AVht2dZddlZddlmZddlmZddlmZddlmZddlm Z ddlm Z dd l m Z dd l m Z dd lmZdd lmZdd lmZdZdZdZdZdZdZdZdZedgddZdZy)z|Gradient checker for functions. The gradient checker verifies numerically that an function properly computes the gradients N)backprop)context)dtypes)indexed_slices)ops)tensor) array_ops)gradients_impl) tf_logging) numpy_compat) tf_exportcFt|tr|Sd}|D]}||z} |S)N) isinstanceint)tyxs Y/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/ops/gradient_checker_v2.py_productr#s33 H A  1fa HcRt|tjrtjrntj |j Dcgc]}|jc}|jDcgc]}|jc}|jS|Scc}wcc}w)aYConverts IndexedSlices to IndexedSlicesValue with numpy indices/values. When eager execution is enabled, converts IndexedSlices to IndexedSlicesValue with numpy indices/values. Args: a: any value. Returns: If a is IndexedSlices and eager execution is enabled, calls numpy() on a's fields. Otherwise returns a unchanged. )indicesvalues dense_shape) rr IndexedSlicesrexecuting_eagerlyIndexedSlicesValuernumpyrr)ars r_eval_indexed_slicesr!-syN001 !  , ,$%II.q.#$88,a ,MM ## (/,s B3B$ ct|tjr|jSt|tj r%tj }|j|St|tjrtj|j}t|jt|jk(s$Jd|jd|jd|t!|j|jD]5\}}d|cxkrt|ksnJd|d|||xx|z cc<7|S|S)a@Converts Tensors, EagerTensors, and IndexedSlicesValue to numpy arrays. Args: a: any value. Returns: If a is EagerTensor or Tensor, returns the evaluation of a by calling numpy() or run(). If a is IndexedSlicesValue, constructs the corresponding dense numpy array. Otherwise returns a unchanged. zIndexedSlicesValue has z value slices but z indices rz%IndexedSlicesValue has invalid index  )rr EagerTensorrrTensorget_default_sessionrunrrnpzerosrlenrrzip)r sessarr values_sliceindexs r _to_numpyr0Cs 3??# 7796==!  " " $D 88A;>445 ((1== !C qxx=C N *" 199a !" * #188QYY7! e % "#c( "F;@! DF " %jL j! J (rc tjrfd}|St||Dcgc]\}}tj||c}} t j  fd}|Scc}}w)a,Return a function that executes 'f'. In TF 2.x, this is the same as `f`. In TF 1.x, returns a Python function that executes the graph defined by `f` in a Session. Args: f: the function. xs_dtypes: dtypes of f's arguments. xs_shapes: shapes of f's arguments. Returns: c>ttj|SN)maprconvert_to_tensor)xs_datafs rdecorated_eagerz!_prepare..decorated_eagerps C))73 44r)shapec |Dcgc] }t|}}jtt|Scc}w)N) feed_dict)r0r'dictr+)r6r r,xsrs rdecorated_graphz!_prepare..decorated_graph{s>%,-y|-G- 88Ac"g&6!78 88.sA)rrr+r placeholderrr&) r7 xs_dtypes xs_shapesr8x_dtypex_shaper>r,r=rs ` @@@r_preparerD`s{  5 ")Y7 '7G73"f! "$9 s!A6c ||}|jjrt|jdzn |j}|jrdnd}t |}t |dd} t ||z} t j | |f|jjj} t j ||j} | jj|jj} tj||gtfd|g|Dcgc]}|jc}zdg|Dcgc]}|jc}z}t| D]}d| |<t|| g|d}t!|}t#|t$j&rNt)|j*|j,D]*\}}|| z}|| z}| |||fxx|j.z cc<,n2|0|jj| j| |ddf<d| |<| dk(rst|| g|d}|j|jk7r%t1d|jd |jt j2|r t1d t5j6dd | | Scc}wcc}w) aYComputes the theoretical Jacobian for f regarding xs[param]. One can think of the relation among f, xs and y as y = f(xs). Args: f: the function. y_shape: the shape of the result. y_dtype: the dtype of the result. xs: a list of tensors. param: the index of the target parameter. Returns: A 2-d numpy array representing the Jacobian. It has "y_size" rows and "x_size" columns where "x_size" is the number of elements in xs[param] and "y_size" is the number of elements in the result. Raises: ValueError: If result is empty but the gradient is nonzero. )rFrNdtypec|d|iS)Ndy)rJr=grad_fn_unpreps rz/_compute_theoretical_jacobian..s^R%?B%?rrz)Empty gradient has wrong shape: expected z, got z#Empty tensor with nonzero gradientszTheoretical Jacobian = %s)rH is_complextupler9rr(r) real_dtypeas_numpy_dtyperavelviewrgradients_functionrDranger0r!rrrr+rrflat ValueErroranyloggingvlog)r7y_shapey_dtyper=paramrrCy_factorx_size x_val_sizey_sizejacobiandy_data dy_data_flatzgrad_fnrowgradivc_beginc_endrLs @r_compute_theoretical_jacobianrms( i!%&WW%7%7E!''NT !QWW'$$Q!( G & $* G x '&XXvv&agg.@.@.O.O P( HHWG$:$: ;'%%g&8&8&G&GH,..q5':. ?Y2!6a!''!66V31qww33 5'6] cL WW*r*1- .D  %D$99:dllDKK0/$!Qj.*$gem#$.