BVh dZddlZddlZddlZddlZddlmZddlmZddlmZddlm Z ddlm Z ddl m Z dd l mZdd lmZdd lmZdd lmZdd lmZddlmZddlmZddlmZddlmZddlmZddlmZddlmZddlm Z ddlm!Z!ddlm"Z"ddl#m$Z$ddl#m%Z%ddl#m&Z&ddl#m'Z'ddl#m(Z(ddl#m)Z*ddl+m,Z,ddl-m.Z/dd l0m1Z1e2ejfejhejjejlejnejpgZ9d!Z:Gd"d#e;Z= 1. indices: A one-dimensional integer `Tensor`, indexing into the first dimension of `values` (as in an IndexedSlices object). Returns: A tuple of (`summed_values`, `unique_indices`) where `unique_indices` is a de-duplicated version of `indices` and `summed_values` contains the sum of `values` slices associated with each unique index. r)runiquerunsorted_segment_sumshape)rindicesunique_indicesnew_index_positions summed_valuess a/home/dcms/DCMS/lib/python3.12/site-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py_deduplicate_indexed_slicesr*=sN)2(8(8(A%.%// !oon%a(*-  ((ceZdZdZdZdZy)NullContextmanagercyN)selfargskwargss r)__init__zNullContextmanager.__init__Sr+cyr/r0r1s r) __enter__zNullContextmanager.__enter__Vr5r+cy)NFr0)r1type_arg value_arg traceback_args r)__exit__zNullContextmanager.__exit__Ys r+N)__name__ __module__ __qualname__r4r8r=r0r+r)r-r-Qs  r+r-chtjstj|St S)a Internal-only entry point for `name_scope*`. Enters a compat.v1.name_scope only when in a function or graph, not when running fully eagerly. Args: name: The name argument that is passed to the op function. Returns: `name_scope*` context manager. )r executing_eagerlyr name_scope_v1r-names r)$name_scope_only_in_function_or_graphrF]s*  " " $   T ""  r+c4eZdZdZdZ d>dZedZedZejdZejdZed Z e jd Z d Z d?d Z d Z dZdZd@dZd>dZ dAdZdZdZdZdZd?dZdZdZfdZfdZfdZdZdBdZd Zd!Z d"Z!d#Z"d$Z#ed%Z$e$jd&Z$d'Z%e&jNd(Z(e)d?d)Z*d*Z+d+Z,ed,Z-d-Z.d.Z/ddde0jbjde0jfjhfd/Z5dCd0Z6d1Z7d2Z8d3Z9d4Z:d5Z;d6Zee?jd9ZAee?jd:ZBd;ZCd<ZDeEjd=ZGxZHS)D OptimizerV2aBase class for Keras optimizers. You should not use this class directly, but instead instantiate one of its subclasses such as `tf.keras.optimizers.SGD`, `tf.keras.optimizers.Adam`, etc. ### Usage ```python # Create an optimizer with the desired parameters. opt = tf.keras.optimizers.SGD(learning_rate=0.1) # `loss` is a callable that takes no argument and returns the value # to minimize. loss = lambda: 3 * var1 * var1 + 2 * var2 * var2 # In graph mode, returns op that minimizes the loss by updating the listed # variables. opt_op = opt.minimize(loss, var_list=[var1, var2]) opt_op.run() # In eager mode, simply call minimize to update the list of variables. opt.minimize(loss, var_list=[var1, var2]) ``` ### Usage in custom training loops In Keras models, sometimes variables are created when the model is first called, instead of construction time. Examples include 1) sequential models without input shape pre-defined, or 2) subclassed models. Pass var_list as callable in these cases. Example: ```python opt = tf.keras.optimizers.SGD(learning_rate=0.1) model = tf.keras.Sequential() model.add(tf.keras.layers.Dense(num_hidden, activation='relu')) model.add(tf.keras.layers.Dense(num_classes, activation='sigmoid')) loss_fn = lambda: tf.keras.losses.mse(model(input), output) var_list_fn = lambda: model.trainable_weights for input, output in data: opt.minimize(loss_fn, var_list_fn) ``` ### Processing gradients before applying them Calling `minimize()` takes care of both computing the gradients and applying them to the variables. If you want to process the gradients before applying them you can instead use the optimizer in three steps: 1. Compute the gradients with `tf.GradientTape`. 