# 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. # ============================================================================== """Utilities for `Model.compile`.""" import copy import tensorflow.compat.v2 as tf from keras import losses as losses_mod from keras import metrics as metrics_mod from keras.saving.experimental import saving_lib from keras.utils import generic_utils from keras.utils import losses_utils from keras.utils import tf_utils class Container: """Base Container class.""" def __init__(self, output_names=None): self._output_names = output_names def build(self, y_pred): if self._output_names is None: # In Subclass API, output names like 'output_1' are used for # `Metric` names. self._output_names = create_pseudo_output_names(y_pred) def _conform_to_outputs(self, outputs, struct): """Convenience method to conform `struct` to `outputs` structure. Mappings performed: (1) Map a dict to a list of outputs, using the output names. (2) Fill missing keys in a dict w/ `None`s. (3) Map a single item to all outputs. Args: outputs: Model predictions. struct: Arbitrary nested structure (e.g. of labels, sample_weights, losses, or metrics). Returns: Mapping of `struct` to `outputs` structure. """ struct = map_to_output_names(outputs, self._output_names, struct) struct = map_missing_dict_keys(outputs, struct) # Allow passing one object that applies to all outputs. if not tf.nest.is_nested(struct) and tf.nest.is_nested(outputs): struct = tf.nest.map_structure(lambda _: struct, outputs) return struct def _maybe_broadcast_to_outputs(self, outputs, objects): """Determines if losses / metrics should be applied to all outputs. NOTE: This method should only be called for Metrics / Losses, not for y_true / sample_weight. Args: outputs: Model predictions. objects: Arbitrary nested structure (e.g. of losses or metrics) Returns: Arbitrary nested structure of objects, maybe copied to each output. Applies a Loss / Metric to all outputs. """ if not self._should_broadcast(objects): return objects # When there is more than one Model output, this is needed to keep # each Metric / Loss separate. When there is only one Model output, # the user-supplied object should be used. should_copy_objects = len(tf.nest.flatten(outputs)) > 1 def _broadcast_fn(): if should_copy_objects: return tf.nest.map_structure(self._copy_object, objects) return objects return tf.nest.map_structure(lambda _: _broadcast_fn(), outputs) def _should_broadcast(self, objects): raise NotImplementedError def _copy_object(self, obj): raise NotImplementedError class LossesContainer(Container): """A container class for losses passed to `Model.compile()`. Args: losses: Struct of loss function(s). See `Model.compile()` doc for more information. loss_weights: Weights of the losses contributions of different model outputs. See `Model.compile()` doc for more information. output_names: List of string. Per-output metric names. total_loss_mean: A `keras.metrics.Mean` instance that is used to track the mean of all losses (including compiled and regularization losses). """ def __init__( self, losses, loss_weights=None, output_names=None, total_loss_mean=None ): super(LossesContainer, self).__init__(output_names=output_names) # Keep user-supplied values untouched for recompiling and serialization. self._user_losses = losses self._user_loss_weights = loss_weights self._losses = losses self._loss_weights = loss_weights self._per_output_metrics = None # Per-output losses become metrics. # Mean of the total loss. self._total_loss_mean = total_loss_mean or metrics_mod.Mean(name="loss") self._built = False def get_config(self): # In case `self._losses` is a single string where we convert it to a # list. self._losses = tf.nest.flatten(self._losses) return { "losses": [ saving_lib.serialize_keras_object(obj) for obj in self._losses if obj is not None ], "total_loss_mean": saving_lib.serialize_keras_object( self._total_loss_mean ), } @classmethod def