$/  **8>>:hsAvL " q[ WW*r*1- .D zzQWW - .. vvd| < == ,,q.9 /9"73s )K K c||j}||j}t||jrdndz}||jrdndz}|jj }|jj }|D cgc]} | j} } |D cgc]} | j} } |D cgc] } t jt| "}} ||} t jd|z|d} tj||f|} t|| | }t|D]}| jj||}| jj||xx|z cc<t||}|| jj||<| jj||xx|zcc<t||}|| jj||<||z | z }|jj|| dd|f< tj dd| | Scc} wcc} wcc} w)aRComputes the numeric Jacobian for f regarding xs[param]. One can think of the relation among f, xs and y as y = f(xs). Args: f: the function. y_size: the number of elements of the result. y_dtype: the dtype of the result. xs: a list of tensors. param: the index of the target parameter. delta: the amount of perturbation we give to the input. Returns: A 2-d numpy array representing the Jacobian. It has "y_size" rows and "x_size" columns where "x_size" is the number of elements in xs[param] and "y_size" is the number of elements in the result. rFrrGrKNzNumeric Jacobian = %s)r9rHrrNrPrQr np_asarrayr0r(r)rDrUrRrSrYrZ)r7rar\r=r]deltarCrBr_rr@rAscalerbcoloriginaly_posy_negdiffs r_compute_numeric_jacobianrws$ uIOO' uIOO' G W%7%7Q ?& ',,Q! 4&    - -'    - -' "#1qww#)# "#1qww#)#8::!  ! -:":i!  ! !!e)7 ;B ?% XXvv&g 6( q)Y'! 6] 2cwwy~~g&s+HGGINN7C E) af E#+AGGINN7C GGINN7C E) af E#+AGGINN7C EMU "Dzz|((1HQV 2 ,,q*H5 /9$#;s>II0%Ic||}|j}tjtjtjtj tj tjg}|j|vsJd|jd||} | j|vsJd| jzt|} t|||||} t|| ||||} | | fS)z0Computes the theoretical and numerical jacobian.z-Cannot compute gradient for unsupported type z of argument z4Cannot compute gradient for unsupported type %s of y) rHrfloat16bfloat16float32float64 complex64 complex128 base_dtypenamerrmrw) r7r[r\r=r]rprr allowed_typest2rajacob_tjacob_ns r_compute_gradientrsi!gg! nnfoov~~v~~ ))-  &:*+&&%*9: &" - 'P+EGIww+OP ' G & )!Wgr5 I' %a"eU K' ' rc|Dcgc]}tj|}}|Dcgc]}|j}}|Dcgc]}|j}}t |||}||}t t tt|Dcgc]9}t||jtj|j|||;c}Scc}wcc}wcc}wcc}w)z)Compute gradients for a list of x values.) rr5rHr9rDrOr+rUr*rras_dtype) r7r=rprr@rAf_temprris r_compute_gradient_listrs+--Qa -"- "#1qww#)# "#1qww#)# Ay) ,& bk!  R>  Aqww(@"a O    .## sC CC>C ztest.compute_gradient)v1c~t|ttfstdt |z|d}t |||S)aComputes the theoretical and numeric Jacobian of `f`. With y = f(x), computes the theoretical and numeric Jacobian dy/dx. Args: f: the function. x: the arguments for the function as a list or tuple of values convertible to a Tensor. delta: (optional) perturbation used to compute numeric Jacobian. Returns: A pair of lists, where the first is a list of 2-d numpy arrays representing the theoretical Jacobians for each argument, and the second list is the numerical ones. Each 2-d array has "y_size" rows and "x_size" columns where "x_size" is the number of elements in the corresponding argument and "y_size" is the number of elements in f(x). Raises: ValueError: If result is empty but the gradient is nonzero. ValueError: If x is not list, but any other type. Example: >>> @tf.function ... def test_func(x): ... return x*x ... >>> >>> class MyTest(tf.test.TestCase): ... ... def test_gradient_of_test_func(self): ... theoretical, numerical = tf.test.compute_gradient(test_func, [1.0]) ... # ((array([[2.]], dtype=float32),), ... # (array([[2.000004]], dtype=float32),)) ... self.assertAllClose(theoretical, numerical) zZ`x` must be a list or tuple of values convertible to a Tensor (arguments to `f`), not a %sgP?)rlistrOrWtyper)r7rrps rcompute_gradientr(sNN Ae} %  ')-a 1 22 ] E 1e ,,rcd}t||D]X\}}|js |jstj|tj||z j }Z|S)aComputes maximum elementwise gap. Computes the maximum elementwise gap between two lists of tensors of the same shape. Args: grad1: a lists of tensors. grad2: a lists of tensors with the same shape as grad1. Returns: The maximum elementwise gap between the two. r)r+sizer(maximumfabsmax)grad1grad2errorj_tj_ns r max_errorr\s] %eU#:hc3 xx388jjc 2 6 6 89e: ,rr3)__doc__rr(tensorflow.python.eagerrrtensorflow.python.frameworkrrrrtensorflow.python.opsr r tensorflow.python.platformr rYtensorflow.python.utilr tensorflow.python.util.tf_exportr rr!r0rDrmrwrrrrrKrrrs~ ,+.6+.+0</6  , :DEP7t(   "r*0-+0-fr