def train(self, data_iterator): """Train a keras model on a worker and send asynchronous updates to parameter server """ feature_iterator, label_iterator = tee(data_iterator, 2) x_train = np.asarray([x for x, y in feature_iterator]) y_train = np.asarray([y for x, y in label_iterator]) if x_train.size == 0: return self.model = model_from_yaml(self.yaml, self.custom_objects) self.model.compile(optimizer=get_optimizer(self.master_optimizer), loss=self.master_loss, metrics=self.master_metrics) self.model.set_weights(self.parameters.value) epochs = self.train_config['epochs'] batch_size = self.train_config.get('batch_size') nb_train_sample = x_train.shape[0] nb_batch = int(np.ceil(nb_train_sample / float(batch_size))) index_array = np.arange(nb_train_sample) batches = [(i * batch_size, min(nb_train_sample, (i + 1) * batch_size)) for i in range(0, nb_batch)] if self.frequency == 'epoch': for epoch in range(epochs): weights_before_training = self.client.get_parameters() self.model.set_weights(weights_before_training) self.train_config['epochs'] = 1 if x_train.shape[0] > batch_size: self.model.fit(x_train, y_train, **self.train_config) self.train_config['epochs'] = epochs weights_after_training = self.model.get_weights() deltas = subtract_params(weights_before_training, weights_after_training) self.client.update_parameters(deltas) elif self.frequency == 'batch': for epoch in range(epochs): if x_train.shape[0] > batch_size: for (batch_start, batch_end) in batches: weights_before_training = self.client.get_parameters() self.model.set_weights(weights_before_training) batch_ids = index_array[batch_start:batch_end] x = slice_arrays(x_train, batch_ids) y = slice_arrays(y_train, batch_ids) self.model.train_on_batch(x, y) weights_after_training = self.model.get_weights() deltas = subtract_params(weights_before_training, weights_after_training) self.client.update_parameters(deltas) else: raise ValueError( 'frequency parameter can be `epoch` or `batch, got {}'.format( self.frequency)) yield []
def slice_arrays(arrays, indices, contiguous=True): """Slices batches out of provided arrays (workaround for eager tensors). Unfortunately eager tensors don't have the same slicing behavior as Numpy arrays (they follow the same slicing behavior as symbolic TF tensors), hence we cannot use `generic_utils.slice_arrays` directly and we have to implement this workaround based on `concat`. This has a performance cost. Args: arrays: Single array or list of arrays. indices: List of indices in the array that should be included in the output batch. contiguous: Boolean flag indicating whether the indices are contiguous. Returns: Slice of data (either single array or list of arrays). """ converted_to_list = False if not isinstance(arrays, list): converted_to_list = True arrays = [arrays] if any(tensor_util.is_tf_type(x) for x in arrays): if not contiguous: entries = [[x[i:i + 1] for i in indices] for x in arrays] slices = [array_ops.concat(x, axis=0) for x in entries] else: slices = [x[indices[0]:indices[-1] + 1] for x in arrays] else: slices = generic_utils.slice_arrays(arrays, indices) if converted_to_list: slices = slices[0] return slices
def slice_arrays(arrays, indices, contiguous=True): """Slices batches out of provided arrays (workaround for eager tensors). Unfortunately eager tensors don't have the same slicing behavior as Numpy arrays (they follow the same slicing behavior as symbolic TF tensors), hence we cannot use `generic_utils.slice_arrays` directly and we have to implement this workaround based on `concat`. This has a performance cost. Arguments: arrays: Single array or list of arrays. indices: List of indices in the array that should be included in the output batch. contiguous: Boolean flag indicating whether the indices are contiguous. Returns: Slice of data (either single array or list of arrays). """ converted_to_list = False if not isinstance(arrays, list): converted_to_list = True arrays = [arrays] if any(tensor_util.is_tensor(x) for x in arrays): if not contiguous: entries = [[x[i:i + 1] for i in indices] for x in arrays] slices = [array_ops.concat(x, axis=0) for x in entries] else: slices = [x[indices[0]:indices[-1] + 1] for x in arrays] else: slices = generic_utils.slice_arrays(arrays, indices) if converted_to_list: slices = slices[0] return slices
def model_iteration(model, inputs, targets=None, sample_weights=None, batch_size=None, epochs=1, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None, validation_freq=1, mode=ModeKeys.TRAIN, validation_in_fit=False, prepared_feed_values_from_dataset=False, steps_name='steps', **kwargs): """Loop function for arrays of data with modes TRAIN/TEST/PREDICT. Arguments: model: Keras Model instance. inputs: Either a list or dictionary of arrays, or a dataset instance. targets: List/dictionary of input arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: 0, 1, or 2. Verbosity mode. 0 = silent, 1 = progress bar, 2 = one line per epoch. Note that the progress bar is not particularly useful when logged to a file, so verbose=2 is recommended when not running interactively (eg, in a production environment). callbacks: List of callbacks to be called during training val_inputs: Either a list or dictionary of arrays, or a dataset instance. val_targets: List/dictionary of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. validation_freq: Only relevant if validation data is provided. Integer or `collections.Container` instance (e.g. list, tuple, etc.). If an integer, specifies how many training epochs to run before a new validation run is performed, e.g. `validation_freq=2` runs validation every 2 epochs. If a Container, specifies the epochs on which to run validation, e.g. `validation_freq=[1, 2, 10]` runs validation at the end of the 1st, 2nd, and 10th epochs. mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT. validation_in_fit: if true, then this method is invoked from within training iteration (for validation). In the case where `val_inputs` is a dataset, this flag indicates that its iterator and feed values are already created so should properly reuse resources. prepared_feed_values_from_dataset: if True, `inputs` is a list of feed tensors returned from `_prepare_feed_values` call on the validation dataset, so do not call it again on `inputs`. Should only be used for inline validation (i.e., only if `validation_in_fit` is also True). steps_name: The string name of the steps argument, either `steps`, `validation_steps`, or `steps_per_epoch`. Only used for error message formatting. **kwargs: Additional arguments for backwards compatibility. Returns: - In TRAIN mode: `History` object. - In TEST mode: Evaluation metrics. - In PREDICT mode: Outputs of the Model called on inputs. Raises: ValueError: in case of invalid arguments. """ # Backwards compatibility. if 'steps' in kwargs: steps_per_epoch = kwargs.pop('steps') if kwargs: raise TypeError('Unknown arguments: %s' % (kwargs, )) # In case we were passed a dataset, we extract symbolic tensors from it. reset_dataset_after_each_epoch = False input_iterator = None is_dataset = isinstance(inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)) # TODO(fchollet): consider moving `steps_per_epoch` inference to # _standardize_user_data and set reset_dataset_after_each_epoch as an # attribute on the dataset instance. if is_dataset: if steps_per_epoch is None: reset_dataset_after_each_epoch = True steps_per_epoch = training_utils_v1.infer_steps_for_dataset( model, inputs, steps_per_epoch, epochs=epochs, steps_name=steps_name) input_iterator = _get_iterator(inputs, model._distribution_strategy) # Enter tf.distribute.Strategy scope. if model._distribution_strategy: scope = distributed_training_utils_v1.distributed_scope( strategy=model._distribution_strategy, learning_phase=(1 if mode == ModeKeys.TRAIN else 0)) scope.