2. Process the gradients as you wish. 3. Apply the processed gradients with `apply_gradients()`. Example: ```python # Create an optimizer. opt = tf.keras.optimizers.SGD(learning_rate=0.1) # Compute the gradients for a list of variables. with tf.GradientTape() as tape: loss = vars = grads = tape.gradient(loss, vars) # Process the gradients, for example cap them, etc. # capped_grads = [MyCapper(g) for g in grads] processed_grads = [process_gradient(g) for g in grads] # Ask the optimizer to apply the processed gradients. opt.apply_gradients(zip(processed_grads, var_list)) ``` ### Use with `tf.distribute.Strategy` This optimizer class is `tf.distribute.Strategy` aware, which means it automatically sums gradients across all replicas. To average gradients, you divide your loss by the global batch size, which is done automatically if you use `tf.keras` built-in training or evaluation loops. See the `reduction` argument of your loss which should be set to `tf.keras.losses.Reduction.SUM_OVER_BATCH_SIZE` for averaging or `tf.keras.losses.Reduction.SUM` for not. To aggregate gradients yourself, call `apply_gradients` with `experimental_aggregate_gradients` set to False. This is useful if you need to process aggregated gradients. If you are not using these and you want to average gradients, you should use `tf.math.reduce_sum` to add up your per-example losses and then divide by the global batch size. Note that when using `tf.distribute.Strategy`, the first component of a tensor's shape is the *replica-local* batch size, which is off by a factor equal to the number of replicas being used to compute a single step. As a result, using `tf.math.reduce_mean` will give the wrong answer, resulting in gradients that can be many times too big. ### Variable Constraints All Keras optimizers respect variable constraints. If constraint function is passed to any variable, the constraint will be applied to the variable after the gradient has been applied to the variable. Important: If gradient is sparse tensor, variable constraint is not supported. ### Thread Compatibility The entire optimizer is currently thread compatible, not thread-safe. The user needs to perform synchronization if necessary. ### Slots Many optimizer subclasses, such as `Adam` and `Adagrad` allocate and manage additional variables associated with the variables to train. These are called Slots. Slots have names and you can ask the optimizer for the names of the slots that it uses. Once you have a slot name you can ask the optimizer for the variable it created to hold the slot value. This can be useful if you want to log debug a training algorithm, report stats about the slots, etc. ### Hyperparameters These are arguments passed to the optimizer subclass constructor (the `__init__` method), and then passed to `self._set_hyper()`. They can be either regular Python values (like 1.0), tensors, or callables. If they are callable, the callable will be called during `apply_gradients()` to get the value for the hyper parameter. Hyperparameters can be overwritten through user code: Example: ```python # Create an optimizer with the desired parameters. opt = tf.keras.optimizers.SGD(learning_rate=0.1) # `loss` is a callable that takes no argument and returns the value # to minimize. loss = lambda: 3 * var1 + 2 * var2 # In eager mode, simply call minimize to update the list of variables. opt.minimize(loss, var_list=[var1, var2]) # update learning rate opt.learning_rate = 0.05 opt.minimize(loss, var_list=[var1, var2]) ``` ### Callable learning rate Optimizer accepts a callable learning rate in two ways. The first way is through built-in or customized `tf.keras.optimizers.schedules.LearningRateSchedule`. The schedule will be called on each iteration with `schedule(iteration)`, a `tf.Variable` owned by the optimizer. Example: >>> var = tf.Variable(np.random.random(size=(1,))) >>> learning_rate = tf.keras.optimizers.schedules.ExponentialDecay( ... initial_learning_rate=.01, decay_steps=20, decay_rate=.1) >>> opt = tf.keras.optimizers.SGD(learning_rate=learning_rate) >>> loss = lambda: 3 * var >>> opt.minimize(loss, var_list=[var]) >> var = tf.Variable(np.random.random(size=(1,))) >>> def lr_callable(): ... return .1 >>> opt = tf.keras.optimizers.SGD(learning_rate=lr_callable) >>> loss = lambda: 3 * var >>> opt.minimize(loss, var_list=[var]) lrdecayclipnorm clipvalueglobal_clipnormz1Unexpected keyword argument passed to optimizer: NrzExpected {} >= 0, received: {}rJz=The `lr` argument is deprecated, use `learning_rate` instead.TrKzdecay cannot be less than 0: {}FrLrNz_Cannot accept both `clipnorm` and `global_clipnorm`, passed `clipnorm` {}, `global_clipnorm` {}rM) TypeErrorstr ValueErrorformatwarningswarn _use_locking_init_set_name_hyper_slots _slot_names_weights _iterations_deferred_slot_restorationspop_initial_decay_hypers_createdr has_strategy get_strategy_distribution_strategyoptimizer_utilsall_reduce_sum_gradientsgradient_aggregatorgradient_transformersrLrNrM)r1rErfrgr3allowed_kwargskrKs r)r4zOptimizerV2.