# Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Version 2 of class Optimizer.""" import abc import contextlib import functools import warnings import tensorflow.compat.v2 as tf from keras import backend from keras import initializers from keras.engine import base_layer_utils from keras.optimizers import utils as optimizer_utils from keras.optimizers.schedules import learning_rate_schedule from keras.utils import generic_utils from keras.utils import layer_utils from keras.utils import tf_inspect from keras.utils import tf_utils # isort: off from tensorflow.python.util.tf_export import keras_export keras_optimizers_gauge = tf.__internal__.monitoring.BoolGauge( "/tensorflow/api/keras/optimizers", "keras optimizer usage", "method" ) _DEFAULT_VALID_DTYPES = frozenset( [ tf.float16, tf.bfloat16, tf.float32, tf.float64, tf.complex64, tf.complex128, ] ) def _deduplicate_indexed_slices(values, indices): """Sums `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. """ unique_indices, new_index_positions = tf.unique(indices) summed_values = tf.math.unsorted_segment_sum( values, new_index_positions, tf.shape(unique_indices)[0] ) return (summed_values, unique_indices) class NullContextmanager: def __init__(self, *args, **kwargs): pass def __enter__(self): pass def __exit__(self, type_arg, value_arg, traceback_arg): return False # False values do not suppress exceptions def name_scope_only_in_function_or_graph(name): """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. """ if not tf.executing_eagerly(): return tf.name_scope(name) else: return NullContextmanager() @keras_export( "keras.optimizers.legacy.Optimizer", v1=["keras.optimizers.Optimizer", "keras.optimizers.legacy.Optimizer"], ) class OptimizerV2(tf.__internal__.tracking.Trackable): """Base class for legacy Keras optimizers. You should not use this class directly, but instead instantiate one of its subclasses such as `tf.keras.optimizers.legacy.SGD`, `tf.keras.optimizers.legacy.Adam`, etc. This is the default Keras optimizer base class until v2.10 (included). In v2.11 and later, `tf.keras.optimizers.Optimizer` points to a new base class implementation. The legacy class won't be deleted in the future and will continue to be available at `tf.keras.optimizers.legacy.Optimizer`. ### Usage ```python # Create an optimizer with the desired parameters. opt = tf.keras.optimizers.legacy.SGD(learning_rate=0.1) # `loss` is a callable that takes no argument and returns the value # to minimize. var1 = tf.Variable(2.0) var2 = tf.Variable(5.0) 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.legacy.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.legacy.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.legacy.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.legacy.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.legacy.SGD(learning_rate=lr_callable) >>> loss = lambda: 3 * var >>> opt.minimize(loss, var_list=[var]) = 0, received: {kwargs[k]}") if k == "lr": warnings.warn( "The `lr` argument is deprecated, " "use `learning_rate` instead.", stacklevel=2, ) self._use_locking = True self._init_set_name(name) self._hyper = {} # dict: {variable name : {slot name : variable}} self._slots = {} self._slot_names = [] self._weights = [] self._iterations = None # For implementing Trackable. Stores information about how to restore # slot variables which have not yet been created # (trackable._CheckpointPosition objects). # {slot_name : # {_var_key(variable_to_train): [checkpoint_position, ... ], ... }, # ... } self._deferred_slot_restorations = {} decay = kwargs.pop("decay", 0.0) if decay < 0.0: raise ValueError( f"decay cannot be less than 0. Received: decay={decay}." ) self._initial_decay = decay self._hypers_created = False # Store the distribution strategy object if the optimizer is created # inside strategy scope, so it could be used to create variables later. if tf.distribute.has_strategy(): self._distribution_strategy = tf.distribute.get_strategy() else: self._distribution_strategy = None # Configure gradient transformations. if gradient_aggregator is None: gradient_aggregator = optimizer_utils.all_reduce_sum_gradients self.gradient_aggregator = gradient_aggregator if gradient_transformers is None: gradient_transformers = [] self.gradient_transformers = gradient_transformers self.clipnorm = kwargs.pop("clipnorm", None) self.global_clipnorm = kwargs.pop("global_clipnorm", None) if self.clipnorm is not None and self.global_clipnorm is not None: raise ValueError( "Cannot accept both `clipnorm` and `global_clipnorm`. " "Received: `clipnorm`={}, `global_clipnorm`={}.".format( self.clipnorm, self.global_clipnorm ) ) self.clipvalue = kwargs.pop("clipvalue", None) @property def clipnorm(self): """`float` or `None`. If set, clips gradients to a maximum norm.""" return self._clipnorm @property def global_clipnorm(self): """`float` or `None`. If set, clips gradients to a maximum norm. Check `tf.clip_by_global_norm` for more details. """ return self._global_clipnorm @clipnorm.setter def clipnorm(self, val): if val is not None and self.gradient_transformers: raise ValueError( "`clipnorm` cannot be set when `gradient_transformers` " "is set. Instead, use the `gradient_transformers` to " "specify clipping and other transformations. Received: " f"val={val}, " f"gradient_transformers={self.gradient_transformers}." ) self._clipnorm = val self._clipnorm_fn = optimizer_utils.make_gradient_clipnorm_fn( self._clipnorm ) @global_clipnorm.setter def global_clipnorm(self, val): if val is not None and self.gradient_transformers: raise ValueError( "`global_clipnorm` cannot be set when " "`gradient_transformers` " "is set. Instead, use the `gradient_transformers` to " "specify clipping and other transformations. Received: " f"val={val}, " f"gradient_transformers={self.gradient_transformers}." ) self._global_clipnorm = val self._global_clipnorm_fn = ( optimizer_utils.make_global_gradient_clipnorm_fn( self._global_clipnorm ) ) @property def clipvalue(self): """`float` or `None`. If set, clips gradients to a maximum value.""" return self._clipvalue @clipvalue.setter def clipvalue(self, val): if val is not None and self.gradient_transformers: raise ValueError( "`clipvalue` cannot be set when `gradient_transformers` " "is set. Instead, use the `gradient_transformers` to " "specify clipping and other transformations. Received: " f"val={val}, " f"gradient_transformers={self.gradient_transformers}." ) self._clipvalue = val self._clipvalue_fn = optimizer_utils.make_gradient_clipvalue_fn( self._clipvalue ) def _transform_loss(self, loss): """Called in `.minimize` to transform loss before computing gradients.""" return loss def _get_gradients(self, tape, loss, var_list, grad_loss=None): """Called in `minimize` to compute gradients from loss.""" grads = tape.gradient(loss, var_list, grad_loss) return list(zip(grads, var_list)) def _transform_unaggregated_gradients(self, grads_and_vars): """Called in `apply_gradients` before gradient aggregation.""" return grads_and_vars def _aggregate_gradients(self, grads_and_vars): """Called 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`. """ return self.gradient_aggregator(grads_and_vars) def _transform_gradients(self, grads_and_vars): """Called in `apply_gradients` after aggregation.""" if self._clipvalue is not None: grads_and_vars = self._clipvalue_fn(grads_and_vars) if self._clipnorm is not None: grads_and_vars = self._clipnorm_fn(grads_and_vars) if self._global_clipnorm is not None: grads_and_vars = self._global_clipnorm_fn(grads_and_vars) for fn in self.gradient_transformers: grads_and_vars = fn(grads_and_vars) return grads_and_vars def minimize(self, loss, var_list, grad_loss=None, name=None, tape=None): """Minimize `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. """ grads_and_vars = self._compute_gradients( loss, var_list=var_list, grad_loss=grad_loss, tape=tape ) return self.apply_gradients(grads_and_vars, name=name) def _compute_gradients(self, loss, var_list, grad_loss=None, tape=None): """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. """ # TODO(joshl): Test that we handle weight decay in a reasonable way. if not callable(loss) and tape is None: raise ValueError( "`tape` is required when a `Tensor` loss is passed. " f"Received: loss={loss}, tape={tape}." ) tape = tape if tape is not None else tf.GradientTape() if callable(loss): with tape: if not callable(var_list): tape.watch(var_list) loss = loss() if callable(var_list): var_list = var_list() with tape: loss = self._transform_loss(loss) var_list = tf.nest.flatten(var_list) with tf.name_scope(self._name + "/gradients"): grads_and_vars = self._get_gradients( tape, loss, var_list, grad_loss ) self._assert_valid_dtypes( [ v for g, v in grads_and_vars if g is not None and v.dtype != tf.resource ] ) return grads_and_vars def apply_gradients( self, grads_and_vars, name=None, experimental_aggregate_gradients=True ): """Apply 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 presence 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. """ grads_and_vars = optimizer_utils.filter_empty_gradients(grads_and_vars) var_list = [v for (_, v) in grads_and_vars] with tf.name_scope(self._name): # Create iteration if necessary. with tf.init_scope(): self._create_all_weights(var_list) if not grads_and_vars: # Distribution strategy does not support reducing an empty list # of gradients return tf.no_op() if tf.distribute.in_cross_replica_context(): raise RuntimeError( "`apply_gradients() cannot be called in cross-replica " "context. Use `tf.distribute.Strategy.run` to enter " "replica context. For more information, please see the " "docstring of `tf.distribute.get_replica_context`." ) strategy = tf.distribute.get_strategy() if ( not experimental_aggregate_gradients and strategy and isinstance( strategy, ( tf.compat.v1.distribute.experimental.ParameterServerStrategy, # noqa: E501 tf.distribute.experimental.ParameterServerStrategy, tf.distribute.experimental.CentralStorageStrategy, tf.compat.v1.distribute.experimental.CentralStorageStrategy, # noqa: E501 ), ) ): raise NotImplementedError( "`experimental_aggregate_gradients=False is not supported " "for ParameterServerStrategy and CentralStorageStrategy. " f"Used: strategy={strategy}." ) apply_state = self._prepare(var_list) if experimental_aggregate_gradients: grads_and_vars = self._transform_unaggregated_gradients( grads_and_vars ) grads_and_vars = self._aggregate_gradients(grads_and_vars) grads_and_vars = self._transform_gradients(grads_and_vars) return tf.__internal__.distribute.interim.maybe_merge_call( functools.partial( self._distributed_apply, apply_state=apply_state ), strategy, grads_and_vars, name=name, ) def _distributed_apply( self, distribution, grads_and_vars, apply_state, name ): """`apply_gradients` using a `DistributionStrategy`.""" def apply_grad_to_update_var(var, grad): """Apply gradient to variable.""" if isinstance(var, tf.Tensor): raise NotImplementedError( "Updating a `Tensor` is not implemented. " f"Received: var={var}." ) apply_kwargs = {} if isinstance(grad, tf.IndexedSlices): if var.constraint is not None: raise RuntimeError( "Cannot use a constraint function on a sparse " f"variable. Received: grad={grad}, " f"var.constraint={var.constraint}." ) if "apply_state" in self._sparse_apply_args: apply_kwargs["apply_state"] = apply_state return self._resource_apply_sparse_duplicate_indices( grad.values, var, grad.indices, **apply_kwargs ) if "apply_state" in self._dense_apply_args: apply_kwargs["apply_state"] = apply_state update_op = self._resource_apply_dense(grad, var, **apply_kwargs) if var.constraint is not None: with tf.control_dependencies([update_op]): return var.assign(var.constraint(var)) else: return update_op eagerly_outside_functions = ( tf.compat.v1.executing_eagerly_outside_functions() ) update_ops = [] with name_scope_only_in_function_or_graph(name or self._name): for grad, var in grads_and_vars: # Colocate the update with variables to avoid unnecessary # communication delays. See b/136304694. with distribution.extended.colocate_vars_with(var): with name_scope_only_in_function_or_graph( "update" if eagerly_outside_functions else "update_" + var.op.name ): update_op = distribution.extended.update( var, apply_grad_to_update_var, args=(grad,), group=False, ) if tf.distribute.in_cross_replica_context(): # In cross-replica context, extended.update returns # a list of update ops from all replicas # (group=False). update_ops.extend(update_op) else: # In replica context, extended.update return the # single update op of current replica. update_ops.append(update_op) any_symbolic = any( isinstance(i, tf.Operation) or tf_utils.is_symbolic_tensor(i) for i in update_ops ) if not tf.executing_eagerly() or any_symbolic: # If the current context is graph mode or any of the update