from_config(cls, config): """Returns the `LossesContainer` instance given the `config`.""" deserialized_config = {} for key, value in config.items(): if isinstance(value, list): deserialized_config[key] = [ saving_lib.deserialize_keras_object(item) for item in value ] else: deserialized_config[key] = saving_lib.deserialize_keras_object( value ) return cls(**deserialized_config) @property def metrics(self): """Per-output loss metrics.""" if not self._built: return [] per_output_metrics = [ metric_obj for metric_obj in tf.nest.flatten(self._per_output_metrics) if metric_obj is not None ] return [self._total_loss_mean] + per_output_metrics def build(self, y_pred): """One-time setup of loss objects.""" super(LossesContainer, self).build(y_pred) self._losses = self._maybe_broadcast_to_outputs(y_pred, self._losses) self._losses = self._conform_to_outputs(y_pred, self._losses) self._losses = tf.nest.map_structure( self._get_loss_object, self._losses ) self._losses = tf.nest.flatten(self._losses) self._loss_weights = self._maybe_broadcast_to_outputs( y_pred, self._loss_weights ) self._loss_weights = self._conform_to_outputs( y_pred, self._loss_weights ) self._loss_weights = tf.nest.flatten(self._loss_weights) self._create_metrics() self._built = True @property def built(self): return self._built def _create_metrics(self): """Creates per-output loss metrics, but only for multi-output Models.""" if len(self._output_names) == 1: self._per_output_metrics = [None] else: self._per_output_metrics = [] for loss_obj, output_name in zip(self._losses, self._output_names): if loss_obj is None: self._per_output_metrics.append(None) else: self._per_output_metrics.append( metrics_mod.Mean(output_name + "_loss") ) def __call__( self, y_true, y_pred, sample_weight=None, regularization_losses=None ): """Computes the overall loss. Args: y_true: An arbitrary structure of Tensors representing the ground truth. y_pred: An arbitrary structure of Tensors representing a Model's outputs. sample_weight: An arbitrary structure of Tensors representing the per-sample loss weights. If one Tensor is passed, it is used for all losses. If multiple Tensors are passed, the structure should match `y_pred`. regularization_losses: Additional losses to be added to the total loss. Returns: The total loss as a `tf.Tensor`, or `None` if no loss results. """ y_true = self._conform_to_outputs(y_pred, y_true) sample_weight = self._conform_to_outputs(y_pred, sample_weight) if not self._built: self.build(y_pred) y_pred = tf.nest.flatten(y_pred) y_true = tf.nest.flatten(y_true) sample_weight = tf.nest.flatten(sample_weight) loss_values = [] # Used for gradient calculation. total_loss_mean_values = [] # Used for loss metric calculation. batch_dim = None zip_args = ( y_true, y_pred, sample_weight, self._losses, self._loss_weights, self._per_output_metrics, ) for y_t, y_p, sw, loss_obj, loss_weight, metric_obj in zip(*zip_args): if ( y_t is None or loss_obj is None ): # Ok to have no loss for an output. continue y_t, y_p, sw = match_dtype_and_rank(y_t, y_p, sw) sw = losses_utils.apply_mask(y_p, sw, losses_utils.get_mask(y_p)) loss_value = loss_obj(y_t, y_p, sample_weight=sw) total_loss_mean_value = loss_value # Correct for the `Mean` loss metrics counting each replica as a # batch. if loss_obj.reduction == losses_utils.ReductionV2.SUM: total_loss_mean_value *= ( tf.distribute.get_strategy().num_replicas_in_sync ) if batch_dim is None: if tf_utils.is_ragged(y_t): batch_dim = y_t.nrows() else: batch_dim = tf.shape(y_t)[0] if metric_obj is not None: metric_obj.update_state( total_loss_mean_value, sample_weight=batch_dim ) if loss_weight is not None: loss_value *= loss_weight total_loss_mean_value *= loss_weight if ( loss_obj.reduction == losses_utils.ReductionV2.SUM_OVER_BATCH_SIZE or loss_obj.reduction == losses_utils.ReductionV2.AUTO ): loss_value = losses_utils.scale_loss_for_distribution( loss_value ) loss_values.append(loss_value) total_loss_mean_values.append(total_loss_mean_value) if