o ۷i@sddlZddlmZddlmZddlmZmZmZmZm Z Gdddej Z Gd d d ej Z Gd d d ej Z    ddedeedeedeeedfdeejdee eejff ddZdS)N) brownian_base)brownian_interval)OptionalScalarTensorTupleUnioncs^eZdZfddZdddZddZed d Zed d Zed dZ eddZ Z S)ReverseBrowniancstt|||_dSN)superr __init__ base_brownian)selfr __class__P/home/ubuntu/vllm_env/lib/python3.10/site-packages/torchsde/_brownian/derived.pyrs zReverseBrownian.__init__NFcCs|j| | ||dSN)return_Ureturn_A)r)rtatbrrrrr__call__szReverseBrownian.__call__cC|jjd|jdS)Nz(base_brownian=))r__name__rrrrr__repr__ zReverseBrownian.__repr__cC|jjSr )rdtyperrrrr"#zReverseBrownian.dtypecCr!r )rdevicerrrrr$'r#zReverseBrownian.devicecCr!r )rshaperrrrr%+r#zReverseBrownian.shapecCr!r )rlevy_area_approximationrrrrr&/r#z'ReverseBrownian.levy_area_approximationNFF) r __module__ __qualname__rrrpropertyr"r$r%r& __classcell__rrrrr s     r csreZdZdZddededeffdd Zdd d Zd d Z e ddZ e ddZ e ddZ e ddZZS) BrownianPathaBrownian path, storing every computed value. Useful for speed, when memory isn't a concern. To use: >>> bm = BrownianPath(t0=0.0, w0=torch.zeros(4, 1)) >>> bm(0., 0.5) tensor([[ 0.0733], [-0.5692], [ 0.1872], [-0.3889]]) t0w0 window_sizecs>|d}||_tj|||j|j|jdd|_tt| dS)zInitialize Brownian path. Arguments: t0: Initial time. w0: Initial state. window_size: Unused; deprecated. rN)r.t1sizer"r$ cache_size) _w0rBrownianIntervalr%r"r$ _intervalr r,r)rr.r/r0r1rrrrBs zBrownianPath.__init__NFcC0|j||||d}|dur|s|s||j}|Srr6r4rtrrroutrrrrO zBrownianPath.__call__cCrNz (interval=rrrr6rrrrrVr zBrownianPath.__repr__cCr!r r6r"rrrrr"Yr#zBrownianPath.dtypecCr!r r6r$rrrrr$]r#zBrownianPath.devicecCr!r r6r%rrrrr%ar#zBrownianPath.shapecCr!r r6r&rrrrr&er#z$BrownianPath.levy_area_approximation)r-r')rr(r)__doc__rrintrrrr*r"r$r%r&r+rrrrr,4s    r,cseZdZdZ       ddededeed eed eed ed ed edeeffdd Z dddZ ddZ e ddZ e ddZe ddZe ddZZS) BrownianTreeaBrownian tree with fixed entropy. Useful when the map from entropy -> Brownian motion shouldn't depend on the locations and order of the query points. (As the usual BrownianInterval does - note that BrownianTree is slower as a result though.) To use: >>> bm = BrownianTree(t0=0.0, w0=torch.zeros(4, 1)) >>> bm(0., 0.5) tensor([[ 0.0733], [-0.5692], [ 0.1872], [-0.3889]], device='cuda:0') Nư> r.r/r1w1entropytol pool_size cache_depthsafetyc sd|dur|d}|durd} n||} ||_tj|||j|j|j|||d| d |_tt| dS)aInitialize the Brownian tree. The random value generation process exploits the parallel random number paradigm and uses `numpy.random.SeedSequence`. The default generator is PCG64 (used by `default_rng`). Arguments: t0: Initial time. w0: Initial state. t1: Terminal time. w1: Terminal state. entropy: Global seed, defaults to `None` for random entropy. tol: Error tolerance before the binary search is terminated; the search depth ~ log2(tol). pool_size: Size of the pooled entropy. This parameter affects the query speed significantly. cache_depth: Unused; deprecated. safety: Unused; deprecated. NrT) r.r1r2r"r$rJrKrL halfway_treeW) r4rr5r%r"r$r6r rEr) rr.r/r1rIrJrKrLrMrNrPrrrrzs$ zBrownianTree.__init__FcCr7rr8r9rrrrr<zBrownianTree.__call__cCrr=r>rrrrrr zBrownianTree.__repr__cCr!r r?rrrrr"r#zBrownianTree.dtypecCr!r r@rrrrr$r#zBrownianTree.devicecCr!r rArrrrr%r#zBrownianTree.shapecCr!r rBrrrrr&r#z$BrownianTree.levy_area_approximation)NNNrFrGrHNr')rr(r)rCrrrrDfloatrrrr*r"r$r%r&r+rrrrrEjsH -   rE?yr.r1r2.r"r$cKsR|dur|jn|}|dur|jn|}|dur|jn|}tjd|||||d|S)zZReturns a BrownianInterval object with the same size, device, and dtype as a given tensor.N)r.r1r2r"r$r)r%r"r$rr5)rTr.r1r2r"r$kwargsrrrbrownian_interval_likesrV)rRrSNNN)torchrrtypesrrrr r BaseBrownianr r,rErDr"strr$rVrrrrs.  6Y