ops # are symbolic then the step update should be carried out under # a graph context. (eager updates execute immediately) with backend._current_graph(update_ops).as_default(): with tf.control_dependencies([tf.group(update_ops)]): return self.iterations.assign_add(1, read_value=False) return self.iterations.assign_add(1) def get_gradients(self, loss, params): """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). """ params = tf.nest.flatten(params) with backend.get_graph().as_default(), backend.name_scope( self._name + "/gradients" ): grads = tf.compat.v1.gradients(loss, params) for grad, param in zip(grads, params): if grad is None: raise ValueError( "Variable {} 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.".format(param) ) return grads def get_updates(self, loss, params): grads = self.get_gradients(loss, params) grads_and_vars = list(zip(grads, params)) self._assert_valid_dtypes( [ v for g, v in grads_and_vars if g is not None and v.dtype != tf.resource ] ) return [self.apply_gradients(grads_and_vars)] def _set_hyper(self, name, value): """set hyper `name` to value. value can be callable, tensor, numeric.""" if isinstance(value, tf.__internal__.tracking.Trackable): self._track_trackable(value, name, overwrite=True) if name not in self._hyper: self._hyper[name] = value else: prev_value = self._hyper[name] if ( callable(prev_value) or isinstance( prev_value, ( tf.Tensor, int, float, learning_rate_schedule.LearningRateSchedule, ), ) or isinstance( value, learning_rate_schedule.LearningRateSchedule ) ): self._hyper[name] = value else: backend.set_value(self._hyper[name], value) def _get_hyper(self, name, dtype=None): if not self._hypers_created: self._create_hypers() value = self._hyper[name] if isinstance(value, learning_rate_schedule.LearningRateSchedule): return value if callable(value): value = value() if dtype: return tf.cast(value, dtype) else: return value def _create_slots(self, var_list): pass def _create_slots_for_sharded_variables(self, var_list): """Add ShardedVariables to slots to later reconstruct for checkpointing. ShardedVariables don't have slot variables created for them; their shards do. This function allows users to call get_slot with a ShardedVariable input and receive a ShardedVariable output containing the appropriate slot vars. Iterate over the variables to find shards, and aggregate the sharded containers in a set. Add these ShardedVariables to _slots so that get_slot can retrieve the proper slot variables for their component shards, and reconstruct those into a ShardedVariable. Args: var_list: list or tuple of `Variable` objects that will be minimized using this optimizer. """ sharded_vars = set() for var in var_list: if getattr(var, "_sharded_container", False): sharded_vars.add(var._sharded_container()) for sharded_var in sharded_vars: sharded_key = _var_key(sharded_var) slot_dict = {} for slot in self.get_slot_names(): slot_dict[slot] = sharded_var self._slots[sharded_key] = slot_dict def _create_all_weights(self, var_list): """Creates 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. """ _ = self.iterations self._create_hypers() self._create_slots(var_list) self._create_slots_for_sharded_variables(var_list) def __getattribute__(self, name): """Overridden to support hyperparameter access.""" try: return super().__getattribute__(name) except AttributeError as e: # Needed to avoid infinite recursion with __setattr__. if name == "_hyper": raise e # Backwards compatibility with Keras optimizers. if name == "lr": name = "learning_rate" if name in self._hyper: return self._get_hyper(name) raise e def __dir__(self): result = set(super().__dir__()) if "_hyper" in result: result |= self._hyper.keys() if "learning_rate" in self._hyper.keys(): result.add("lr") return list(result) def __setattr__(self, name, value): """Override setattr to support dynamic hyperparameter setting.""" # Backwards compatibility with Keras optimizers. if name == "lr": name = "learning_rate" if hasattr(self, "_hyper") and name in self._hyper: self._set_hyper(name, value) else: super().