a .Sicq@sdZddlZddlZddlmZddlmZddlmZddlmZ ddl m Z ddl m Z dd l mZdd l mZdd l mZdd lmZdd lmZddlmZddlmZddlmZddlmZddlmZddlmZddlmZddlmZddlm Z ddl!m"Z"ddZ#ddZ$ddZ%e&ej'Gddde(Z)Gd d!d!e)Z*Gd"d#d#e)Z+Gd$d%d%e)Z,Gd&d'd'e)Z-d(d)Z.e"d*gd+Gd,d-d-ej/Z0dS).zBase class for optimizers.N)distribute_lib)distribute_utils)distribution_strategy_context) reduce_util)backprop)context)dtypes)indexed_slices)ops) array_ops)control_flow_ops) gradients)math_ops)resource_variable_ops) state_ops)variable_scope) variables)base) slot_creator)nest) tf_exportcsfdd}|S)Ncsdd|i|DS)NcSs g|]\}}|dur||fqSN.0gvrr`/var/www/html/django/DPS/env/lib/python3.9/site-packages/tensorflow/python/training/optimizer.py 9zBget_filtered_grad_fn..filtered_grad_fn..rargskwargsgrad_fnrrfiltered_grad_fn8sz.get_filtered_grad_fn..filtered_grad_fnr)r$r%rr#rget_filtered_grad_fn.s r&cCs.t|\}}t||t|d}||fS)aSums `values` associated with any non-unique `indices`. Args: values: A `Tensor` with rank >= 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)r uniquerunsorted_segment_sumshape)valuesindicesunique_indicesZnew_index_positions summed_valuesrrr_deduplicate_indexed_slices>s  r.cCsPt|dr|}t|r0ts0|j|jfSt|drJ|jj|jj fS|j S)zReturns slot key for `var`._distributed_containerop) hasattrr/ris_distributed_variabler #executing_eagerly_outside_functionsgraph _shared_namer0name _unique_id)varrrr_var_keyQs    r9c@s,eZdZdZejddZejddZdS)_OptimizableVariablez;Interface for abstracting over variables in the optimizers.cCs tddS)z2Returns the optimization target for this variable.Calling an abstract method.NNotImplementedErrorselfrrrtargetcsz_OptimizableVariable.targetcCs tddS)z1Returns the update ops for updating the variable.r;Nr<r? optimizerrrrr update_ophsz_OptimizableVariable.update_opN)__name__ __module__ __qualname____doc__abcabstractmethodr@rCrrrrr:_s  r:c@s0eZdZdZddZddZddZdd Zd S) _RefVariableProcessorzProcessor for Variable.cCs ||_dSr_vr?rrrr__init__qsz_RefVariableProcessor.__init__cCs d|jS)Nz<_RefVariableProcessor(%s)>rKr>rrr__str__tsz_RefVariableProcessor.__str__cCs |jSr)rL_refr>rrrr@wsz_RefVariableProcessor.targetcCst|tjrp|||j}|jjdurjt|g&|j|j|jWdS1s^0Yq|Snrrrr@sz*_DenseReadResourceVariableProcessor.targetcCsn|||jjjd}|jjdurft|g&|j|j|jWdS1sZ0Yn|SdSNr)_resource_apply_denserLr0inputsrUr rVrWr[rrrrCs  6z-_DenseReadResourceVariableProcessor.update_opNrDrErFrGrNr@rCrrrrr\sr\c@s(eZdZdZddZddZddZdS) _DenseResourceVariableProcessorr]cCs ||_dSrrKrMrrrrNsz(_DenseResourceVariableProcessor.__init__cCs|jSrrKr>rrrr@sz&_DenseResourceVariableProcessor.targetcCst|tjr4|jjdur td||j|j|jS| ||j}|jjdurt |g&|j |j|jWdS1s0Yn|SdS)NrQ) rRr rXrLrUrY(_resource_apply_sparse_duplicate_indicesr*r+r_r rVrWr[rrrrCs    6z)_DenseResourceVariableProcessor.update_opNrarrrrrbsrbc@s(eZdZdZddZddZddZdS) _TensorProcessorzProcessor for ordinary Tensors. Even though a Tensor can't really be updated, sometimes it is useful to compute the gradients with respect to a Tensor using the optimizer. Updating the Tensor is, of course, unsupported. cCs ||_dSrrKrMrrrrNsz_TensorProcessor.