regularization_losses: regularization_losses = losses_utils.cast_losses_to_common_dtype( regularization_losses ) reg_loss = tf.add_n(regularization_losses) total_loss_mean_values.append(reg_loss) loss_values.append( losses_utils.scale_loss_for_distribution(reg_loss) ) if loss_values: total_loss_mean_values = losses_utils.cast_losses_to_common_dtype( total_loss_mean_values ) total_total_loss_mean_value = tf.add_n(total_loss_mean_values) self._total_loss_mean.update_state( total_total_loss_mean_value, sample_weight=batch_dim ) loss_values = losses_utils.cast_losses_to_common_dtype(loss_values) total_loss = tf.add_n(loss_values) return total_loss else: return None def reset_state(self): """Resets the state of loss metrics.""" if not self._built: return metrics = [self._total_loss_mean] + tf.nest.flatten( self._per_output_metrics ) for metric_obj in metrics: if metric_obj is not None: metric_obj.reset_state() def _get_loss_object(self, loss): """Returns a `Loss` object. Converts the user-supplied loss to a `Loss` object. Also allows `SUM_OVER_BATCH_SIZE` reduction to be used for this loss. Args: loss: A string, function, or `Loss` object. Returns: A `Loss` object. """ if loss is None: return None # Ok to have no loss for an output. loss = losses_mod.get(loss) if not isinstance(loss, losses_mod.Loss): loss_name = get_custom_object_name(loss) if loss_name is None: raise ValueError(f"Loss should be a callable, received: {loss}") loss = losses_mod.LossFunctionWrapper(loss, name=loss_name) loss._allow_sum_over_batch_size = True return loss def _should_broadcast(self, obj): return not tf.nest.is_nested(obj) def _copy_object(self, obj): return obj # Losses don't need to be copied. class MetricsContainer(Container): """A container class for metrics passed to `Model.compile`.""" def __init__( self, metrics=None, weighted_metrics=None, output_names=None, from_serialized=False, ): """Initializes a container for metrics. Arguments: metrics: see the `metrics` argument from `tf.keras.Model.compile`. weighted_metrics: see the `weighted_metrics` argument from `tf.keras.Model.compile`. output_names: A list of strings of names of outputs for the model. from_serialized: Whether the model being compiled is from a serialized model. Used to avoid redundantly applying pre-processing renaming steps. """ super(MetricsContainer, self).__init__(output_names=output_names) self._check_duplicated_metrics(metrics, weighted_metrics) # Keep user-supplied values untouched for recompiling and serialization. self._user_metrics = metrics self._user_weighted_metrics = weighted_metrics self._metrics = metrics self._weighted_metrics = weighted_metrics self._built = False self._from_serialized = from_serialized def _check_duplicated_metrics(self, metrics, weighted_metrics): """Check and raise error when user provided metrics has any duplications. Note that metrics are stateful container, a shared metric instance between model.metric and model.weighted_metric will make the same intance to be udpated twice, and report wrong value. Args: metrics: User provided metrics list. weighted_metrics: User provided weighted metrics list. Raises: ValueError, when duplicated metrics instance discovered in user provided metrics and weighted metrics. """ seen = set() duplicated = [] for x in tf.nest.flatten(metrics) + tf.nest.flatten(weighted_metrics): # We only check metrics object. The string and function objects # will be converted to unique Metric instance. if not isinstance(x, metrics_mod.Metric): continue if x in seen: duplicated.append(x) seen.add(x) if duplicated: raise ValueError( "Found duplicated metrics object in the user provided " "metrics and weighted metrics. This will cause the same " "metric object to be updated multiple times, and report " "wrong results. \n" f"Duplicated items: {duplicated}" ) @property def metrics(self): """All metrics in this container.""" if not self._built: return [] return self._metrics_in_order @property