# Copyright 2019 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. # ============================================================================== """Contains the base ProcessingLayer and a subclass that uses Combiners.""" import abc import collections import numpy as np from tensorflow.python.eager import context from tensorflow.python.eager import def_function from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.framework import sparse_tensor from tensorflow.python.keras import backend from tensorflow.python.keras.engine import data_adapter from tensorflow.python.keras.engine.base_layer import Layer from tensorflow.python.keras.utils import tf_utils from tensorflow.python.keras.utils import version_utils from tensorflow.python.ops import math_ops from tensorflow.python.ops import sparse_ops from tensorflow.python.ops import variables from tensorflow.python.ops.ragged import ragged_tensor from tensorflow.python.trackable import base as trackable from tensorflow.python.util.tf_export import keras_export @keras_export('keras.layers.experimental.preprocessing.PreprocessingLayer') class PreprocessingLayer(Layer, metaclass=abc.ABCMeta): """Base class for Preprocessing Layers. **Don't use this class directly: it's an abstract base class!** You may be looking for one of the many built-in [preprocessing layers](https://keras.io/guides/preprocessing_layers/) instead. Preprocessing layers are layers whose state gets computed before model training starts. They do not get updated during training. Most preprocessing layers implement an `adapt()` method for state computation. The `PreprocessingLayer` class is the base class you would subclass to implement your own preprocessing layers. Attributes: streaming: Whether a layer can be adapted multiple times without resetting the state of the layer. """ _must_restore_from_config = True def __init__(self, streaming=True, **kwargs): super(PreprocessingLayer, self).__init__(**kwargs) self._streaming = streaming self._is_compiled = False self._is_adapted = False # Sets `is_adapted=False` when `reset_state` is called. self._reset_state_impl = self.reset_state self.reset_state = self._reset_state_wrapper self._adapt_function = None @property def streaming(self): """Whether `adapt` can be called twice without resetting the state.""" return self._streaming @property def is_adapted(self): """Whether the layer has been fit to data already.""" return self._is_adapted def update_state(self, data): """Accumulates statistics for the preprocessing layer. Arguments: data: A mini-batch of inputs to the layer. """ raise NotImplementedError def reset_state(self): # pylint: disable=method-hidden """Resets the statistics of the preprocessing layer.""" raise NotImplementedError def merge_state(self, layers): """Merge the statistics of multiple preprocessing layers. This layer will contain the merged state. Arguments: layers: Layers whose statistics should be merge with the statistics of this layer. """ raise NotImplementedError def finalize_state(self): """Finalize the statistics for the preprocessing layer. This method is called at the end of `adapt` or after restoring a serialized preprocessing layer's state. This method handles any one-time operations that should occur on the layer's state before `Layer.__call__`. """ pass def make_adapt_function(self): """Creates a function to execute one step of `adapt`. This method can be overridden to support custom adapt logic. This method is called by `PreprocessingLayer.adapt`. Typically, this method directly controls `tf.function` settings, and delegates the actual state update logic to `PreprocessingLayer.update_state`. This function is cached the first time `PreprocessingLayer.adapt` is called. The cache is cleared whenever `PreprocessingLayer.compile` is called. Returns: Function. The function created by this method should accept a `tf.data.Iterator`, retrieve a batch, and update the state of the layer. """ if self._adapt_function is not None: return self._adapt_function def adapt_step(iterator): data = next(iterator) self._adapt_maybe_build(data) self.update_state(data) if self._steps_per_execution.numpy().item() == 1: adapt_fn = adapt_step else: def adapt_fn(iterator): for _ in math_ops.range(self._steps_per_execution): adapt_step(iterator) if not self._run_eagerly: adapt_fn = def_function.function(adapt_fn) self._adapt_function = adapt_fn return self._adapt_function def compile(self, run_eagerly=None, steps_per_execution=None): """Configures the layer for `adapt`. Arguments: run_eagerly: Bool. Defaults to `False`. If `True`, this `Model`'s logic will not be wrapped in a `tf.function`. Recommended to leave this as `None` unless your `Model` cannot be run inside a `tf.function`. steps_per_execution: Int. Defaults to 1. The number of batches to run during each `tf.function` call. Running multiple batches inside a single `tf.function` call can greatly improve performance on TPUs or small models with a large Python overhead. """ if steps_per_execution is None: steps_per_execution = 1 self._configure_steps_per_execution(steps_per_execution) if run_eagerly is None: run_eagerly = self.dynamic self._run_eagerly = run_eagerly self._is_compiled = True def adapt(self, data, batch_size=None, steps=None, reset_state=True): """Fits the state of the preprocessing layer to the data being passed. After calling `adapt` on a layer, a preprocessing layer's state will not update during training. In order to make preprocessing layers efficient in any distribution context, they are kept constant with respect to any compiled `tf.Graph`s that call the layer. This does not affect the layer use when adapting each layer only once, but if you adapt a layer multiple times you will need to take care to re-compile any compiled functions as follows: * If you are adding a preprocessing layer to a `keras.Model`, you need to call `model.compile` after each subsequent call to `adapt`. * If you are calling a preprocessing layer inside `tf.data.Dataset.map`, you should call `map` again on the input `tf.data.Dataset` after each `adapt`. * If you are using a `tf.function` directly which calls a preprocessing layer, you need to call `tf.function` again on your callable after each subsequent call to `adapt`. `tf.keras.Model` example with multiple adapts: >>> layer = tf.keras.layers.experimental.preprocessing.Normalization( ... axis=None) >>> layer.adapt([0, 2]) >>> model = tf.keras.Sequential(layer) >>> model.predict([0, 1, 2]) array([-1., 0., 1.], dtype=float32) >>> layer.adapt([-1, 1]) >>> model.compile() # This is needed to re-compile model.predict! >>> model.predict([0, 1, 2]) array([0., 1., 2.], dtype=float32) `tf.data.Dataset` example with multiple adapts: >>> layer = tf.keras.layers.experimental.preprocessing.Normalization( ... axis=None) >>> layer.adapt([0, 2]) >>> input_ds = tf.data.Dataset.range(3) >>> normalized_ds = input_ds.map(layer) >>> list(normalized_ds.as_numpy_iterator()) [array([-1.], dtype=float32), array([0.], dtype=float32), array([1.], dtype=float32)] >>> layer.adapt([-1, 1]) >>> normalized_ds = input_ds.map(layer) # Re-map over the input dataset. >>> list(normalized_ds.as_numpy_iterator()) [array([0.], dtype=float32), array([1.], dtype=float32), array([2.], dtype=float32)] Arguments: data: The data to train on. It can be passed either as a tf.data Dataset, or as a numpy array. batch_size: Integer or `None`. Number of samples per state update. If unspecified, `batch_size` will default to 32. Do not specify the `batch_size` if your data is in the form of datasets, generators, or `keras.utils.Sequence` instances (since they generate batches). steps: Integer or `None`. Total number of steps (batches of samples) When training with input tensors such as TensorFlow data tensors, the default `None` is equal to the number of samples in your dataset divided by the batch size, or 1 if that cannot be determined. If x is a `tf.data` dataset, and 'steps' is None, the epoch will run until the input dataset is exhausted. When passing an infinitely repeating dataset, you must specify the `steps` argument. This argument is not supported with array inputs. reset_state: Optional argument specifying whether to clear the state of the layer at the start of the call to `adapt`, or whether to start from the existing state. This argument may not be relevant to all preprocessing layers: a subclass of PreprocessingLayer may choose to throw if 'reset_state' is set to False. """ _disallow_inside_tf_function('adapt') if not version_utils.should_use_v2(): raise RuntimeError('`adapt` is only supported in tensorflow v2.') # pylint: disable=g-doc-exception if not self.streaming and self._is_adapted and not reset_state: raise ValueError('{} does not supporting calling `adapt` twice without ' 'resetting the state.'.format(self.__class__.__name__)) if not self._is_compiled: self.compile() # Compile with defaults. if self.built and reset_state: self.reset_state() data_handler = data_adapter.DataHandler( data, batch_size=batch_size, steps_per_epoch=steps, epochs=1, steps_per_execution=self._steps_per_execution, distribute=False) self._adapt_function = self.make_adapt_function() for _, iterator in data_handler.enumerate_epochs(): with data_handler.catch_stop_iteration(): for _ in data_handler.steps(): self._adapt_function(iterator) if data_handler.should_sync: context.async_wait() self.finalize_state() self._is_adapted = True def _reset_state_wrapper(self): """Calls `reset_state` and sets `adapted` to `False`.""" self._reset_state_impl() self._is_adapted = False @trackable.no_automatic_dependency_tracking def _configure_steps_per_execution(self, steps_per_execution): self._steps_per_execution = variables.Variable( steps_per_execution, dtype='int64', aggregation=variables.VariableAggregationV2.ONLY_FIRST_REPLICA) # TODO(omalleyt): Unify this logic with `Layer._maybe_build`. def _adapt_maybe_build(self, data): if not self.built: try: # If this is a Numpy array or tensor, we can get shape from .shape. # If not, an attribute error will be thrown. data_shape = data.shape data_shape_nones = tuple([None] * len(data.shape)) except AttributeError: # The input has an unknown number of dimensions. data_shape = None data_shape_nones = None # TODO (b/159261555): move