__enter__() use_steps = is_dataset or steps_per_epoch is not None do_validation = val_inputs is not None # Prepare input data. inputs = input_iterator or inputs if validation_in_fit and prepared_feed_values_from_dataset: # When invoking validation in training loop, avoid creating iterator and # list of feed values for the same validation dataset multiple times (which # essentially would call `iterator.get_next()` that slows down execution and # leads to OOM errors eventually. ins = inputs else: ins = _prepare_feed_values(model, inputs, targets, sample_weights, mode) # `ins` is a function when a distribute strategy is used in Eager mode. In # that case `is_dataset` is True. The code branches that have requirements # about the type of `ins` do not trigger in the distributed case. if not is_dataset: num_samples_or_steps = _get_num_samples_or_steps( ins, batch_size, steps_per_epoch) else: num_samples_or_steps = steps_per_epoch # Update sample_weight_mode of the model if sample_weights is specified by the # user. We need to call this function after we have a handle on the inputs # (both numpy arrays and datasets) in order to determine if the user has # specified sample_weights. _update_sample_weight_mode(model, mode, ins) # Get step function and loop type. As part of building the execution # function we recompile the metrics based on the updated # sample_weight_mode value. f = _make_execution_function(model, mode) # Prepare validation data. Hold references to the iterator and the input list # to properly reinitialize and reuse in multiple validation passes. val_iterator = None if isinstance(val_inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)): if validation_steps is None: # Because we pass an iterator feed instead of a Dataset to the eval # model_iteration() call, it will not trigger the dataset-input path # that determines the number of steps required. To avoid this issue, # set validation_steps here if validation_steps is None. validation_steps = training_utils_v1.infer_steps_for_dataset( model, val_inputs, validation_steps, epochs=epochs, steps_name='validation_steps') val_iterator = _get_iterator(val_inputs, model._distribution_strategy) val_inputs = _prepare_feed_values(model, val_iterator, val_targets, val_sample_weights, ModeKeys.TEST) # Get num steps for printing. val_samples_or_steps = validation_steps else: # Get num samples for printing. val_samples_or_steps = val_inputs and nest.flatten( val_inputs)[0].shape[0] or None if mode == ModeKeys.TRAIN and verbose: _print_train_info(num_samples_or_steps, val_samples_or_steps, is_dataset) # Configure callbacks. count_mode = 'steps' if use_steps else 'samples' callbacks = cbks.configure_callbacks(callbacks, model, do_validation=do_validation, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_samples_or_steps, count_mode=count_mode, verbose=verbose, mode=mode) # Find beforehand arrays that need sparse-to-dense conversion. if issparse is not None and not use_steps: indices_for_conversion_to_dense = [] feed = _get_model_feed(model, mode) for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)): if issparse(input_data) and not K.is_sparse(feed_tensor): indices_for_conversion_to_dense.append(i) # Select aggregation method. if mode == ModeKeys.PREDICT: aggregator = training_utils_v1.OutputsAggregator( use_steps, num_samples=None if steps_per_epoch else num_samples_or_steps, steps=steps_per_epoch) else: aggregator = training_utils_v1.MetricsAggregator( use_steps, num_samples=None if steps_per_epoch else num_samples_or_steps, steps=steps_per_epoch) if model._compile_distribution: distributed_training_utils_v1._copy_weights_to_distributed_model( model, mode) callbacks.model.stop_training = False callbacks._call_begin_hook(mode) initial_epoch = model._maybe_load_initial_epoch_from_ckpt( initial_epoch, mode) for epoch in range(initial_epoch, epochs): if callbacks.model.stop_training: break # Setup work for each epoch epoch_logs = {} if mode != ModeKeys.PREDICT: # Collecting and resetting metrics has non-zero cost and will needlessly # slow down model.predict. model.reset_metrics() if mode == ModeKeys.TRAIN: callbacks.on_epoch_begin(epoch, epoch_logs) if use_steps: # Step-wise loop. if steps_per_epoch is None: # Loop over dataset until `OutOfRangeError` is raised. target_steps = np.inf else: # Loop over dataset for the specified number of steps. target_steps = steps_per_epoch step = 0 while step < target_steps: batch_logs = {'batch': step, 'size': 1} callbacks._call_batch_hook(mode, 'begin', step, batch_logs) # Get outputs. try: # `ins` can be callable in tf.distribute.Strategy + eager case. if not callable(ins) or ( model._distribution_strategy and not distributed_training_utils_v1. is_distributing_by_cloning(model)): actual_inputs = ins else: actual_inputs = ins() batch_outs = f(actual_inputs) except errors.OutOfRangeError: if is_dataset: # The dataset passed by the user ran out of batches. # Now we know the cardinality of the dataset. # If steps_per_epoch was specified, then running out of data is # unexpected, so we stop training and inform the user. if steps_per_epoch: callbacks.model.stop_training = True logging.warning( 'Your dataset ran out of data; interrupting training. ' 'Make sure that your dataset can generate at least ' '`%s * epochs` batches (in this case, %d batches). ' 'You may need to use the repeat() function when ' 'building your dataset.' % (steps_name, steps_per_epoch * epochs)) elif step > 0: steps_per_epoch = step aggregator.steps = steps_per_epoch else: # We ran out of batches while the user passed an iterator (legacy). callbacks.model.stop_training = True logging.warning( 'Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your iterator ' 'can generate at least `%s * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % (steps_name, steps_per_epoch * epochs)) break if not isinstance(batch_outs, list): batch_outs = [batch_outs] if model._distribution_strategy: batch_outs = (distributed_training_utils_v1. _per_replica_aggregate_batch( model._distribution_strategy, batch_outs, model, mode)) # Aggregate results. if step == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', step, batch_logs) step += 1 if callbacks.model.stop_training: break else: # Sample-wise loop. index_array = np.arange(num_samples_or_steps) if shuffle == 'batch': index_array = training_utils_v1.batch_shuffle( index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_samples_or_steps, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] # Slice into a batch. if len(batches) == 1: # If we only have one batch, do not slice. This takes care of # composite tensors in non-Dataset modes; we currently don't support # slicing them. # TODO(b/133517906): Add slicing support. ins_batch = ins else: try: if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') # Sparse to dense conversion. if issparse is not None: for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() # Callbacks batch_begin. batch_logs = {'batch': batch_index, 'size': len(batch_ids)} callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs) # Get outputs. batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if batch_index == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs, batch_start, batch_end) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs) if callbacks.model.stop_training: break aggregator.finalize() results = aggregator.results epoch_logs = cbks.make_logs(model, epoch_logs, results, mode) if len(results) == 1: results = results[0] # Run the test loop every `validation_freq` epochs during training. if (do_validation and training_utils_v1.should_run_validation( validation_freq, epoch) and not callbacks.model.stop_training): if model._compile_distribution: # Since we create a new clone from the original model we need to copy # the weights back to the original model before we can run validation. distributed_training_utils_v1._copy_weights_to_original_model( model, ModeKeys.TRAIN) val_results = model_iteration( model, val_inputs, targets=val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps_per_epoch=validation_steps, callbacks=callbacks, verbose=0, mode=ModeKeys.TEST, validation_in_fit=True, prepared_feed_values_from_dataset=(val_iterator is not None), steps_name='validation_steps') if not isinstance(val_results, list): val_results = [val_results] epoch_logs = cbks.make_logs(model, epoch_logs, val_results, mode, prefix='val_') if val_iterator and epoch < epochs - 1: _reinitialize_iterator(val_iterator, model._distribution_strategy) if mode == ModeKeys.TRAIN: # Epochs only apply to `fit`. callbacks.on_epoch_end(epoch, epoch_logs) # Reinitialize dataset iterator for the next epoch. if reset_dataset_after_each_epoch and epoch < epochs - 1: _reinitialize_iterator(input_iterator, model._distribution_strategy) model._successful_loop_finish = True callbacks._call_end_hook(mode) if model._distribution_strategy: if model._compile_distribution: # TODO(priyag, psv): Copy back metrics to the original model as well? distributed_training_utils_v1._copy_weights_to_original_model( model, mode) scope.__exit__(None, None, None) if mode == ModeKeys.TRAIN: return model.history return results