__setattr__(name, value) def get_slot_names(self): """A list of names for this optimizer's slots.""" return self._slot_names def add_slot(self, var, slot_name, initializer="zeros", shape=None): """Add 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. """ if slot_name not in self._slot_names: self._slot_names.append(slot_name) var_key = _var_key(var) slot_dict = self._slots.setdefault(var_key, {}) weight = slot_dict.get(slot_name, None) if weight is None: if isinstance(initializer, str) or callable(initializer): initializer = initializers.get(initializer) if isinstance( initializer, tf.__internal__.tracking.CheckpointInitialValueCallable, ) or (shape is not None): slot_shape = shape else: slot_shape = var.shape initial_value = functools.partial( initializer, shape=slot_shape, dtype=var.dtype ) else: initial_value = initializer with self._distribution_strategy_scope(): strategy = tf.distribute.get_strategy() if not strategy.extended.variable_created_in_scope(var): raise ValueError( "Trying 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.".format(strategy, var) ) with strategy.extended.colocate_vars_with(var): weight = tf.Variable( name=f"{var._shared_name}/{slot_name}", dtype=var.dtype, trainable=False, initial_value=initial_value, ) backend.track_variable(weight) slot_dict[slot_name] = weight self._restore_slot_variable( slot_name=slot_name, variable=var, slot_variable=weight ) self._weights.append(weight) return weight def get_slot(self, var, slot_name): var_key = _var_key(var) slot_dict = self._slots[var_key] slot_variable = slot_dict[slot_name] if isinstance( slot_variable, tf.__internal__.distribute.ShardedVariable ): # Construct a ShardedVariable that points to the input # ShardedVariable's component shard's slot variables. shard_vars = [] for shard in slot_variable.variables: slot_shard = self.get_slot(shard, slot_name) shard_vars.append(slot_shard) slot_variable = tf.__internal__.distribute.ShardedVariable( shard_vars, name=slot_variable.name ) return slot_variable def _prepare(self, var_list): keys = set() for var in var_list: if isinstance(var, tf.distribute.DistributedValues): var_devices = var._devices else: var_devices = [var.device] var_dtype = var.dtype.base_dtype for var_device in var_devices: keys.add((var_device, var_dtype)) apply_state = {} for var_device, var_dtype in keys: apply_state[(var_device, var_dtype)] = {} with tf.device(var_device): self._prepare_local(var_device, var_dtype, apply_state) return apply_state def _prepare_local(self, var_device, var_dtype, apply_state): if "learning_rate" in self._hyper: lr_t = tf.identity(self._decayed_lr(var_dtype)) apply_state[(var_device, var_dtype)]["lr_t"] = lr_t def _fallback_apply_state(self, var_device, var_dtype): """Compatibility for subclasses that don't pass apply_state through.""" apply_state = {(var_device, var_dtype): {}} self._prepare_local(var_device, var_dtype, apply_state) return apply_state[(var_device, var_dtype)] def _create_hypers(self): if self._hypers_created: return with self._distribution_strategy_scope(): # Iterate hyper values deterministically. for name, value in sorted(self._hyper.items()): if isinstance(value, (tf.Tensor, tf.Variable)) or callable( value ): # The check for `callable` covers the usage when `value` is # a `LearningRateSchedule`, in which case it does not need # to create a variable. continue else: self._hyper[name] = self.add_weight( name, shape=[], trainable=False, initializer=value, aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA, ) self._hypers_created = True @property def iterations(self): """Variable. The number of training steps this Optimizer has run.""" if self._iterations is None: with self._distribution_strategy_scope(): self._iterations = self.add_weight( "iter", shape=[], dtype=tf.int64, trainable=False, aggregation=tf.VariableAggregation.ONLY_FIRST_REPLICA, ) self._weights.append(self._iterations) return self._iterations @iterations.setter def iterations(self, variable): if self._iterations is not None: raise RuntimeError( "Cannot set `iterations` to a new Variable after " "the Optimizer weights have been created. Here it is " f"attempting to set `iterations` to {variable}." ) self._iterations = variable self._weights.append(self._iterations) def _decayed_lr(self, var_dtype): """Get decayed learning rate as a Tensor with dtype=var_dtype.""" lr_t = self._get_hyper("learning_rate", var_dtype) if isinstance(lr_t, learning_rate_schedule.LearningRateSchedule): local_step = tf.cast(self.iterations, var_dtype) lr_t = tf.cast(lr_t(local_step), var_dtype) if self._initial_decay > 0.0: local_step = tf.cast(self.iterations, var_dtype) decay_t = tf.cast(self._initial_decay, var_dtype) lr_t = lr_t / (1.0 + decay_t * local_step) return lr_t @abc.abstractmethod def get_config(self): """Returns the config of the optimizer. An optimizer config is a Python dictionary (serializable) containing the configuration of an optimizer. The same optimizer can be reinstantiated later (without any saved state) from this configuration. Returns: Python dictionary. """ config = {"name": self._name} if self.clipnorm is not None: config["clipnorm"] = self.clipnorm if self.clipvalue is not None: config["clipvalue"] = self.clipvalue if self.global_clipnorm is not None: config["global_clipnorm"] = self.global_clipnorm return config @classmethod def from_config(cls, config, custom_objects=None): """Creates an optimizer from its config. This method is the reverse of `get_config`, capable of instantiating the same optimizer from the config dictionary. Args: config: A Python dictionary, typically the output of get_config. custom_objects: A Python dictionary mapping names to additional Python objects used to create this optimizer, such as a function used for a hyperparameter. Returns: An optimizer instance. """ if "lr" in config: config["learning_rate"] = config.pop("lr") if "learning_rate" in config: if isinstance(config["learning_rate"], dict): config["learning_rate"] = learning_rate_schedule.deserialize( config["learning_rate"], custom_objects=custom_objects ) return cls(**config) def _serialize_hyperparameter(self, hyperparameter_name): """Serialize a hyperparameter that can be a float, callable, or Tensor.""" value = self._hyper[hyperparameter_name] if isinstance(value, learning_rate_schedule.LearningRateSchedule): return learning_rate_schedule.serialize(value) if callable(value): return value() if tf.is_tensor(value): return backend.get_value(value) return value def variables(self): """Returns variables of this Optimizer based on the order created.""" return self._weights @property def weights(self): """Returns variables of this Optimizer based on the order created.""" return self._weights def get_weights(self): """Returns the current weights of the optimizer. The weights of an optimizer are its state (ie, variables). This function returns 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 were created. The returned list can in turn be used to load state into similarly parameterized optimizers. For example, the RMSprop optimizer for this simple model returns 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.legacy.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. """ params = self.weights return backend.batch_get_value(params) # TODO(tanzheny): Maybe share this logic with base_layer. def set_weights(self, weights): """Set 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.legacy.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. """ params = self.weights if len(params) != len(weights): raise ValueError( f"You called `set_weights(weights)` on optimizer {self._name} " f"with a weight list of length {str(len(weights))}, " f"but the optimizer was expecting {str(len(params))} " f"weights. Provided weights: {str(weights)[:50]}..." ) if not params: return weight_value_tuples = [] param_values = backend.batch_get_value(params) for pv, p, w in zip(param_values, params, weights): if pv.shape != w.shape: raise ValueError( f"Optimizer weight shape {str(pv.shape)} " "not compatible with " f"provided weight shape {str(w.shape)}." ) weight_value_tuples.append((p, w)) backend.batch_set_value(weight_value_tuples) def add_weight( self, name, shape, dtype=None, initializer="zeros", trainable=None, synchronization=tf.VariableSynchronization.AUTO, aggregation=tf.VariableAggregation.NONE, ): if dtype is None: dtype = tf.float32 if isinstance(initializer, str) or callable(initializer): initializer = initializers.get(initializer) if synchronization == tf.VariableSynchronization.ON_READ: if trainable: raise ValueError( "Synchronization value can be set to " "VariableSynchronization.ON_READ only for non-trainable " "variables. You have specified trainable=True and " "synchronization=VariableSynchronization.ON_READ." ) else: # Set trainable to be false when variable is to be synced on # read. trainable = False elif trainable is None: trainable = True variable = self._add_variable_with_custom_getter( name=name, shape=shape, getter=base_layer_utils.make_variable, overwrite=True, initializer=initializer, dtype=dtype, trainable=trainable, use_resource=True, synchronization=synchronization, aggregation=aggregation, ) backend.track_variable(variable) return variable def _init_set_name(self, name, zero_based=True): if not name: self._name = backend.unique_object_name( generic_utils.to_snake_case(self.