# 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. # ============================================================================== """Adapter module that convert different input data objects into tf.dataset.""" import abc import contextlib import functools import itertools import math import random import numpy as np from tensorflow.python.data.experimental.ops import cardinality from tensorflow.python.data.ops import dataset_ops from tensorflow.python.data.ops import iterator_ops from tensorflow.python.data.ops import options as options_lib from tensorflow.python.distribute import distribution_strategy_context as ds_context from tensorflow.python.distribute import input_lib from tensorflow.python.eager import context from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors from tensorflow.python.framework import ops from tensorflow.python.framework import smart_cond from tensorflow.python.framework import sparse_tensor from tensorflow.python.framework import tensor_shape from tensorflow.python.keras import backend from tensorflow.python.keras.engine import training_utils from tensorflow.python.keras.utils import data_utils from tensorflow.python.keras.utils import dataset_creator from tensorflow.python.keras.utils import tf_utils from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import random_ops from tensorflow.python.ops import script_ops from tensorflow.python.platform import tf_logging as logging from tensorflow.python.util import nest from tensorflow.python.util.tf_export import keras_export class DataAdapter(object, metaclass=abc.ABCMeta): """Base class for input data adapter. In TF 2.0, tf.data is the preferred API for user to feed in data. In order to simplify the training code path, all the input data object will be converted to `tf.data.Dataset` if possible. Note that since this class is mainly targeted for TF 2.0, it might have a lot of assumptions under the hood, eg eager context by default, distribution strategy, etc. In the meantime, some legacy feature support might be dropped, eg, Iterator from dataset API in v1, etc. The sample usage of this class is like: ``` x = tf.data.Dataset.range(100) adapter_cls = [NumpyArrayDataAdapter, ..., DatasetAdapter] applicable_adapters = [cls for cls in adapter_cls if cls.can_handle(x)] if len(applicable_adapters) != 1: raise ValueError("Expect only one adapter class to handle the input") dataset = applicable_adapters[0](x).get_dataset() for data in dataset: # training ``` """ @staticmethod def can_handle(x, y=None): """Whether the current DataAdapter could handle the input x and y. Structure wise, x and y can be single object, or list of objects if there multiple input/output, or dictionary of objects when the intput/output are named. Args: x: input features. y: target labels. Note that y could be None in the case of prediction. Returns: boolean """ raise NotImplementedError @abc.abstractmethod def __init__(self, x, y=None, **kwargs): """Create a DataAdapter based on data inputs. The caller must make sure to call `can_handle()` first before invoking this method. Provide unsupported data type will result into unexpected behavior. Args: x: input features. y: target labels. Note that y could be None in the case of prediction. **kwargs: Other keyword arguments for DataAdapter during the construction of the tf.dataset.Dataset. For example: - Numpy data might have `sample_weights` which will be used for weighting the loss function during training. - Numpy data might need to have `batch_size` parameter when constructing the dataset and iterator. - Certain input might need to be distribution strategy aware. When `distribution_strategy` is passed, the created dataset need to respect the strategy. DataAdapter might choose to ignore any keyword argument if it doesn't use it, or raise exception if any required argument is not provide. """ if not self.can_handle(x, y): raise ValueError("{} Cannot handle input {}, {}".format( self.__class__, x, y)) @abc.abstractmethod def get_dataset(self): """Get a dataset instance for the current DataAdapter. Note that the dataset returned does not repeat for epoch, so caller might need to create new iterator for the same dataset at the beginning of the epoch. This behavior might change in future. Returns: An tf.dataset.Dataset. Caller might use the dataset in different context, eg iter(dataset) in eager to get the value directly, or in graph mode, provide the iterator tensor to Keras model function. """ raise NotImplementedError @abc.abstractmethod def get_size(self): """Return the size (number of batches) for the dataset created. For certain type of the data input, the number of batches is known, eg for Numpy data, the size is same as (number_of_element / batch_size). Whereas for dataset or python generator, the size is unknown since it may or may not have a end state. Returns: int, the number of batches for the dataset, or None if it is unknown. The caller could use this to control the loop of training, show progress bar, or handle unexpected StopIteration error. """ raise NotImplementedError @abc.abstractmethod def batch_size(self): """Return the batch size of the dataset created. For certain type of the data input, the batch size is known, and even required, like numpy array. Where as for dataset, the batch is unknown unless we take a peek. Returns: int, the batch size of the dataset, or None if it is unknown. """ raise NotImplementedError def representative_batch_size(self): """Return a representative size for batches in the dataset. This is not guaranteed to be the batch size for all batches in the dataset. It just needs to be a rough approximation for batch sizes in the dataset. Returns: int, a representative size for batches found in the dataset, or None if it is unknown. """ return self.batch_size() @abc.abstractmethod def has_partial_batch(self): """Whether the dataset has partial batch at the end.""" raise NotImplementedError @abc.abstractmethod def partial_batch_size(self): """The size of the final partial batch for dataset. Will return None if has_partial_batch is False or batch_size is None. """ raise NotImplementedError @abc.abstractmethod def should_recreate_iterator(self): """Returns whether a new iterator should be created every epoch.""" raise NotImplementedError def get_samples(self): """Returns