__init__5sbQN  M . 025a&9: :  6!9q=9@@F1INOO d K M MDDKDKDDMD(*D$ JJw $E rz 8??F GGD D""$$2$?$?$Ad!$(d!"+DD2D$ !6DJJz40DM!::&7>D }} T%9%9%E DDJF==$*>*>E@ AAZZ T2DNr+c|jSz=`float` or `None`. If set, clips gradients to a maximum norm.) _clipnormr7s r)rLzOptimizerV2.clipnorms >>r+c|jSrk)_global_clipnormr7s r)rNzOptimizerV2.global_clipnorms   r+c||jr td||_tj|j|_yNz`clipnorm` cannot be set when `gradient_transformers` is set. Instead, use the `gradient_transformers` to specify clipping and other transformations.)rgrRrlrdmake_gradient_clipnorm_fn _clipnorm_fnr1vals r)rLzOptimizerV2.clipnormsG 455 E FFDN'AA Dr+c||jr td||_tj|j|_yrp)rgrRrnrd make_global_gradient_clipnorm_fn_global_clipnorm_fnrss r)rNzOptimizerV2.global_clipnormsJ 455 E FF D.OO  Dr+c|jS)z>`float` or `None`. If set, clips gradients to a maximum value.) _clipvaluer7s r)rMzOptimizerV2.clipvalues ??r+c||jr td||_tj|j|_y)Nz`clipvalue` cannot be set when `gradient_transformers` is set. Instead, use the `gradient_transformers` to specify clipping and other transformations.)rgrRryrdmake_gradient_clipvalue_fn _clipvalue_fnrss r)rMzOptimizerV2.clipvaluesG 455 E FFDO(CC Dr+c|S)zCCalled in `.minimize` to transform loss before computing gradients.r0)r1losss r)_transform_losszOptimizerV2._transform_losss Kr+cR|j|||}tt||S)z4Called in `minimize` to compute gradients from loss.)gradientlistzip)r1taper~var_list grad_lossgradss r)_get_gradientszOptimizerV2._get_gradientss& MM$) 4E E8$ %%r+c|S)z8Called in `apply_gradients` before gradient aggregation.r0r1grads_and_varss r)!_transform_unaggregated_gradientsz-OptimizerV2._transform_unaggregated_gradientss r+c$|j|S)aiCalled in `apply_gradients` to aggregate gradients across devices. Note that user subclasses may override this, so the interface should not be changed. Args: grads_and_vars: List of (gradient, variable) pairs. Returns: A list of (aggregrated_gradient, variable) pairs. By default, this calls `self.gradient_aggregator`. )rfrs r)_aggregate_gradientsz OptimizerV2._aggregate_gradientss  # #N 33r+c|j|j|}|j|j|}|j|j |}|j D] }||} |S)z.Called in `apply_gradients` after aggregation.)ryr|rlrrrnrwrg)r1rfns r)_transform_gradientsz OptimizerV2._transform_gradientssw ")).9n ~~!((8n (//?n((*.)n* r+cR|j||||}|j||S)a9Minimize `loss` by updating `var_list`. This method simply computes gradient using `tf.GradientTape` and calls `apply_gradients()`. If you want to process the gradient before applying then call `tf.GradientTape` and `apply_gradients()` explicitly instead of using this function. Args: loss: `Tensor` or callable. If a callable, `loss` should take no arguments and return the value to minimize. If a `Tensor`, the `tape` argument must be passed. var_list: list or tuple of `Variable` objects to update to minimize `loss`, or a callable returning the list or tuple of `Variable` objects. Use callable when the variable list would otherwise be incomplete before `minimize` since the variables are created at the first time `loss` is called. grad_loss: (Optional). A `Tensor` holding the gradient computed for `loss`. name: (Optional) str. Name for the returned operation. tape: (Optional) `tf.GradientTape`. If `loss` is provided as a `Tensor`, the tape that computed the `loss` must be provided. Returns: An `Operation` that updates the variables in `var_list`. The `iterations` will be automatically increased by 1. Raises: ValueError: If some of the variables are not `Variable` objects. )rrrrD)_compute_gradientsapply_gradients)r1r~rrrErrs r)minimizezOptimizerV2.minimizes9>,, x94-AN   T  ::r+ct|s | td||ntj}t|r@|5t|s|j ||}t|r|}ddd|5|j |}dddt j|}tj|jdz5|j||||}ddd|jDcgc]&\}}||jtjk7r|(c}}|S#1swYxYw#1swYxYw#1swYfxYwcc}}w)a[Compute gradients of `loss` for the variables in `var_list`. This is the first part of `minimize()`. It returns a list of (gradient, variable) pairs where "gradient" is the gradient for "variable". Note that "gradient" can be a `Tensor`, an `IndexedSlices`, or `None` if there is no gradient for the given variable. Args: loss: `Tensor` or callable. If a callable, `loss` should take no arguments and return the value to minimize. If a `Tensor`, the `tape` argument must be passed. var_list: list or tuple of `Variable` objects to update to minimize `loss`, or a callable returning the list or tuple of `Variable` objects. Use callable when the variable list would otherwise be incomplete before `minimize` and the variables are created at the first time when `loss` is called. grad_loss: Optional. A `Tensor` holding the gradient computed for `loss`. tape: (Optional) `tf.GradientTape`. If `loss` is provided as a `Tensor`, the tape that computed the `loss` must be provided. Returns: A list of (gradient, variable) pairs. Variable is always present, but gradient can be `None`. Raises: TypeError: If `var_list` contains anything else than `Variable` objects. ValueError: If some arguments are invalid, or var_list is None. Nz2`tape` is required when a `Tensor` loss is passed. /gradients)callablerRr GradientTapewatchrr flattenr name_scope_v2_namer_assert_valid_dtypesdtyper resource)r1r~rrrrgvs r)rzOptimizerV2._compute_gradientssC> D>dl K LL#4)>)>)@D~  ! **X v H Z(   (  ! !$ 'd(||H%H  4:: 4 5L**4xKnL $a =QWW7  '  ((LLs)6D1>D=E <+E 1D:=E Ectj|}|Dcgc]\}}| }}}tj|j5tj 5|j |ddd|stjcdddStjr tdtj}|sU|rSt|tjt j"t$j&t$j(fr t+d|j-|}|r"|j/|}|j1|}|j3|}tj4r|j7||||cdddStj8j;t=j>|j6||fd|icdddScc}}w#1swYxYw#1swYyxYw)aLApply gradients to variables. This is the second part of `minimize()`. It returns an `Operation` that applies gradients. The method sums gradients from all replicas in the presence of `tf.distribute.Strategy` by default. You can aggregate gradients yourself by passing `experimental_aggregate_gradients=False`. Example: ```python grads = tape.gradient(loss, vars) grads = tf.distribute.get_replica_context().all_reduce('sum', grads) # Processing aggregated gradients. optimizer.apply_gradients(zip(grads, vars), experimental_aggregate_gradients=False) ``` Args: grads_and_vars: List of (gradient, variable) pairs. name: Optional name for the returned operation. Default to the name passed to the `Optimizer` constructor. experimental_aggregate_gradients: Whether to sum gradients from different replicas in the presense of `tf.distribute.Strategy`. If False, it's user responsibility to aggregate the gradients. Default to True. Returns: An `Operation` that applies the specified gradients. The `iterations` will be automatically increased by 1. Raises: TypeError: If `grads_and_vars` is malformed. ValueError: If none of the variables have gradients. RuntimeError: If called in a cross-replica context. Nzx`apply_gradients() cannot be called in cross-replica context. Use `tf.distribute.Strategy.run` to enter replica context.zo`experimental_aggregate_gradients=False is not supported for ParameterServerStrategy and CentralStorageStrategy) apply_staterE)r2r3) rdfilter_empty_gradientsr rr init_scope_create_all_weightsrno_oprin_cross_replica_context RuntimeErrorrb isinstancerParameterServerStrategyV1rParameterServerStrategyV2rCentralStorageStrategyCentralStorageStrategyV1NotImplementedError_preparerrrstrategy_supports_no_merge_call_distributed_applyget_replica_context merge_call functoolspartial) r1rrE experimental_aggregate_gradients_rrstrategyrs r)rzOptimizerV2.apply_gradientsQsR%;;NKN ./fq!/H/  4:: &* >> +   *+ %%'**  0 0 2    ,,.h.8 X/II2LL.EE.GGI J " AB BMM(+k )??O22>B00@n  8 8 :&&x'24E**J113>>   d55; O "?K**0++**s7 G) G<G/0G<C7G<AG</G9 4G<<Hc  fd}tj}g}t|xs j5|D]\}} |jj | 5t|rdnd| j jz5|jj| ||fd} tjr|j| n|j| ddddddtd|D} tjr| rt!j"|j%5tj&t)j*|g5j,j/dd cdddcdddcdddSj,j/dcdddS#1swYxYw#1swYxYw#1swYnxYw dddS#1swY\xYw#1swYyxYw) z1`apply_gradients` using a `DistributionStrategy`.c0t|tjr td|i}t|tj rS|j tddjvr|d<j|j||jfi|Sdjvr|d<j||fi|}|j @tj|g5|j!