"""Layer is an Operation with state. Takes care of: - Weights / variables (and tracking thereof) - deferred build - trainable argument value inference - masking - autocasting And some more magic: - add_loss - metric tracking - RNG seed tracking - activity regularization """ import collections import inspect import warnings from functools import wraps from keras.src import backend from keras.src import constraints from keras.src import dtype_policies from keras.src import initializers from keras.src import regularizers from keras.src import tree from keras.src import utils from keras.src.api_export import keras_export from keras.src.backend import KerasTensor from keras.src.backend.common import global_state from keras.src.backend.common.name_scope import current_path from keras.src.backend.common.symbolic_scope import in_symbolic_scope from keras.src.distribution import distribution_lib from keras.src.dtype_policies import DTypePolicyMap from keras.src.layers import input_spec from keras.src.metrics.metric import Metric from keras.src.ops.operation import Operation from keras.src.saving.keras_saveable import KerasSaveable from keras.src.utils import python_utils from keras.src.utils import summary_utils from keras.src.utils import traceback_utils from keras.src.utils import tracking if backend.backend() == "tensorflow": from keras.src.backend.tensorflow.layer import TFLayer as BackendLayer elif backend.backend() == "jax": from keras.src.backend.jax.layer import JaxLayer as BackendLayer elif backend.backend() == "torch": from keras.src.backend.torch.layer import TorchLayer as BackendLayer elif backend.backend() == "numpy": from keras.src.backend.numpy.layer import NumpyLayer as BackendLayer else: raise RuntimeError( f"Backend '{backend.backend()}' must implement a layer mixin class." ) @keras_export(["keras.Layer", "keras.layers.Layer"]) class Layer(BackendLayer, Operation, KerasSaveable): """This is the class from which all layers inherit. A layer is a callable object that takes as input one or more tensors and that outputs one or more tensors. It involves *computation*, defined in the `call()` method, and a *state* (weight variables). State can be created: * in `__init__()`, for instance via `self.add_weight()`; * in the optional `build()` method, which is invoked by the first `__call__()` to the layer, and supplies the shape(s) of the input(s), which may not have been known at initialization time. Layers are recursively composable: If you assign a Layer instance as an attribute of another Layer, the outer layer will start tracking the weights created by the inner layer. Nested layers should be instantiated in the `__init__()` method or `build()` method. Users will just instantiate a layer and then treat it as a callable. Args: trainable: Boolean, whether the layer's variables should be trainable. name: String name of the layer. dtype: The dtype of the layer's computations and weights. Can also be a `keras.DTypePolicy`, which allows the computation and weight dtype to differ. Defaults to `None`. `None` means to use `keras.config.dtype_policy()`, which is a `float32` policy unless set to different value (via `keras.config.set_dtype_policy()`). Attributes: name: The name of the layer (string). dtype: Dtype of the layer's weights. Alias of `layer.variable_dtype`. variable_dtype: Dtype of the layer's weights. compute_dtype: The dtype of the layer's computations. Layers automatically cast inputs to this dtype, which causes the computations and output to also be in this dtype. When mixed precision is used with a `keras.DTypePolicy`, this will be different than `variable_dtype`. trainable_weights: List of variables to be included in backprop. non_trainable_weights: List of variables that should not be included in backprop. weights: The concatenation of the lists trainable_weights and non_trainable_weights (in this order). trainable: Whether the layer should be trained (boolean), i.e. whether its potentially-trainable weights should be returned as part of `layer.trainable_weights`. input_spec: Optional (list of) `InputSpec` object(s) specifying the constraints on inputs that can be accepted by the layer. We recommend that descendants of `Layer` implement the following methods: * `__init__()`: Defines custom layer attributes, and creates layer weights that do not depend on input shapes, using `add_weight()`, or other state. * `build(self, input_shape)`: This method can be used to create weights that depend on the shape(s) of the input(s), using `add_weight()`, or other state. `__call__()` will automatically build the layer (if it has not been built yet) by calling `build()`. * `call(self, *args, **kwargs)`: Called in `__call__` after making sure `build()` has been called. `call()` performs the logic of applying the layer to the input arguments. Two reserved keyword arguments you can optionally use in `call()` are: 1. `training` (boolean, whether the call is in inference mode or training mode). 2. `mask` (boolean tensor encoding masked timesteps in the input, used e.g. in RNN layers). A typical signature for this method is `call(self, inputs)`, and user could optionally add `training` and `mask` if the layer need them. * `get_config(self)`: Returns a dictionary containing the configuration used to initialize this layer. If the keys differ from the arguments in `__init__()`, then override `from_config(self)` as well. This method is used when saving the layer or a model that contains this layer. Examples: Here's a basic example: a layer with two variables, `w` and `b`, that returns `y = w . x + b`. It shows how to implement `build()` and `call()`. Variables set as attributes of a layer are tracked as weights of the layers (in `layer.weights`). ```python class SimpleDense(Layer): def __init__(self, units=32): super().__init__() self.units = units # Create the state of the layer (weights) def build(self, input_shape): self.kernel = self.add_weight( shape=(input_shape[-1], self.units), initializer="glorot_uniform", trainable=True, name="kernel", ) self.bias = self.add_weight( shape=(self.units,), initializer="zeros", trainable=True, name="bias", ) # Defines the computation def call(self, inputs): return ops.matmul(inputs, self.kernel) + self.bias # Instantiates the layer. linear_layer = SimpleDense(4) # This will also call `build(input_shape)` and create the weights. y = linear_layer(ops.ones((2, 2))) assert len(linear_layer.weights) == 2 # These weights are trainable, so they're listed in `trainable_weights`: assert len(linear_layer.trainable_weights) == 2 ``` Besides trainable weights, updated via backpropagation during training, layers can also have non-trainable weights. These weights are meant to be updated manually during `call()`. Here's a example layer that computes the running sum of its inputs: ```python class ComputeSum(Layer): def __init__(self, input_dim): super(ComputeSum, self).