number of samples in the data, or `None`.""" if not self.get_size() or not self.batch_size(): return None total_sample = self.get_size() * self.batch_size() if self.has_partial_batch(): total_sample -= (self.batch_size() - self.partial_batch_size()) return total_sample def on_epoch_end(self): """A hook called after each epoch.""" pass class TensorLikeDataAdapter(DataAdapter): """Adapter that handles Tensor-like objects, e.g. EagerTensor and NumPy.""" @staticmethod def can_handle(x, y=None): # TODO(kaftan): Check performance implications of using a flatten # here for other types of inputs. flat_inputs = nest.flatten(x) if y is not None: flat_inputs += nest.flatten(y) tensor_types = _get_tensor_types() def _is_tensor(v): if isinstance(v, tensor_types): return True return False return all(_is_tensor(v) for v in flat_inputs) def __init__(self, x, y=None, sample_weights=None, sample_weight_modes=None, batch_size=None, epochs=1, steps=None, shuffle=False, **kwargs): super(TensorLikeDataAdapter, self).__init__(x, y, **kwargs) x, y, sample_weights = _process_tensorlike((x, y, sample_weights)) sample_weight_modes = broadcast_sample_weight_modes( sample_weights, sample_weight_modes) # If sample_weights are not specified for an output use 1.0 as weights. (sample_weights, _, _) = training_utils.handle_partial_sample_weights( y, sample_weights, sample_weight_modes, check_all_flat=True) inputs = pack_x_y_sample_weight(x, y, sample_weights) num_samples = set(int(i.shape[0]) for i in nest.flatten(inputs)).pop() _check_data_cardinality(inputs) # If batch_size is not passed but steps is, calculate from the input data. # Default to 32 for backwards compat. if not batch_size: batch_size = int(math.ceil(num_samples / steps)) if steps else 32 self._size = int(math.ceil(num_samples / batch_size)) self._batch_size = batch_size num_full_batches = int(num_samples // batch_size) self._partial_batch_size = num_samples % batch_size if isinstance(shuffle, str): shuffle = shuffle.lower() self._shuffle = shuffle # Vectorized version of shuffle. # This is a performance improvement over using `from_tensor_slices`. # The indices of the data are shuffled and batched, and these indices # are then zipped with the data and used to extract a batch of the data # at each step. The performance improvements here come from: # 1. vectorized batch using gather # 2. parallelized map # 3. pipelined permutation generation # 4. optimized permutation batching # 5. disabled static optimizations indices_dataset = dataset_ops.DatasetV2.range(1) if shuffle != "batch": indices_dataset = indices_dataset.repeat(epochs) def permutation(_): # It turns out to be more performant to make a new set of indices rather # than reusing the same range Tensor. (presumably because of buffer # forwarding.) indices = math_ops.range(num_samples, dtype=dtypes.int64) if shuffle and shuffle != "batch": indices = random_ops.random_shuffle(indices) return indices # We prefetch a single element. Computing large permutations can take quite # a while so we don't want to wait for prefetching over an epoch boundary to # trigger the next permutation. On the other hand, too many simultaneous # shuffles can contend on a hardware level and degrade all performance. indices_dataset = indices_dataset.map(permutation).prefetch(1) def slice_batch_indices(indices): """Convert a Tensor of indices into a dataset of batched indices. This step can be accomplished in several ways. The most natural is to slice the Tensor in a Dataset map. (With a condition on the upper index to handle the partial batch.) However it turns out that coercing the Tensor into a shape which is divisible by the batch size (and handling the last partial batch separately) allows for a much more favorable memory access pattern and improved performance. Args: indices: Tensor which determines the data order for an entire epoch. Returns: A Dataset of batched indices. """ num_in_full_batch = num_full_batches * batch_size first_k_indices = array_ops.slice(indices, [0], [num_in_full_batch]) first_k_indices = array_ops.reshape( first_k_indices, [num_full_batches, batch_size]) flat_dataset = dataset_ops.DatasetV2.from_tensor_slices(first_k_indices) if self._partial_batch_size: index_remainder = dataset_ops.DatasetV2.from_tensors(array_ops.slice( indices, [num_in_full_batch], [self._partial_batch_size])) flat_dataset = flat_dataset.concatenate(index_remainder) if shuffle == "batch": # 1024 is a magic constant that has not been properly evaluated flat_dataset = flat_dataset.shuffle(1024).repeat(epochs) return flat_dataset indices_dataset = indices_dataset.flat_map(slice_batch_indices) dataset = self.slice_inputs(indices_dataset, inputs) if shuffle == "batch": def shuffle_batch(*batch): return nest.map_structure(random_ops.random_shuffle, batch) dataset = dataset.map(shuffle_batch) self._dataset = dataset def slice_inputs(self, indices_dataset, inputs): """Slice inputs into a Dataset of batches. Given a Dataset of batch indices and the unsliced inputs, this step slices the inputs in a parallelized fashion and produces a dataset of input batches. Args: indices_dataset: A Dataset of batched indices inputs: A python data structure that contains the inputs, targets, and possibly sample weights. Returns: A Dataset of input batches matching the batch indices. """ dataset = dataset_ops.DatasetV2.zip(( indices_dataset, dataset_ops.DatasetV2.from_tensors(inputs).repeat() )) def grab_batch(i, data): return nest.map_structure(lambda d: array_ops.gather(d, i, axis=0), data) dataset = dataset.map( grab_batch, num_parallel_calls=dataset_ops.AUTOTUNE) # Default optimizations are disabled to avoid the overhead of (unnecessary) # input pipeline graph serialization and deserialization options = options_lib.Options() options.experimental_optimization.apply_default_optimizations = False if self._shuffle: # See b/141490660 for more details. options.experimental_external_state_policy = ( options_lib.ExternalStatePolicy.IGNORE) dataset = dataset.with_options(options) return dataset def get_dataset(self): return self._dataset def get_size(self): return self._size def batch_size(self): return self._batch_size def has_partial_batch(self): return self._partial_batch_size > 0 def partial_batch_size(self): return self._partial_batch_size or None def should_recreate_iterator(self): # An