BVhYbdZddlmZddlmZddlmZddlmZddlmZddlmZ ddl m Z dd l m Z dd lmZdd lmZdd lmZd ZGddeZej-ej.GddeZGddeZej5ej6dej5ej6dej5ej6dej5ej6dej5ej6dGdde j8ZGddeZGddeZej.efej@efej6efeefgZ!dZ"d Z#d!Z$d$d#Z%y")%z9Keras Input Tensor used to track functional API Topology.) constant_op)dtypes) sparse_tensor)tensor) tensor_shape) type_spec)object_identity) array_ops)ragged_operators) ragged_tensor)nestceZdZdZddZedZedZedZ eddZ dZ d e jfd Zd Zed Zd ZdZdZedZdZdZdZedZdZdZedZedZedZy) KerasTensoraA representation of a Keras in/output during Functional API construction. `KerasTensor`s are tensor-like objects that represent the symbolic inputs and outputs of Keras layers during Functional model construction. They are comprised of the `tf.TypeSpec` of the (Composite)Tensor that will be consumed/produced in the corresponding location of the Functional model. KerasTensors are intended as a private API, so users should never need to directly instantiate `KerasTensor`s. **Building Functional Models with KerasTensors** `tf.keras.Input` produces `KerasTensor`s that represent the symbolic inputs to your model. Passing a `KerasTensor` to a `tf.keras.Layer` `__call__` lets the layer know that you are building a Functional model. The layer __call__ will infer the output signature and return `KerasTensor`s with `tf.TypeSpec`s corresponding to the symbolic outputs of that layer call. These output `KerasTensor`s will have all of the internal KerasHistory metadata attached to them that Keras needs to construct a Functional Model. Currently, layers infer the output signature by: * creating a scratch `FuncGraph` * making placeholders in the scratch graph that match the input typespecs * Calling `layer.call` on these placeholders * extracting the signatures of the outputs before clearing the scratch graph (Note: names assigned to KerasTensors by this process are not guaranteed to be unique, and are subject to implementation details). `tf.nest` methods are used to insure all of the inputs/output data structures get maintained, with elements swapped between KerasTensors and placeholders. In rare cases (such as when directly manipulating shapes using Keras layers), the layer may be able to partially infer the value of the output in addition to just inferring the signature. When this happens, the returned KerasTensor will also contain the inferred value information. Follow-on layers can use this information. during their own output signature inference. E.g. if one layer produces a symbolic `KerasTensor` that the next layer uses as the shape of its outputs, partially knowing the value helps infer the output shape. **Automatically converting TF APIs to layers**: If you passing a `KerasTensor` to a TF API that supports dispatching, Keras will automatically turn that API call into a lambda layer in the Functional model, and return KerasTensors representing the symbolic outputs. Most TF APIs that take only tensors as input and produce output tensors will support dispatching. Calling a `tf.function` does not support dispatching, so you cannot pass `KerasTensor`s as inputs to a `tf.function`. Higher-order APIs that take methods which produce tensors (e.g. `tf.while`, `tf.map_fn`, `tf.cond`) also do not currently support dispatching. So, you cannot