def model_iteration(model, inputs, targets=None, sample_weights=None, batch_size=None, epochs=1, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None, mode='train', **kwargs): """Loop function for arrays of data with modes 'train'/'test'/'predict'. Arguments: model: Keras Model instance. inputs: Either a list of arrays or a dictionary. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: List of input arrays. val_targets: List of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. mode: One of 'train'/'test'/'predict'. **kwargs: Additional arguments for backwards compatibility. Returns: - In 'train' mode: `History` object. - In 'test' mode: Evaluation metrics. - In 'predict' mode: Outputs of the Model called on inputs. Raises: ValueError: in case of invalid arguments. """ # Backwards compatibility. if 'steps' in kwargs: steps_per_epoch = kwargs['steps'] _validate_arguments(steps_per_epoch, validation_steps, kwargs) if mode == 'train': _print_train_info(inputs, val_inputs, steps_per_epoch, verbose) # Get step function and loop type. f = model._get_execution_function(mode) use_steps = steps_per_epoch is not None do_validation = val_inputs is not None # Prepare input data. inputs = training_utils.ModelInputs(inputs).as_list() targets = targets or [] sample_weights = sample_weights or [] learning_phase_input = [] if not isinstance(K.symbolic_learning_phase(), int): learning_phase_input = [True] if mode == 'train' else [False] ins = inputs + targets + sample_weights + learning_phase_input num_samples_or_steps = _get_num_samples_or_steps(ins, batch_size, steps_per_epoch) # Configure callbacks. count_mode = 'steps' if use_steps else 'samples' callbacks = cbks.configure_callbacks( callbacks, model, do_validation=do_validation, val_inputs=val_inputs, val_targets=val_targets, val_sample_weights=val_sample_weights, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_samples_or_steps, validation_steps=validation_steps, verbose=0, # Handle ProgBarLogger separately in this loop. count_mode=count_mode, mode=mode) # TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready. progbar = _get_progbar(model, count_mode) progbar.params = callbacks.params progbar.params['verbose'] = verbose # Find beforehand arrays that need sparse-to-dense conversion. if issparse is not None: indices_for_conversion_to_dense = [] feed = _get_model_feed(model, mode) for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)): if issparse(input_data) and not K.is_sparse(feed_tensor): indices_for_conversion_to_dense.append(i) # Select aggregation method. if mode == 'predict': aggregator = OutputsAggregator(use_steps, num_samples_or_steps) else: aggregator = MetricsAggregator(use_steps, num_samples_or_steps) callbacks.model.stop_training = False callbacks._call_begin_hook(mode) progbar.on_train_begin() for epoch in range(initial_epoch, epochs): if callbacks.model.stop_training: break # Setup work for each epoch results = [] epoch_logs = {} if hasattr(model, 'stateful_metric_functions'): for m in model.stateful_metric_functions: m.reset_states() callbacks.on_epoch_begin(epoch, epoch_logs, mode=mode) progbar.on_epoch_begin(epoch, epoch_logs) if use_steps: # Step-wise loop. for step in range(steps_per_epoch): batch_logs = {'batch': step, 'size': 1} callbacks._call_batch_hook(mode, 'begin', step, batch_logs) progbar.on_batch_begin(step, batch_logs) # Get outputs. try: batch_outs = f(ins) except errors.OutOfRangeError: logging.warning('Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your dataset ' 'can generate at least `steps_per_epoch * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % steps_per_epoch * epochs) break if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if step == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs) # Callbacks batch end. batch_logs.update(_make_logs(model, batch_outs, mode)) callbacks._call_batch_hook(mode, 'end', step, batch_logs) progbar.on_batch_end(step, batch_logs) if callbacks.model.stop_training: break else: # Sample-wise loop. index_array = np.arange(num_samples_or_steps) if shuffle == 'batch': index_array = training_utils.batch_shuffle(index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_samples_or_steps, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] # Slice into a batch. try: if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') # Sparse to dense conversion. for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() # Callbacks batch_begin. batch_logs = {'batch': batch_index, 'size': len(batch_ids)} callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs) progbar.on_batch_begin(batch_index, batch_logs) # Get outputs. batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if batch_index == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs, batch_start, batch_end) # Callbacks batch end. batch_logs.update(_make_logs(model, batch_outs, mode)) callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs) progbar.on_batch_end(batch_index, batch_logs) if callbacks.model.stop_training: break aggregator.finalize() results = aggregator.results epoch_logs.update(_make_logs(model, results, mode)) if len(results) == 1: results = results[0] # Run the test loop every epoch during training. if do_validation and not callbacks.model.stop_training: val_results = model_iteration( model, val_inputs, targets=val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps_per_epoch=validation_steps, callbacks=callbacks, verbose=0, mode='test') if not isinstance(val_results, list): val_results = [val_results] epoch_logs.update(_make_logs(model, val_results, mode, prefix='val_')) callbacks.on_epoch_end(epoch, epoch_logs, mode=mode) progbar.on_epoch_end(epoch, epoch_logs) callbacks._call_end_hook(mode) if mode == 'train': return model.history return results
def model_iteration(model, inputs, targets=None, sample_weights=None, batch_size=None, epochs=1, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None, mode='train', **kwargs): """Loop function for arrays of data with modes 'train'/'test'/'predict'. Arguments: model: Keras Model instance. inputs: Either a list of arrays or a dictionary. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: List of input arrays. val_targets: List of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. mode: One of 'train'/'test'/'predict'. **kwargs: Additional arguments for backwards compatibility. Returns: - In 'train' mode: `History` object. - In 'test' mode: Evaluation metrics. - In 'predict' mode: Outputs of the Model called on inputs. Raises: ValueError: in case of invalid arguments. """ # Backwards compatibility. if 'steps' in kwargs: steps_per_epoch = kwargs['steps'] _validate_arguments(steps_per_epoch, validation_steps, kwargs) if mode == 'train': _print_train_info(inputs, val_inputs, steps_per_epoch, verbose) # Get step function and loop type. f = model._get_execution_function(mode) use_steps = steps_per_epoch is not None do_validation = val_inputs is not None # Prepare input data. inputs = training_utils.ModelInputs(inputs).as_list() targets = targets or [] sample_weights = sample_weights or [] learning_phase_input = [] if not isinstance(K.symbolic_learning_phase(), int): learning_phase_input = [True] if mode == 'train' else [False] ins = inputs + targets + sample_weights + learning_phase_input num_samples_or_steps = _get_num_samples_or_steps(ins, batch_size, steps_per_epoch) # Configure callbacks. count_mode = 'steps' if use_steps else 'samples' callbacks = cbks.configure_callbacks( callbacks, model, do_validation=do_validation, val_inputs=val_inputs, val_targets=val_targets, val_sample_weights=val_sample_weights, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_samples_or_steps, validation_steps=validation_steps, verbose=0, # Handle ProgBarLogger separately in this loop. count_mode=count_mode, mode=mode) # TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready. progbar = _get_progbar(model, count_mode) progbar.params = callbacks.params progbar.params['verbose'] = verbose # Find beforehand arrays that need sparse-to-dense conversion. if issparse is not None: indices_for_conversion_to_dense = [] feed = _get_model_feed(model, mode) for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)): if issparse(input_data) and not K.is_sparse(feed_tensor): indices_for_conversion_to_dense.append(i) # Select aggregation method. if mode == 'predict': aggregator = OutputsAggregator(use_steps, num_samples_or_steps) else: aggregator = MetricsAggregator(use_steps, num_samples_or_steps) callbacks.model.stop_training = False callbacks._call_begin_hook(mode) progbar.on_train_begin() for epoch in range(initial_epoch, epochs): if callbacks.model.stop_training: break # Setup work for each epoch results = [] epoch_logs = {} if hasattr(model, 'stateful_metric_functions'): for m in model.stateful_metric_functions: m.reset_states() callbacks.on_epoch_begin(epoch, epoch_logs, mode=mode) progbar.on_epoch_begin(epoch, epoch_logs) if use_steps: # Step-wise loop. for step in range(steps_per_epoch): batch_logs = {'batch': step, 'size': 