__init__() # Create a non-trainable weight. self.total = self.add_weight( shape=(), initializer="zeros", trainable=False, name="total", ) def call(self, inputs): self.total.assign(self.total + ops.sum(inputs)) return self.total my_sum = ComputeSum(2) x = ops.ones((2, 2)) y = my_sum(x) assert my_sum.weights == [my_sum.total] assert my_sum.non_trainable_weights == [my_sum.total] assert my_sum.trainable_weights == [] ``` """ def __new__(cls, *args, **kwargs): obj = super().__new__(cls, *args, **kwargs) # Wrap the user-provided `build` method in the `build_wrapper` # to add name scope support and serialization support. original_build_method = obj.build @wraps(original_build_method) def build_wrapper(*args, **kwargs): with obj._open_name_scope(): obj._path = current_path() original_build_method(*args, **kwargs) # Record build config. signature = inspect.signature(original_build_method) obj._build_shapes_dict = signature.bind(*args, **kwargs).arguments # Set built, post build actions, and lock state. obj.built = True obj._post_build() obj._lock_state() obj.build = build_wrapper # Wrap the user-provided `quantize` method in the `quantize_wrapper` # to add tracker support. original_quantize_method = obj.quantize @wraps(original_quantize_method) def quantize_wrapper(mode, **kwargs): obj._check_quantize_args(mode, obj.compute_dtype) obj._tracker.unlock() try: original_quantize_method(mode, **kwargs) except Exception: raise finally: obj._tracker.lock() obj.quantize = quantize_wrapper return obj def __init__( self, *, activity_regularizer=None, trainable=True, dtype=None, autocast=True, name=None, **kwargs, ): BackendLayer.__init__(self) self._lock = False Operation.__init__(self, dtype=dtype, name=name) self.activity_regularizer = regularizers.get(activity_regularizer) input_dim_arg = kwargs.pop("input_dim", None) if input_dim_arg is not None: input_shape_arg = (input_dim_arg,) else: input_shape_arg = kwargs.pop("input_shape", None) if input_shape_arg is not None: warnings.warn( "Do not pass an `input_shape`/`input_dim` argument to " "a layer. When using Sequential models, " "prefer using an `Input(shape)` object as the " "first layer in the model instead.", stacklevel=2, ) self._input_shape_arg = input_shape_arg if kwargs: raise ValueError( "Unrecognized keyword arguments " f"passed to {self.__class__.__name__}: {kwargs}" ) self._path = None # Will be determined in `build_wrapper` self.built = False self.autocast = autocast self._input_spec = None self._called = False self.supports_jit = True self._trainable = trainable self._losses = [] self._loss_ids = set() self._losses_override = [] self._call_signature = inspect.signature(self.call) call_signature_parameters = [ p.name for p in self._call_signature.parameters.values() ] self._call_has_training_arg = "training" in call_signature_parameters self._call_has_mask_arg = "mask" in call_signature_parameters self._supports_masking = not utils.is_default(self.compute_mask) # Whether to automatically convert (+ auto-cast) inputs to `call()`. self._convert_input_args = True # Whether to allow non-tensors as positional arguments in `call()`. self._allow_non_tensor_positional_args = False # Dict of shapes that were used to call `build()`. self._build_shapes_dict = None # Parent path self._parent_path = None self._initialize_tracker() @tracking.no_automatic_dependency_tracking def _initialize_tracker(self): if hasattr(self, "_tracker"): return trainable_variables = [] non_trainable_variables = [] layers = [] metrics = [] seed_generators = [] self._tracker = tracking.Tracker( { "trainable_variables": ( lambda x: isinstance(x, backend.Variable) and x.trainable, trainable_variables, ), "non_trainable_variables": ( lambda x: isinstance(x, backend.Variable) and not x.trainable, non_trainable_variables, ), "metrics": (lambda x: isinstance(x, Metric), metrics), "layers": ( lambda x: isinstance(x, Layer) and not isinstance(x, Metric), layers, ), "seed_generators": ( lambda x: isinstance(x, backend.random.SeedGenerator), seed_generators, ), }, exclusions={"non_trainable_variables": ["trainable_variables"]}, ) if backend.backend() == "tensorflow": # Remove attribute tracking for lists (TF-specific attribute) _self_setattr_tracking = getattr( self, "_self_setattr_tracking", True ) self._self_setattr_tracking = False self._trainable_variables = trainable_variables self._non_trainable_variables = non_trainable_variables self._layers = layers self._metrics = metrics self._seed_generators = seed_generators if backend.backend() == "tensorflow": # Reset attribute tracking (TF-specific) self._self_setattr_tracking = _self_setattr_tracking @property def path(self): """The path of the layer. If the layer has not been built yet, it will be `None`. """ return self._path @property def input_spec(self): return self._input_spec @input_spec.setter def input_spec(self, value): self._input_spec = value @utils.default def build(self, input_shape): self._check_super_called() if utils.is_default(self.build) and might_have_unbuilt_state(self): warnings.warn( f"`build()` was called on layer '{self.name}', however " "the layer does not have a `build()` method implemented " "and it looks like it has unbuilt state. This will cause " "the layer to be marked as built, despite not being " "actually built, which may cause failures down the line. " "Make sure to implement a proper `build()` method." ) self.built = True def _lock_state(self): """Prevent further state updates, called automatically in `build()`.""" if not self._tracker.locked: self._tracker.lock( msg=( "You cannot add new elements of state " "(variables or sub-layers) " "to a layer that is already built. All state " "must be created in the `__init__()` method or " "in the `build()` method." ) ) def get_build_config(self): """Returns a dictionary with the layer's input shape. This method returns a config dict that can be used by `build_from_config(config)` to create all states (e.g. Variables and Lookup tables) needed by the layer. By