directly pass KerasTensors as inputs to these APIs either. If you want to use these APIs inside of a Functional model, you must put them inside of a custom layer. Args: type_spec: The `tf.TypeSpec` for the symbolic input created by `tf.keras.Input`, or symbolically inferred for the output during a symbolic layer `__call__`. inferred_value: (Optional) a non-symbolic static value, possibly partially specified, that could be symbolically inferred for the outputs during a symbolic layer `__call__`. This will generally only happen when grabbing and manipulating `tf.int32` shapes directly as tensors. Statically inferring values in this way and storing them in the KerasTensor allows follow-on layers to infer output signatures more effectively. (e.g. when using a symbolic shape tensor to later construct a tensor with that shape). name: (optional) string name for this KerasTensor. Names automatically generated by symbolic layer `__call__`s are not guaranteed to be unique, and are subject to implementation details. Ncxt|tjs td||_||_||_y)zConstructs a KerasTensor.z6KerasTensors must be constructed with a `tf.TypeSpec`.N) isinstancetype_spec_moduleTypeSpec ValueError _type_spec_inferred_value_name)selfrinferred_valuenames [/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/keras/engine/keras_tensor.py__init__zKerasTensor.__init__ws5 i!1!:!: ; O PPDO)DDJc|jS)zFReturns the `tf.TypeSpec` symbolically inferred for this Keras output.)rrs rrzKerasTensor.type_specs ??rc.|jjS)zFReturns the `TensorShape` symbolically inferred for this Keras output.)r_shaper s rshapezKerasTensor.shape ?? ! !!rc0t|tjrt|dd}t j |}d}|j tjk(r|jjl|jjdkrStj|j}|jr%|j}t|t kDrd}nd}t#|||St|dd}t j |}|||S)CConvert a traced (composite)tensor to a representative KerasTensor.rNr#)rrr)r tensor_libTensorgetattrrtype_spec_from_valuedtyperint32r#rankr onesdimsas_listlen_MAX_TENSOR_RANKr)clsrrrrs r from_tensorzKerasTensor.from_tensors&*++, VVT *d"77?in //V\\ )ioo.B.B.Noo""Q&&#f5;;   )113.  #3 3!N. > MMVVT *d"77?i  &&rc|||S)N)rrr6rrs rfrom_type_speczKerasTensor.from_type_specs  ..rc<|jltjtj|jtj }|j jjdk(r|d}|Sd}tj||j dS)z5Convert this KerasTensor to a placeholder in a graph.)r#r.rcVtj|j|jSN)r placeholderr.r#) components rcomponent_to_placeholderz=KerasTensor._to_placeholder..component_to_placeholders  " "9??IOO DDrT)expand_composites) rr r#r?rr/rr0r map_structure)rrrAs r_to_placeholderzKerasTensor._to_placeholders '!   ((  >?n    " "a '(* E    $..D JJrreturnc|jSr>r(r s r get_shapezKerasTensor.get_shapes ::rctd)Na_Keras symbolic inputs/outputs do not implement `__len__`. You may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model. This error will also get raised if you try asserting a symbolic input/output directly. TypeErrorr s r__len__zKerasTensor.__len__s M NNrctd)NaKeras symbolic inputs/outputs do not implement `op`. You may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model.rIr s ropzKerasTensor.ops / 00rctd|z)NzATensors are unhashable. (%s)Instead, use tensor.ref() as the key.rIr s r__hash__zKerasTensor.__hash__s <>BC DDrdctd)Na~Cannot convert a symbolic Keras input/output to a numpy array. This error may indicate that you're trying to pass a symbolic value to a NumPy call, which is not supported. Or, you may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model.rIr s r __array__zKerasTensor.