1} callbacks._call_batch_hook(mode, 'begin', step, batch_logs) progbar.on_batch_begin(step, batch_logs) # Get outputs. try: batch_outs = f(ins) except errors.OutOfRangeError: logging.warning( 'Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your dataset ' 'can generate at least `steps_per_epoch * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % steps_per_epoch * epochs) break if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if step == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs) # Callbacks batch end. batch_logs.update(_make_logs(model, batch_outs, mode)) callbacks._call_batch_hook(mode, 'end', step, batch_logs) progbar.on_batch_end(step, batch_logs) if callbacks.model.stop_training: break else: # Sample-wise loop. index_array = np.arange(num_samples_or_steps) if shuffle == 'batch': index_array = training_utils.batch_shuffle( index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_samples_or_steps, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] # Slice into a batch. try: if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') # Sparse to dense conversion. if issparse is not None: for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() # Callbacks batch_begin. batch_logs = {'batch': batch_index, 'size': len(batch_ids)} callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs) progbar.on_batch_begin(batch_index, batch_logs) # Get outputs. batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if batch_index == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs, batch_start, batch_end) # Callbacks batch end. batch_logs.update(_make_logs(model, batch_outs, mode)) callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs) progbar.on_batch_end(batch_index, batch_logs) if callbacks.model.stop_training: break aggregator.finalize() results = aggregator.results epoch_logs.update(_make_logs(model, results, mode)) if len(results) == 1: results = results[0] # Run the test loop every epoch during training. if do_validation and not callbacks.model.stop_training: val_results = model_iteration(model, val_inputs, targets=val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps_per_epoch=validation_steps, callbacks=callbacks, verbose=0, mode='test') if not isinstance(val_results, list): val_results = [val_results] epoch_logs.update( _make_logs(model, val_results, mode, prefix='val_')) callbacks.on_epoch_end(epoch, epoch_logs, mode=mode) progbar.on_epoch_end(epoch, epoch_logs) callbacks._call_end_hook(mode) if mode == 'train': return model.history return results
def fit_loop(model, inputs, targets, sample_weights=None, batch_size=None, epochs=100, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, callback_metrics=None, initial_epoch=0, steps_per_epoch=None, validation_steps=None): """Abstract fit function for arrays of data. Arguments: model: Keras Model instance. inputs: List of input arrays. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: List of input arrays. val_targets: List of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch callback_metrics: List of strings, the display names of the metrics passed to the callbacks. They should be the concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. Returns: `History` object. Raises: ValueError: in case of invalid arguments. """ model._make_train_function() f = model.train_function sample_weights = sample_weights or [] val_sample_weights = val_sample_weights or [] if model.uses_learning_phase and not isinstance(K.learning_phase(), int): ins = inputs + targets + sample_weights + [1] if val_inputs: val_ins = val_inputs + val_targets + val_sample_weights + [1] else: ins = inputs + targets + sample_weights if val_inputs: val_ins = val_inputs + val_targets + val_sample_weights if not val_inputs: val_ins = [] do_validation = False if val_inputs: do_validation = True if (steps_per_epoch is None and verbose and inputs and hasattr(inputs[0], 'shape') and hasattr(val_inputs[0], 'shape')): print('Train on %d samples, validate on %d samples' % (inputs[0].shape[0], val_inputs[0].shape[0])) if validation_steps: do_validation = True if steps_per_epoch is None: raise ValueError('Can only use `validation_steps` ' 'when doing step-wise ' 'training, i.e. `steps_per_epoch` ' 'must be set.') out_labels = model.metrics_names if do_validation: callback_metrics = copy.copy(out_labels) + [ 'val_' + n for n in out_labels ] else: callback_metrics = copy.copy(out_labels) num_train_samples = training_utils.check_num_samples( ins, batch_size, steps_per_epoch, 'steps_per_epoch') if num_train_samples is not None: index_array = np.arange(num_train_samples) model.history = cbks.History() all_callbacks = [ cbks.BaseLogger(stateful_metrics=model.stateful_metric_names) ] if verbose: if steps_per_epoch is not None: count_mode = 'steps' else: count_mode = 'samples' all_callbacks.append( cbks.ProgbarLogger(count_mode, stateful_metrics=model.stateful_metric_names)) all_callbacks += (callbacks or []) + [model.history] callbacks = cbks.CallbackList(all_callbacks) out_labels = out_labels or [] # it's possible to callback a different model than self # (used by Sequential models) if hasattr(model, 'callback_model') and model.callback_model: callback_model = model.callback_model else: callback_model = model callbacks.set_model(callback_model) callbacks.set_params({ 'batch_size': batch_size, 'epochs': epochs, 'steps': steps_per_epoch, 'samples': num_train_samples, 'verbose': verbose, 'do_validation': do_validation, 'metrics': callback_metrics or [], }) callbacks.on_train_begin() callback_model.stop_training = False for cbk in callbacks: cbk.validation_data = val_ins # To prevent a slowdown, we find beforehand the arrays that need conversion. feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights indices_for_conversion_to_dense = [] for i in range(len(feed)): if issparse is not None and issparse( ins[i]) and not K.is_sparse(feed[i]): indices_for_conversion_to_dense.append(i) for epoch in range(initial_epoch, epochs): # Reset stateful metrics for m in model.stateful_metric_functions: m.reset_states() # Update callbacks callbacks.on_epoch_begin(epoch) epoch_logs = {} if steps_per_epoch is not None: for step_index in range(steps_per_epoch): batch_logs = {} batch_logs['batch'] = step_index batch_logs['size'] = 1 callbacks.on_batch_begin(step_index, batch_logs) try: outs = f(ins) except errors.OutOfRangeError: logging.warning( 'Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your dataset ' 'can generate at least `steps_per_epoch * epochs` ' 'batches (in this case, %d batches).' % steps_per_epoch * epochs) break if not isinstance(outs, list): outs = [outs] for l, o in zip(out_labels, outs): batch_logs[l] = o callbacks.on_batch_end(step_index, batch_logs) if callback_model.stop_training: break if do_validation: val_outs = test_loop(model, val_inputs, val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps=validation_steps, verbose=0) if not isinstance(val_outs, list): val_outs = [val_outs] # Same labels assumed. for l, o in zip(out_labels, val_outs): epoch_logs['val_' + l] = o else: if shuffle == 'batch': index_array = training_utils.batch_shuffle( index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_train_samples, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] try: if isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') batch_logs = {} batch_logs['batch'] = batch_index batch_logs['size'] = len(batch_ids) callbacks.on_batch_begin(batch_index, batch_logs) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() outs = f(ins_batch) if not isinstance(outs, list): outs = [outs] for l, o in zip(out_labels, outs): batch_logs[l] = o callbacks.on_batch_end(batch_index, batch_logs) if callback_model.stop_training: break if batch_index == len(batches) - 1: # Last batch. if do_validation: val_outs = test_loop(model, val_inputs, val_targets, sample_weights=val_sample_weights, batch_size=batch_size, verbose=0) if not isinstance(val_outs, list): val_outs = [val_outs] # Same labels assumed. for l, o in zip(out_labels, val_outs): epoch_logs['val_' + l] = o callbacks.on_epoch_end(epoch, epoch_logs) if callback_model.stop_training: break callbacks.on_train_end() return model.history