__init__cCs|jSrrKr>rrrr@sz_TensorProcessor.targetcCstd|jdS)NzTrying to update a Tensor )r=rLrArrrrCsz_TensorProcessor.update_opNrarrrrrdsrdcCstr$t|tjrt|St|St|r<|j sZ&d?d@Z'dAdBZ(dCdDZ)dEdFZ*dGdHZ+Z,S)L OptimizeraBase class for optimizers. This class defines the API to add Ops to train a model. You never use this class directly, but instead instantiate one of its subclasses such as `GradientDescentOptimizer`, `AdagradOptimizer`, or `MomentumOptimizer`. ### Usage ```python # Create an optimizer with the desired parameters. opt = GradientDescentOptimizer(learning_rate=0.1) # Add Ops to the graph to minimize a cost by updating a list of variables. # "cost" is a Tensor, and the list of variables contains tf.Variable # objects. opt_op = opt.minimize(cost, var_list=) ``` In the training program you will just have to run the returned Op. ```python # Execute opt_op to do one step of training: opt_op.run() ``` ### 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 `compute_gradients()`. 2. Process the gradients as you wish. 3. Apply the processed gradients with `apply_gradients()`. Example: ```python # Create an optimizer. opt = GradientDescentOptimizer(learning_rate=0.1) # Compute the gradients for a list of variables. grads_and_vars = opt.compute_gradients(loss, ) # grads_and_vars is a list of tuples (gradient, variable). Do whatever you # need to the 'gradient' part, for example cap them, etc. capped_grads_and_vars = [(MyCapper(gv[0]), gv[1]) for gv in grads_and_vars] # Ask the optimizer to apply the capped gradients. opt.apply_gradients(capped_grads_and_vars) ``` ### Gating Gradients Both `minimize()` and `compute_gradients()` accept a `gate_gradients` argument that controls the degree of parallelism during the application of the gradients. The possible values are: `GATE_NONE`, `GATE_OP`, and `GATE_GRAPH`. `GATE_NONE`: Compute and apply gradients in parallel. This provides the maximum parallelism in execution, at the cost of some non-reproducibility in the results. For example the two gradients of `matmul` depend on the input values: With `GATE_NONE` one of the gradients could be applied to one of the inputs _before_ the other gradient is computed resulting in non-reproducible results. `GATE_OP`: For each Op, make sure all gradients are computed before they are used. This prevents race conditions for Ops that generate gradients for multiple inputs where the gradients depend on the inputs. `GATE_GRAPH`: Make sure all gradients for all variables are computed before any one of them is used. This provides the least parallelism but can be useful if you want to process all gradients before applying any of them. ### Slots Some optimizer subclasses, such as `MomentumOptimizer` and `AdagradOptimizer` 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. @compatibility(TF2) `tf.compat.v1.train.Optimizer` can be used in eager mode and `tf.function`, but it is not recommended. Please use the subclasses of `tf.keras.optimizers.Optimizer` instead in TF2. Please see [Basic training loops](https://www.tensorflow.org/guide/basic_training_loops) or [Writing a training loop from scratch] (https://www.tensorflow.org/guide/keras/writing_a_training_loop_from_scratch) for examples. If your TF1 code contains a `tf.compat.v1.train.Optimizer` symbol, whether it is used with or without a `tf.estimator.Estimator`, you cannot simply replace that with the corresponding `tf.keras.optimizers.Optimizer`s. To migrate to TF2, it is advised the whole training program used with `Estimator` to be migrated to Keras `Model.fit` based or TF2 custom training loops. #### Structural