|j |cdddS|S#1swYyxYw)zApply gradient to variable.zTrying to update a Tensor Nz6Cannot use a constraint function on a sparse variable.r)rr Tensorrr IndexedSlices constraintr_sparse_apply_args(_resource_apply_sparse_duplicate_indicesrr%_dense_apply_args_resource_apply_denser control_dependenciesassign)vargrad apply_kwargs update_oprr1s r)apply_grad_to_update_varz@OptimizerV2._distributed_apply..apply_grad_to_update_vars C '!">DDl D.66 7 >> %FH H D33 3(3,} %z1OptimizerV2._distributed_apply..s=M<=$As}}58!44Q78Ms;=) read_value)r #executing_eagerly_outside_functionsrFrextendedcolocate_vars_withoprErrrextendappendanyr rBr_current_graph as_defaultrrrr\ assign_add) r1 distributionrrErreagerly_outside_functions update_opsrrr any_symbolics ` ` r)rzOptimizerV2._distributed_applys0!$ G G IJ -d.@djj A,%+)$ " " 5 5c : +33hffkk: +%--44-TG55JI668 * * + + ++$MAKMMl  & & (L # #J / : : < D'')9)?)? )K(LM D##..qU.C D D D D3,,:   ( ( +;,, + + + +, D D D D D D3,,sy$G:'G AF? G #AG:8*G."G? G. G:G:?G G  G G:G! G.% G:.G7 3G::Hctj|}tjj 5tj |j dz5tj||}t||D]!\}}| tdj| ddddddS#1swYxYw#1swYSxYw)a[Returns gradients of `loss` with respect to `params`. Should be used only in legacy v1 graph mode. Args: loss: Loss tensor. params: List of variables. Returns: List of gradient tensors. Raises: ValueError: In case any gradient cannot be computed (e.g. if gradient function not implemented). rNzVariable {} has `None` for gradient. Please make sure that all of your ops have a gradient defined (i.e. are differentiable). Common ops without gradient: K.argmax, K.round, K.eval.) r rr get_graphr name_scoperrrrRrS)r1r~paramsrrparams r) get_gradientszOptimizerV2.get_gradientss \\& !F     ' ' ) G7+=+=djj>J?K,L G!!$/eUF+G+$ <89?u G GG G G L G G G Ls)#C+B5B5#C5B> :CC c|j||}tt||}|j|Dcgc]&\}}||jt j k7r|(c}}|j|gScc}}wr/)rrrrrr rr)r1r~rrrrrs r) get_updateszOptimizerV2.get_updatess|   tV ,E#eV,-N$a =QWW7    0 11 s+A= ct|tjr|j||d||jvr||j|<y|j|}t |sNt|t jtttjfst|tjr||j|<ytj|j||y)zBset hyper `name` to value. value can be callable, tensor, numeric.T) overwriteN)r trackable Trackable_track_trackablerXrr rintfloatrLearningRateScheduler set_value)r1rEvalue prev_values r) _set_hyperzOptimizerV2._set_hyper s%,,- E448 4;;dkk$;;t$j :   U/DDFG5JJ K! D$++d+U3r+c|js|j|j|}t|tj r|St |r|}|rtj||S|Sr/) r`_create_hypersrXrrrrrcast)r1rErrs r) _get_hyperzOptimizerV2._get_hypersb     KK E%/DDE lge ]]5% (( lr+cyr/r0)r1rs r) _create_slotszOptimizerV2._create_slots)r5r+c^|j}|j|j|y)aCreates all weights, including iterations, hyperparameters and slot vars. This will add newly created variables to `optimizer.weights`. New variables are only created when this method is called the first time, or when called with different variables in the var_list. Args: var_list: list or tuple of `Variable` objects that will be minimized using this optimizer. N) iterationsrr)r1rrs r)rzOptimizerV2._create_all_weights,s' Ax r+c tt| |S#t$r9}|dk(r||dk(rd}||jvr|j |cYd}~S|d}~wwxYw)z,Overridden to support hyperparameter access.rXrJ learning_rateN)superrH__getattribute__AttributeErrorrXr)r1rEe __class__s r)rzOptimizerV2.__getattribute__=sc  ; 6t <<       t$$ g s  A,A AAActtt|}d|vrJ||jj z}d|jj vr|j dt|S)NrXrrJ)setrrH__dir__rXkeysaddr)r1resultrs r)r zOptimizerV2.__dir__Lsa {D13 4F6    ""f DKK,,. . 4 <r+c|dk(rd}t|dr!||jvr|j||ytt|||y)z;Override setattr to support dynamic hyperparameter setting.rJrrXN)hasattrrXrrrH __setattr__)r1rErrs r)rzOptimizerV2.