channel; 0 disables buffering.

property buffer_skip_batches¶: Number of initial batches during which no samples are buffered.

property buffer_unweighting_quantile¶: Quantile used to estimate the max weight for buffer unweighting.

property buffered_steps_fraction¶: Fraction of training steps done on buffered samples, reached once the buffers are full.

property channel_dropping_interval¶: Batches between two channel-dropping checks.

property channel_dropping_threshold¶: Minimum integral fraction below which a channel is dropped.

property compressed_channel_weight_count¶: Number of channel weights kept per event when compressing the multi-channel weight loss.

property cpu_generator_batch_size¶: Generator batch size on a CPU device.

property fixed_cwnet_fraction¶: Fraction of training spent with the channel-weight network frozen, before it starts training too.

property generator_target_size_factor¶: Multiple of the training batch size kept buffered per channel.

property gpu_generator_batch_granularity¶: Granularity generator batches on a GPU device are rounded to.

property gpu_generator_batch_size¶: Generator batch size on a GPU device.

property grad_clip_threshold¶: Maximum gradient norm; 0 disables clipping.

property integration_history_length¶: Number of integration-history entries kept per channel.

property learning_rate¶: Adam learning rate.

property log_interval¶: Batches between two status-history entries.

property lr_schedule¶: Adam learning-rate schedule.

property minimum_buffer_size¶: Minimum buffered samples before buffered steps are taken.

property softclip_threshold¶: Softclip threshold applied to the channel weights; 0 disables it.

property uniform_channel_ratio¶: Fraction of the training batch drawn uniformly across channels rather than proportionally to their integral.