Mapping to Native TF2 Before: ```python sgd_op = tf.compat.v1.train.GradientDescentOptimizer(3.0) opt_op = sgd_op.minimize(cost, global_step, [var0, var1]) opt_op.run(session=session) ``` After: ```python sgd = tf.keras.optimizers.SGD(3.0) sgd.minimize(cost_fn, [var0, var1]) ``` #### How to Map Arguments | TF1 Arg Name | TF2 Arg Name | Note | | :-------------------- | :-------------- | :------------------------- | | `use_locking` | Not supported | - | | `name` | `name. ` | - | #### Before & After Usage Example Before: >>> g = tf.compat.v1.Graph() >>> with g.as_default(): ... var0 = tf.compat.v1.Variable([1.0, 2.0]) ... var1 = tf.compat.v1.Variable([3.0, 4.0]) ... cost = 5 * var0 + 3 * var1 ... global_step = tf.compat.v1.Variable( ... tf.compat.v1.zeros([], tf.compat.v1.int64), name='global_step') ... init_op = tf.compat.v1.initialize_all_variables() ... sgd_op = tf.compat.v1.train.GradientDescentOptimizer(3.0) ... opt_op = sgd_op.minimize(cost, global_step, [var0, var1]) >>> session = tf.compat.v1.Session(graph=g) >>> session.run(init_op) >>> opt_op.run(session=session) >>> print(session.run(var0)) [-14. -13.] After: >>> var0 = tf.Variable([1.0, 2.0]) >>> var1 = tf.Variable([3.0, 4.0]) >>> cost_fn = lambda: 5 * var0 + 3 * var1 >>> sgd = tf.keras.optimizers.SGD(3.0) >>> sgd.minimize(cost_fn, [var0, var1]) >>> print(var0.numpy()) [-14. -13.] @end_compatibility rcCs.|s td||_||_i|_i|_i|_dS)abCreate a new Optimizer. This must be called by the constructors of subclasses. Args: use_locking: Bool. If True apply use locks to prevent concurrent updates to variables. name: A non-empty string. The name to use for accumulators created for the optimizer. Raises: ValueError: If name is malformed. zMust specify the optimizer nameN) ValueErrorZ _use_locking_name_slots_non_slot_dict_deferred_slot_restorations)r? use_lockingr6rrrrNszOptimizer.__init__cCs|jSr)rrr>rrrget_nameszOptimizer.get_nameNFc CsR|j||||||d} dd| D} | sBtddd| D|f|j| ||dS)aAdd operations to minimize `loss` by updating `var_list`. This method simply combines calls `compute_gradients()` and `apply_gradients()`. If you want to process the gradient before applying them call `compute_gradients()` and `apply_gradients()` explicitly instead of using this function. Args: loss: A `Tensor` containing the value to minimize. global_step: Optional `Variable` to increment by one after the variables have been updated. var_list: Optional list or tuple of `Variable` objects to update to minimize `loss`. Defaults to the list of variables collected in the graph under the key `GraphKeys.TRAINABLE_VARIABLES`. gate_gradients: How to gate the computation of gradients. Can be `GATE_NONE`, `GATE_OP`, or `GATE_GRAPH`. aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class `AggregationMethod`. colocate_gradients_with_ops: If True, try colocating gradients with the corresponding op. name: Optional name for the returned operation. grad_loss: Optional. A `Tensor` holding the gradient computed for `loss`. Returns: An Operation that updates the variables in `var_list`. If `global_step` was not `None`, that operation also increments `global_step`. Raises: ValueError: If some of the variables are not `Variable` objects. @compatibility(eager) When eager execution is enabled, `loss` should be a Python function that takes