__array__s  # $$rcy)NTr9r s ris_tensor_likezKerasTensor.is_tensor_likes rc&t|tjstj|}|j|jDcgc]}|j}}t t |D];}|| |jj |jj|||<=tj|}|jj|std|jd|||j_ ycc}w)zFUpdates the shape of this KerasTensor. Mimics `tf.Tensor.set_shape()`.Nz$Keras symbolic input/output's shape z& is notcompatible with supplied shape ) rr TensorShaper2valueranger4r#is_compatible_withrrr")rr#dimdim_lists r set_shapezKerasTensor.set_shape s e\55 6&&u-e zz',zz2#))2h2s8}%/# C= TZZ__%@**//#.(3-/&&x0e :: ( ( /  ::u   %doo3s Dc d}d}d}t|dr&|jj}d|jd}|jd|jz}|jd|j z}d|j |||dS) N_keras_historyz!, description="created by layer 'z'"z, inferred_value=%sz , name='%s'zKerasTensor(type_spec=))hasattrr_layerrrrr)rsymbolic_descriptioninferred_value_string name_stringrbs r__str__zKerasTensor.__str__sKt%&!!''e7) rrr* TensorSpecr#r.rrar_rbr)rrcrdtype_spec_stringrbs r__repr__zKerasTensor.__repr__/s$..*"7"7804 DJJOOL'$..8t%&!!''e=BZZI ' !5!5 5 /1E GGrc.|jjS)z@Returns the `dtype` symbolically inferred for this Keras output.)r_dtyper s rr.zKerasTensor.dtype@r$rc,tj|S)aReturns a hashable reference object to this KerasTensor. The primary use case for this API is to put KerasTensors in a set/dictionary. We can't put tensors in a set/dictionary as `tensor.__hash__()` is not available and tensor equality (`==`) is supposed to produce a tensor representing if the two inputs are equal. See the documentation of `tf.Tensor.ref()` for more info. )r Referencer s rrefzKerasTensor.refIs  $ $T **rcd}|jj-|jjDcgc]}|j}}| t d|s t d|d t dt ||dScc}w)Nz0Cannot iterate over a Tensor with unknown shape.zCannot iterate over a scalar.rz:Cannot iterate over a Tensor with unknown first dimension.)r#ndimsr2rWrJ_KerasTensorIterator)rr#rZs r__iter__zKerasTensor.__iter__Us E zz#$(JJOO4Ssyy4e4 } H II  5 66 Qx  F HH eAh //5sA?c|jS)zEReturns the (non-unique, optional) name of this symbolic Keras value.)rr s rrzKerasTensor.namecs ::rctjjD]}|j||t |dr|j|dyy)z%Register overloads for all operators.experimental_refN)r*r+OVERLOADABLE_OPERATORS_overload_operatorra)r6 tensor_classoperators r_overload_all_operatorsz#KerasTensor._overload_all_operatorshsR%%<<5 \845 |/0 \+=>1rcPt||}t|d|}t|||y)aOverload an operator with the same implementation as a base Tensor class. We pull the operator out of the class dynamically to avoid ordering issues. Args: tensor_class: The (Composite)Tensor to get the method from. operator: string. The operator name. __func__N)r,setattr)r6rzr{ tensor_opers rryzKerasTensor._overload_operatorss,,1K +z;?K C;'r)NNr>)__name__ __module__ __qualname____doc__rpropertyrr# classmethodr7r;rDrrVrGrKrMrO__array_priority__rRrTr\rfrkr.rprtrr|ryr9rrrr's"M^   " "''''R//J811N 0 0D $  %$+$G" " " + 0  ??((rrceZdZdZdZy)SparseKerasTensorzA specialized KerasTensor representation for `tf.sparse.SparseTensor`s. Specifically, it specializes the conversion to a placeholder in order to maintain dense shape information. cp|j}tj|j|jS)N)r.r#)rr sparse_placeholderr.r#)rspecs rrDz!SparseKerasTensor._to_placeholders( >>D  ' 'djj KKrN)rrrrrDr9rrrrs  Lrrc2eZdZdZfdZedZxZS)RaggedKerasTensoraA specialized KerasTensor representation for `tf.RaggedTensor`s. Specifically, it: 1. Specializes the conversion to a placeholder in order to maintain shape information for non-ragged dimensions. 