\;j+ddlZddlZddlZddlmZmZmZddlmZddl m Z m Z ejd dZ ejd dZejeeddfdZy) N)backwardcore framework) prim_config)primxutilsc, tjs tdt|tj t jfstdt|dt|tj t jfstdt|dtj|tj|tj|}}}t jj t fd||zDr tdtj tj |dj"}|j%|||\}}t|tj r|dS|S)aSForward mode of automatic differentiation. Note: **ONLY available in the static graph mode and primitive operators.** Args: outputs(Tensor|Sequence[Tensor]): The output tensor or tensors. inputs(Tensor|Sequence[Tensor]): The input tensor or tensors. grad_inputs(Tensor|Sequence[Tensor]): Optional, the gradient Tensor or Tensors of inputs which has the same shape with inputs, Defaults to None, in this case is equivalent to all ones. Returns: grad_outputs(Tensor|Sequence[Tensor]): The gradients for outputs. Examples: .. code-block:: python >>> import numpy as np >>> import paddle >>> paddle.enable_static() >>> paddle.incubate.autograd.enable_prim() >>> startup_program = paddle.static.Program() >>> main_program = paddle.static.Program() >>> with paddle.static.program_guard(main_program, startup_program): ... x = paddle.static.data('x', shape=[1], dtype='float32') ... y = x * x ... y_grad = paddle.incubate.autograd.forward_grad(y, x) ... paddle.incubate.autograd.prim2orig() ... >>> exe = paddle.static.Executor() >>> exe.run(startup_program) >>> y_grad = exe.run(main_program, feed={'x': np.array([2.]).astype('float32')}, fetch_list=[y_grad]) >>> print(y_grad) [array([4.], dtype=float32)] >>> paddle.incubate.autograd.disable_prim() >>> paddle.disable_static() zRforward_grad must be running on primitiveoperators, use enable_prim to turn it on.5Expected outputs is Tensor|Sequence[Tesnor], but got .4Expected inputs is Tensor|Sequence[Tesnor], but got c3<K|]}|jk7ywNblock.0xrs iG:\00. PROJECTS\API\Inventory\templateJSON\kerjaOCR\Lib\site-packages\paddle/incubate/autograd/primapi.py zforward_grad..^s -W177e WszMVariable in inputs and targets should exist in current block of main program.r)r prim_enabled RuntimeError isinstancerVariabletypingSequence TypeErrortype as_tensorsdefault_main_program current_blockanyr orig2prim Transformr linearize) outputsinputs grad_inputsysxsxs_dotad_ys_dotrs @r forward_gradr.shZ     8  g 2 2FOOD EG}oQ (  fy116??C DF|nA '  !   % B  * * , : : ,>?6!9KVKc"tjs`tj|||}t |t j r-t |tjrt|dkDr|dS|St |t j tjfstdt|dt |t j tjfstdt|dtj|tj|tj|}}}t jjtfd||zDr t!dt#j$t#j&}|j)||\}} td| Dr t!d|j+| ||\}} g} |D]O} | j,j/| j0} | dkrt3d | d| j5| Q|j7t9| |j;|t |t j r| dS| S) anReverse mode of automatic differentiation. Note: **ONLY available in the static graph mode and primitive operators** Args: outputs(Tensor|Sequence[Tensor]): The output Tensor or Tensors. inputs(Tensor|Sequence[Tensor]): The input Tensor or Tensors. grad_outputs(Tensor|Sequence[Tensor]): Optional, the gradient Tensor or Tensors of outputs which has the same shape with outputs, Defaults to None, in this case is equivalent to all ones. Returns: grad_inputs(Tensor|Tensors): The gradients for inputs. Examples: .. code-block:: python >>> import numpy as np >>> import paddle >>> paddle.enable_static() >>> paddle.incubate.autograd.enable_prim() >>> startup_program = paddle.static.Program() >>> main_program = paddle.static.Program() >>> with paddle.static.program_guard(main_program, startup_program): ... x = paddle.static.data('x', shape=[1], dtype='float32') ... x.stop_gradients = False ... y = x * x ... x_grad = paddle.incubate.autograd.grad(y, x) ... paddle.incubate.autograd.prim2orig() ... >>> exe = paddle.static.Executor() >>> exe.run(startup_program) >>> x_grad = exe.run(main_program, feed={'x': np.array([2.]).astype('float32')}, fetch_list=[x_grad]) >>> print(x_grad) [array([4.], dtype=float32)] >>> paddle.incubate.autograd.disable_prim() >>> paddle.disable_static() rr r r c3HK|]}|duxr|jk7ywrrrs rrzgrad..s' AAATM .agg. .s"zQVariable in inputs and outputs should be None or in current block of main programc3$K|]}|du ywr)rvars rrzgrad..s )&33$;&szEGrads cannot be computed. The given outputs does not depend on inputsz !  g 2 2FOOD EG}oQ (  fy116??C DF|nA '  !   & B  * * , : : 6!9JFJr/ctjsyt|tjj j r"tjd|j}nt|tjr]|D]H}t|tjj j r2tdt|d|dj}ntdt|dtttfstdtdtttfstdtdt d zt j"|5tjd t%dkDrt%dkDrfd }nGt%dkDrt%dk(rfd }n%t%dk(rt%dkDrfd }nd}t'j(||||t d}tjd|dddy#1swYyxYw)aSearch nonbasic ops which have be registered composite rules and replace them with primitive ops. The operators in blacklist will be excluded from program when lowering into primitives, and only the operators in whitelist will be lowering. The priority of blacklist is higher than whitelist, it means an operator both in blacklist and whitelist will not be lowering. The finally set that will be lowering is: (blocks.ops & ops have decomposite rule & whitelist) - blacklist Args: blacklist(frozenset): The Operators that will be exclude when lowering into primitives. whitelist(frozenset): Only the operators in whitelist will be lowering into primitives. start_idx(int): If start_idx exceeds -1, ops[start_idx:] will be processed. Default: -1. backward_length(int): If backward_length exceeds -1, ops[:-backward_length] will be processed. Default: -1. Nz,Atomize composite op to primitive ops begin.z:Expect block or sequence of blocks, but sequence contains r rz,Expect block or sequence of blocks, but got z6Expected type of blacklisst is set|frozenset, but got z6Expected type of whiltelist is set|frozenset, but got forward_blacklistz'Lowering composite forward ops begin...c@|jvxr|jvSrr)r blacklist whitelists rzto_prim..s) 3 Oi8O Or/c |jvSrrJ)rrKs rrMzto_prim..si 7r/c |jvSrrJ)rrLs rrMzto_prim..s) 3r/cy)NTr3)rs rrMzto_prim..sr/) start_idxbackward_lengthcomposite_ops_recordz'Lowering composite forward ops finish: )r_is_fwd_prim_enabledrpaddlebaserBlocklogginginfoprogramrrrrset frozensetr program_guardr6r_lower_composite) blocksrKrLrQrR main_programitemfilter_ replace_opss `` rto_primrds,  $ $ &&&++//556 CD~~ FFOO ,DdFKK$9$9$?$?@PQUVZQ[P\\]^ ay(( :4<. J   i#y!1 2DT)_DUUV W   i#y!1 2DT)_DUUV W  /09? y>A #i.1"4OG ^a C Na$77G ^q S^a%73G$G   +  ""89  >{mLM% / . .s 4B:H77Ir)rXrrU paddle.baserrrpaddle.base.corerpaddle.incubate.autogradrr static_onlyr.rEr\rdr3r/rris 11(1 OLOLd lKlK^ kk BNBNr/