def test_loop(model, inputs, targets, sample_weights=None, batch_size=None, verbose=0, steps=None): """Abstract method to loop over some data in batches. Arguments: model: Keras Model instance. inputs: List of input arrays. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: integer batch size or `None`. verbose: verbosity mode. steps: Total number of steps (batches of samples) before declaring predictions finished. Ignored with the default value of `None`. Returns: Scalar loss (if the model has a single output and no metrics) or list of scalars (if the model has multiple outputs and/or metrics). The attribute `model.metrics_names` will give you the display labels for the scalar outputs. """ model._make_test_function() f = model.test_function sample_weights = sample_weights or [] if model.uses_learning_phase and not isinstance(K.learning_phase(), int): ins = inputs + targets + sample_weights + [0] else: ins = inputs + targets + sample_weights if hasattr(model, 'metrics'): for m in model.stateful_metric_functions: m.reset_states() stateful_metric_indices = [ i for i, name in enumerate(model.metrics_names) if str(name) in model.stateful_metric_names ] else: stateful_metric_indices = [] num_samples = training_utils.check_num_samples(ins, batch_size, steps, 'steps') outs = [] if verbose == 1: if steps is not None: progbar = Progbar(target=steps) else: progbar = Progbar(target=num_samples) # To prevent a slowdown, we find beforehand the arrays that need conversion. feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights indices_for_conversion_to_dense = [] for i in range(len(feed)): if issparse is not None and issparse( ins[i]) and not K.is_sparse(feed[i]): indices_for_conversion_to_dense.append(i) if steps is not None: for step in range(steps): batch_outs = f(ins) if isinstance(batch_outs, list): if step == 0: for _ in enumerate(batch_outs): outs.append(0.) for i, batch_out in enumerate(batch_outs): if i in stateful_metric_indices: outs[i] = batch_out else: outs[i] += batch_out else: if step == 0: outs.append(0.) outs[0] += batch_outs if verbose == 1: progbar.update(step + 1) for i in range(len(outs)): if i not in stateful_metric_indices: outs[i] /= steps else: batches = make_batches(num_samples, batch_size) index_array = np.arange(num_samples) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] if isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() batch_outs = f(ins_batch) if isinstance(batch_outs, list): if batch_index == 0: for batch_out in enumerate(batch_outs): outs.append(0.) for i, batch_out in enumerate(batch_outs): if i in stateful_metric_indices: outs[i] = batch_out else: outs[i] += batch_out * len(batch_ids) else: if batch_index == 0: outs.append(0.) outs[0] += batch_outs * len(batch_ids) if verbose == 1: progbar.update(batch_end) for i in range(len(outs)): if i not in stateful_metric_indices: outs[i] /= num_samples if len(outs) == 1: return outs[0] return outs
def predict_loop(model, inputs, batch_size=32, verbose=0, steps=None): """Abstract method to loop over some data in batches. Arguments: model: Keras Model instance. inputs: list of tensors to be fed to `f`. batch_size: integer batch size. verbose: verbosity mode. steps: Total number of steps (batches of samples) before declaring `_predict_loop` finished. Ignored with the default value of `None`. Returns: Array of predictions (if the model has a single output) or list of arrays of predictions (if the model has multiple outputs). """ model._make_predict_function() f = model.predict_function if model.uses_learning_phase and not isinstance(K.learning_phase(), int): ins = inputs + [0] else: ins = inputs num_samples = training_utils.check_num_samples(inputs, batch_size, steps, 'steps') if verbose == 1: if steps is not None: progbar = Progbar(target=steps) else: progbar = Progbar(target=num_samples) indices_for_conversion_to_dense = [] for i in range(len(model._feed_inputs)): if (issparse is not None and issparse(inputs[i]) and not K.is_sparse(model._feed_inputs[i])): indices_for_conversion_to_dense.append(i) if steps is not None: # Step-based predictions. # Since we do not know how many samples # we will see, we cannot pre-allocate # the returned Numpy arrays. # Instead, we store one array per batch seen # and concatenate them upon returning. unconcatenated_outs = [] for step in range(steps): batch_outs = f(ins) if not isinstance(batch_outs, list): batch_outs = [batch_outs] if step == 0: for batch_out in batch_outs: unconcatenated_outs.append([]) for i, batch_out in enumerate(batch_outs): unconcatenated_outs[i].append(batch_out) if verbose == 1: progbar.update(step + 1) if len(unconcatenated_outs) == 1: return np.concatenate(unconcatenated_outs[0], axis=0) return [ np.concatenate(unconcatenated_outs[i], axis=0) for i in range(len(unconcatenated_outs)) ] else: # Sample-based predictions. outs = [] batches = make_batches(num_samples, batch_size) index_array = np.arange(num_samples) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] if batch_index == 0: # Pre-allocate the results arrays. for batch_out in batch_outs: shape = (num_samples, ) + batch_out.shape[1:] outs.append(np.zeros(shape, dtype=batch_out.dtype)) for i, batch_out in enumerate(batch_outs): outs[i][batch_start:batch_end] = batch_out if verbose == 1: progbar.update(batch_end) if len(outs) == 1: return outs[0] return outs
def model_iteration(model, inputs, targets=None, sample_weights=None, batch_size=None, epochs=1, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None, validation_freq=1, mode=ModeKeys.TRAIN, validation_in_fit=False, prepared_feed_values_from_dataset=False, steps_name='steps', **kwargs): """Loop function for arrays of data with modes TRAIN/TEST/PREDICT. Arguments: model: Keras Model instance. inputs: Either a list or dictionary of arrays, or a dataset instance. targets: List/dictionary of input arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: Either a list or dictionary of arrays, or a dataset instance. val_targets: List/dictionary of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. validation_freq: Only relevant if validation data is provided. Integer or `collections.Container` instance (e.g. list, tuple, etc.). If an integer, specifies how many training epochs to run before a new validation run is performed, e.g. `validation_freq=2` runs validation every 2 epochs. If a Container, specifies the epochs on which to run validation, e.g. `validation_freq=[1, 2, 10]` runs validation at the end of the 1st, 2nd, and 10th epochs. mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT. validation_in_fit: if true, then this method is invoked from within training iteration (for validation). In the case where `val_inputs` is a dataset, this flag indicates that its iterator and feed values are already created so should properly reuse resources. prepared_feed_values_from_dataset: if True, `inputs` is a list of feed tensors returned from `_prepare_feed_values` call on the validation dataset, so do not call it again on `inputs`. Should only be used for inline validation (i.e., only if `validation_in_fit` is also True). steps_name: The string name of the steps argument, either `steps`, `validation_steps`, or `steps_per_epoch`. Only used for error message formatting. **kwargs: Additional arguments for backwards compatibility. Returns: - In TRAIN mode: `History` object. - In TEST mode: Evaluation metrics. - In PREDICT mode: Outputs of the Model called on inputs. Raises: ValueError: in case of invalid arguments. """ # Backwards compatibility. if 'steps' in kwargs: steps_per_epoch = kwargs.pop('steps') if kwargs: raise TypeError('Unknown arguments: %s' % (kwargs,)) # In case we were passed a dataset, we extract symbolic tensors from it. reset_dataset_after_each_epoch = False input_iterator = None is_dataset = isinstance(inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)) # TODO(fchollet): consider moving `steps_per_epoch` inference to # _standardize_user_data and set reset_dataset_after_each_epoch as an # attribute on the dataset instance. if is_dataset: if steps_per_epoch is None: reset_dataset_after_each_epoch = True steps_per_epoch = training_utils.infer_steps_for_dataset( inputs, steps_per_epoch, epochs=epochs, steps_name=steps_name) input_iterator = _get_iterator(inputs, model._distribution_strategy) if mode == ModeKeys.TRAIN: _print_train_info(inputs, val_inputs, steps_per_epoch, verbose) # Enter DistributionStrategy scope. if model._distribution_strategy: scope = distributed_training_utils.distributed_scope( strategy=model._distribution_strategy, learning_phase=(1 if mode == ModeKeys.TRAIN else 0)) scope.