2. Overloads the KerasTensor's operators with the RaggedTensor versions when they don't match the `tf.Tensor` versions 3. Exposes some of the instance method/attribute that are unique to the RaggedTensor API (such as ragged_rank). c|j}|jdk(s|jjtt |S|j|jd}tj|j|}g}d}|jD],}|| t|ddd}n||z}|j|.t|jddD]}|j|}| t|ddS||dz }||dz}tj|j|g} tjj!|| d}tt#j$||j} tjj'|| d}|S)NrrWTF)validate)r ragged_rankr#r0superrrDr r?r.r,appendrXrow_splits_dtyper RaggedTensorfrom_row_splitsrconstantfrom_uniform_row_length) r ragged_spec flat_shaperesultknown_num_splitsprod axis_sizeaxis num_splitssplitsrowlen __class__s rrDz!RaggedKerasTensor._to_placeholders..K!#{'8'8'='='E $d ; =="";#:#:#;)rrr s rrzRaggedKerasTensor.ragged_ranks >> % %%r)rrrrrDrr __classcell__rs@rrrs" !F & &rr __getitem____add____radd____mul____rmul__c4eZdZdZdZdZdZdZdZdZ y) UserRegisteredSpecz7TypeSpec to represent user-registered symbolic objects.c.||_||_||_yr>)r#rmr.)rr#r.s rrzUserRegisteredSpec.__init__sDJDKDJrctr>NotImplementedErrorr s r_component_specsz#UserRegisteredSpec._component_specs rctr>r)r componentss r_from_componentsz#UserRegisteredSpec._from_componentsrrctr>rr s r _serializezUserRegisteredSpec._serializerrctr>r)rrWs r_to_componentsz!UserRegisteredSpec._to_componentsrrctr>rr s r value_typezUserRegisteredSpec.value_typerrN) rrrrrrrrrrr9rrrrs#? rrcJeZdZdZfdZedZeddZdZxZ S)UserRegisteredTypeKerasTensorzAKerasTensor that represents legacy register_symbolic_tensor_type.c|}||_t|j|j}t |dd}t t |||y)Nr) _user_registered_symbolic_objectrr#r.r,rrr)ruser_registered_symbolic_objectxrrrs rrz&UserRegisteredTypeKerasTensor.__init__sD'A,-D)"177AGG4I 1fd #D '7 4Hrc||Sr>r9)r6rs rr7z)UserRegisteredTypeKerasTensor.from_tensor s v;rctd|z)Nz?You cannot instantiate a KerasTensor directly from TypeSpec: %srr:s rr;z,UserRegisteredTypeKerasTensor.from_type_spec s ;=FG HHrc|jSr>)rr s rrDz-UserRegisteredTypeKerasTensor._to_placeholders  0 00rr>) rrrrrrr7r;rDrrs@rrrs;IIHH1rrc"eZdZdZdZdZdZy)rszHIterates over the leading dim of a KerasTensor. Performs 0 error checks.c.||_d|_||_y)Nr)_tensor_index_limit)rrdim0s rrz_KerasTensorIterator.__init__sDLDKDKrc|Sr>r9r s rrtz_KerasTensorIterator.__iter__s Krc|j|jk(rt|j|j}|xjdz c_|S)Nr)rr StopIterationr)rrs r__next__z_KerasTensorIterator.__next__!s= {{dkk!  \\$++ &FKK1K MrN)rrrrrrtrr9rrrsrssP rrsc4tjd||fy)z>Register a specialized KerasTensor subclass for a Tensor type.rN)keras_tensor_classesinsert)r6keras_tensor_subclasss r$register_keras_tensor_specializationr9sb3(=">?rcFt|tr|jS|S)zFConstruct a graph placeholder to represent a KerasTensor when tracing.)rrrD)rs rkeras_tensor_to_placeholderr?s ;    Hrcd}tD]\}}t||s|}n|j|}t|drt |j |_|S)r&N _keras_mask)rrr7rakeras_tensor_from_tensorr)rkeras_tensor_cls tensor_typer6outs rrrGsd. k3&+&  $$V,# V]#.v/A/ABCO *rNcd}|j}tD]\}}t||s|}n|j||S)z3Convert a TypeSpec to a representative KerasTensor.Nr))rr issubclassr;)rrrrrr6s rkeras_tensor_from_type_specrXsT##*. k3*k*   ( ( ( >>rr>)&rtensorflow.python.frameworkrrrrr*rrrtensorflow.python.keras.utilsr tensorflow.python.opsr tensorflow.python.ops.raggedr r tensorflow.python.utilr r5objectrr|r+rrryrrrrrs SparseTensorrrrrrr9rrrsd@3.5<4E9+96'](&](@  ##J$5$56L L$2& 2&j$$]%?%?O$$]%?%?K$$]%?%?L$$]%?%?K$$]%?%?L)22<1K1068 $!23!23 [ @   " ?r