uki*dZddlmZddlZddlmZddlZddlmZddl m Z ddl m Z ddlmZddlmZdd lmZdd lmZmZdd lmZdd lmZmZdd l mZeZeZdede j<fdZej@dZ!dZ"dZ#dZ$dZ%dZ&dZ' d$dZ(dde jRe jRddZ*eejVddefdZ,eejZd d!Z.d"Z/d#Z0e.jce/e0y)%a JAX-based Dormand-Prince ODE integration with adaptive stepsize. Integrate systems of ordinary differential equations (ODEs) using the JAX autograd/diff library and the Dormand-Prince method for adaptive integration stepsize calculation. Provides improved integration accuracy over fixed stepsize integration methods. For details of the mixed 4th/5th order Runge-Kutta integration method, see https://doi.org/10.1090/S0025-5718-1986-0815836-3 Adjoint algorithm based on Appendix C of https://arxiv.org/pdf/1806.07366.pdf )partialN)Callable)api_util)core)custom_derivatives)lax)promote_dtypes_inexact)safe_mapsafe_zip ravel_pytree) tree_leavestree_map) linear_utilf debug_infocXttj|||jS)N)r)ravel_first_arg_lu wrap_init call_wrapped)runravelrs O/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/jax/experimental/ode.pyravel_first_argr1s$ ",,qZ@! ##/<0cD||}||g|}t|\}}|SNr )rry_flatargsyansans_flat_s rrr5s+ fo! ! d #S!+(A /rc tjgd|j}||j|jtj||zz}tj t ||||d|d|S)N)ge g?rgG?gg$Wgl,?gh ygHn^z?dtyper)jnparrayr&astypedotasarrayfit_4th_order_polynomial)y0y1kdt dps_c_midy_mids rinterp_fit_doprir4<sqii?GIhhP) ryy"SWWY%:: :% -b"eQqT1R5"M NNrc |j|j}d|z|zd|z|zzd|zz d|zz d|zz}d|z|zd|z|zz d|zzd|zzd |zz }d |z|z||zzd |zz d|zz d|zz}||z} |} |||| | fS) Ngg@g @g0@@g@g2@g,@g@@gg&@)r*r&) r.r/r3dy0dy1r1abcdes rr-r-Es yy" "fSj2b592%B.U:! "uSy2b59s2v%B.U:! "fSjbfs2v%B.U:!3h!! Aq!Qrc t||\}}|j}|tj||zz}tjj ||j |z } tjj ||j |z } tj| dk| dkzdd| z| z } || j ||zz} || || z} tjj | |z |j |z | z }tj| dk|dkztjd| dzdtj| |z d|dzz z}tjd| z|S)Ngh㈵>gư>g{Gz?gV瞯%gr!@gE]t?g/pѿrr)gUUUUUU?rgVI?gUUUUU?gϡԿg1 0 ?r)gdjJ?rg9gg>>?g8\gOc?gc |dz zz} jjtj|dz ddf|zz} ||}|j|ddfj |S)N)r*r&r(r+atset) ir0tiyiftalphabetar1rKfuncrGr.s rbody_funz"runge_kutta_step..body_funvsr b51:o B bii!CGGD1aL!$<< >" rrrWr') r(r)r&zerosshaperXrYr fori_loopr*r+)r`r.rKrGr1c_solc_errorrar0r/y1_errorrPr^r_s````` @@rrunge_kutta_steprids; ))rIrJmxstephmaxc ~t|D]A}t|tjrtj|r4t d|dt j|jt jst d|dtj|||dg|\} } t| ||||||g|| S)aAdaptive stepsize (Dormand-Prince) Runge-Kutta odeint implementation. Args: func: function to evaluate the time derivative of the solution `y` at time `t` as `func(y, t, *args)`, producing the same shape/structure as `y0`. y0: array or pytree of arrays representing the initial value for the state. t: array of float times for evaluation, like `jnp.linspace(0., 10., 101)`, in which the values must be strictly increasing. *args: tuple of additional arguments for `func`, which must be arrays scalars, or (nested) standard Python containers (tuples, lists, dicts, namedtuples, i.e. pytrees) of those types. rtol: float, relative