Event Graph Mamba model¶
The active model uses continuous message history and independent recording streams on one GPU. See event architecture.
event_graph_mamba ¶
Message-resolution Mamba and learned causal CAN-ID graph readout.
Every observed frame enters Mamba before any learned compression. Graph keys are latest preceding per-ID temporal states, with learned attention rather than count/change correlation edges. This is a separate, incompatible checkpoint type.
EventGraphMamba ¶
EventGraphMamba(*, width: int = 128, layers: int = 2, d_state: int = 16, d_conv: int = 4, expand: int = 2, dropout: float = 0.1, backend: str = 'official', learning_rate: float = 0.001, weight_decay: float = 0.01, payload_weight: float = 1.0, positive_weight: float = 1.0, node_capacity: int = 512, max_graph_tokens: int = defaults.MAX_GRAPH_TOKENS, use_id_gap: bool | None = None, timing_min_support: int = 32, schema_version: int = 2, **legacy_options)
Bases: LightningModule
Source code in graphids/models/event_graph_mamba.py
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 | |
batched_loss_sums ¶
scan ¶
Differentiable equal-length independent streams, with explicit state.
Detach only returned learned history for truncated backpropagation; current chunk outputs retain their complete parameter gradients.
Source code in graphids/models/event_graph_mamba.py
config ¶
Typed configuration for the shared graph-Mamba detector.
EventGraphMambaConfig ¶
Bases: BaseModel
Validated settings shared by YAML construction and checkpoint loading.
read_selected_and_legacy_checkpoints
classmethod
¶
Strip retired neutral flags; reject checkpoints for removed variants.