no arguments and computes the value to be minimized. Minimization (and gradient computation) is done with respect to the elements of `var_list` if not None, else with respect to any trainable variables created during the execution of the `loss` function. `gate_gradients`, `aggregation_method`, `colocate_gradients_with_ops` and `grad_loss` are ignored when eager execution is enabled. @end_compatibility )var_listgate_gradientsaggregation_methodcolocate_gradients_with_ops grad_losscSsg|]\}}|dur|qSrrrrrrrrz&Optimizer.minimize..zNo gradients provided for any variable, check your graph for ops that do not support gradients, between variables %s and loss %s.cSsg|]\}}t|qSrstrr_rrrrrr) global_stepr6)compute_gradientsrqapply_gradients) r?lossrrxryrzr{r6r|grads_and_varsZvars_with_gradrrrminimizes ,zOptimizer.minimizec Cst|rt2}|dur$|||}||}Wdn1sH0Y|durb|}t|g||||} Wdn1s0Yt t | |St rt d||}|tjtjtjfvrtd|||g|dur||g|dur(tttjj}n t|}|ttjj7}dd|D} |s`tddd| D} tj|| ||tjk||d} |tjkrt| } t t | |} |d d| D| S) 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. @compatibility(TF2) `tf.keras.optimizers.Optimizer` in TF2 does not provide a `compute_gradients` method, and you should use a `tf.GradientTape` to obtain the gradients: ```python @tf.function def train step(inputs): batch_data, labels = inputs with tf.GradientTape() as tape: predictions = model(batch_data, training=True) loss = tf.keras.losses.CategoricalCrossentropy( reduction=tf.keras.losses.Reduction.NONE)(labels, predictions) gradients = tape.gradient(loss, model.trainable_variables) optimizer.apply_gradients(zip(gradients, model.trainable_variables)) ``` Args: loss: A Tensor containing the value to minimize or a callable taking no arguments which returns the value to minimize. When eager execution is enabled it must be a callable. var_list: Optional list or tuple of `tf.Variable` to update to minimize `loss`. Defaults to the list of variables collected in the graph under the key `GraphKeys.TRAINABLE_VARIABLES`. gate_gradients: How to gate the computation of gradients. Can be `GATE_NONE`, `GATE_OP`, or `GATE_GRAPH`. aggregation_method: Specifies the method used to combine gradient terms. Valid values are defined in the class `AggregationMethod`. colocate_gradients_with_ops: If True, try colocating gradients with the corresponding op. grad_loss: Optional. A `Tensor` holding the gradient computed for `loss`. 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. RuntimeError: If called with eager execution enabled and `loss` is not callable. @compatibility(eager) When eager execution is enabled, `gate_gradients`, `aggregation_method`, and `colocate_gradients_with_ops` are ignored. @end_compatibility Nzb`loss` passed to Optimizer.compute_gradients should be a function when eager execution is enabled.zdgate_gradients must be one of: Optimizer.GATE_NONE, Optimizer.GATE_OP, Optimizer.GATE_GRAPH. Not %scSsg|] }t|qSr)rlrrrrrrWrz/Optimizer.compute_gradients..zNo variables to optimize.cSsg|] }|qSr)r@)rprrrrZr)grad_ysryrzr{cSs(g|] \}}|dur|jtjkr|qSr)dtyperresourcerrrrrds )callabler GradientTapewatch _scale_losswatched_variablesr rVgradientlistziprrfrYrn GATE_NONEGATE_OP GATE_GRAPHrq_assert_valid_dtypesrtrainable_variablesget_collection GraphKeysTRAINABLE_RESOURCE_VARIABLESrflatten_STREAMING_MODEL_PORTSr r tuple) r?rrxryrzr{r|tape