reservoir opconn_graph × matrix → matriximport fullseye as fs; fs.ledger.reservoir_states(W: 'Any', U: 'Any', in_scale: 'float' = 1.0, leak: 'float' = 1.0, nonlinearity: 'str' = 'tanh', seed: 'int' = 0, washout: 'int' = 0, W_in: 'Any' = None) -> 'np.ndarray' (実装を直接呼ぶなら import conngraph; conngraph.reservoir_states(W: 'Any', U: 'Any', in_scale: 'float' = 1.0, leak: 'float' = 1.0, nonlinearity: 'str' = 'tanh', seed: 'int' = 0, washout: 'int' = 0, W_in: 'Any' = None) -> 'np.ndarray'、台帳から引くなら opsconngraph.get("reservoir_states"))reservoir の状態列: x_{t+1} = (1−leak) x_t + leak · f(Wᵀ x_t + W_in u_t)。返りは (T − washout, n)。
U は (T, d) の入力列(1-D は (T, 1))。W_in は seed で決まる一様 (−in_scale, in_scale)
の (n, d) 行列。nonlinearity は tanh / linear。x_0 = 0 から始め、各ステップの更新後の
状態を並べる。washout 行を先頭から捨てる(T 以上は拒否)。
W_in を渡すと乱数の代わりにその (n, d) 行列を使う(in_scale / seed は無視)——
決まったノード群に刺激を入れる(列 = 刺激するノードの指示子)のはこちら。
py -3.11 examples/poc_malecns_activity_wave.pypy -3.11 examples/poc_microns_brain_wave.pymatrix を入力に取れる)reservoir_encode · ridge_readout · ridge_predict · graph_activation_latency · graph_activity_spread · points_activity_video · graph_layer_propagate · states_participation_ratio
reservoir)reservoir_from_graph · reservoir_encode · ridge_readout · ridge_predict
Provenance: conngraph.py — CONNGRAPH operator registry. この per-op ノートは tools/opdocs.py md が自動生成(手編集しない)。
© 2026 Kazufumi Furuse — Fullseye operator documentation. Licensed under Apache-2.0.