infinite dataset is always created here. return False class GenericArrayLikeDataAdapter(TensorLikeDataAdapter): """Adapter that handles array-like data without forcing it into memory. This adapter handles array-like datasets that may be too big to fully fit into memory. Specifically, this adapter handles any Python class which implements: `__get_item__`, `__len__`, `shape`, and `dtype` with the same meanings as Numpy, but it ignores any case where all the inputs are Tensors or Numpy arrays (because that case is handled by the base TensorLikeDataAdapter). It ignores scipy sparse matrices and Composite Tensors because those are handled by the CompositeTensorDataAdapter. It also does not handle lists/tuples of scalars, because those are handled by the ListsOfScalarsDataAdapter. """ @staticmethod def can_handle(x, y=None): flat_inputs = nest.flatten(x) if y is not None: flat_inputs += nest.flatten(y) def _is_array_like(v): """Return True if v is a Tensor, array, or is array-like.""" return ( hasattr(v, "__getitem__") and hasattr(v, "shape") and hasattr(v, "dtype") and hasattr(v, "__len__") ) if (not TensorLikeDataAdapter.can_handle(x, y) and not CompositeTensorDataAdapter.can_handle(x, y)): return all(_is_array_like(v) for v in flat_inputs) else: return False def __init__(self, *args, **kwargs): logging.warning( "Keras is training/fitting/evaluating on array-like data. Keras may " "not be optimized for this format, so if your input data format is " "supported by TensorFlow I/O (https://github.com/tensorflow/io) we " "recommend using that to load a Dataset instead.") super(GenericArrayLikeDataAdapter, self).__init__(*args, **kwargs) def slice_inputs(self, indices_dataset, inputs): """Slice inputs into a Dataset of batches. Given a Dataset of batch indices and the unsliced inputs, this step slices the inputs in a parallelized fashion and produces a dataset of input batches. Args: indices_dataset: A Dataset of batched indices inputs: A python data structure that contains the inputs, targets, and possibly sample weights. Returns: A Dataset of input batches matching the batch indices. """ flat_inputs = nest.flatten(inputs) def dynamic_shape_like(t): shape = list(t.shape) shape[0] = None return tuple(shape) flat_dtypes = [inp.dtype for inp in flat_inputs] contiguous = True if self._shuffle and self._shuffle != "batch": contiguous = False def grab_batch(indices): """Grab a batch of data from the inputs.""" # This uses a py_function to avoid converting the array-like # into a Tensor before slicing it, because converting the array-like # to a Tensor may force it into memory.. def py_method(ind): def slice_array(data): return training_utils.slice_arrays(data, ind.numpy(), contiguous=contiguous) return [slice_array(inp) for inp in flat_inputs] flat_out = script_ops.eager_py_func(py_method, [indices], flat_dtypes) for v, original_inp in zip(flat_out, flat_inputs): v.set_shape(dynamic_shape_like(original_inp)) return nest.pack_sequence_as(inputs, flat_out) dataset = indices_dataset.map( grab_batch, num_parallel_calls=dataset_ops.AUTOTUNE) return dataset class DatasetCreatorAdapter(DataAdapter): """Adapter that handles dataset functions.""" def __init__(self, x, y, steps=None, distribution_strategy=None, **kwargs): super(DatasetCreatorAdapter, self).__init__(x, **kwargs) if not isinstance(x, dataset_creator.DatasetCreator): raise TypeError("The input of a `DatasetCreatorAdapter` should be a " "`DatasetCreator` but it received type {}.".format( type(x))) if steps is None: raise ValueError("When using a " "`tf.keras.utils.experimental.DatasetCreator`, " "`steps_per_epoch`, `validation_steps` or `steps` " "argument must be provided in `Model.fit`, " "`Model.evaluate`, or `Model.predict`.") self.dataset_creator = x self.steps = steps self.strategy = distribution_strategy @staticmethod def can_handle(x, y=None): if isinstance(x, dataset_creator.DatasetCreator): assert y is None return True def should_recreate_iterator(self): # We expect users to shuffle the dataset in their `dataset_fn` supplied to # `DatasetCreator`. Since that is a buffered shuffle, we intend to not reset # the dataset so the batches that are not shuffled can still be pulled. return False def get_size(self): return None # To be inferred by `DataHandler`. def get_dataset(self): return self.strategy.distribute_datasets_from_function( self.dataset_creator, options=self.dataset_creator.input_options) def batch_size(self): raise NotImplementedError() def has_partial_batch(self): raise NotImplementedError() def partial_batch_size(self): raise NotImplementedError() class CompositeTensorDataAdapter(DataAdapter): """Adapter that handles composite tensor.""" @staticmethod def can_handle(x, y=None): flat_inputs = nest.flatten(x) if y is not None: flat_inputs += nest.flatten(y) def _is_composite(v): # Dataset/iterator/DistributedDataset inherits from CompositeTensor but # should be handled by DatasetAdapter and GeneratorAdapter. if (tf_utils.is_extension_type(v) and not isinstance(v, (dataset_ops.DatasetV2, iterator_ops.IteratorBase)) and not _is_distributed_dataset(v)): return True # Support Scipy sparse tensors if scipy is installed return _is_scipy_sparse(v) def _is_tensor_or_composite(v): if isinstance(v, (ops.Tensor, np.ndarray)): return True return _is_composite(v) return (any(_is_composite(v) for v in flat_inputs) and all(_is_tensor_or_composite(v) for v in flat_inputs)) def __init__(self, x, y=None, sample_weights=None, sample_weight_modes=None, batch_size=None, steps=None, shuffle=False, **kwargs): super(CompositeTensorDataAdapter, self).