loss_valuegrads processorsZvar_refsrrrrrsb<  (,           zOptimizer.compute_gradientscCsFdt_ttjjkrBt j }|dkrB|d|9}dt_|S)NFrog?T) r get_default_graph_is_loss_scaled_by_optimizerrZget_loss_reductionds_reduce_utilReduceOpMEANdistribute_ctx get_strategyZnum_replicas_in_sync)r num_replicasrrrrhs    zOptimizer._scale_lossc s>trDtrtdtfddtj|j||fdStsXt dg}D]t\}}|durzt |}Wnt yt d|Yn0t |t jtjfst d|t|}||||fq`t|}d d |D}|s t d d d |Dft ||Wdn1s40Yg} t j||jd d}||D]\} } } | dur|qdtst| r| jsd} n| jj} t jd| d dNt | "| | || Wdn1s0YWdn1s0Yqd|dur6|!| |}nt "|!| dgzt |Nt |tj#rtj$|j%t j&d|j'd|d}nt(j)|d|d}Wdn1s0YWdn1s0Ytst |t jr|j}t *t j+j,}||vr|||WdS1s00YdS)aApply gradients to variables. This is the second part of `minimize()`. It returns an `Operation` that applies gradients. @compatibility(TF2) #### How to Map Arguments | TF1 Arg Name | TF2 Arg Name | Note | | :-------------------- | :-------------- | :------------------------- | | `grads_and_vars` | `grads_and_vars`| - | | `global_step` | Not supported. | Use `optimizer.iterations` | | `name` | `name. ` | - | Args: grads_and_vars: List of (gradient, variable) pairs as returned by `compute_gradients()`. global_step: Optional `Variable` to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the `Optimizer` constructor. Returns: An `Operation` that applies the specified gradients. If `global_step` was not None, that operation also increments `global_step`. Raises: TypeError: If `grads_and_vars` is malformed. ValueError: If none of the variables have gradients. RuntimeError: If you should use `_distributed_apply()` instead. zUUse `_distributed_apply()` instead of `apply_gradients()` in a cross-replica context.csSrrrrrrrz+Optimizer.apply_gradients..)r!zNo variables provided.NFGradient must be convertible to a Tensor or IndexedSlices, or None: %s5Gradient must be a Tensor, IndexedSlices, or None: %scSsg|]\}}}|dur|qSrr)rrrrrrrrrz-Optimizer.apply_gradients..z+No gradients provided for any variable: %s.cSsg|]\}}}t|qSrr}rrrrrrF) skip_on_eagerupdate_updatero)rr6)-r has_strategyin_cross_replica_contextrYr&Zget_replica_contextZ merge_call_distributed_applyrrqr #convert_to_tensor_or_indexed_slices TypeErrorrRrSr rXrlappend init_scope _create_slots name_scoperr_preparerrfrrgrhr0r6 colocate_withrC_finishrVBaseResourceVariableassign_add_variable_ophandleconvert_to_tensorrr assign_addget_collection_refrTRAIN_OP)r?rrr6Zconverted_grads_and_varsrrrrx update_opsgradr8 processor scope_name apply_updatestrain_oprrrrrs'     *  V  P   zOptimizer.apply_gradientsc sjtjj|}dd|D}t||}t|Wdn1sR0Yfddt |j } fdd|D}dd}j |} jj | ||fd d } |durڈj| |d } nFt| *jj|tjd d |id} Wdn1s0Yts`t| tjr>| j} ttjj} | | vr`| | | WdS1sz0YdS)atA version of `apply_gradients` for cross-replica context. This is a version of `apply_gradients()` for when you are using a `DistributionStrategy` and are in a cross-replica context. If in a replica context, use `apply_gradients()` as normal. Args: distribution: A `DistributionStrategy` object. grads_and_vars: List of (gradient, variable) pairs as returned by `compute_gradients()`, and then aggregated across replicas. global_step: Optional (mirrored) `Variable` to increment by one after the variables have been updated. name: Optional name for the returned operation. Default to the name passed to the `Optimizer` constructor. Returns: An `Operation` that applies the specified gradients across all replicas. If `global_step` was not None, that operation also increments `global_step` cSsg|] \}}|qSrrrrrrrrz0Optimizer._distributed_apply..Ncs|dus Jzt|}Wnty8td|Yn0t|tjtjfsXtd|t|}t sxt |r|j s|j dd}n|jj }td|||WdS1s0YdS)z&Apply gradients to a replica variable.Nrr:rr)r rrrRrSr rXrlrrfrrgrhr6splitr0rrC)rrrrr>rrr s*   z,Optimizer._distributed_apply..updatecs0g|](\}}jj||fddD]}|q"qS)Fr!group)extendedr)rrr8r0) distributionrrrr+s  cSs ||dS)Nr)r)r?rrrrfinish2sz,Optimizer._distributed_apply..finishFrr)ror6r )rZbatch_reduce_torrSUMrr rrrrrrnon_slot_devicesZupdate_non_slotrrVrrrrrfrRrSr0rrrr) r?rrrr6Z reduced_gradsrxrrrZfinish_updatesrrr)rr?rrrs@  (     &   zOptimizer._distributed_applycCsF|j|d}|sdS|t|d}t|rBt|sB|}|S)a>Return a slot named `name` created for `var` by the Optimizer. Some `Optimizer` subclasses use additional variables. For example `Momentum` and `Adagrad` use variables to accumulate updates. This method gives access to these `Variable` objects if for some reason you need them. Use `get_slot_names()` to get the list of slot names created by the `Optimizer`. Args: var: A variable passed to `minimize()` or `apply_gradients()`. name: A string. Returns: The `Variable` for the slot if it was created, `None` otherwise. N)rsgetr9rr2Z_get_on_device_or_primary)r?r8r6 named_slotsslotrrrget_slotIs zOptimizer.get_slotcCst|jS)zReturn a list of the names of slots created by the `Optimizer`. See `get_slot()`. Returns: A list of strings. )sortedrskeysr>rrrget_slot_nameseszOptimizer.get_slot_namescsrtfddfdd|D}|jD],\}}|D]\}}|rD||qDq4t|dddS)zA list of variables which encode the current state of `Optimizer`. Includes slot variables and additional global variables created by the optimizer in the current default graph. Returns: A list of variables. cs"|jr|jjuS|jjkSdSr)rhr0r4 _graph_key)variable) current_graphrr_from_current_graphzs z0Optimizer.variables.._from_current_graphcsg|]}|r|qSrrr)rrrrsz'Optimizer.variables..cSs|jSrrrkrrrrrz%Optimizer.variables..key)r r_non_slot_variablesrsitemsrr)r?Zoptimizer_variablesrZ variable_dictZslot_for_variabler)rrrros  zOptimizer.variablesc Cst}|rdn|j}||f}|j|d}|dur|t}|j |D|rn|j |d} | durn| }t j ||dt |d}Wdn1s0Y|j||d||j|<|S)z2Add an extra variable, not associated with a slot.NrF)r6 trainable use_resource)r6 trackable)r r3r4rtr_maybe_initialize_trackablerrrZcolocate_vars_with_preload_simple_restorationrrrrg_handle_deferred_dependencies) r? initial_valuer6reagerr4rrdistribution_strategyZrestored_initial_valuerrr_create_non_slot_variables.$  z#Optimizer._create_non_slot_variablec sfi}tj}t|jdddD]\\}}}|j|kr$|||<q$|tt|j |fi||S)zAFrom Trackable. Gather graph-specific non-slot variables to save.cSs |ddSr^r)itemrrrrrz/Optimizer._trackable_children..r) r rrrrtrrsuperrn_trackable_children)r? save_typer"Z current_graph_non_slot_variablesZcurrent_graph_keyr6rZvariable_object __class__rrrs    zOptimizer._trackable_childrencs>tt||}|dur|Str(dnt}|j||dS)z>From Trackable. Find a non-slot variable in the current graph.N)r4)rrn_lookup_dependencyrrfr r_get_non_slot_variable)r?r6Z unconditionalr4rrrrs zOptimizer._lookup_dependencycCs,|j||fd}t|dr$|S|SdS)Nr/)rtrr1)r?r6r4Znon_slotrrrrs z Optimizer._get_non_slot_variablecCs |jS)zgAdditional variables created by the `Optimizer`. Returns: A list