__class__.__name__), zero_based=zero_based, ) else: self._name = name def _assert_valid_dtypes(self, tensors): """Asserts tensors are all valid types (see `_valid_dtypes`). Args: tensors: Tensors to check. Raises: ValueError: If any tensor is not a valid type. """ valid_dtypes = self._valid_dtypes() for t in tensors: dtype = t.dtype.base_dtype if dtype not in valid_dtypes: raise ValueError( "Invalid type {} for {}, expected: {}.".format( dtype, t.name, [v for v in valid_dtypes] ) ) def _valid_dtypes(self): """Valid types for loss, variables and gradients. Subclasses should override to allow other float types. Returns: Valid types for loss, variables and gradients. """ return _DEFAULT_VALID_DTYPES def _call_if_callable(self, param): """Call the function if param is callable.""" return param() if callable(param) else param def _resource_apply_dense(self, grad, handle, apply_state): """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. apply_state: A dict which is used across multiple apply calls. Returns: An `Operation` which updates the value of the variable. """ raise NotImplementedError( "`_resource_apply_dense` must be implemented in subclasses." ) def _resource_apply_sparse_duplicate_indices( self, grad, handle, indices, **kwargs ): """Add 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. """ summed_grad, unique_indices = _deduplicate_indexed_slices( values=grad, indices=indices ) return self._resource_apply_sparse( summed_grad, handle, unique_indices, **kwargs ) def _resource_apply_sparse(self, grad, handle, indices, apply_state): """Add 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. """ raise NotImplementedError( "`_resource_apply_sparse` Must be implemented in subclasses." ) def _resource_scatter_add(self, x, i, v): with tf.control_dependencies( [ tf.raw_ops.ResourceScatterAdd( resource=x.handle, indices=i, updates=v ) ] ): return x.value() def _resource_scatter_update(self, x, i, v): with tf.control_dependencies( [ tf.raw_ops.ResourceScatterUpdate( resource=x.handle, indices=i, updates=v ) ] ): return x.value() @property @layer_utils.cached_per_instance def _dense_apply_args(self): return tf_inspect.getfullargspec(self._resource_apply_dense).args @property @layer_utils.cached_per_instance def _sparse_apply_args(self): return tf_inspect.getfullargspec(self._resource_apply_sparse).args # --------------- # For implementing the trackable interface # --------------- def _restore_slot_variable(self, slot_name, variable, slot_variable): """Restore a newly created slot variable's value.""" variable_key = _var_key(variable) deferred_restorations = self._deferred_slot_restorations.get( slot_name, {} ).pop(variable_key, []) # Iterate over restores, highest restore UID first to minimize the # number of assignments. deferred_restorations.sort( key=lambda position: position.restore_uid, reverse=True ) for checkpoint_position in deferred_restorations: checkpoint_position.restore(slot_variable) def _create_or_restore_slot_variable( self, slot_variable_position, slot_name, variable ): """Returns the slot variable that should have a value restored into it. It is up to the caller to restore the value into the slot variable if a valid slot variable is returned. 