default, the config only contains the input shape that the layer was built with. If you're writing a custom layer that creates state in an unusual way, you should override this method to make sure this state is already created when Keras attempts to load its value upon model loading. Returns: A dict containing the input shape associated with the layer. """ if self._build_shapes_dict is not None: if len(self._build_shapes_dict) == 1: return { "input_shape": tuple(self._build_shapes_dict.values())[0], } else: return {"shapes_dict": self._build_shapes_dict} def build_from_config(self, config): """Builds the layer's states with the supplied config dict. By default, this method calls the `build(config["input_shape"])` method, which creates weights based on the layer's input shape in the supplied config. If your config contains other information needed to load the layer's state, you should override this method. Args: config: Dict containing the input shape associated with this layer. """ if config: if "input_shape" in config: self.build(config["input_shape"]) elif "shapes_dict" in config: self.build(**config["shapes_dict"]) self.built = True def _obj_type(self): return "Layer" def add_variable( self, shape, initializer, dtype=None, trainable=True, autocast=True, regularizer=None, constraint=None, name=None, ): """Add a weight variable to the layer. Alias of `add_weight()`. """ return self.add_weight( shape=shape, initializer=initializer, dtype=dtype, trainable=trainable, autocast=autocast, regularizer=regularizer, constraint=constraint, name=name, ) def add_weight( self, shape=None, initializer=None, dtype=None, trainable=True, autocast=True, regularizer=None, constraint=None, aggregation="mean", name=None, ): """Add a weight variable to the layer. Args: shape: Shape tuple for the variable. Must be fully-defined (no `None` entries). Defaults to `()` (scalar) if unspecified. initializer: Initializer object to use to populate the initial variable value, or string name of a built-in initializer (e.g. `"random_normal"`). If unspecified, defaults to `"glorot_uniform"` for floating-point variables and to `"zeros"` for all other types (e.g. int, bool). dtype: Dtype of the variable to create, e.g. `"float32"`. If unspecified, defaults to the layer's variable dtype (which itself defaults to `"float32"` if unspecified). trainable: Boolean, whether the variable should be trainable via backprop or whether its updates are managed manually. Defaults to `True`. autocast: Boolean, whether to autocast layers variables when accessing them. Defaults to `True`. regularizer: Regularizer object to call to apply penalty on the weight. These penalties are summed into the loss function during optimization. Defaults to `None`. constraint: Contrainst object to call on the variable after any optimizer update, or string name of a built-in constraint. Defaults to `None`. aggregation: String, one of `'mean'`, `'sum'`, `'only_first_replica'`. Annotates the variable with the type of multi-replica aggregation to be used for this variable when writing custom data parallel training loops. name: String name of the variable. Useful for debugging purposes. """ self._check_super_called() if shape is None: shape = () if dtype is not None: dtype = backend.standardize_dtype(dtype) else: dtype = self.variable_dtype if initializer is None: if "float" in dtype: initializer = "glorot_uniform" else: initializer = "zeros" initializer = initializers.get(initializer) with backend.name_scope(self.name, caller=self): variable = backend.Variable( initializer=initializer, shape=shape, dtype=dtype, trainable=trainable, autocast=autocast, aggregation=aggregation, name=name, ) # Will be added to layer.losses variable.regularizer = regularizers.get(regularizer) variable.constraint = constraints.get(constraint) self._track_variable(variable) return variable @property def trainable(self): """Settable boolean, whether this layer should be trainable or not.""" return self._trainable @trainable.setter def trainable(self, value): """Sets trainable attribute for the layer and its sublayers. When this value is changed during training (e.g. with a `Callback`) you need to call the parent `Model.make_train_function` with `force=True` in order to recompile the training graph. Args: value: Boolean with the desired state for the layer's trainable attribute. """ value = bool(value) self._trainable = value for v in self._trainable_variables: v.trainable = value for layer in self._layers: layer.trainable = value @property def variables(self): """List of all layer state, including random seeds. This extends `layer.weights` to include all state used by the layer including `SeedGenerator`s. Note that metrics variables are not included here, use `metrics_variables` to visit all the metric variables. """ # Return all `Variables` associate with the layer including metrics # and random seeds. Also deduplicate them. variables = [] seen_ids = set() for v in self._trainable_variables + self._non_trainable_variables: if id(v) not in seen_ids: variables.append(v) seen_ids.add(id(v)) for sg in self._seed_generators: variables.append(sg.state) for layer in self._layers: for v in layer.variables: if id(v) not in seen_ids: variables.append(v) seen_ids.add(id(v)) return variables @property def trainable_variables(self): """List of all trainable layer state. This is equivalent to `layer.trainable_weights`. """ if not self.trainable: return [] return [v for v in self.variables if v.trainable] @property def non_trainable_variables(self): """List of all non-trainable layer state. This extends `layer.non_trainable_weights` to include all state used by the layer including state for metrics and `SeedGenerator`s. """ if not self.trainable: return self.variables return [v for v in self.variables if not v.trainable] @property def weights(self): """List of all weight variables of the layer. Unlike, `layer.variables` this excludes metric state and random seeds. """ # Return only `Variables` directly owned by layers and sub-layers. # Also deduplicate them. weights = [] seen_ids = set() for w in self._trainable_variables + self._non_trainable_variables: if id(w) not in seen_ids: weights.append(w) seen_ids.add(id(w)) for layer in self._layers: for w in layer.weights: if id(w) not in seen_ids: weights.append(w) seen_ids.add(id(w)) return weights @property def trainable_weights(self): """List of all trainable weight variables of the layer. These are