or tuple of variables. )rtr*r>rrrrszOptimizer._non_slot_variablescCsD|}|D]2}|jj}||vr td||jdd|Dfq dS)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 %r for %s, expected: %s.cSsg|]}|qSrrrrrrrrz2Optimizer._assert_valid_dtypes..N) _valid_dtypesr base_dtyperqr6)r?tensorsZ valid_dtypestrrrrrs zOptimizer._assert_valid_dtypescCsttjtjtjtjgS)zValid types for loss, variables and gradients. Subclasses should override to allow other float types. Returns: Valid types for loss, variables and gradients. )setrfloat16bfloat16float32float64r>rrrrszOptimizer._valid_dtypescCsdS)zgCreate all slots needed by the variables. Args: var_list: A list of `Variable` objects. Nr)r?rxrrrrszOptimizer._create_slotscCsdS)zCreate all needed tensors before applying gradients. This is called with the name_scope using the "name" that users have chosen for the application of gradients. Nrr>rrrrszOptimizer._preparecCs tdS)zAdd ops to apply dense gradients to `var`. Args: grad: A `Tensor`. var: A `Variable` object. Returns: An `Operation`. Nr<r?rr8rrrrTs zOptimizer._apply_densecCs tdS)a(Add 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. Returns: An `Operation` which updates the value of the variable. Nr<)r?rrrrrr_s zOptimizer._resource_apply_densecCst||d\}}||||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. Returns: An `Operation` which updates the value of the variable. r*r+)r._resource_apply_sparse)r?rrr+Z summed_gradr,rrrrcs z2Optimizer._resource_apply_sparse_duplicate_indicescCs tdS)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. Returns: An `Operation` which updates the value of the variable. Nr<)r?rrr+rrrr6sz Optimizer._resource_apply_sparsecCs2t|j|jd\}}tj|||jd}|||S)aAdd ops to apply sparse gradients to `var`, with repeated sparse indices. Optimizers which override this method must deal with IndexedSlices objects such as the following: IndexedSlicesValue(values=[1, 1], indices=[0, 0], dense_shape=[1]) The correct interpretation is: IndexedSlicesValue(values=[2], indices=[0], dense_shape=[1]) Many optimizers deal incorrectly with repeated indices when updating based on sparse gradients (e.g. summing squares rather than squaring the sum, or applying momentum terms multiple times). Adding first is always the correct behavior, so this is enforced here by reconstructing the IndexedSlices to have only unique indices, then calling _apply_sparse. Optimizers which deal correctly with repeated indices may instead override this method to avoid the overhead of summing indices. Args: grad: `IndexedSlices`. var: A `Variable` object. Returns: An `Operation`. r)r+r* dense_shape)r.r*r+r rXr _apply_sparse)r?rr8r-r,Zgradient_no_duplicate_indicesrrrrZJs z)Optimizer._apply_sparse_duplicate_indicescCs tdS)aBAdd ops to apply sparse gradients to `var`. The IndexedSlices object passed to `grad` in this function is by default pre-processed in `_apply_sparse_duplicate_indices` to remove duplicate indices (see its docstring for details). Optimizers which can tolerate or have correct special cases for duplicate sparse indices may override `_apply_sparse_duplicate_indices` instead of this function, avoiding that overhead. Args: grad: `IndexedSlices`, with no repeated indices. var: A `Variable` object. Returns: An `Operation`. Nr<rrrrrnszOptimizer._apply_sparsecCstj|d|iS)aDo what is