quantdata.dysanos.grid.gendysanos.case_c_partition#
Parallel fixed-notional direction models for conditional-arbitrage searches.
- class quantdata.dysanos.grid.gendysanos.case_c_partition.GatedParallelFeedForwardDirections(env, *, config, state_mean, state_whitener)[source]#
Bases:
_ParallelDirectionBaseIndependent statewise directions with a learned zero-to-one activity.
- class quantdata.dysanos.grid.gendysanos.case_c_partition.LearnedPartitionDirections(env, *, config, state_mean, state_whitener)[source]#
Bases:
_ParallelDirectionBaseLearned soft state partitions with independent local portfolio directions.
- class quantdata.dysanos.grid.gendysanos.case_c_partition.ParallelDirectionConfig(*, label, source_cache_uid, architecture, state_features, notionals, partitions=16, kernels=64, deterministic_time=0, conditional_time=6, expiries=7, instruments_per_expiry=21, learn_oce_translation=False, action_radius=1.0, group='', user_version='0.1', network_config=FeedForwardConfig(activation='SiLU', num_depth=2, num_width=96, use_bias=True))[source]#
Bases:
objectLogical geometry shared by the conditional-direction models.
- network_config: FeedForwardConfig = FeedForwardConfig(activation='SiLU', num_depth=2, num_width=96, use_bias=True)#
- class quantdata.dysanos.grid.gendysanos.case_c_partition.ParallelFeedForwardDirections(env, *, config, state_mean, state_whitener)[source]#
Bases:
_ParallelDirectionBaseIndependent feed-forward conditional directions for every notional.
- class quantdata.dysanos.grid.gendysanos.case_c_partition.ParallelKernelDirections(env, *, config, state_mean, state_whitener, centers, inverse_bandwidth)[source]#
Bases:
_ParallelDirectionBaseIndependent local Gaussian-kernel portfolios for every direction head.
- class quantdata.dysanos.grid.gendysanos.case_c_partition.ParallelRadialFeedForwardDirections(env, *, config, state_mean, state_whitener)[source]#
Bases:
_ParallelDirectionBaseIndependent zero-capable feed-forward actions for every objective head.
The network emits one signed position per instrument. Dividing by one plus its gross norm keeps the action strictly inside the unit L1 ball while retaining the identity derivative at the zero action. This avoids both a no-trade softmax coordinate and a hard clipping boundary.
- quantdata.dysanos.grid.gendysanos.case_c_partition.project_rows_to_l1_ball_(value)[source]#
Euclidean-project direct nonnegative rows onto the unit L1 ball.