the weights that get updated by the optimizer during training. """ if not self.trainable: return [] return [v for v in self.weights if v.trainable] @property def non_trainable_weights(self): """List of all non-trainable weight variables of the layer. These are the weights that should not be updated by the optimizer during training. Unlike, `layer.non_trainable_variables` this excludes metric state and random seeds. """ if not self.trainable: return self.weights return [v for v in self.weights if not v.trainable] @property def metrics(self): """List of all metrics.""" metrics = list(self._metrics) for layer in self._layers: metrics.extend(layer.metrics) return metrics @property def metrics_variables(self): """List of all metric variables.""" vars = [] for metric in self.metrics: vars.extend(metric.variables) return vars def get_weights(self): """Return the values of `layer.weights` as a list of NumPy arrays.""" return [v.numpy() for v in self.weights] def set_weights(self, weights): """Sets the values of `layer.weights` from a list of NumPy arrays.""" layer_weights = self.weights if len(layer_weights) != len(weights): raise ValueError( f"You called `set_weights(weights)` on layer '{self.name}' " f"with a weight list of length {len(weights)}, but the layer " f"was expecting {len(layer_weights)} weights." ) for variable, value in zip(layer_weights, weights): if variable.shape != value.shape: raise ValueError( f"Layer {self.name} weight shape {variable.shape} " "is not compatible with provided weight " f"shape {value.shape}." ) variable.assign(value) @property def dtype_policy(self): return self._dtype_policy @dtype_policy.setter def dtype_policy(self, value): policy = dtype_policies.get(value) if isinstance(self._dtype_policy, DTypePolicyMap) and self.path: if self.path in self._dtype_policy: del self._dtype_policy[self.path] self._dtype_policy[self.path] = policy else: self._dtype_policy = policy if policy.quantization_mode is not None: if self.built and not getattr(self, "_is_quantized", False): self.quantize(policy.quantization_mode) @property def dtype(self): """Alias of `layer.variable_dtype`.""" return self.variable_dtype @property def compute_dtype(self): """The dtype of the computations performed by the layer.""" if isinstance(self._dtype_policy, DTypePolicyMap) and self.path: policy = self._dtype_policy[self.path] else: policy = self._dtype_policy return policy.compute_dtype @property def variable_dtype(self): """The dtype of the state (weights) of the layer.""" if isinstance(self._dtype_policy, DTypePolicyMap) and self.path: policy = self._dtype_policy[self.path] else: policy = self._dtype_policy return policy.variable_dtype @property def quantization_mode(self): """The quantization mode of this layer, `None` if not quantized.""" if isinstance(self._dtype_policy, DTypePolicyMap) and self.path: policy = self._dtype_policy[self.path] else: policy = self._dtype_policy return policy.quantization_mode @property def input_dtype(self): """The dtype layer inputs should be converted to.""" return self.compute_dtype @property def supports_masking(self): """Whether this layer supports computing a mask using `compute_mask`.""" return self._supports_masking @supports_masking.setter def supports_masking(self, value): self._supports_masking = value @utils.default def compute_mask(self, inputs, previous_mask): return previous_mask @traceback_utils.filter_traceback def __call__(self, *args, **kwargs): self._check_super_called() self._called = True ##################################### # 1. Convert any array arguments to tensors of correct dtype. def maybe_convert(x): return self.dtype_policy.convert_input( x, self.autocast, self.input_dtype ) # Used to avoid expensive `tree` operations in the most common case. if ( kwargs or len(args) != 1 or not backend.is_tensor(args[0]) or backend.standardize_dtype(args[0].dtype) != self.input_dtype ) and self._convert_input_args: args = tree.map_structure(maybe_convert, args) kwargs = tree.map_structure(maybe_convert, kwargs) ########################################################## # 2. Enforce that only tensors can be passed positionally. if not self._allow_non_tensor_positional_args: for arg in tree.flatten(args): if ( not isinstance(arg, KerasTensor) and not backend.is_tensor(arg) and arg is not None ): raise ValueError( "Only input tensors may be passed as " "positional arguments. The following argument value " f"should be passed as a keyword argument: {arg} " f"(of type {type(arg)})" ) # Caches info about `call()` signature, args, kwargs. call_spec = CallSpec(self._call_signature, args, kwargs) ############################################ # 3. Check input spec for 1st positional arg. # TODO: consider extending this to all args and kwargs. self._assert_input_compatibility(call_spec.first_arg) ################ # 4. Call build with self._open_name_scope(): self._maybe_build(call_spec) ########################## # 5. Infer training value # Training phase for `Layer.call` is set via (in order of priority): # (1) The `training` argument passed to this `Layer.call`, if not None # (2) The training argument of an outer `Layer.call`. # (4) Any non-None default value for `training` in the call signature # (5) False (treating the layer as if it's in inference) # Maintains info about the `Layer.call` stack # across nested calls. call_context = self._get_call_context() # This is the value explicitly passed by the user training = call_spec.user_arguments_dict.get("training", None) if training is None: # Wasn't passed explicitly: use context value training = call_context.training if training is None: # Get signature default value training = call_spec.arguments_dict.get("training", None) call_context.training = training if self._call_has_training_arg and training is not None: # Only populate arg if it has a concrete value kwargs["training"] = training ############################## # 6. Populate mask argument(s) if len(call_spec.tensor_arguments_dict) == 1: if ( "mask" in call_spec.argument_names and call_spec.arguments_dict["mask"] is None ): arg_name = list(call_spec.tensor_arguments_dict.keys())[0] only_tensor_arg = call_spec.tensor_arguments_dict[arg_name] mask = tree.map_structure( backend.get_keras_mask, only_tensor_arg, ) kwargs["mask"] = mask elif len(call_spec.tensor_arguments_dict) > 1: for k, v in call_spec.tensor_arguments_dict.items(): expected_mask_arg_name = f"{k}_mask" if expected_mask_arg_name in call_spec.argument_names: if call_spec.arguments_dict[expected_mask_arg_name] is None: mask = tree.map_structure(backend.get_keras_mask, v) kwargs[expected_mask_arg_name] = mask #################### # 7. Call the layer. try: with self._open_name_scope(): current_scope = backend.get_autocast_scope() new_scope = None if current_scope is not None: # Clear or update the current scope if necessary. if not self.autocast: new_scope = backend.AutocastScope(None) elif not backend.is_float_dtype(self.compute_dtype): # Some preprocessing layers might have a non-float # dtype, we should not autocast in this case. new_scope = backend.AutocastScope(None) elif current_scope.dtype != self.compute_dtype: new_scope = backend.AutocastScope(self.compute_dtype) elif self.compute_dtype != self.variable_dtype: # Enter a new scope if our dtypes are "mixed". new_scope = backend.AutocastScope(self.compute_dtype) if new_scope is not None: with new_scope: outputs = super().