needed to finish the update. This is called with the `name_scope` using the "name" that users have chosen for the application of gradients. Args: update_ops: List of `Operation` objects to update variables. This list contains the values returned by the `_apply_dense()` and `_apply_sparse()` calls. name_scope: String. Name to use for the returned operation. Returns: The operation to apply updates. r6)r r)r?rrrrrrszOptimizer._finishcCs(|j|d}|dur$i}||j|<|S)zReturns a dict for caching slots created under the given name. Args: slot_name: Name for the slot. Returns: A dict that maps primary `Variable` objects to the slot created for that variable, under the given slot name. N)rsr)r? slot_namerrrr _slot_dicts  zOptimizer._slot_dictcCsL||}t||vr@t|||}|j|||d||t|<|t|S)a=Find or create a slot for a variable. Args: var: A `Variable` object. val: A `Tensor`. The initial value of the slot. slot_name: Name for the slot. op_name: Name to use when scoping the Variable that needs to be created for the slot. Returns: A `Variable` object. rr slot_variable)rr9rZ create_slot_restore_slot_variable)r?r8valrop_namernew_slot_variablerrr_get_or_make_slots   zOptimizer._get_or_make_slotc CsP||}t||vrDt|||||}|j|||d||t|<|t|S)aFind or create a slot for a variable, using an Initializer. Args: var: A `Variable` object. initializer: An `Initializer`. The initial value of the slot. shape: Shape of the initial value of the slot. dtype: Type of the value of the slot. slot_name: Name for the slot. op_name: Name to use when scoping the Variable that needs to be created for the slot. Returns: A `Variable` object. r)rr9rZcreate_slot_with_initializerr ) r?r8 initializerr)rrr rr rrr"_get_or_make_slot_with_initializers    z,Optimizer._get_or_make_slot_with_initializercCsN||}t||vrBtj||dd}|j|||d||t|<|t|S)a Find or create a slot initialized with 0.0. Args: var: A `Variable` object. slot_name: Name for the slot. op_name: Name to use when scoping the Variable that needs to be created for the slot. Returns: A `Variable` object. T)Zcopy_xla_shardingr)rr9rZcreate_zeros_slotr )r?r8rr rr rrr _zeros_slots   zOptimizer._zeros_slotcCsHt|}|j|i|g}|jdddd|D]}||q4dS)z.Restore a newly created slot variable's value.cSs|jSr) restore_uid)positionrrrrrz2Optimizer._restore_slot_variable..T)rreverseN)r9rurpopsortrestore)r?rrr variable_keyZdeferred_restorationscheckpoint_positionrrrr s z Optimizer._restore_slot_variablecCs||}t|}||d}|durhtrh|rhtjsht j |d}|j |||j |j ||jd}|dur|||n|j|i|g|dS)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)r8rr)rrr )rr9rrrfis_simple_variabler r_variable_creator_stackrCheckpointInitialValueCallablerr)rrrrru setdefaultr)r?Zslot_variable_positionrrrrrrrrr _create_or_restore_slot_variables8     z*Optimizer._create_or_restore_slot_variablecCst|r|S|S)z'Call the function if param is callable.)r)r?paramrrr_call_if_callable>szOptimizer._call_if_callable)NN)NN)N)-rDrErFrGrrrrNrwrr staticmethodrrrrrrrrSaveType CHECKPOINTrrrrrrrrrTr_rcrrZrrrr rrr rr __classcell__rrrrrnsf!' < { { _ !      $ Brn)1rGrHsixtensorflow.python.distributerrrrrrtensorflow.python.eagerrrtensorflow.python.frameworkrr r tensorflow.python.opsr r r rrrrrtensorflow.python.trackablerrtensorflow.python.trainingrtensorflow.python.utilr tensorflow.python.util.tf_exportrr&r.r9 add_metaclassABCMetaobjectr:rJr\rbrdrl TrackablernrrrrsJ