__init__(x, y, **kwargs) x, y, sample_weights = _process_tensorlike((x, y, sample_weights)) sample_weight_modes = broadcast_sample_weight_modes( sample_weights, sample_weight_modes) # If sample_weights are not specified for an output use 1.0 as weights. (sample_weights, _, _) = training_utils.handle_partial_sample_weights( y, sample_weights, sample_weight_modes, check_all_flat=True) inputs = pack_x_y_sample_weight(x, y, sample_weights) dataset = dataset_ops.DatasetV2.from_tensor_slices(inputs) num_samples = int(nest.flatten(x)[0].shape[0]) if shuffle: dataset = dataset.shuffle(num_samples) # If batch_size is not passed but steps is, calculate from the input data. # Default to 32 for backwards compat. if not batch_size: batch_size = int(math.ceil(num_samples / steps)) if steps else 32 dataset = dataset.batch(batch_size) self._size = int(math.ceil(num_samples / batch_size)) self._batch_size = batch_size self._has_partial_batch = (self._size != (num_samples // batch_size)) self._partial_batch_size = None if self._has_partial_batch: self._partial_batch_size = ( num_samples - (self._size - 1) * self._batch_size) self._dataset = dataset def get_dataset(self): return self._dataset def get_size(self): return self._size def batch_size(self): return self._batch_size def has_partial_batch(self): return self._has_partial_batch def partial_batch_size(self): return self._partial_batch_size def should_recreate_iterator(self): return True class ListsOfScalarsDataAdapter(DataAdapter): """Adapter that handles lists of scalars and lists of lists of scalars.""" @staticmethod def can_handle(x, y=None): handles_x = ListsOfScalarsDataAdapter._is_list_of_scalars(x) handles_y = True if y is not None: handles_y = ListsOfScalarsDataAdapter._is_list_of_scalars(y) return handles_x and handles_y @staticmethod def _is_list_of_scalars(inp): if isinstance(inp, (float, int, str, bytes, bytearray)): return True if isinstance(inp, (list, tuple)) and inp: return ListsOfScalarsDataAdapter._is_list_of_scalars(inp[0]) return False def __init__(self, x, y=None, sample_weights=None, sample_weight_modes=None, batch_size=None, shuffle=False, **kwargs): super(ListsOfScalarsDataAdapter, self).__init__(x, y, **kwargs) x = np.asarray(x) if y is not None: y = np.asarray(y) if sample_weights is not None: sample_weights = np.asarray(sample_weights) sample_weight_modes = broadcast_sample_weight_modes( sample_weights, sample_weight_modes) self._internal_adapter = TensorLikeDataAdapter( x, y=y, sample_weights=sample_weights, sample_weight_modes=sample_weight_modes, batch_size=batch_size, shuffle=shuffle, **kwargs) def get_dataset(self): return self._internal_adapter.get_dataset() def get_size(self): return self._internal_adapter.get_size() def batch_size(self): return self._internal_adapter.batch_size() def has_partial_batch(self): return self._internal_adapter.has_partial_batch() def partial_batch_size(self): return self._internal_adapter.partial_batch_size() def should_recreate_iterator(self): return True class DatasetAdapter(DataAdapter): """Adapter that handles `tf.data.Dataset`.""" @staticmethod def can_handle(x, y=None): return (isinstance(x, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)) or _is_distributed_dataset(x)) def __init__(self, x, y=None, sample_weights=None, steps=None, **kwargs): super(DatasetAdapter, self).__init__(x, y, **kwargs) # Note that the dataset instance is immutable, its fine to reuse the user # provided dataset. self._dataset = x # The user-provided steps. self._user_steps = steps self._validate_args(y, sample_weights, steps) def get_dataset(self): return self._dataset def get_size(self): return # Inferred in `DataHandler`. def batch_size(self): return None def has_partial_batch(self): return False def partial_batch_size(self): return None def should_recreate_iterator(self): # Since DistributedDatasets have no cardinality, the user must provide # all steps that need to be run, calling `.repeat()` as needed. if _is_distributed_dataset(self._dataset): return False # If user doesn't supply `steps`, or if they supply `steps` that # exactly equals the size of the `Dataset`, create a new iterator # each epoch. return (self._user_steps is None or cardinality.cardinality(self._dataset).numpy() == self._user_steps) def _validate_args(self, y, sample_weights, steps): """Validates `__init__` arguments.""" # Arguments that shouldn't be passed. if not is_none_or_empty(y): raise ValueError("`y` argument is not supported when using " "dataset as input.") if not is_none_or_empty(sample_weights): raise ValueError("`sample_weight` argument is not supported when using " "dataset as input.") if steps is None: if _is_distributed_dataset(self._dataset): raise ValueError("When providing a distributed dataset, you must " "specify the number of steps to run.") size = cardinality.cardinality(self._dataset).numpy() if size == cardinality.INFINITE and steps is None: raise ValueError( "When providing an infinite dataset, you must specify " "the number of steps to run (if you did not intend to " "create an infinite dataset, make sure to not call " "`repeat()` on the dataset).") class GeneratorDataAdapter(DataAdapter): """Adapter that handles python generators and iterators.""" @staticmethod def can_handle(x, y=None): return ((hasattr(x, "__next__") or hasattr(x, "next")) and hasattr(x, "__iter__") and not isinstance(x, data_utils.Sequence)) def __init__(self, x, y=None, sample_weights=None, workers=1, use_multiprocessing=False, max_queue_size=10, model=None, **kwargs): # Generators should never shuffle as exhausting the generator in order to # shuffle the batches is inefficient. kwargs.pop("shuffle", None) if not is_none_or_empty(y): raise ValueError("`y` argument is not supported when using " "python generator as input.") if not is_none_or_empty(sample_weights): raise ValueError("`sample_weight` argument is not supported when using " "python generator as input.") super(GeneratorDataAdapter, self).__init__(x, y, **kwargs) # Since we have to know the dtype of the python generator when we build the # dataset, we have to look at a batch to infer the structure. peek, x = self._peek_and_restore(x) peek = self._standardize_batch(peek) peek = _process_tensorlike(peek) # Need to build the Model on concrete input shapes. if model is not None and not model.built: concrete_x, _, _ = unpack_x_y_sample_weight(peek) model.distribute_strategy.run( lambda x: model(x, training=False), args=(concrete_x,)) self._first_batch_size = int(nest.flatten(peek)[0].shape[0]) def _get_dynamic_shape(t): shape = t.shape # Unknown number of dimensions, `as_list` cannot be called. if shape.rank is None: return shape return tensor_shape.TensorShape([None for _ in shape.as_list()]) output_shapes = nest.map_structure(_get_dynamic_shape, peek) output_types = nest.map_structure(lambda t: t.dtype, peek) # Note that dataset API takes a callable that creates a generator object, # rather than generator itself, which is why we define a function here. generator_fn = self._handle_multiprocessing(x, workers, use_multiprocessing, max_queue_size) def wrapped_generator(): for data in generator_fn(): yield self._standardize_batch(data) dataset = dataset_ops.DatasetV2.from_generator( wrapped_generator, output_types, output_shapes=output_shapes) if workers == 1 and not use_multiprocessing: dataset = dataset.prefetch(1) self._dataset = dataset def _standardize_batch(self, data): """Standardizes a batch output by a generator.""" # Removes `None`s. x, y, sample_weight = unpack_x_y_sample_weight(data) data = pack_x_y_sample_weight(x, y, sample_weight) data = nest.list_to_tuple(data) def _convert_dtype(t): if (isinstance(t, np.ndarray) and issubclass(t.dtype.type, np.floating)): return np.array(t, dtype=backend.floatx()) return t data = nest.map_structure(_convert_dtype, data) return data @staticmethod def _peek_and_restore(x): peek = next(x) return peek, itertools.chain([peek], x) def _handle_multiprocessing(self, x, workers, use_multiprocessing, max_queue_size): """Create a callable, possibly including an Enqueuer.""" if workers > 1 or (workers > 0 and use_multiprocessing): def generator_fn(): enqueuer = data_utils.GeneratorEnqueuer( x, use_multiprocessing=use_multiprocessing) enqueuer.start(workers=workers, max_queue_size=max_queue_size) return enqueuer.get() else: generator_fn = lambda: x return generator_fn def get_dataset(self): return self._dataset def get_size(self): return None def batch_size(self): return None def representative_batch_size(self): return self._first_batch_size def has_partial_batch(self): return False def partial_batch_size(self): return def should_recreate_iterator(self): return False class KerasSequenceAdapter(GeneratorDataAdapter): """Adapter that handles `keras.utils.Sequence`.""" @staticmethod def can_handle(x, y=None): return isinstance(x, data_utils.Sequence) def __init__(self, x, y=None, sample_weights=None, shuffle=False, workers=1, use_multiprocessing=False, max_queue_size=10, model=None, **kwargs): if not is_none_or_empty(y): raise ValueError("`y` argument is not supported when using " "`keras.utils.Sequence` as input.") if not is_none_or_empty(sample_weights): raise ValueError("`sample_weight` argument is not supported when using " "`keras.utils.Sequence` as input.") self._size = len(x) self._shuffle_sequence = shuffle self._keras_sequence = x self._enqueuer = None super(KerasSequenceAdapter, self).__init__( x, shuffle=False, # Shuffle is handed in the _make_callable override. workers=workers, use_multiprocessing=use_multiprocessing, max_queue_size=max_queue_size, model=model, **kwargs) @staticmethod def _peek_and_restore(x): return x[0], x def _handle_multiprocessing(self, x, workers, use_multiprocessing, max_queue_size): if workers > 1 or (workers > 0 and use_multiprocessing): def generator_fn(): self._enqueuer = data_utils.OrderedEnqueuer( x, use_multiprocessing=use_multiprocessing, shuffle=self._shuffle_sequence) self._enqueuer.start(workers=workers, max_queue_size=max_queue_size) return self._enqueuer.get() else: def generator_fn(): order = range(len(x)) if self._shuffle_sequence: # Match the shuffle convention in OrderedEnqueuer. order = list(order) random.shuffle(order) for i in order: yield x[i] return generator_fn def get_size(self): return self._size def should_recreate_iterator(self): return True def on_epoch_end(self): if self._enqueuer: self._enqueuer.stop() self._keras_sequence.on_epoch_end() ALL_ADAPTER_CLS = [ ListsOfScalarsDataAdapter, TensorLikeDataAdapter, GenericArrayLikeDataAdapter, DatasetAdapter, GeneratorDataAdapter, KerasSequenceAdapter, CompositeTensorDataAdapter, DatasetCreatorAdapter ] def select_data_adapter(x, y): """Selects a data adapter than can handle a given x and y.""" adapter_cls = [cls for cls in ALL_ADAPTER_CLS if cls.can_handle(x, y)] if not adapter_cls: # TODO(scottzhu): This should be a less implementation-specific error. raise ValueError( "Failed to find data adapter that can handle " "input: {}, {}".format( _type_name(x), _type_name(y))) elif len(adapter_cls) > 1: raise RuntimeError( "Data adapters should be mutually exclusive for " "handling inputs. Found multiple adapters {} to handle " "input: {}, {}".format( adapter_cls, _type_name(x), _type_name(y))) return adapter_cls[0] def _type_name(x): """Generates a description of the type of an object.""" if isinstance(x, dict): key_types = set(_type_name(key) for key in x.keys()) val_types = set(_type_name(key) for key in x.values()) return "({} containing {} keys and {} values)".format( type(x), key_types, val_types) if isinstance(x, (list, tuple)): types = set(_type_name(val) for val in x) return "({} containing values of types {})".format( type(x), types) return str(type(x)) def _process_tensorlike(inputs): """Process tensor-like inputs. This function: (1) Converts `Numpy` arrays to `Tensor`s. (2) Converts `Scipy` sparse matrices to `SparseTensor`s. (2) Converts `list`s to `tuple`s (for `tf.data` support). Args: inputs: Structure of `Tensor`s, `NumPy` arrays, or tensor-like. Returns: Structure of `Tensor`s or tensor-like. """ def _convert_numpy_and_scipy(x): if isinstance(x, np.ndarray): dtype = None if issubclass(x.dtype.type, np.floating): dtype = backend.floatx() return ops.convert_to_tensor_v2_with_dispatch(x, dtype=dtype) elif _is_scipy_sparse(x): return _scipy_sparse_to_sparse_tensor(x) return x inputs = nest.map_structure(_convert_numpy_and_scipy, inputs) return nest.list_to_tuple(inputs) def is_none_or_empty(inputs): # util method to check if the input is a None or a empty list. # the python "not" check will raise an error like below if the input is a # numpy array # "The truth value of an array with more than one element is ambiguous. # Use a.any() or a.all()" return inputs is None or not nest.flatten(inputs) def broadcast_sample_weight_modes(target_structure, sample_weight_modes): """Match sample_weight_modes structure with output structure.""" if target_structure is None or not nest.flatten(target_structure): return sample_weight_modes if isinstance(sample_weight_modes, str): if isinstance(target_structure, dict): return {key: sample_weight_modes for key in target_structure.keys()} return [sample_weight_modes for _ in target_structure] if sample_weight_modes: try: nest.assert_same_structure( training_utils.list_to_tuple(target_structure), training_utils.list_to_tuple(sample_weight_modes)) except (ValueError, TypeError): target_str = str(nest.map_structure(lambda _: "...", target_structure)) mode_str = str(nest.map_structure(lambda _: "...", sample_weight_modes)) # Attempt to coerce sample_weight_modes to the target structure. This # implicitly depends on the fact that Model flattens outputs for its # internal representation. try: sample_weight_modes = nest.pack_sequence_as( target_structure, nest.flatten(sample_weight_modes)) logging.warning( "sample_weight modes were coerced from\n {}\n to \n {}" .format(target_str, mode_str)) except (ValueError, TypeError): raise ValueError( "Unable to match target structure and sample_weight_modes " "structure:\n {}\n to \n {}".format(target_str, mode_str)) return sample_weight_modes class DataHandler(object): """Handles iterating over epoch-level `tf.data.Iterator` objects.""" def __init__(self, x, y=None, sample_weight=None, batch_size=None, steps_per_epoch=None, initial_epoch=0, epochs=1, shuffle=False, class_weight=None, max_queue_size=10, workers=1, use_multiprocessing=False, model=None, steps_per_execution=None, distribute=True): """Initializes a `DataHandler`. Arguments: x: See `Model.fit`. y: See `Model.fit`. sample_weight: See `Model.fit`. batch_size: See `Model.fit`. steps_per_epoch: See `Model.fit`. initial_epoch: See `Model.fit`. epochs: See `Model.fit`. shuffle: See `Model.fit`. class_weight: See `Model.fit`. max_queue_size: See `Model.fit`. workers: See `Model.fit`. use_multiprocessing: See `Model.fit`. model: The `Model` instance. Needed in order to correctly `build` the `Model` using generator-like inputs (see `GeneratorDataAdapter`). steps_per_execution: See `Model.compile`. distribute: Whether to distribute the `tf.dataset`. `PreprocessingLayer.adapt` does not support distributed datasets, `Model` should always set this to `True`. """ self._initial_epoch = initial_epoch self._epochs = epochs self._insufficient_data = False self._model = model # `steps_per_execution_value` is the cached initial value. # `steps_per_execution` is mutable and may be changed by the DataAdapter # to handle partial executions. if steps_per_execution is None: self._steps_per_execution = 1 self._steps_per_execution_value = 1 else: self._steps_per_execution = steps_per_execution self._steps_per_execution_value = steps_per_execution.numpy().item() adapter_cls = select_data_adapter(x, y) self._adapter = adapter_cls( x, y, batch_size=batch_size, steps=steps_per_epoch, epochs=epochs - initial_epoch, sample_weights=sample_weight, shuffle=shuffle, max_queue_size=max_queue_size, workers=workers, use_multiprocessing=use_multiprocessing, distribution_strategy=ds_context.get_strategy(), model=model) strategy = ds_context.get_strategy() self._current_step = 0 self._step_increment = self._steps_per_execution_value - 1 self._insufficient_data = False self._configure_dataset_and_inferred_steps(strategy, x, steps_per_epoch, class_weight, distribute) def _configure_dataset_and_inferred_steps(self, strategy, x, steps_per_epoch, class_weight, distribute): """Configure the `_dataset` and `_inferred_steps` attributes.""" del x dataset = self._adapter.get_dataset() if class_weight: dataset = dataset.map(_make_class_weight_map_fn(class_weight)) self._inferred_steps = self._infer_steps(steps_per_epoch, dataset) # `PreprocessingLayer.adapt` does not currently support distributed # datasets, so we pass `distribute=False` there. if distribute and not _is_distributed_dataset(dataset): dataset = strategy.experimental_distribute_dataset(dataset) self._dataset = dataset self._validate_data_handler() def enumerate_epochs(self): """Yields `(epoch, tf.data.Iterator)`.""" with self._truncate_execution_to_epoch(): data_iterator = iter(self._dataset) for epoch in range(self._initial_epoch, self._epochs): if self._insufficient_data: # Set by `catch_stop_iteration`. break if self._adapter.should_recreate_iterator(): data_iterator = iter(self._dataset) yield epoch, data_iterator self._adapter.on_epoch_end() @contextlib.contextmanager def _truncate_execution_to_epoch(self): """Truncates steps per execution to at most one epoch.""" should_truncate = ( self._inferred_steps is not None and self._steps_per_execution_value > self._inferred_steps) original_value = self._steps_per_execution_value try: if should_truncate: self._steps_per_execution.assign(self._inferred_steps) self._steps_per_execution_value = self._inferred_steps yield finally: if should_truncate: self._steps_per_execution.assign(original_value) self._steps_per_execution_value = original_value def sync(self): context.async_wait() @contextlib.contextmanager def catch_stop_iteration(self): """Catches errors when an iterator runs out of data.""" try: yield self.sync() except (StopIteration, errors.OutOfRangeError): if self._inferred_steps is None: self._inferred_steps = self._current_step else: self._insufficient_data = True total_epochs = self._epochs - self._initial_epoch logging.warning( "Your input ran out of data; interrupting training. " "Make sure that your dataset or generator can generate at " "least `steps_per_epoch * epochs` batches (in this case, " "{} batches). You may need to use the repeat() function " "when building your dataset.".format(total_epochs * self._inferred_steps)) def steps(self): """Yields steps for the current epoch.""" self._current_step = 0 # `self._inferred_steps` can be changed by `catch_stop_iteration`. while (self._inferred_steps is None or self._current_step < self._inferred_steps): if self._insufficient_data: # Set by `catch_stop_iteration`. break can_run_full_execution = ( self._steps_per_execution_value == 1 or self._inferred_steps is None or self._inferred_steps - self._current_step >= self._steps_per_execution_value) if can_run_full_execution: self._step_increment = self._steps_per_execution_value - 1 yield self._current_step self._current_step += self._steps_per_execution_value else: # Last partial execution. steps_remaining = self._inferred_steps - self._current_step self._steps_per_execution.assign(steps_remaining) self._step_increment = steps_remaining - 1 yield self._current_step self._current_step += steps_remaining self._steps_per_execution.assign(self._steps_per_execution_value) @property def step_increment(self): """The number to increment the step for `on_batch_end` methods.""" return