__enter__() # Get step function and loop type. f = _make_execution_function(model, mode) use_steps = is_dataset or steps_per_epoch is not None do_validation = val_inputs is not None # Convert Eager Tensors to NumPy arrays to support batching/shuffling. inputs, targets, sample_weights = training_utils. \ convert_eager_tensors_to_numpy((inputs, targets, sample_weights)) # Prepare input data. inputs = input_iterator or inputs if validation_in_fit and prepared_feed_values_from_dataset: # When invoking validation in training loop, avoid creating iterator and # list of feed values for the same validation dataset multiple times (which # essentially would call `iterator.get_next()` that slows down execution and # leads to OOM errors eventually. ins = inputs else: ins = _prepare_feed_values(model, inputs, targets, sample_weights, mode) if not is_dataset: num_samples_or_steps = _get_num_samples_or_steps(ins, batch_size, steps_per_epoch) else: num_samples_or_steps = steps_per_epoch # Prepare validation data. Hold references to the iterator and the input list # to properly reinitialize and reuse in multiple validation passes. val_iterator = None if isinstance(val_inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)): if validation_steps is None: # Because we pass an iterator feed instead of a Dataset to the eval # model_iteration() call, it will not trigger the dataset-input path # that determines the number of steps required. To avoid this issue, # set validation_steps here if validation_steps is None. validation_steps = training_utils.infer_steps_for_dataset( val_inputs, validation_steps, epochs=epochs, steps_name='validation_steps') val_iterator = _get_iterator(val_inputs, model._distribution_strategy) val_inputs = _prepare_feed_values( model, val_iterator, val_targets, val_sample_weights, ModeKeys.TEST) # Configure callbacks. count_mode = 'steps' if use_steps else 'samples' callbacks = cbks.configure_callbacks( callbacks, model, do_validation=do_validation, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_samples_or_steps, verbose=0, # Handle ProgBarLogger separately in this loop. mode=mode) # TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready. progbar = training_utils.get_progbar(model, count_mode) progbar.params = callbacks.params progbar.params['verbose'] = verbose # Find beforehand arrays that need sparse-to-dense conversion. if issparse is not None and not use_steps: indices_for_conversion_to_dense = [] feed = _get_model_feed(model, mode) for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)): if issparse(input_data) and not K.is_sparse(feed_tensor): indices_for_conversion_to_dense.append(i) # Select aggregation method. if mode == ModeKeys.PREDICT: aggregator = training_utils.OutputsAggregator(use_steps, num_samples_or_steps) else: aggregator = training_utils.MetricsAggregator(use_steps, num_samples_or_steps) if model._compile_distribution: distributed_training_utils._copy_weights_to_distributed_model(model, mode) callbacks.model.stop_training = False callbacks._call_begin_hook(mode) progbar.on_train_begin() for epoch in range(initial_epoch, epochs): if callbacks.model.stop_training: break # Setup work for each epoch epoch_logs = {} model.reset_metrics() if mode == ModeKeys.TRAIN: callbacks.on_epoch_begin(epoch, epoch_logs) progbar.on_epoch_begin(epoch, epoch_logs) if use_steps: # Step-wise loop. if steps_per_epoch is None: # Loop over dataset until `OutOfRangeError` is raised. target_steps = np.inf else: # Loop over dataset for the specified number of steps. target_steps = steps_per_epoch step = 0 while step < target_steps: batch_logs = {'batch': step, 'size': 1} callbacks._call_batch_hook(mode, 'begin', step, batch_logs) progbar.on_batch_begin(step, batch_logs) # Get outputs. try: # `ins` can be callable in DistributionStrategy + eager case. actual_inputs = ins() if callable(ins) else ins batch_outs = f(actual_inputs) except errors.OutOfRangeError: if is_dataset: # The dataset passed by the user ran out of batches. # Now we know the cardinality of the dataset. # If steps_per_epoch was specified, then running out of data is # unexpected, so we stop training and inform the user. if steps_per_epoch: callbacks.model.stop_training = True logging.warning( 'Your dataset ran out of data; interrupting training. ' 'Make sure that your dataset can generate at least ' '`%s * epochs` batches (in this case, %d batches). ' 'You may need to use the repeat() function when ' 'building your dataset.' % (steps_name, steps_per_epoch * epochs)) elif step > 0: steps_per_epoch = step aggregator.num_samples_or_steps = steps_per_epoch if mode == ModeKeys.TRAIN: progbar.params['steps'] = steps_per_epoch progbar.progbar.target = steps_per_epoch else: # We ran out of batches while the user passed an iterator (legacy). callbacks.model.stop_training = True logging.warning( 'Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your iterator ' 'can generate at least `%s * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % (steps_name, steps_per_epoch * epochs)) break if not isinstance(batch_outs, list): batch_outs = [batch_outs] if model._distribution_strategy: batch_outs = distributed_training_utils._per_device_aggregate_batch( batch_outs, model, mode) # Aggregate results. if step == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', step, batch_logs) progbar.on_batch_end(step, batch_logs) step += 1 if callbacks.model.stop_training: break else: # Sample-wise loop. index_array = np.arange(num_samples_or_steps) if shuffle == 'batch': index_array = training_utils.batch_shuffle(index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_samples_or_steps, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] # Slice into a batch. try: if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') # Sparse to dense conversion. if issparse is not None: for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() # Callbacks batch_begin. batch_logs = {'batch': batch_index, 'size': len(batch_ids)} callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs) progbar.on_batch_begin(batch_index, batch_logs) # Get outputs. batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if batch_index == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs, batch_start, batch_end) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs) progbar.on_batch_end(batch_index, batch_logs) if callbacks.model.stop_training: break aggregator.finalize() results = aggregator.results epoch_logs = cbks.make_logs(model, epoch_logs, results, mode) if len(results) == 1: results = results[0] # Run the test loop every `validation_freq` epochs during training. if (do_validation and training_utils.should_run_validation(validation_freq, epoch) and not callbacks.model.stop_training): if model._compile_distribution: # Since we create a new clone from the original model we need to copy # the weights back to the original model before we can run validation. distributed_training_utils._copy_weights_to_original_model( model, ModeKeys.TRAIN) val_results = model_iteration( model, val_inputs, targets=val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps_per_epoch=validation_steps, callbacks=callbacks, verbose=0, mode=ModeKeys.TEST, validation_in_fit=True, prepared_feed_values_from_dataset=(val_iterator is not None), steps_name='validation_steps') if not isinstance(val_results, list): val_results = [val_results] epoch_logs = cbks.make_logs( model, epoch_logs, val_results, mode, prefix='val_') if val_iterator and epoch < epochs - 1: _reinitialize_iterator(val_iterator, model._distribution_strategy) if mode == ModeKeys.TRAIN: # Epochs only apply to `fit`. callbacks.on_epoch_end(epoch, epoch_logs) progbar.on_epoch_end(epoch, epoch_logs) # Reinitialize dataset iterator for the next epoch. if reset_dataset_after_each_epoch and epoch < epochs - 1: _reinitialize_iterator(input_iterator, model._distribution_strategy) callbacks._call_end_hook(mode) if model._distribution_strategy: if model._compile_distribution: # TODO(priyag, psv): Copy back metrics to the original model as well? distributed_training_utils._copy_weights_to_original_model(model, mode) scope.__exit__(None, None, None) if mode == ModeKeys.TRAIN: return model.history return results