__call__(*args, **kwargs) else: outputs = super().__call__(*args, **kwargs) # Change the layout for the layer output if needed. # This is useful for relayout intermediate tensor in the model # to achieve the optimal performance. distribution = distribution_lib.distribution() if distribution is not None: current_layer_path = current_path() current_layer_path += "/output" layout = distribution.get_tensor_layout(current_layer_path) if layout: outputs = distribution_lib.distribute_tensor( outputs, layout ) if not self.built: self.built = True # Record activity regularizer loss. if self.activity_regularizer is not None: for output in tree.flatten(outputs): if backend.is_tensor(output): self.add_loss(self.activity_regularizer(output)) # Set masks on outputs, # provided only the first positional input arg and its mask. # TODO: consider extending this to all args and kwargs. previous_mask = tree.map_structure( backend.get_keras_mask, call_spec.first_arg ) if self.supports_masking: self._set_mask_metadata( call_spec.first_arg, outputs, previous_mask ) elif any(m is not None for m in tree.flatten(previous_mask)): warnings.warn( f"Layer '{self.name}' (of type {self.__class__.__name__}) " "was passed an input with a mask attached to it. " "However, this layer does not support masking and will " "therefore destroy the mask information. Downstream " "layers will not see the mask." ) finally: # Destroy call context if we created it self._maybe_reset_call_context() return outputs def call(self, *args, **kwargs): raise self._not_implemented_error(self.call) @traceback_utils.filter_traceback def stateless_call( self, trainable_variables, non_trainable_variables, *args, return_losses=False, **kwargs, ): """Call the layer without any side effects. Args: trainable_variables: List of trainable variables of the model. non_trainable_variables: List of non-trainable variables of the model. *args: Positional arguments to be passed to `call()`. return_losses: If `True`, `stateless_call()` will return the list of losses created during `call()` as part of its return values. **kwargs: Keyword arguments to be passed to `call()`. Returns: A tuple. By default, returns `(outputs, non_trainable_variables)`. If `return_losses = True`, then returns `(outputs, non_trainable_variables, losses)`. Note: `non_trainable_variables` include not only non-trainable weights such as `BatchNormalization` statistics, but also RNG seed state (if there are any random operations part of the layer, such as dropout), and `Metric` state (if there are any metrics attached to the layer). These are all elements of state of the layer. Example: ```python model = ... data = ... trainable_variables = model.trainable_variables non_trainable_variables = model.non_trainable_variables # Call the model with zero side effects outputs, non_trainable_variables = model.stateless_call( trainable_variables, non_trainable_variables, data, ) # Attach the updated state to the model # (until you do this, the model is still in its pre-call state). for ref_var, value in zip( model.non_trainable_variables, non_trainable_variables ): ref_var.assign(value) ``` """ self._check_super_called() if not self.built: raise ValueError( f"To call stateless_call, {self.__class__.__name__} must be " "built (i.e. its variables must have been already created). " "You can build it by calling it on some data." ) if len(trainable_variables) != len(self.trainable_variables): raise ValueError( "Argument `trainable_variables` must be a list of tensors " "corresponding 1:1 to " f"{self.__class__.__name__}().trainable_variables. " f"Received list with length {len(trainable_variables)}, " f"but expected {len(self.trainable_variables)} variables." ) if len(non_trainable_variables) != len(self.non_trainable_variables): raise ValueError( "Argument `non_trainable_variables` must be a list of tensors " "corresponding 1:1 to " f"{self.__class__.__name__}().non_trainable_variables. " f"Received list with length {len(non_trainable_variables)}, " f"but expected {len(self.non_trainable_variables)} variables." ) # Gather variable mapping trainable_mapping = zip(self.trainable_variables, trainable_variables) non_trainable_mapping = zip( self.non_trainable_variables, non_trainable_variables ) mapping = list(trainable_mapping) + list(non_trainable_mapping) # Call in stateless scope losses = None with backend.StatelessScope( state_mapping=mapping, collect_losses=return_losses ) as scope: if self.dtype_policy.quantization_mode is not None: outputs = self.quantized_call(*args, **kwargs) else: outputs = self.call(*args, **kwargs) if return_losses: losses = self.losses # Gather updated non-trainable variables non_trainable_variables = [] for v in self.non_trainable_variables: new_v = scope.get_current_value(v) non_trainable_variables.append(new_v) if return_losses: return outputs, non_trainable_variables, losses return outputs, non_trainable_variables def compute_output_spec(self, *args, **kwargs): if utils.is_default(self.compute_output_shape): return super().compute_output_spec(*args, **kwargs) else: # Use compute_output_shape() to return the right output spec call_spec = CallSpec(self._call_signature, args, kwargs) shapes_dict = get_shapes_dict(call_spec) shapes_dict = update_shapes_dict_for_target_fn( self.compute_output_shape, shapes_dict=shapes_dict, call_spec=call_spec, class_name=self.__class__.