def unweighted_metrics(self): """Metrics in the container that should not be passed sample_weight.""" if not self._built: return None return tf.nest.flatten(self._metrics) @property def weighted_metrics(self): """Metrics in this container that should be passed `sample_weight`.""" if not self._built: return None return tf.nest.flatten(self._weighted_metrics) def build(self, y_pred, y_true): """One-time setup of metric objects.""" super(MetricsContainer, self).build(y_pred) self._metrics = self._maybe_broadcast_to_outputs(y_pred, self._metrics) self._metrics = self._conform_to_outputs(y_pred, self._metrics) self._weighted_metrics = self._maybe_broadcast_to_outputs( y_pred, self._weighted_metrics ) self._weighted_metrics = self._conform_to_outputs( y_pred, self._weighted_metrics ) # Standardize on tuple since `tf.data` turns lists into `Tensor`s. y_pred = tf.__internal__.nest.list_to_tuple(y_pred) y_true = tf.__internal__.nest.list_to_tuple(y_true) self._metrics = tf.__internal__.nest.list_to_tuple(self._metrics) self._weighted_metrics = tf.__internal__.nest.list_to_tuple( self._weighted_metrics ) # Convert to `Metric` objects, potentially disambiguating based on # output properties. self._metrics = tf.__internal__.nest.map_structure_up_to( y_pred, self._get_metric_objects, self._metrics, y_true, y_pred ) self._weighted_metrics = tf.__internal__.nest.map_structure_up_to( y_pred, self._get_metric_objects, self._weighted_metrics, y_true, y_pred, ) self._metrics = tf.__internal__.nest.flatten_up_to( y_pred, self._metrics, check_types=False ) self._weighted_metrics = tf.__internal__.nest.flatten_up_to( y_pred, self._weighted_metrics, check_types=False ) # Assumes metrics, weighted_metrics have been flattened up to outputs. # # If we are loading a model that has been already serialized, we do not # want to re-apply any pre-processing metric renaming steps. if not self._from_serialized: self._set_metric_names() self._create_ordered_metrics() self._built = True @property def built(self): return self._built def _set_metric_names(self): """Sets unique metric names.""" # For multi-output models, prepend the output name to the metric name. # For weighted metrics, prepend "weighted_" if the name would be # non-unique. metric_names = set() is_multi_output = len(self._output_names) > 1 zip_args = (self._output_names, self._metrics, self._weighted_metrics) for output_name, output_metrics, weighted_output_metrics in zip( *zip_args ): for m in output_metrics: if m is None: continue if is_multi_output: m._name = output_name + "_" + m._name if m._name in metric_names: raise ValueError( f"Found two metrics with the same name: {m._name}. " "All the metrics added to the model need to have " "unique names." ) metric_names.add(m._name) for wm in weighted_output_metrics: if wm is None: continue if is_multi_output: if output_name + "_" + wm._name in metric_names: wm._name = output_name + "_weighted_" + wm._name else: wm._name = output_name + "_" + wm._name elif wm._name in metric_names: wm._name = "weighted_" + wm._name if wm._name in metric_names: raise ValueError( "Found two weighted metrics with the same name: " f"{wm._name}.All the metrics added to the model need " "to have unique names." ) metric_names.add(wm._name) def _create_ordered_metrics(self): """Cache the flat order needed when return metrics, for backcompat.""" self._metrics_in_order = [] for output_metrics, output_weighted_metrics in zip( self._metrics, self._weighted_metrics ): for m in tf.nest.flatten(output_metrics): if m is not None: self._metrics_in_order.append(m) for wm in tf.nest.flatten(output_weighted_metrics): if wm is not None: self._metrics_in_order.append(wm) def update_state(self, y_true, y_pred, sample_weight=None): """Updates the state of per-output metrics.""" y_true = self._conform_to_outputs(y_pred, y_true) sample_weight = self._conform_to_outputs(y_pred, sample_weight) if not self._built: self.build(y_pred, y_true) y_pred = tf.nest.flatten(y_pred) y_true = tf.nest.flatten(y_true) if y_true is not None else [] sample_weight = tf.nest.flatten(sample_weight) zip_args = ( y_true, y_pred, sample_weight, self._metrics, self._weighted_metrics, ) for y_t, y_p, sw, metric_objs, weighted_metric_objs in zip(*zip_args): # Ok to have no metrics for an output. if y_t is None or ( all(m is None for m in metric_objs) and all(wm is None for wm in weighted_metric_objs) ): continue y_t, y_p, sw = match_dtype_and_rank(y_t, y_p, sw) mask = losses_utils.get_mask(y_p) sw = losses_utils.apply_mask(y_p, sw, mask) for metric_obj in metric_objs: if metric_obj is None: continue metric_obj.update_state(y_t, y_p, sample_weight=mask) for weighted_metric_obj in weighted_metric_objs: if weighted_metric_obj is None: continue weighted_metric_obj.update_state(y_t, y_p, sample_weight=sw) def reset_state(self): """Resets the state of all `Metric`s in this container.""" if self._built: metrics = self._metrics_in_order else: # If the user supplied `Metric` objects directly, we should # reset those. This could also contain `str`s or `function`s # though. metrics = tf.nest.flatten(self._user_metrics) + tf.nest.flatten( self._user_weighted_metrics ) for metric_obj in metrics: if isinstance(metric_obj, metrics_mod.Metric): metric_obj.reset_state() def _get_metric_objects(self, metrics, y_t, y_p): """Convert user-supplied metrics to `Metric` objects.""" metrics = tf.nest.flatten(metrics) return [self._get_metric_object(m, y_t, y_p) for m in metrics] def _get_metric_object(self, metric, y_t, y_p): """Converts user-supplied metric to a `Metric` object. Args: metric: A string, function, or `Metric` object. y_t: Sample of label. y_p: Sample of output. Returns: A `Metric` object. """ if metric is None: return None # Ok to have no metric for an output. # Convenience feature for selecting b/t binary, categorical, # and sparse categorical. if str(metric).lower() not in ["accuracy", "acc", "crossentropy", "ce"]: metric_obj = metrics_mod.get(metric) else: y_t_rank = len(y_t.shape.as_list()) y_p_rank = len(y_p.shape.as_list()) y_t_last_dim = y_t.shape.as_list()[-1] y_p_last_dim = y_p.shape.as_list()[-1] is_binary = y_p_last_dim == 1 is_sparse_categorical = ( y_t_rank < y_p_rank or y_t_last_dim == 1 and y_p_last_dim > 1 ) if str(metric).lower() in ["accuracy", "acc"]: if is_binary: metric_obj = metrics_mod.binary_accuracy elif is_sparse_categorical: metric_obj = metrics_mod.sparse_categorical_accuracy else: metric_obj = metrics_mod.categorical_accuracy else: if is_binary: metric_obj = metrics_mod.binary_crossentropy elif is_sparse_categorical: metric_obj = metrics_mod.sparse_categorical_crossentropy else: metric_obj = metrics_mod.categorical_crossentropy if isinstance(metric_obj, losses_mod.Loss): metric_obj._allow_sum_over_batch_size = True if not isinstance(metric_obj, metrics_mod.Metric): if isinstance(metric, str): metric_name = metric else: metric_name = get_custom_object_name(metric) if metric_name is None: raise ValueError( f"Metric should be a callable, received: {metric}" ) metric_obj = metrics_mod.MeanMetricWrapper( metric_obj, name=metric_name ) return metric_obj def _should_broadcast(self, obj): # e.g. 'mse'. if not tf.nest.is_nested(obj): return True # e.g. ['mse'] or ['mse', 'mae']. return isinstance(obj, (list, tuple)) and not any( tf.nest.is_nested(o) for o in obj ) def _copy_object(self, obj): if isinstance(obj, metrics_mod.Metric): return obj.__class__.from_config(obj.get_config()) return obj # Can be a function or `None`. def create_pseudo_output_names(outputs): """Create pseudo output names for a subclassed Model.""" return _create_pseudo_names(outputs, prefix="output_") def create_pseudo_input_names(inputs): """Create pseudo input names for a subclassed Model.""" return _create_pseudo_names(inputs, prefix="input_") def _create_pseudo_names(tensors, prefix): """Creates pseudo {input | output} names for subclassed Models. Warning: this function should only be used to define default names for `Metics` and `SavedModel`. No other use cases should rely on a `Model`'s input or output names. Example with dict: `{'a': [x1, x2], 'b': x3}` becomes: `['a_1', 'a_2', 'b']` Example with list: `[x, y]` becomes: `['output_1', 'output_2']` Args: tensors: `Model`'s outputs or inputs. prefix: 'output_' for outputs, 'input_' for inputs. Returns: Flattened list of pseudo names. """ def one_index(ele): # Start with "output_1" instead of "output_0". if isinstance(ele, int): return ele + 1 return ele flat_paths = list(tf.