local error tolerance for solver (optional). atol: float, absolute local error tolerance for solver (optional). mxstep: int, maximum number of steps to take for each timepoint (optional). hmax: float, maximum step size allowed (optional). Returns: Values of the solution `y` (i.e. integrated system values) at each time point in `t`, represented as an array (or pytree of arrays) with the same shape/structure as `y0` except with a new leading axis of length `len(t)`. z@The contents of odeint *args must be arrays or scalars, but got .z&t must be an array of floats, but got r) r isinstancerTracer valid_jaxtype TypeErrorr( issubdtyper&floatingrclosure_convert_odeint_wrapper) r`r.trIrJrrrarg convertedconstss rodeintrs, Sc c4;; '0B0B30G  J3%qQ SSS  . ># ' 3 r)nondiff_argnumsc Vfdfd}||d} tjt|d|d| d} tj|gdz} || |d| |d| g} t j || |dd\} }tj |d|fS) Nc||gSr)r rrr`s rz_odeint..stAq(4(rc fd} fd}tj||dg|z^}}|\}}}}}}|z ||z z }tj||j |j } || fS)Nc:|\}}}}}}}|k|kz|dkDzS)Nrr)staterZr#rr1rtarget_ts rcond_funz+_odeint..scan_fun..cond_funs3"aAq"a(lq6z *b1f 55rcH|\}}}}}}}t||||\}} } } ||z} t| ||} t||| |}tjt || d}|dz|| | |||g}|dz||||||g}t ttj| dk||S)NminmaxrWr?) rirwr4r(clipr~maprrC)rrZr rrr1last_t interp_coeffnext_ynext_f next_y_errorr0next_t error_rationew_interp_coeffnewoldrJfunc_rrIs rraz+_odeint..scan_fun..body_funs-2*aAq"fl(81a(L%fflA2vf$\4q&Ik)!VQ; 88%b+6BD Ib UFFFBQ8H Ic UBL Ic K2$56S AArr)r while_loopr(polyvalr*r&)carryrrrar#rrrrelative_output_timey_targetrJrrrrIs ` rscan_funz_odeint..scan_funs~6 BxA3;?IA',$Aq!Q $v-!f*={{<)=)D)D\EWEW)XYH (?rrrrrrW)r(rrSr)rscan concatenate)r`rIrJrrr.rrrrKr1r init_carryr#ysrs````` ` @rrrs (%0 RA" xx!%AAtT2FBTXY"B4!8$,B1r2a5,7* ((8ZAB 0%!R "T(B ((rc 4t|||||||g|}||||ffSr)r) r`rIrJrrr.rrrs r _odeint_fwdrs.tT4r2==" b"d^ rc |\fdd}g}d} f d} ddttjf} tj| | tj t dz dd\\}} } } tjtj| g| dddg}||g| S)Nc`|^}}}tj|| g|\}}| g||S)z9Original system augmented with vjp_y, vjp_t and vjp_args.)rvjp) augmented_staterrr y_barr#y_dotvjpfunr`s r aug_dynamicsz!_odeint_rev..aug_dynamicssA"LAuqGGD!aR/$/ME6 F #VE] ##rr'rc l |\}}}tj ||g |j}||z }t ||||ftj| |dz  gg d\}}}}t t jd|||f\}}}| |dz z}|||f|fS)NrWr)r(r+rmrr)rop itemgetter)rrZrt0_barargs_bart_barr#rrJrr`grrrIrrs rrz_odeint_rev..scan_funs#E68 GGDA1--qt 4 9 9E e^F!'r!ueVX6 BqE6Bq1uI:&'"? "?tF"?Aufh'r}}Q'7%9RSE68 Aa!eH E 68 $e ++rrWr) rr( zeros_likerrarangelenrr))r`rIrJrrresrrts_barrrrr rev_ts_barrrrrs````` ` @@@@r _odeint_revrs,"b$$ B%% & &,, "r8CNND9:**-(( CJJs2w{Ar:+<'5&(Z ??CIIvh/DbD1AB C&  #( ##r)g?g$@rUr6)2__doc__ functoolsroperatorrcollections.abcrrr jax.numpynumpyr(jax._srcrrrjax._src.numpy.utilr jax._src.utilr r jax.flatten_utilr jax.tree_utilrrrrrzip DebugInfortransformation2rr4r-rSrirprwr~infrjitr custom_vjprrrdefvjprrrrs $ "6,)/&0x0dnn0 O$,@ - HL),S%+cggU@ 1 ( 2  9 ): )D#$J{K(r