this to base layer build. batch_input_shape = getattr(self, '_batch_input_shape', None) if batch_input_shape is None: # Set the number of dimensions. self._batch_input_shape = data_shape_nones self.build(data_shape) self.built = True # TODO(omalleyt): This class will be gradually replaced. class CombinerPreprocessingLayer(PreprocessingLayer): """Base class for PreprocessingLayers that do computation using a Combiner. This class provides several helper methods to make creating a PreprocessingLayer easier. It assumes that the core of your computation will be done via a Combiner object. Subclassing this class to create a PreprocessingLayer allows your layer to be compatible with distributed computation. This class is compatible with Tensorflow 2.0+. """ def __init__(self, combiner, **kwargs): super(CombinerPreprocessingLayer, self).__init__(**kwargs) self.state_variables = collections.OrderedDict() self._combiner = combiner self._adapt_accumulator = None def reset_state(self): # pylint: disable=method-hidden self._adapt_accumulator = None @trackable.no_automatic_dependency_tracking def update_state(self, data): if self._adapt_accumulator is None: self._adapt_accumulator = self._get_accumulator() self._adapt_accumulator = self._combiner.compute(data, self._adapt_accumulator) def merge_state(self, layers): accumulators = ([self._get_accumulator()] + [l._get_accumulator() for l in layers]) # pylint: disable=protected-access merged_accumulator = self._combiner.merge(accumulators) self._set_accumulator(merged_accumulator) def finalize_state(self): if self._adapt_accumulator is not None: self._set_accumulator(self._adapt_accumulator) def compile(self, run_eagerly=None, steps_per_execution=None): # TODO(omalleyt): Remove this once sublayers are switched to new APIs. if run_eagerly is None: run_eagerly = True super(CombinerPreprocessingLayer, self).compile( run_eagerly=run_eagerly, steps_per_execution=steps_per_execution) def adapt(self, data, batch_size=None, steps=None, reset_state=True): if not reset_state: self._adapt_accumulator = self._combiner.restore(self._restore_updates()) super(CombinerPreprocessingLayer, self).adapt( data, batch_size=batch_size, steps=steps, reset_state=reset_state) def _add_state_variable(self, name, shape, dtype, initializer=None, partitioner=None, use_resource=None, **kwargs): """Add a variable that can hold state which is updated during adapt(). Args: name: Variable name. shape: Variable shape. Defaults to scalar if unspecified. dtype: The type of the variable. Defaults to `self.dtype` or `float32`. initializer: initializer instance (callable). partitioner: Partitioner to be passed to the `Trackable` API. use_resource: Whether to use `ResourceVariable` **kwargs: Additional keyword arguments. Accepted values are `getter` and `collections`. Returns: The created variable. """ weight = self.add_weight( name=name, shape=shape, dtype=dtype, initializer=initializer, regularizer=None, trainable=False, constraint=None, partitioner=partitioner, use_resource=use_resource, **kwargs) # TODO(momernick): Do not allow collisions here. self.state_variables[name] = weight return weight def _restore_updates(self): """Recreates a dict of updates from the layer's weights.""" data_dict = {} for name, var in self.state_variables.items(): data_dict[name] = var.numpy() return data_dict def _get_accumulator(self): if self._is_adapted: return self._combiner.restore(self._restore_updates()) else: return None def _set_accumulator(self, accumulator): updates = self._combiner.extract(accumulator) self._set_state_variables(updates) self._adapt_accumulator = None # Reset accumulator from adapt. def _set_state_variables(self, updates): """Directly update the internal state of this Layer. This method expects a string-keyed dict of {state_variable_name: state}. The precise nature of the state, and the names associated, are describe by the subclasses of CombinerPreprocessingLayer. Args: updates: A string keyed dict of weights to update. Raises: RuntimeError: if 'build()' was not called before 'set_processing_state'. """ # TODO(momernick): Do we need to do any more input sanitization? if not self.built: raise RuntimeError('_set_state_variables() must be called after build().') with ops.init_scope(): for var_name, value in updates.items(): self.state_variables[var_name].assign(value) def convert_to_list(values, sparse_default_value=None): """Convert a TensorLike, CompositeTensor, or ndarray into a Python list.""" if tf_utils.is_ragged(values): # There is a corner case when dealing with ragged tensors: if you get an # actual RaggedTensor (not a RaggedTensorValue) passed in non-eager mode, # you can't call to_list() on it without evaluating it first. However, # because we don't yet fully support composite tensors across Keras, # backend.get_value() won't evaluate the tensor. # TODO(momernick): Get Keras to recognize composite tensors as Tensors # and then replace this with a call to backend.get_value. if (isinstance(values, ragged_tensor.RaggedTensor) and not context.executing_eagerly()): values = backend.get_session(values).run(values) values = values.to_list() if