__name__, ) output_shape = self.compute_output_shape(**shapes_dict) if ( isinstance(output_shape, list) and output_shape and isinstance(output_shape[0], (int, type(None))) ): output_shape = tuple(output_shape) if not isinstance(output_shape, (list, tuple, dict)): try: output_shape = tuple(output_shape) except: raise ValueError( "Method `compute_output_shape()` of layer " f"{self.__class__.__name__} is returning " "a type that cannot be interpreted as a shape. " "It should return a shape tuple. " f"Received: {output_shape}" ) if ( isinstance(output_shape, tuple) and output_shape and isinstance(output_shape[0], (int, type(None))) ): return KerasTensor(output_shape, dtype=self.compute_dtype) # Case: nested. Could be a tuple/list of shapes, or a dict of # shapes. Could be deeply nested. return tree.map_shape_structure( lambda s: KerasTensor(s, dtype=self.compute_dtype), output_shape ) @utils.default def compute_output_shape(self, *args, **kwargs): raise self._not_implemented_error( self.compute_output_shape, "Should implement `def compute_output_shape(self, input_shape)`.", ) def add_loss(self, loss): """Can be called inside of the `call()` method to add a scalar loss. Example: ```python class MyLayer(Layer): ... def call(self, x): self.add_loss(ops.sum(x)) return x ``` """ # Eager only. losses = tree.flatten(loss) for x in losses: if not backend.is_tensor(x): raise ValueError( "`add_loss()` can only be called from inside `build()` or " f"`call()`, on a tensor input. Received invalid value: {x}" ) if backend.in_stateless_scope(): scope = backend.get_stateless_scope() if scope.collect_losses: for x in losses: scope.add_loss(loss) self._loss_ids.add(id(loss)) else: self._losses.extend(losses) def _get_own_losses(self): if backend.in_stateless_scope(): losses = [] scope = backend.get_stateless_scope() for loss in scope.losses: if id(loss) in self._loss_ids: losses.append(loss) return losses else: return self._losses[:] def _get_regularization_losses(self): weight_regularization_losses = [] for variable in self.trainable_weights: if variable.regularizer is None: continue if backend.in_stateless_scope() and not in_symbolic_scope(): # If in symbolic scope, we might get `None` from # `get_current_value` in `backend.compute_output_spec`. So we # assign `variable` instead. v = backend.get_stateless_scope().get_current_value(variable) else: v = variable weight_regularization_losses.append(variable.regularizer(v)) return weight_regularization_losses @property def losses(self): """List of scalar losses from `add_loss`, regularizers and sublayers.""" if self._losses_override: return self._losses_override losses = self._get_own_losses() for layer in self._flatten_layers(include_self=False): losses.extend(layer._get_own_losses()) weight_regularization_losses = self._get_regularization_losses() losses.extend(weight_regularization_losses) return losses def _clear_losses(self): if backend.in_stateless_scope(): scope = backend.get_stateless_scope() if scope.collect_losses: for x in scope.losses: if id(x) in self._loss_ids: scope.losses.remove(x) self._losses.clear() self._loss_ids.clear() for layer in self._layers: layer._clear_losses() # Quantization-related (int8 and float8) methods def quantized_build(self, input_shape, mode): raise self._not_implemented_error(self.quantized_build) def quantize(self, mode, type_check=True): raise self._not_implemented_error(self.quantize) def _check_quantize_args(self, mode, compute_dtype): if not self.built: raise ValueError( "Cannot quantize a layer that isn't yet built. " f"Layer '{self.name}' (of type '{self.__class__.__name__}') " "is not built yet." ) if getattr(self, "_is_quantized", False): raise ValueError( f"Layer '{self.name}' is already quantized with " f"dtype_policy='{self.dtype_policy.name}'. " f"Received: mode={mode}" ) if mode not in dtype_policies.QUANTIZATION_MODES: raise ValueError( "Invalid quantization mode. " f"Expected one of {dtype_policies.QUANTIZATION_MODES}. " f"Received: mode={mode}" ) if mode == "int8" and compute_dtype == "float16": raise ValueError( f"Quantization mode='{mode}' doesn't work well with " "compute_dtype='float16'. Consider loading model/layer with " "another dtype policy such as 'mixed_bfloat16' or " "'mixed_float16' before calling `quantize()`." ) def quantized_call(self, *args, **kwargs): if self.quantization_mode == "int8": return self._int8_call(*args, **kwargs) elif self.quantization_mode == "float8": return self._float8_call(*args, **kwargs) else: raise self._quantization_mode_error(self.quantization_mode) def _int8_call(self, *args, **kwargs): raise self._not_implemented_error(self._int8_call) def _float8_call(self, *args, **kwargs): raise self._not_implemented_error(self._float8_call) def _not_implemented_error(self, attr, msg=None): if callable(attr): attr_name = attr.__name__ attr_type = "method" else: attr_name = str(attr) attr_type = "attribute" msg = " " + msg if msg is not None else "" return NotImplementedError( f"Layer {self.__class__.__name__} does not have a `{attr_name}` " f"{attr_type} implemented.{msg}" ) def _quantization_mode_error(self, mode): return NotImplementedError( "Invalid quantization mode. Expected one of " f"{dtype_policies.QUANTIZATION_MODES}. " f"Received: quantization_mode={mode}" ) def save_own_variables(self, store): """Saves the state of the layer. You can override this method to take full control of how the state of the layer is saved upon calling `model.save()`. Args: store: Dict where the state of the model will be saved. """ all_vars = self._trainable_variables + self._non_trainable_variables for i, v in enumerate(all_vars): store[f"{i}"] = v def load_own_variables(self, store): """Loads the state of the layer. You can override this method to take full control of how the state of the layer is loaded upon calling `keras.models.load_model()`. Args: store: Dict from which the state of the model will be loaded. """ all_vars = self._trainable_variables + self._non_trainable_variables if len(store.keys()) != len(all_vars): if len(all_vars) == 0 and not self.built: raise ValueError( f"Layer '{self.name}' was never built " "and thus it doesn't have any variables. " f"However the weights file lists {len(store.keys())} " "variables for this layer.\n" "In most cases, this error indicates that either:\n\n" "1. The layer is owned by a parent layer that " "implements a `build()` method, but calling the " "parent's `build()` method did NOT create the state of " f"the child layer '{self.name}'. A `build()` method " "must create ALL state for the layer, including " "the state of any children layers.