__setattr__TsE t| dtX44;;#6 oodE" K*47r+c|jS)z+A list of names for this optimizer's slots.)rZr7s r)get_slot_nameszOptimizerV2.get_slot_names^s   r+zerosc||jvr|jj|t|}|jj |i}|j |d}|t |ts t|rctj |}t |tjs||}n |j}tj|||j} n|} |j!5t#j$} | j&j)|st+dj-| || j&j/|5t1j2|j4d||jd| }ddddddt7j8||||<|j;||||j<j||S#1swY\xYw#1swY`xYw)aAdd a new slot variable for `var`. A slot variable is an additional variable associated with `var` to train. It is allocated and managed by optimizers, e.g. `Adam`. Args: var: a `Variable` object. slot_name: name of the slot variable. initializer: initializer of the slot variable shape: (Optional) shape of the slot variable. If not set, it will default to the shape of `var`. Returns: A slot variable. N)r$ra0Trying to create optimizer slot variable under the scope for tf.distribute.Strategy ({}), which is different from the scope used for the original variable ({}). Make sure the slot variables are created under the same strategy scope. This may happen if you're restoring from a checkpoint outside the scope/F)rEr trainable initial_value) slot_namevariable slot_variable)rZr_var_keyrY setdefaultgetrrQrrrCheckpointInitialValueCallabler$rrr_distribution_strategy_scoperrbrvariable_created_in_scoperRrSr tf_variablesVariable _shared_namertrack_variable_restore_slot_variabler[) r1rr initializerr$var_key slot_dictweight slot_shaperrs r)add_slotzOptimizerV2.add_slotbs ((( i(smG &&w3I ]]9d +F ~ K %+)>"&&{3    4 4 6:?:K*yy*!)) z< $  , , .+!..0  ::3?O vh$ & &   1 1# 6 +((!.. :II) +& ++"V$#i  !!"  mm6" M + +++s%6A&G63G*G6*G3 /G66G?c@t|}|j|}||Sr/)rrY)r1rrr&r's r)get_slotzOptimizerV2.get_slots$smG G$I Y r+ct}|D]f}t|tjr |j}n |j g}|j j}|D]}|j||fhi}|D]=\}}i|||f<tj |5|j|||ddd?|S#1swYKxYwr/) rr ds_valuesDistributedValues_devicesdevicer base_dtyper r _prepare_local)r1rr r var_devices var_dtype var_devicers r)rzOptimizerV2._prepares 5D* C44 5ll zzl ))&&i#** *i()* *K!%@ I-/k:y)* ::j !@ J ;?@@@ @@s B<<C c~d|jvr/tj|j|}||||fd<yy)Nrlr_t)rXridentity _decayed_lr)r1r6r5rr8s r)r3zOptimizerV2._prepare_localsB$++%    0 0 ; Fuo "oo&::MM .O$++d  OO  D!O  Ds BC  CcH|j|j5|jdgtjdt j j|_ddd|jj|j|jS#1swY:xYw)z>Variable. The number of training steps this Optimizer has run.NiterF)r$rrr>) r\rrAr int64r rBrCr[rr7s r)rzOptimizerV2.iterationss   , , .M?? ,,$88KK +MM mm4++,   MMs ABB!c|j td||_|jj|jy)NzWCannot set `iterations` to a new Variable after the Optimizer weights have been created)r\rr[r)r1rs r)rzOptimizerV2.iterationssC # C DDDMM))*r+c|jd|}t|tjr>> opt = tf.keras.optimizers.RMSprop() >>> m = tf.keras.models.Sequential([tf.keras.layers.Dense(10)]) >>> m.compile(opt, loss='mse') >>> data = np.arange(100).reshape(5, 20) >>> labels = np.zeros(5) >>> results = m.fit(data, labels) # Training. >>> len(opt.get_weights()) 3 Returns: Weights values as a list of numpy arrays. )r^rbatch_get_value)r1rs r) get_weightszOptimizerV2.get_weights:s4\\F  " "6 **r+cZ|j}t|t|k7r]td|jzdzt t|zdzt t|zdzt |ddzdz|syg}t j |}t|||D]l\}}}|j|jk7r:tdt |jzd zt |jz|j||fnt j|y) aSet the weights of the optimizer. The weights of an optimizer are its state (ie, variables). This function takes the weight values associated with this optimizer as a list of Numpy arrays. The first value is always the iterations count of the optimizer, followed by the optimizer's state variables in the order they are created. The passed values are used to set the new state of the optimizer. For example, the RMSprop optimizer for this simple model takes a list of three values-- the iteration count, followed by the root-mean-square value of the kernel and bias of the single Dense layer: >>> opt = tf.keras.optimizers.RMSprop() >>> m = tf.keras.models.Sequential([tf.keras.layers.Dense(10)]) >>> m.compile(opt, loss='mse') >>> data = np.arange(100).reshape(5, 20) >>> labels = np.zeros(5) >>> results = m.fit(data, labels) # Training. >>> new_weights = [np.array(10), np.ones([20, 10]), np.zeros([10])] >>> opt.set_weights(new_weights) >>> opt.iterations Args: weights: weight values as a list of numpy arrays. z/You called `set_weights(weights)` on optimizer z with a weight list of length z", but the optimizer was expecting z weights. Provided weights: N2z...zOptimizer weight shape z+ not compatible with provided weight shape ) r^lenrRrrQrr`rr$rbatch_set_value)r1r^rweight_value_tuples param_valuespvpws r) set_weightszOptimizerV2.set_weightsXsC8\\F 6{c'l"  ;djj H + ,.1#g,.? @ . /14S[1A B ) ),/w<+< =@E E FF  **62L fg6)Aq QWW 2S]B2247LAB B  !Q( )  /0r+c h|tj}t|ts t |rt j |}|tjjk(r|r tdd}n|d}|j||tjd|||d|| }tj||S)NzSynchronization value can be set to VariableSynchronization.ON_READ only for non-trainable variables. You have specified trainable=True and synchronization=VariableSynchronization.ON_READ.FT) rEr$getterrr%rr use_resourcesynchronizationr>)r float32rrQrrrr VariableSynchronizationON_READrR _add_variable_with_custom_getterr make_variablerr#) r1rEr$rr%rror>rs r)rAzOptimizerV2.add_weights }nne+s#x '< $$[1k,>>FFF  ?