isinstance(values, (sparse_tensor.SparseTensor, sparse_tensor.SparseTensorValue)): if sparse_default_value is None: if dtypes.as_dtype(values.values.dtype) == dtypes.string: sparse_default_value = '' else: sparse_default_value = -1 dense_tensor = sparse_ops.sparse_tensor_to_dense( values, default_value=sparse_default_value) values = backend.get_value(dense_tensor) if isinstance(values, ops.Tensor): values = backend.get_value(values) # We may get passed a ndarray or the code above may give us a ndarray. # In either case, we want to force it into a standard python list. if isinstance(values, np.ndarray): values = values.tolist() return values # TODO(omalleyt): This class will be gradually replaced. class Combiner(object): """Functional object that defines a shardable computation. This object defines functions required to create and manipulate data objects. These data objects, referred to below as 'accumulators', are computation- specific and may be implemented alongside concrete subclasses of Combiner (if necessary - some computations may be simple enough that standard Python types can be used as accumulators). The intent for this class is that by describing computations in this way, we can arbitrarily shard a dataset, perform computations on a subset, and then merge the computation into a final result. This enables distributed computation. The combiner itself does not own any state - all computational state is owned by the accumulator objects. This is so that we can have an arbitrary number of Combiners (thus sharding the computation N ways) without risking any change to the underlying computation. These accumulator objects are uniquely associated with each Combiner; a Combiner defines what the accumulator object should be and will only work with accumulators of that type. """ __metaclass__ = abc.ABCMeta def __repr__(self): return '<{}>'.format(self.__class__.__name__) @abc.abstractmethod def compute(self, batch_values, accumulator=None): """Compute a step in this computation, returning a new accumulator. This method computes a step of the computation described by this Combiner. If an accumulator is passed, the data in that accumulator is also used; so compute(batch_values) results in f(batch_values), while compute(batch_values, accumulator) results in merge(f(batch_values), accumulator). Args: batch_values: A list of ndarrays representing the values of the inputs for this step of the computation. accumulator: the current accumulator. Can be None. Returns: An accumulator that includes the passed batch of inputs. """ pass @abc.abstractmethod def merge(self, accumulators): """Merge several accumulators to a single accumulator. This method takes the partial values in several accumulators and combines them into a single accumulator. This computation must not be order-specific (that is, merge([a, b]) must return the same result as merge([b, a]). Args: accumulators: the accumulators to merge, as a list. Returns: A merged accumulator. """ pass @abc.abstractmethod def extract(self, accumulator): """Convert an accumulator into a dict of output values. Args: accumulator: The accumulator to convert. Returns: A dict of ndarrays representing the data in this accumulator. """ pass @abc.abstractmethod def restore(self, output): """Create an accumulator based on 'output'. This method creates a new accumulator with identical internal state to the one used to create the data in 'output'. This means that if you do output_data = combiner.extract(accumulator_1) accumulator_2 = combiner.restore(output_data) then accumulator_1 and accumulator_2 will have identical internal state, and computations using either of them will be equivalent. Args: output: The data output from a previous computation. Should be in the same form as provided by 'extract_output'. Returns: A new accumulator. """ pass @abc.abstractmethod def serialize(self, accumulator): """Serialize an accumulator for a remote call. This function serializes an accumulator to be sent to a remote process. Args: accumulator: The accumulator to serialize. Returns: A byte string representing the passed accumulator. """ pass @abc.abstractmethod def deserialize(self, encoded_accumulator): """Deserialize an accumulator received from 'serialize()'. This function deserializes an accumulator serialized by 'serialize()'. Args: encoded_accumulator: A byte string representing an accumulator. Returns: The accumulator represented by the passed byte_string. """ pass def _disallow_inside_tf_function(method_name): """Disallow calling a method inside a `tf.function`.""" if ops.inside_function(): error_msg = ( 'Detected a call to `PreprocessingLayer.{method_name}` inside a ' '`tf.function`. `PreprocessingLayer.{method_name} is a high-level ' 'endpoint that manages its own `tf.function`. Please move the call ' 'to `PreprocessingLayer.{method_name}` outside of all enclosing ' '`tf.function`s. Note that you can call a `PreprocessingLayer` ' 'directly on `Tensor`s inside a `tf.function` like: `layer(x)`, ' 'or update its state like: `layer.update_state(x)`.').format( method_name=method_name) raise RuntimeError(error_msg)