\n\n" "2. You need to implement " "the `def build_from_config(self, config)` method " f"on layer '{self.name}', to specify how to rebuild " "it during loading. " "In this case, you might also want to implement the " "method that generates the build config at saving time, " "`def get_build_config(self)`. " "The method `build_from_config()` is meant " "to create the state " "of the layer (i.e. its variables) upon deserialization.", ) raise ValueError( f"Layer '{self.name}' expected {len(all_vars)} variables, " "but received " f"{len(store.keys())} variables during loading. " f"Expected: {[v.name for v in all_vars]}" ) for i, v in enumerate(all_vars): v.assign(store[f"{i}"]) def _track_variable(self, variable): if variable.trainable: self._tracker.add_to_store("trainable_variables", variable) else: self._tracker.add_to_store("non_trainable_variables", variable) if not self.trainable: variable.trainable = False self._post_track_variable(variable) def _untrack_variable(self, variable): previous_lock_state = self._tracker.locked self._tracker.unlock() self._tracker.untrack(variable) if previous_lock_state is True: self._tracker.lock() self._post_untrack_variable(variable) def add_metric(self, *args, **kwargs): # Permanently disabled raise NotImplementedError( "Layer `add_metric()` method is deprecated" " add your metric in `Model.compile(metrics=[...]).`" ) def count_params(self): """Count the total number of scalars composing the weights. Returns: An integer count. """ if not self.built: raise ValueError( "You tried to call `count_params` " f"on layer '{self.name}', " "but the layer isn't built. " "You can build it manually via: " f"`layer.build(input_shape)`." ) return summary_utils.count_params(self.weights) def _maybe_build(self, call_spec): if self.built: return shapes_dict = get_shapes_dict(call_spec) first_shape = next(iter(shapes_dict.values()), None) # If the layer has a build method, call it with our input shapes. if not utils.is_default(self.build): shapes_dict = update_shapes_dict_for_target_fn( self.build, shapes_dict=shapes_dict, call_spec=call_spec, class_name=self.__class__.__name__, ) self.build(**shapes_dict) # Check input spec again (after build, since self.input_spec # may have been updated self._assert_input_compatibility(call_spec.first_arg) return # Otherwise, attempt to build the layer by calling it on symbolic input. if might_have_unbuilt_state(self): try: backend.compute_output_spec( self.call, **call_spec.arguments_dict ) except Exception as e: if call_spec.eager: # Will let the actual eager call do state-building return warnings.warn( f"Layer '{self.name}' looks like it has unbuilt state, but " "Keras is not able to trace the layer `call()` in order to " "build it automatically. Possible causes:\n" "1. The `call()` method of your layer may be crashing. Try " "to `__call__()` the layer eagerly on some test input " "first to see if it works. " "E.g. `x = np.random.random((3, 4)); y = layer(x)`\n" "2. If the `call()` method is correct, then you may need " "to implement the `def build(self, input_shape)` method on " "your layer. It should create all variables used by the " "layer (e.g. by calling `layer.build()` on all its " "children layers).\n" f"Exception encountered: ''{e}''" ) self.build(first_shape) def _build_by_run_for_single_pos_arg(self, input_shape): # Case: all inputs are in the first arg (possibly nested). input_tensors = tree.map_shape_structure( lambda s: backend.KerasTensor(s), input_shape ) try: backend.compute_output_spec(self.call, input_tensors) return True except: return False def _build_by_run_for_kwargs(self, shapes_dict): # Case: inputs were recorded as multiple keyword arguments. if all(is_shape_tuple(s) for s in shapes_dict.values()): # Case: all input keyword arguments were plain tensors. input_tensors = { # We strip the `_shape` suffix to recover kwarg names. utils.removesuffix(k, "_shape"): backend.KerasTensor(shape) for k, shape in shapes_dict.items() } try: backend.compute_output_spec(self.call, **input_tensors) return True except: return False else: # Not supported: nested input keyword arguments. return False def __repr__(self): return ( f"<{self.__class__.__name__} " f"name={self.name}, built={self.built}>" ) def __str__(self): return self.__repr__() def __setattr__(self, name, value): # Track Variables, Layers, Metrics, SeedGenerators. name, value = self._setattr_hook(name, value) if name != "_tracker": if not hasattr(self, "_tracker"): self._initialize_tracker() value = self._tracker.track(value) return super().__setattr__(name, value) def __delattr__(self, name): obj = getattr(self, name) if isinstance(obj, backend.Variable): import gc # It will take a short amount of time for the corresponding buffer # to be actually removed from the device. # https://stackoverflow.com/a/74631949 self._untrack_variable(obj) super().__delattr__(name) gc.collect() else: super().__delattr__(name) def _check_super_called(self): if getattr(self, "_lock", True): raise RuntimeError( f"In layer '{self.__class__.__name__}', you forgot to call " "`super().__init__()` as the first statement " "in the `__init__()` method. Go add it!" ) def _assert_input_compatibility(self, arg_0): if self.input_spec: input_spec.assert_input_compatibility( self.input_spec, arg_0, layer_name=self.name ) def _get_call_context(self): """Returns currently active `CallContext`.""" layer_call_ctx = global_state.get_global_attribute("current_call_ctx") if layer_call_ctx is None: # Enter new call context. layer_call_ctx = CallContext(entry_layer=self) global_state.set_global_attribute( "current_call_ctx", layer_call_ctx ) self._clear_losses() return layer_call_ctx def _maybe_reset_call_context(self): layer_call_ctx = global_state.get_global_attribute("current_call_ctx") if layer_call_ctx is None or layer_call_ctx.entry_layer == self: global_state.set_global_attribute("current_call_ctx", None) def _flatten_layers(self, include_self=True, recursive=True): layers = [] if include_self: layers.append(self) seen_object_ids = set() deque = collections.deque(self._layers) while deque: layer = deque.popleft() if id(layer) in seen_object_ids: continue seen_object_ids.add(id(layer)) layers.append(layer) # Introspect recursively through sublayers. if recursive: deque.extendleft(layer._layers) return layers def _set_mask_metadata(self, inputs, outputs, previous_mask): flat_outputs = tree.flatten(outputs) mask_already_computed = all( backend.get_keras_mask(x) is not None for x in flat_outputs ) if mask_already_computed: return output_masks = self.compute_mask(inputs, previous_mask) if output_masks is None: return flat_masks = tree.flatten(output_masks) for tensor, mask in zip(flat_outputs, flat_masks): if backend.get_keras_mask(tensor) is None and mask is not None: if backend.backend() == "numpy": warnings.warn( "The NumPy backend does not support masking at this" "time. Masks will be ignored." ) else: backend.set_keras_mask(tensor, mask) @python_utils.default def get_config(self): self._check_super_called() base_config = super().get_config() config = { "trainable": self.trainable, "dtype": dtype_policies.serialize(self.dtype_policy), } if self.activity_regularizer is not None: config["activity_regularizer"] = regularizers.serialize( self.activity_regularizer ) return {**base_config, **config} def _open_name_scope(self): if self._parent_path is None: self._parent_path = current_path() return backend.name_scope(self.name, caller=self) def is_backend_tensor_or_symbolic(x, allow_none=False): if allow_none and x is None: return True return backend.is_tensor(x) or isinstance(x, backend.KerasTensor) class CallSpec: def __init__(self, signature, args, kwargs): # `training` and `mask` are special kwargs that are always available in # a layer, if user specifies them in their call without adding to spec, # we remove them to be able to bind variables. User is not using # `training` anyway so we can ignore. # TODO: If necessary use workaround for `mask` if "training" in kwargs and "training" not in signature.parameters: kwargs.pop("training") bound_args = signature.bind(*args, **kwargs) else: bound_args = signature.bind(*args, **kwargs) self.user_arguments_dict = { k: v for k, v in bound_args.arguments.items() } bound_args.apply_defaults() arg_dict = {} arg_names = [] tensor_arg_dict = {} tensor_args = [] tensor_arg_names = [] nested_tensor_arg_names = [] for name, value in bound_args.arguments.items(): arg_dict[name] = value arg_names.append(name) if is_backend_tensor_or_symbolic(value): tensor_args.append(value) tensor_arg_names.append(name) tensor_arg_dict[name] = value elif tree.is_nested(value) and len(value) > 0: flat_values = tree.flatten(value) if all( is_backend_tensor_or_symbolic(x, allow_none=True) for x in flat_values ): tensor_args.append(value) tensor_arg_names.append(name) tensor_arg_dict[name] = value nested_tensor_arg_names.append(name) elif any(is_backend_tensor_or_symbolic(x) for x in flat_values): raise ValueError( "In a nested call() argument, " "you cannot mix tensors and non-tensors. " "Received invalid mixed argument: " f"{name}={value}" ) self.arguments_dict = arg_dict self.argument_names = arg_names self.tensor_arguments_dict = tensor_arg_dict self.tensor_arguments_names = tensor_arg_names self.nested_tensor_argument_names = nested_tensor_arg_names self.first_arg = arg_dict[arg_names[0]] if all( backend.is_tensor(x) for x in self.tensor_arguments_dict.values() ): self.eager = True else: self.eager = False def get_arguments_dict(fn, args, kwargs): """Return a dict mapping argument names to their values.""" sig = inspect.signature(fn) bound_args = sig.bind(*args, **kwargs) arg_dict = {} for name, value in bound_args.arguments.items(): arg_dict[name] = value return arg_dict def get_shapes_dict(call_spec): """Convert the call() arguments dict into a dict of input shape arguments. Example: ``` >>> get_shapes_dict(call_spec) {"input_a_shape": (2, 3)} ``` """ shapes_dict = {} for k, v in call_spec.tensor_arguments_dict.items(): if k == "mask" or k.endswith("_mask"): # Do not include mask tensors in shapes dict continue if k == "kwargs" or k == "args": # Do not include catch-alls in shapes dict continue if k in call_spec.nested_tensor_argument_names: shapes_dict[f"{k}_shape"] = tree.map_structure( lambda x: backend.standardize_shape(x.shape), v ) else: shapes_dict[f"{k}_shape"] = backend.standardize_shape(v.shape) return shapes_dict def update_shapes_dict_for_target_fn( target_fn, shapes_dict, call_spec, class_name, ): """Updates a `shapes_dict` for `build()` or `compute_output_shape()`. This function will align a dictionary of the shapes of all tensor passed to `call`, with the signatures of `build()` or `compute_output_shape()`. The alignment is a follows: - If `build()` or `compute_output_shape()` accept only one argument, forward the shape of the first positional argument from call without checking any argument names. - If `build()` or `compute_output_shape()` accept multiple arguments, enforce that all argument names match a call argument name, e.g. `foo_shape` would match call argument `foo`. Returns: An updated `shapes_dict` that can be used to invoke `target_fn(**shapes_dict)`. """ if utils.is_default(target_fn): return None sig = inspect.signature(target_fn) expected_names = [] for name, param in sig.parameters.items(): if param.kind in ( param.POSITIONAL_OR_KEYWORD, param.POSITIONAL_ONLY, param.KEYWORD_ONLY, ): expected_names.append(name) # Single arg: don't check names, pass first shape. if len(expected_names) == 1: key = expected_names[0] values = tuple(shapes_dict.values()) if values: input_shape = values[0] else: input_shape = None return {key: input_shape} # Multiple args: check that all names line up. kwargs = {} for name in expected_names: method_name = target_fn.__name__ error_preamble = ( f"For a `{method_name}()` method with more than one argument, all " "arguments should have a `_shape` suffix and match an argument " f"from `call()`. E.g. `{method_name}(self, foo_shape, bar_shape)` " ) if not name.endswith("_shape"): raise ValueError( f"{error_preamble} For layer '{class_name}', " f"Received `{method_name}()` argument " f"`{name}`, which does not end in `_shape`." ) expected_call_arg = utils.removesuffix(name, "_shape") if expected_call_arg not in call_spec.arguments_dict: raise ValueError( f"{error_preamble} For layer '{class_name}', " f"received `{method_name}()` argument " f"`{name}`, but `call()` does not have argument " f"`{expected_call_arg}`." ) if name in shapes_dict: kwargs[name] = shapes_dict[name] return kwargs class CallContext: def __init__(self, entry_layer): self.entry_layer = entry_layer self.training = None def is_shape_tuple(s): return isinstance(s, (list, tuple)) and all( d is None or isinstance(d, int) for d in s ) def might_have_unbuilt_state(layer): return any(not lr.built for lr in layer._layers)