self._step_increment @property def inferred_steps(self): """The inferred steps per epoch of the created `Dataset`. This will be `None` in the case where: (1) A `Dataset` of unknown cardinality was passed to the `DataHandler`, and (2) `steps_per_epoch` was not provided, and (3) The first epoch of iteration has not yet completed. Returns: The inferred steps per epoch of the created `Dataset`. """ return self._inferred_steps @property def should_sync(self): # Catch OutOfRangeError for Datasets of unknown size. # This blocks until the batch has finished executing. # TODO(b/150292341): Allow multiple async steps here. return self._inferred_steps is None def _log_indefinite_training_warning(self): logging.warning("The training loop will run indefinitely since you have " "set `steps_per_epoch=-1`. Please use batch-level " "callbacks to save checkpoints or log training progress, " "etc") def _infer_steps(self, steps, dataset): """Infers steps_per_epoch needed to loop through a dataset.""" if steps == -1: self._log_indefinite_training_warning() return None if steps is not None: return steps adapter_steps = self._adapter.get_size() if adapter_steps is not None: return adapter_steps size = cardinality.cardinality(dataset) if size == cardinality.INFINITE and steps is None: raise ValueError( "When passing an infinitely repeating dataset, please specify a " "`steps_per_epoch` value so that epoch level " "callbacks continue to work. The value can be arbitrary, or a number " "that you think correctly defines the size of an epoch. " "Epoch-level callbacks will then be called at this interval.") if size >= 0: return size.numpy().item() return None @property def _samples(self): return self._adapter.get_samples() def _validate_data_handler(self): # TODO(b/152094471): Support this with DistIter.get_next_as_optional. if self._steps_per_execution_value > 1 and self._inferred_steps is None: raise ValueError( "Could not infer the size of the data. With " "`steps_per_execution > 1`, you must specify the number of steps " "to run.") class _ClusterCoordinatorDataHandler(DataHandler): """A `DataHandler` that is compatible with `ClusterCoordinator`.""" def __init__(self, x, y=None, **kwargs): if not isinstance(x, dataset_creator.DatasetCreator): x = self._convert_to_dataset_creator(x, y, **kwargs) super().__init__(x=x, **kwargs) def _convert_to_dataset_creator(self, x, y, **kwargs): """Converts non-tf.data.Dataset to `DatasetCreator` instances.""" def _dataset_fn(input_context): del input_context data_adapter_cls = select_data_adapter(x, y) return data_adapter_cls(x=x, y=y, **kwargs).get_dataset() # This check is needed because types like `tf.data.Dataset` don't work with # PSS yet. So only apply this logic to the types we can support. if (isinstance(x, _get_tensor_types()) and isinstance(y, _get_tensor_types())): return dataset_creator.DatasetCreator(_dataset_fn) else: raise NotImplementedError( "Only `tf.keras.utils.experimental.DatasetCreator`, `tf.Tensor`, " "numpy arrays and pandas dataframes are supported types at this " "time.") def _configure_dataset_and_inferred_steps(self, strategy, x, steps_per_epoch, class_weight, distribute): if not isinstance(x, dataset_creator.DatasetCreator): raise TypeError("When using `ParameterServerStrategy`, `x` must be a " "`DatasetCreator`.") def per_worker_dataset_fn(): return strategy.distribute_datasets_from_function( x, options=x.input_options) self._dataset = self._model._cluster_coordinator.create_per_worker_dataset( # pylint: disable=protected-access per_worker_dataset_fn) if steps_per_epoch == -1: self._inferred_steps = None self._log_indefinite_training_warning() else: self._inferred_steps = steps_per_epoch def sync(self): self._model._cluster_coordinator.join() # pylint: disable=protected-access def get_data_handler(*args, **kwargs): if getattr(kwargs["model"], "_cluster_coordinator", None): return _ClusterCoordinatorDataHandler(*args, **kwargs) return DataHandler(*args, **kwargs) def _make_class_weight_map_fn(class_weight): """Applies class weighting to a `Dataset`. The `Dataset` is assumed to be in format `(x, y)` or `(x, y, sw)`, where `y` must be a single `Tensor`. Args: class_weight: A map where the keys are integer class ids and values are the class weights, e.g. `{0: 0.2, 1: 0.6, 2: 0.3}` Returns: A function that can be used with `tf.data.Dataset.map` to apply class weighting. """ class_ids = list(sorted(class_weight.keys())) expected_class_ids = list(range(len(class_ids))) if class_ids != expected_class_ids: error_msg = ( "Expected `class_weight` to be a dict with keys from 0 to one less " "than the number of classes, found {}").format(class_weight) raise ValueError(error_msg) class_weight_tensor = ops.convert_to_tensor_v2_with_dispatch( [class_weight[int(c)] for c in class_ids]) def _class_weights_map_fn(*data): """Convert `class_weight` to `sample_weight`.""" x, y, sw = unpack_x_y_sample_weight(data) if nest.is_nested(y): raise ValueError( "`class_weight` is only supported for Models with a single output.") if y.shape.rank > 2: raise ValueError("`class_weight` not supported for " "3+ dimensional targets.") y_classes = smart_cond.smart_cond( y.shape.rank == 2 and backend.shape(y)[1] > 1, lambda: backend.argmax(y, axis=1), lambda: math_ops.cast(backend.reshape(y, (-1,)), dtypes.int64)) cw = array_ops.gather_v2(class_weight_tensor, y_classes) if sw is not None: cw = math_ops.cast(cw, sw.dtype) sw, cw = expand_1d((sw, cw)) # `class_weight` and `sample_weight` are multiplicative. sw = sw * cw else: sw = cw return x, y, sw return _class_weights_map_fn def expand_1d(data): """Expands 1-dimensional `Tensor`s into 2-dimensional `Tensor`s.""" def _expand_single_1d_tensor(t): # Leaves `CompositeTensor`s as-is. if (isinstance(t, ops.Tensor) and isinstance(t.shape, tensor_shape.TensorShape) and t.shape.rank == 1): return array_ops.expand_dims_v2(t, axis=-1) return t return nest.map_structure(_expand_single_1d_tensor, data) def train_validation_split(arrays, validation_split): """Split arrays into train and validation subsets in deterministic order. The last part of data will become validation data. Args: arrays: Tensors to split. Allowed inputs are arbitrarily nested structures of Tensors and NumPy arrays. validation_split: Float between 0 and 1. The proportion of the dataset to