a d% @sddlZddlmZddlmZddlmZddlmZddl m Z ddl m Z m Z mZdd lmZmZdd lmZdd lmZdd lmZmZeeee ee e eeeeed ZddddddddZdedddZGdddejjZdS)N)vjp) Dopri5Solver) Bosh3Solver)AdaptiveHeunSolver) Fehlberg2)EulerMidpointRK4)AdamsBashforthAdamsBashforthMoulton) Dopri8Solver)ScipyWrapperODESolver) _check_inputs_flat_to_shape) dopri8dopri5bosh3 fehlberg2 adaptive_heunZeulerZmidpointZrk4Zexplicit_adamsZimplicit_adams fixed_adamsZ scipy_solvergHz>g& .>)rtolatolmethodoptionsevent_fnc Cst||||||||t \ }}}}}}}}}} t|f||||d|} |dur\| |} n(| |d|\} } | |} | r| } |durt| t|f|} |dur| S| | fSdS)aIntegrate a system of ordinary differential equations. Solves the initial value problem for a non-stiff system of first order ODEs: ``` dy/dt = func(t, y), y(t[0]) = y0 ``` where y is a Tensor or tuple of Tensors of any shape. Output dtypes and numerical precision are based on the dtypes of the inputs `y0`. Args: func: Function that maps a scalar Tensor `t` and a Tensor holding the state `y` into a Tensor of state derivatives with respect to time. Optionally, `y` can also be a tuple of Tensors. y0: N-D Tensor giving starting value of `y` at time point `t[0]`. Optionally, `y0` can also be a tuple of Tensors. t: 1-D Tensor holding a sequence of time points for which to solve for `y`, in either increasing or decreasing order. The first element of this sequence is taken to be the initial time point. rtol: optional float64 Tensor specifying an upper bound on relative error, per element of `y`. atol: optional float64 Tensor specifying an upper bound on absolute error, per element of `y`. method: optional string indicating the integration method to use. options: optional dict of configuring options for the indicated integration method. Can only be provided if a `method` is explicitly set. event_fn: Function that maps the state `y` to a Tensor. The solve terminates when event_fn evaluates to zero. If this is not None, all but the first elements of `t` are ignored. Returns: y: Tensor, where the first dimension corresponds to different time points. Contains the solved value of y for each desired time point in `t`, with the initial value `y0` being the first element along the first dimension. Raises: ValueError: if an invalid `method` is provided. )funcy0rrNr)rSOLVERSZ integrateZintegrate_until_eventtorlen) rrtrrrrrshapesZ t_is_reversedZsolversolutionevent_tr%a/var/www/html/stable-diffusion-webui/venv/lib/python3.9/site-packages/torchdiffeq/_impl/odeint.pyodeints),  r'F) reverse_timeodeint_interfacec Ks6|r(t|d|ddg}n"t|d|ddg}||||fd|i|\}} t|||dddd|t \ } } } }} } } } }} | durtdd| D} n| d} |r| }t| ||| \}} |r| }| durt| d| } t d d t | | D} ntj| dd| dgd d } || fS) z8Automatically links up the gradient from the event time.g?rgNcSsg|]}|ddqS)r*)reshape).0sr%r%r& kz odeint_event..r%css0|](\}}tj|dd|dgddVqdS)Nr*rZdim)torchcat)r,r-Zs_tr%r%r& {r/zodeint_event..rr0) r1r2r+detachrrImplicitFnGradientReroutingapplyrtuplezip)rrt0rr(r)kwargsr!r$r#r"Z_func_state_tr%r%r& odeint_event]s$$",  r=c@s$eZdZeddZeddZdS)r5cCs(||_||_|||||fS)z% event_t is the solution to event_fn )rrZsave_for_backwardr4)ctxrrr$r<r%r%r&forwards z#ImplicitFnGradientRerouting.forwardc Cs|j}|j}|j\}}|d}|d}|||}t&t|||f\}\} } Wdn1sx0Y| t | |} |t ||}| | | d |} || }ddd|fS)NTg-q=) rrZ saved_tensorsr4cloneZrequires_grad_r1Z enable_gradrsumZ reshape_as) r>Zgrad_tZ grad_staterrr$r<Zf_valcZpar_dtZdstateZdcdtr%r%r&backwards   4z$ImplicitFnGradientRerouting.backwardN)__name__ __module__ __qualname__ staticmethodr?rCr%r%r%r&r5s r5)r1Ztorch.autograd.functionalrrrrrrrrrZ fixed_gridrr r rr r rr Z scipy_wrapperrmiscrrrr'r=ZautogradZFunctionr5r%r%r%r&s4       >%