def fit_loop(model, inputs, targets, sample_weights=None, batch_size=None, epochs=100, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None): """Abstract fit function for arrays of data. Arguments: model: Keras Model instance. inputs: List of input arrays. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: List of input arrays. val_targets: List of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. Returns: `History` object. Raises: ValueError: in case of invalid arguments. """ model._make_train_function() f = model.train_function sample_weights = sample_weights or [] val_sample_weights = val_sample_weights or [] if model.uses_learning_phase and not isinstance(K.learning_phase(), int): ins = inputs + targets + sample_weights + [1] else: ins = inputs + targets + sample_weights do_validation = False if val_inputs: do_validation = True if (steps_per_epoch is None and verbose and inputs and hasattr(inputs[0], 'shape') and hasattr(val_inputs[0], 'shape')): print('Train on %d samples, validate on %d samples' % (inputs[0].shape[0], val_inputs[0].shape[0])) if validation_steps: do_validation = True if steps_per_epoch is None: raise ValueError('Can only use `validation_steps` ' 'when doing step-wise ' 'training, i.e. `steps_per_epoch` ' 'must be set.') num_train_samples = training_utils.check_num_samples( ins, batch_size, steps_per_epoch, 'steps_per_epoch') count_mode = 'steps' if steps_per_epoch else 'samples' callbacks = cbks.configure_callbacks( callbacks, model, do_validation=do_validation, val_inputs=val_inputs, val_targets=val_targets, val_sample_weights=val_sample_weights, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_train_samples, validation_steps=validation_steps, verbose=verbose, count_mode=count_mode) if num_train_samples is not None: index_array = np.arange(num_train_samples) # To prevent a slowdown, we find beforehand the arrays that need conversion. feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights indices_for_conversion_to_dense = [] for i in range(len(feed)): if issparse is not None and issparse(ins[i]) and not K.is_sparse(feed[i]): indices_for_conversion_to_dense.append(i) callbacks.on_train_begin() for epoch in range(initial_epoch, epochs): # Reset stateful metrics for m in model.stateful_metric_functions: m.reset_states() # Update callbacks callbacks.on_epoch_begin(epoch) epoch_logs = {} if steps_per_epoch is not None: # Step-wise fit loop. for step_index in range(steps_per_epoch): batch_logs = {'batch': step_index, 'size': 1} callbacks.on_batch_begin(step_index, batch_logs) try: outs = f(ins) except errors.OutOfRangeError: logging.warning('Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your dataset ' 'can generate at least `steps_per_epoch * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % steps_per_epoch * epochs) break if not isinstance(outs, list): outs = [outs] for l, o in zip(model.metrics_names, outs): batch_logs[l] = o callbacks.on_batch_end(step_index, batch_logs) if callbacks.model.stop_training: break if do_validation: val_outs = test_loop( model, val_inputs, val_targets, sample_weights=val_sample_weights, steps=validation_steps, verbose=0) if not isinstance(val_outs, list): val_outs = [val_outs] # Same labels assumed. for l, o in zip(model.metrics_names, val_outs): epoch_logs['val_' + l] = o else: # Sample-wise fit loop. if shuffle == 'batch': index_array = training_utils.batch_shuffle(index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_train_samples, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] try: if isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') batch_logs = {} batch_logs['batch'] = batch_index batch_logs['size'] = len(batch_ids) callbacks.on_batch_begin(batch_index, batch_logs) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() outs = f(ins_batch) if not isinstance(outs, list): outs = [outs] for l, o in zip(model.metrics_names, outs): batch_logs[l] = o callbacks.on_batch_end(batch_index, batch_logs) if callbacks.model.stop_training: break if batch_index == len(batches) - 1: # Last batch. if do_validation: val_outs = test_loop( model, val_inputs, val_targets, sample_weights=val_sample_weights, batch_size=batch_size, verbose=0) if not isinstance(val_outs, list): val_outs = [val_outs] # Same labels assumed. for l, o in zip(model.metrics_names, val_outs): epoch_logs['val_' + l] = o callbacks.on_epoch_end(epoch, epoch_logs) if callbacks.model.stop_training: break callbacks.on_train_end() return model.history
def test_loop(model, inputs, targets, sample_weights=None, batch_size=None, verbose=0, steps=None): """Abstract method to loop over some data in batches. Arguments: model: Keras Model instance. inputs: List of input arrays. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: integer batch size or `None`. verbose: verbosity mode. steps: Total number of steps (batches of samples) before declaring predictions finished. Ignored with the default value of `None`. Returns: Scalar loss (if the model has a single output and no metrics) or list of scalars (if the model has multiple outputs and/or metrics). The attribute `model.metrics_names` will give you the display labels for the scalar outputs. """ model._make_test_function() f = model.test_function sample_weights = sample_weights or [] if model.uses_learning_phase and not isinstance(K.learning_phase(), int): ins = inputs + targets + sample_weights + [0] else: ins = inputs + targets + sample_weights if hasattr(model, 'metrics'): for m in model.stateful_metric_functions: m.reset_states() stateful_metric_indices = [ i for i, name in enumerate(model.metrics_names) if str(name) in model.stateful_metric_names ] else: stateful_metric_indices = [] num_samples = training_utils.check_num_samples( ins, batch_size, steps, 'steps') outs = [] if verbose == 1: if steps is not None: progbar = Progbar(target=steps) else: progbar = Progbar(target=num_samples) # To prevent a slowdown, we find beforehand the arrays that need conversion. feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights indices_for_conversion_to_dense = [] for i in range(len(feed)): if issparse is not None and issparse(ins[i]) and not K.is_sparse(feed[i]): indices_for_conversion_to_dense.append(i) if steps is not None: for step in range(steps): batch_outs = f(ins) if isinstance(batch_outs, list): if step == 0: for _ in enumerate(batch_outs): outs.append(0.) for i, batch_out in enumerate(batch_outs): if i in stateful_metric_indices: outs[i] = batch_out else: outs[i] += batch_out else: if step == 0: outs.append(0.) outs[0] += batch_outs if verbose == 1: progbar.update(step + 1) for i in range(len(outs)): if i not in stateful_metric_indices: outs[i] /= steps else: batches = make_batches(num_samples, batch_size) index_array = np.arange(num_samples) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] if isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() batch_outs = f(ins_batch) if isinstance(batch_outs, list): if batch_index == 0: outs.extend([0.] * len(batch_outs)) for i, batch_out in enumerate(batch_outs): if i in stateful_metric_indices: outs[i] = batch_out else: outs[i] += batch_out * len(batch_ids) else: if batch_index == 0: outs.append(0.) outs[0] += batch_outs * len(batch_ids) if verbose == 1: progbar.update(batch_end) for i in range(len(outs)): if i not in stateful_metric_indices: outs[i] /= num_samples if len(outs) == 1: return outs[0] return outs
def predict_loop(model, inputs, batch_size=32, verbose=0, steps=None): """Abstract method to loop over some data in batches. Arguments: model: Keras Model instance. inputs: list of tensors to be fed to `f`. batch_size: integer batch size. verbose: verbosity mode. steps: Total number of steps (batches of samples) before declaring `_predict_loop` finished. Ignored with the default value of `None`. Returns: Array of predictions (if the model has a single output) or list of arrays of predictions (if the model has multiple outputs). """ model._make_predict_function() f = model.predict_function if model.uses_learning_phase and not isinstance(K.learning_phase(), int): ins = inputs + [0] else: ins = inputs num_samples = training_utils.check_num_samples( inputs, batch_size, steps, 'steps') if verbose == 1: if steps is not None: progbar = Progbar(target=steps) else: progbar = Progbar(target=num_samples) indices_for_conversion_to_dense = [] for i in range(len(model._feed_inputs)): if (issparse is not None and issparse(inputs[i]) and not K.is_sparse(model._feed_inputs[i])): indices_for_conversion_to_dense.append(i) if steps is not None: # Step-based predictions. # Since we do not know how many samples # we will see, we cannot pre-allocate # the returned Numpy arrays. # Instead, we store one array per batch seen # and concatenate them upon returning. unconcatenated_outs = [] for step in range(steps): batch_outs = f(ins) if not isinstance(batch_outs, list): batch_outs = [batch_outs] if step == 0: for batch_out in batch_outs: unconcatenated_outs.append([]) for i, batch_out in enumerate(batch_outs): unconcatenated_outs[i].append(batch_out) if verbose == 1: progbar.update(step + 1) if len(unconcatenated_outs) == 1: return np.concatenate(unconcatenated_outs[0], axis=0) return [ np.concatenate(unconcatenated_outs[i], axis=0) for i in range(len(unconcatenated_outs)) ] else: # Sample-based predictions. outs = [] batches = make_batches(num_samples, batch_size) index_array = np.arange(num_samples) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] if batch_index == 0: # Pre-allocate the results arrays. for batch_out in batch_outs: shape = (num_samples,) + batch_out.shape[1:] outs.append(np.zeros(shape, dtype=batch_out.dtype)) for i, batch_out in enumerate(batch_outs): outs[i][batch_start:batch_end] = batch_out if verbose == 1: progbar.update(batch_end) if len(outs) == 1: return outs[0] return outs