include in the validation split. The rest of the dataset will be included in the training split. Returns: `(train_arrays, validation_arrays)` """ def _can_split(t): tensor_types = _get_tensor_types() return isinstance(t, tensor_types) or t is None flat_arrays = nest.flatten(arrays) unsplitable = [type(t) for t in flat_arrays if not _can_split(t)] if unsplitable: raise ValueError( "`validation_split` is only supported for Tensors or NumPy " "arrays, found following types in the input: {}".format(unsplitable)) if all(t is None for t in flat_arrays): return arrays, arrays first_non_none = None for t in flat_arrays: if t is not None: first_non_none = t break # Assumes all arrays have the same batch shape or are `None`. batch_dim = int(first_non_none.shape[0]) split_at = int(math.floor(batch_dim * (1. - validation_split))) if split_at == 0 or split_at == batch_dim: raise ValueError( "Training data contains {batch_dim} samples, which is not sufficient " "to split it into a validation and training set as specified by " "`validation_split={validation_split}`. Either provide more data, or a " "different value for the `validation_split` argument." .format( batch_dim=batch_dim, validation_split=validation_split)) def _split(t, start, end): if t is None: return t return t[start:end] train_arrays = nest.map_structure( functools.partial(_split, start=0, end=split_at), arrays) val_arrays = nest.map_structure( functools.partial(_split, start=split_at, end=batch_dim), arrays) return train_arrays, val_arrays @keras_export("keras.utils.unpack_x_y_sample_weight", v1=[]) def unpack_x_y_sample_weight(data): """Unpacks user-provided data tuple. This is a convenience utility to be used when overriding `Model.train_step`, `Model.test_step`, or `Model.predict_step`. This utility makes it easy to support data of the form `(x,)`, `(x, y)`, or `(x, y, sample_weight)`. Standalone usage: >>> features_batch = tf.ones((10, 5)) >>> labels_batch = tf.zeros((10, 5)) >>> data = (features_batch, labels_batch) >>> # `y` and `sample_weight` will default to `None` if not provided. >>> x, y, sample_weight = tf.keras.utils.unpack_x_y_sample_weight(data) >>> sample_weight is None True Example in overridden `Model.train_step`: ```python class MyModel(tf.keras.Model): def train_step(self, data): # If `sample_weight` is not provided, all samples will be weighted # equally. x, y, sample_weight = tf.keras.utils.unpack_x_y_sample_weight(data) with tf.GradientTape() as tape: y_pred = self(x, training=True) loss = self.compiled_loss( y, y_pred, sample_weight, regularization_losses=self.losses) trainable_variables = self.trainable_variables gradients = tape.gradient(loss, trainable_variables) self.optimizer.apply_gradients(zip(gradients, trainable_variables)) self.compiled_metrics.update_state(y, y_pred, sample_weight) return {m.name: m.result() for m in self.metrics} ``` Args: data: A tuple of the form `(x,)`, `(x, y)`, or `(x, y, sample_weight)`. Returns: The unpacked tuple, with `None`s for `y` and `sample_weight` if they are not provided. """ if not isinstance(data, tuple): return (data, None, None) elif len(data) == 1: return (data[0], None, None) elif len(data) == 2: return (data[0], data[1], None) elif len(data) == 3: return (data[0], data[1], data[2]) else: error_msg = ("Data is expected to be in format `x`, `(x,)`, `(x, y)`, " "or `(x, y, sample_weight)`, found: {}").format(data) raise ValueError(error_msg) @keras_export("keras.utils.pack_x_y_sample_weight", v1=[]) def pack_x_y_sample_weight(x, y=None, sample_weight=None): """Packs user-provided data into a tuple. This is a convenience utility for packing data into the tuple formats that `Model.fit` uses. Standalone usage: >>> x = tf.ones((10, 1)) >>> data = tf.keras.utils.pack_x_y_sample_weight(x) >>> isinstance(data, tf.Tensor) True >>> y = tf.ones((10, 1)) >>> data = tf.keras.utils.pack_x_y_sample_weight(x, y) >>> isinstance(data, tuple) True >>> x, y = data Args: x: Features to pass to `Model`. y: Ground-truth targets to pass to `Model`. sample_weight: Sample weight for each element. Returns: Tuple in the format used in `Model.fit`. """ if y is None: # For single x-input, we do no tuple wrapping since in this case # there is no ambiguity. This also makes NumPy and Dataset # consistent in that the user does not have to wrap their Dataset # data in an unecessary tuple if not nest.is_nested(x): return x else: return (x,) elif sample_weight is None: return (x, y) else: return (x, y, sample_weight) def single_batch_iterator(strategy, x, y=None, sample_weight=None, class_weight=None): """Creates a single-batch dataset.""" x, y, sample_weight = _process_tensorlike((x, y, sample_weight)) if y is None: data = (x,) elif sample_weight is None: data = (x, y) else: data = (x, y, sample_weight) _check_data_cardinality(data) dataset = dataset_ops.DatasetV2.from_tensors(data) if class_weight: dataset = dataset.map(_make_class_weight_map_fn(class_weight)) dataset = strategy.experimental_distribute_dataset(dataset) return iter(dataset) def _check_data_cardinality(data): num_samples = set(int(i.shape[0]) for i in nest.flatten(data)) if len(num_samples) > 1: msg = "Data cardinality is ambiguous:\n" for label, single_data in zip(["x", "y", "sample_weight"], data): msg += " {} sizes: {}\n".format( label, ", ".join(str(i.shape[0]) for i in nest.flatten(single_data))) msg += "Make sure all arrays contain the same number of samples." raise ValueError(msg) def _get_tensor_types(): try: import pandas as pd # pylint: disable=g-import-not-at-top return (ops.Tensor, np.ndarray, pd.Series, pd.DataFrame) except ImportError: return (ops.Tensor, np.ndarray) def _is_scipy_sparse(x): try: from scipy.sparse import issparse # pylint: disable=g-import-not-at-top return issparse(x) except ImportError: return False def _scipy_sparse_to_sparse_tensor(t): """Converts a SciPy sparse matrix to a SparseTensor.""" sparse_coo = t.tocoo() row, col = sparse_coo.row, sparse_coo.col data, shape = sparse_coo.data, sparse_coo.shape if issubclass(data.dtype.type, np.floating): data = data.astype(backend.floatx()) indices = np.concatenate( (np.expand_dims(row, axis=1), np.expand_dims(col, axis=1)), axis=1) return sparse_tensor.SparseTensor(indices, data, shape) def _is_distributed_dataset(ds): return isinstance(ds, input_lib.DistributedDatasetInterface)