@ @  i44 --'5 !H 8$ Or+c|sDtjtj|jj ||_y||_y)N) zero_based)runique_object_namer to_snake_caserr>r)r1rErvs r)rWzOptimizerV2._init_set_names; --  % %dnn&=&= >!djdjr+c |j}|D]I}|jj}||vstd|d|jd|Dcgc]}|c}dycc}w)zAsserts tensors are all valid types (see `_valid_dtypes`). Args: tensors: Tensors to check. Raises: ValueError: If any tensor is not a valid type. z Invalid type z for z , expected: .N) _valid_dtypesrr2rRrE)r1tensors valid_dtypestrrs r)rz OptimizerV2._assert_valid_dtypessk%%'L Egg  e l "\)B!)BDE EE*Cs A ctS)zValid types for loss, variables and gradients. Subclasses should override to allow other float types. Returns: Valid types for loss, variables and gradients. )_DEFAULT_VALID_DTYPESr7s r)r{zOptimizerV2._valid_dtypess  ! r+c*t|r|S|S)z'Call the function if param is callable.)r)r1rs r)_call_if_callablezOptimizerV2._call_if_callablesuo57050r+ctd)anAdd ops to apply dense gradients to the variable `handle`. Args: grad: a `Tensor` representing the gradient. handle: a `Tensor` of dtype `resource` which points to the variable to be updated. apply_state: A dict which is used across multiple apply calls. Returns: An `Operation` which updates the value of the variable. "Must be implemented in subclasses.r)r1rhandlers r)rz!OptimizerV2._resource_apply_denses B CCr+c Lt||\}}|j|||fi|S)aAdd ops to apply sparse gradients to `handle`, with repeated indices. Optimizers which override this method must deal with repeated indices. See the docstring of `_apply_sparse_duplicate_indices` for details. By default the correct behavior, to sum non-unique indices and their associated gradients, is enforced by first pre-processing `grad` and `indices` and passing them on to `_resource_apply_sparse`. Optimizers which deal correctly with duplicate indices may instead override this method to avoid the overhead of summing. Args: grad: a `Tensor` representing the gradient for the affected indices. handle: a `Tensor` of dtype `resource` which points to the variable to be updated. indices: a `Tensor` of integral type representing the indices for which the gradient is nonzero. Indices may be repeated. **kwargs: May optionally contain `apply_state` Returns: An `Operation` which updates the value of the variable. )rr%)r*_resource_apply_sparse)r1rrr%r3 summed_gradr&s r)rz4OptimizerV2._resource_apply_sparse_duplicate_indicess;.#>W#&K &4 & &{FN 1)/ 11r+ctd)aAdd ops to apply sparse gradients to the variable `handle`. Similar to `_apply_sparse`, the `indices` argument to this method has been de-duplicated. Optimizers which deal correctly with non-unique indices may instead override `_resource_apply_sparse_duplicate_indices` to avoid this overhead. Args: grad: a `Tensor` representing the gradient for the affected indices. handle: a `Tensor` of dtype `resource` which points to the variable to be updated. indices: a `Tensor` of integral type representing the indices for which the gradient is nonzero. Indices are unique. apply_state: A dict which is used across multiple apply calls. Returns: An `Operation` which updates the value of the variable. rr)r1rrr%rs r)rz"OptimizerV2._resource_apply_sparses& B CCr+ctjtj|j||g5|j cdddS#1swYyxYwN)rr%updates)r rrResourceScatterAddrrr1xrrs r)_resource_scatter_addz!OptimizerV2._resource_scatter_addsN ! !!44XXq! 5# WWY  AActjtj|j||g5|j cdddS#1swYyxYwr)r rrResourceScatterUpdaterrrs r)_resource_scatter_updatez$OptimizerV2._resource_scatter_updatesN ! ! " 8 8XXq! 5 6 7WWYrcTtj|jjSr/)rgetfullargspecrr2r7s r)rzOptimizerV2._dense_apply_args#s!  $ $T%?%? @ E EEr+cTtj|jjSr/)rrrr2r7s r)rzOptimizerV2._sparse_apply_args(s!  $ $T%@%@ A F FFr+ct|}|jj|ij|g}|j dd|D]}|j |y)z.Restore a newly created slot variable's value.c|jSr/) restore_uid)positions r)z4OptimizerV2._restore_slot_variable..8s H4H4Hr+T)keyreverseN)rr]rr^sortrestore)r1rrr variable_keydeferred_restorationscheckpoint_positions r)r$z"OptimizerV2._restore_slot_variable1soH%L <<@@2s<,#H'+-41!!-01r+c&t|}|jj|i}|j|d}|tjrs|j rct jjr |jr9tj|}|j||||j}||j|y|jj!|ij!|gj#|y)aRestore a slot variable's value, possibly creating it. Called when a variable which has an associated slot variable is created or restored. When executing eagerly, we create the slot variable with a restoring initializer. No new variables are created when graph building. Instead, _restore_slot_variable catches these after normal creation and adds restore ops to the graph. This method is nonetheless important when graph building for the case when a slot variable has already been created but `variable` has just been added to a dependency graph (causing us to realize that the slot variable needs to be restored). Args: slot_variable_position: A `trackable._CheckpointPosition` object indicating the slot variable `Trackable` object to be restored. slot_name: The name of this `Optimizer`'s slot to restore into. variable: The variable object this slot is being created for. N)r)rr%rr$)rrYrr rBis_simple_variabler get_default_graph_variable_creator_stackrcrrr* value_shaperr]rr)r1slot_variable_positionrrrr'rr%s r) _create_or_restore_slot_variablez,OptimizerV2._create_or_restore_slot_variable=s*H%L  b1IMM)T2M'";";"=113&&(@@  ( (<<46kmm!&224 $6m $$]3  &&11 R#L"5ff$7&r+c#K|jrTtjs@|jj5|jjdddydy#1swYyxYww)zFReturns the `tf.distribute.Strategy` this optimizer was created under.N)rcrrascoper7s r)rz(OptimizerV2._distribution_strategy_scopesa "">+F+F+H  & & , , .2))//1122 22s;A4A(A4(A1-A4)NNr/)NNN)NT)rN)T)Ir>r?r@__doc___HAS_AGGREGATE_GRADr4propertyrLrNsetterrMrrrrrrrrrrrrrrrrr rrr*r,rr3r<rrr:abcabstractmethodrM classmethodrSrYrr^rarkr rqAUTOrBNONErArWrr{rrrrrrrcached_per_instancerrr$r contextlibcontextmanagerr __classcell__rs@r)rHrHoszF$(%)h3T   ! ! ??   &  4 !;F7v 7;Vp:,x:24"  !" 8:x &: 0  *    ++ (0   +<-1d$!-!E!E!J!J)==BB(TE !1 D18D*  ""F# F ""G# G 1B&H  r+rHct|dr|j}|jr |jS|jS)aPKey for representing a primary variable, for looking up slots. In graph mode the name is derived from the var shared name. In eager mode the name is derived from the var unique id. If distribution strategy exists, get the primary variable first. Args: var: the variable. Returns: the unique name of the variable. _distributed_container)rr_in_graph_moder" _unique_id)rs r)rrs<  S*+ $ $ &C    r+c(t|}|dz|zS)z6Get the slot key for the variable: var_name/slot_name.r)r)rrrEs r)_get_slot_key_from_varrs #$ i r+c(eZdZdZfdZdZxZS)RestoredOptimizeraqA non-functional Optimizer implementation for checkpoint compatibility. Holds slot variables and hyperparameters when an optimizer is restored from a SavedModel. These variables may be referenced in functions along with ops created by the original optimizer, but currently we do not support using the optimizer object iself (e.g. through `apply_gradients`). c:tt| dd|_y)NrT)rrr4r`)r1rs r)r4zRestoredOptimizer.__init__s T+,?@Dr+ctd)NzRestoring functional Optimizers from SavedModels is not currently supported. Please file a feature request if this limitation bothers you.rr7s r)rMzRestoredOptimizer.get_configs   r+)r>r?r@rr4rMrrs@r)rrs r+rtf_deprecated_optimizerc"t|tSr/)rrH)objs r)rrs 3 ,r+ctSr/)r)protos r)rrs %6%8r+r)object_factoryversionmin_producer_versionmin_consumer_versionr)versions)FrrrrrTtensorflow.python.distributerrrrrr.tensorflow.python.eagerrr tensorflow.python.frameworkr r r r rtensorflow.python.kerasrrtensorflow.python.keras.enginer$tensorflow.python.keras.optimizer_v2rrrdtensorflow.python.keras.utilsrrrrtensorflow.python.opsrrrrrrr tensorflow.python.saved_modelrtensorflow.python.trackablerrtensorflow.python.utilr frozensetfloat16bfloat16rpfloat64 complex64 complex128rr*objectr-rFrrHrrrregister_revived_typeVersionedTypeRegistrationrr0r+r)rs1$ A7BE<,+.6+.3+0;GI7542+2;+*;79'" NNFOOV^^V^^ f''# )(   $Y )%%Y x(.  ,$ ##,5m558 ++  r+