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. Returns: A slot variable that should have a value restored into it, or None if a slot variable should not be restored at this time. """ variable_key = _var_key(variable) slot_dict = self._slots.get(variable_key, {}) slot_variable = slot_dict.get(slot_name, None) if ( slot_variable is None and tf.executing_eagerly() and slot_variable_position.is_simple_variable() # Defer slot variable creation if there is an active variable # creator scope. Generally we'd like to eagerly create/restore slot # variables when possible, but this may mean that scopes intended to # catch `variable` also catch its eagerly created slot variable # unintentionally (specifically make_template would add a dependency # on a slot variable if not for this case). Deferring is mostly # harmless (aside from double initialization), and makes variable # creator scopes behave the same way they do when graph building. # # One notable case is with distribution strategy, which uses # variable creator scope but always desires the `variable` and the # slot to use the same scope, thus we can safely eagerly # create/restore slot variables. and ( not tf.compat.v1.get_default_graph()._variable_creator_stack or self._distribution_strategy ) ): initializer = ( tf.__internal__.tracking.CheckpointInitialValueCallable( checkpoint_position=slot_variable_position ) ) slot_variable = self.add_slot( var=variable, initializer=initializer, slot_name=slot_name, shape=slot_variable_position.value_shape(), ) # Slot variables are not owned by any one object (because we don't # want to save the slot variable if the optimizer is saved without # the non-slot variable, or if the non-slot variable is saved # without the optimizer; it's a dependency hypergraph with edges of # the form (optimizer, non-slot variable, variable)). So we don't # _track_ slot variables anywhere, and instead special-case this # dependency and otherwise pretend it's a normal graph. if slot_variable is not None: # For sharded variables, we need the logic in get_slot to combine # slot variables for its shards if (slot_variable is variable) and ( isinstance(variable, tf.__internal__.distribute.ShardedVariable) ): return self.get_slot(variable, slot_name) # If we've either made this slot variable, or if we've pulled out an # existing slot variable, we should restore it. return slot_variable else: # We didn't make the slot variable. Defer restoring until it gets # created normally. We keep a list rather than the one with the # highest restore UID in case slot variables have their own # dependencies, in which case those could differ between restores. self._deferred_slot_restorations.setdefault( slot_name, {} ).setdefault(variable_key, []).append(slot_variable_position) return None @contextlib.contextmanager def _distribution_strategy_scope(self): """Returns the `tf.distribute.Strategy` this optimizer was created under.""" if self._distribution_strategy and not tf.distribute.has_strategy(): with self._distribution_strategy.scope(): yield self._distribution_strategy.scope() else: yield def _var_key(var): """Key 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. """ # Get the distributed variable if it exists. if hasattr(var, "_distributed_container"): var = var._distributed_container() elif ( tf_utils.is_extension_type(var) and hasattr(var, "handle") and hasattr(var.handle, "_distributed_container") ): # For ResourceVariables, the _distributed_container attribute # is added to their handle tensors. var = var.handle._distributed_container() if getattr(var, "_in_graph_mode", False): return var._shared_name return var._unique_id def _get_slot_key_from_var(var, slot_name): """Get the slot key for the variable: var_name/slot_name.""" name = _var_key(var) return name + "/" + slot_name class RestoredOptimizer(OptimizerV2): """A 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 itself (e.g. through `apply_gradients`). """ # TODO(allenl): Make the restored optimizer functional by tracing its apply # methods. def __init__(self): super().__init__("RestoredOptimizer") self._hypers_created = True def get_config(self): # TODO(allenl): Save and restore the Optimizer's config raise NotImplementedError( "Restoring functional Optimizers from SavedModels is not currently " "supported. Please file a feature request if this limitation " "bothers you." ) tf.__internal__.saved_model.load.register_revived_type( "optimizer", lambda obj: isinstance(obj, OptimizerV2), versions=[ tf.__internal__.saved_model.load.VersionedTypeRegistration( object_factory=lambda proto: RestoredOptimizer(), version=2, min_producer_version=1, min_consumer_version=1, setter=RestoredOptimizer._set_hyper, ) ], )