__internal__.nest.yield_flat_paths(tensors)) flat_paths = tf.nest.map_structure(one_index, flat_paths) names = [] for path in flat_paths: if not path: name = prefix + "1" # Single output. else: name = "_".join(str(p) for p in path) if isinstance(path[0], int): name = prefix + name names.append(name) return names def map_to_output_names(y_pred, output_names, struct): """Maps a dict to a list using `output_names` as keys. This is a convenience feature only. When a `Model`'s outputs are a list, you can specify per-output losses and metrics as a dict, where the keys are the output names. If you specify per-output losses and metrics via the same structure as the `Model`'s outputs (recommended), no mapping is performed. For the Functional API, the output names are the names of the last layer of each output. For the Subclass API, the output names are determined by `create_pseudo_output_names` (For example: `['output_1', 'output_2']` for a list of outputs). This mapping preserves backwards compatibility for `compile` and `fit`. Args: y_pred: Sample outputs of the Model, to determine if this convenience feature should be applied (`struct` is returned unmodified if `y_pred` isn't a flat list). output_names: List. The names of the outputs of the Model. struct: The structure to map. Returns: `struct` mapped to a list in same order as `output_names`. """ single_output = not tf.nest.is_nested(y_pred) outputs_are_flat_list = ( not single_output and isinstance(y_pred, (list, tuple)) and not any(tf.nest.is_nested(y_p) for y_p in y_pred) ) if (single_output or outputs_are_flat_list) and isinstance(struct, dict): output_names = output_names or create_pseudo_output_names(y_pred) struct = copy.copy(struct) new_struct = [struct.pop(name, None) for name in output_names] if struct: raise ValueError( "Found unexpected losses or metrics that do not correspond " f"to any Model output: {struct.keys()}. " f"Valid mode output names: {output_names}. " f"Received struct is: {struct}." ) if len(new_struct) == 1: return new_struct[0] return new_struct else: return struct def map_missing_dict_keys(y_pred, struct): """Replaces missing dict keys in `struct` with `None` placeholders.""" if not isinstance(y_pred, dict) or not isinstance(struct, dict): return struct struct = copy.copy(struct) for k in y_pred.keys(): if k not in struct: struct[k] = None return struct def match_dtype_and_rank(y_t, y_p, sw): """Match dtype and rank of predictions.""" if y_t.shape.rank == 1 and y_p.shape.rank == 2: y_t = tf.expand_dims(y_t, axis=-1) if sw is not None: if sw.shape.rank == 1 and y_p.shape.rank == 2: sw = tf.expand_dims(sw, axis=-1) # Dtype. # This is required mainly for custom loss functions which do not take care # casting dtypes. if (y_t.dtype.is_floating and y_p.dtype.is_floating) or ( y_t.dtype.is_integer and y_p.dtype.is_integer ): y_t = tf.cast(y_t, y_p.dtype) if sw is not None: sw = tf.cast(sw, y_p.dtype) return y_t, y_p, sw def get_custom_object_name(obj): """Returns the name to use for a custom loss or metric callable. Args: obj: Custom loss of metric callable Returns: Name to use, or `None` if the object was not recognized. """ if hasattr(obj, "name"): # Accept `Loss` instance as `Metric`. return obj.name elif hasattr(obj, "__name__"): # Function. return obj.__name__ elif hasattr(obj, "__class__"): # Class instance. return generic_utils.to_snake_case(obj.__class__.__name__) else: # Unrecognized object. return None