def model_iteration(model, inputs, targets=None, sample_weights=None, batch_size=None, epochs=1, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None, validation_freq=1, mode=ModeKeys.TRAIN, validation_in_fit=False, prepared_feed_values_from_dataset=False, steps_name='steps', **kwargs): """Loop function for arrays of data with modes TRAIN/TEST/PREDICT. Arguments: model: Keras Model instance. inputs: Either a list or dictionary of arrays, or a dataset instance. targets: List/dictionary of input arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: Either a list or dictionary of arrays, or a dataset instance. val_targets: List/dictionary of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. validation_freq: Only relevant if validation data is provided. Integer or `collections.Container` instance (e.g. list, tuple, etc.). If an integer, specifies how many training epochs to run before a new validation run is performed, e.g. `validation_freq=2` runs validation every 2 epochs. If a Container, specifies the epochs on which to run validation, e.g. `validation_freq=[1, 2, 10]` runs validation at the end of the 1st, 2nd, and 10th epochs. mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT. validation_in_fit: if true, then this method is invoked from within training iteration (for validation). In the case where `val_inputs` is a dataset, this flag indicates that its iterator and feed values are already created so should properly reuse resources. prepared_feed_values_from_dataset: if True, `inputs` is a list of feed tensors returned from `_prepare_feed_values` call on the validation dataset, so do not call it again on `inputs`. Should only be used for inline validation (i.e., only if `validation_in_fit` is also True). steps_name: The string name of the steps argument, either `steps`, `validation_steps`, or `steps_per_epoch`. Only used for error message formatting. **kwargs: Additional arguments for backwards compatibility. Returns: - In TRAIN mode: `History` object. - In TEST mode: Evaluation metrics. - In PREDICT mode: Outputs of the Model called on inputs. Raises: ValueError: in case of invalid arguments. """ # Backwards compatibility. if 'steps' in kwargs: steps_per_epoch = kwargs.pop('steps') if kwargs: raise TypeError('Unknown arguments: %s' % (kwargs, )) # In case we were passed a dataset, we extract symbolic tensors from it. reset_dataset_after_each_epoch = False input_iterator = None is_dataset = isinstance(inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)) # TODO(fchollet): consider moving `steps_per_epoch` inference to # _standardize_user_data and set reset_dataset_after_each_epoch as an # attribute on the dataset instance. if is_dataset: if steps_per_epoch is None: reset_dataset_after_each_epoch = True steps_per_epoch = training_utils.infer_steps_for_dataset( inputs, steps_per_epoch, epochs=epochs, steps_name=steps_name) input_iterator = _get_iterator(inputs, model._distribution_strategy) if mode == ModeKeys.TRAIN: _print_train_info(inputs, val_inputs, steps_per_epoch, verbose) # Enter DistributionStrategy scope. if model._distribution_strategy: scope = model._distribution_strategy.scope() scope.__enter__() # Get step function and loop type. f = _make_execution_function(model, mode) use_steps = is_dataset or steps_per_epoch is not None do_validation = val_inputs is not None # Convert Eager Tensors to NumPy arrays to support batching/shuffling. inputs, targets, sample_weights = training_utils. \ convert_eager_tensors_to_numpy((inputs, targets, sample_weights)) # Prepare input data. inputs = input_iterator or inputs if validation_in_fit and prepared_feed_values_from_dataset: # When invoking validation in training loop, avoid creating iterator and # list of feed values for the same validation dataset multiple times (which # essentially would call `iterator.get_next()` that slows down execution and # leads to OOM errors eventually. ins = inputs else: ins = _prepare_feed_values(model, inputs, targets, sample_weights, mode) if not is_dataset: num_samples_or_steps = _get_num_samples_or_steps( ins, batch_size, steps_per_epoch) else: num_samples_or_steps = steps_per_epoch # Prepare validation data. Hold references to the iterator and the input list # to properly reinitialize and reuse in multiple validation passes. val_iterator = None if isinstance(val_inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)): val_iterator = _get_iterator(val_inputs, model._distribution_strategy) val_inputs = _prepare_feed_values(model, val_iterator, val_targets, val_sample_weights, ModeKeys.TEST) # Configure callbacks. count_mode = 'steps' if use_steps else 'samples' callbacks = cbks.configure_callbacks( callbacks, model, do_validation=do_validation, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_samples_or_steps, verbose=0, # Handle ProgBarLogger separately in this loop. mode=mode) # TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready. progbar = training_utils.get_progbar(model, count_mode) progbar.params = callbacks.params progbar.params['verbose'] = verbose # Find beforehand arrays that need sparse-to-dense conversion. if issparse is not None and not use_steps: indices_for_conversion_to_dense = [] feed = _get_model_feed(model, mode) for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)): if issparse(input_data) and not K.is_sparse(feed_tensor): indices_for_conversion_to_dense.append(i) # Select aggregation method. if mode == ModeKeys.PREDICT: aggregator = training_utils.OutputsAggregator(use_steps, num_samples_or_steps) else: aggregator = training_utils.MetricsAggregator(use_steps, num_samples_or_steps) if model._compile_distribution and not validation_in_fit: distributed_training_utils._copy_weights_to_distributed_model( model, model._distributed_model) callbacks.model.stop_training = False callbacks._call_begin_hook(mode) progbar.on_train_begin() for epoch in range(initial_epoch, epochs): if callbacks.model.stop_training: break # Setup work for each epoch epoch_logs = {} model.reset_metrics() if mode == ModeKeys.TRAIN: callbacks.on_epoch_begin(epoch, epoch_logs) progbar.on_epoch_begin(epoch, epoch_logs) if use_steps: # Step-wise loop. if steps_per_epoch is None: # Loop over dataset until `OutOfRangeError` is raised. target_steps = np.inf else: # Loop over dataset for the specified number of steps. target_steps = steps_per_epoch step = 0 while step < target_steps: batch_logs = {'batch': step, 'size': 1} callbacks._call_batch_hook(mode, 'begin', step, batch_logs) progbar.on_batch_begin(step, batch_logs) # Get outputs. try: # `ins` can be callable in DistributionStrategy + eager case. actual_inputs = ins() if callable(ins) else ins batch_outs = f(actual_inputs) except errors.OutOfRangeError: if not is_dataset: # We ran out of batches while the user passed an iterator (legacy). logging.warning( 'Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your iterator ' 'can generate at least `%s * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % (steps_name, steps_per_epoch * epochs)) callbacks.model.stop_training = True else: # The dataset passed by the user ran out of batches. # Now we know the cardinality of the dataset. if step > 0: steps_per_epoch = step aggregator.num_samples_or_steps = steps_per_epoch progbar.params['steps'] = steps_per_epoch progbar.progbar.target = steps_per_epoch break if not isinstance(batch_outs, list): batch_outs = [batch_outs] if model._distribution_strategy: batch_outs = distributed_training_utils._per_device_aggregate_batch( batch_outs, model, mode) # Aggregate results. if step == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', step, batch_logs) progbar.on_batch_end(step, batch_logs) step += 1 if callbacks.model.stop_training: break else: # Sample-wise loop. index_array = np.arange(num_samples_or_steps) if shuffle == 'batch': index_array = training_utils.batch_shuffle( index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_samples_or_steps, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] # Slice into a batch. try: if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') # Sparse to dense conversion. if issparse is not None: for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() # Callbacks batch_begin. batch_logs = {'batch': batch_index, 'size': len(batch_ids)} callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs) progbar.on_batch_begin(batch_index, batch_logs) # Get outputs. batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if batch_index == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs, batch_start, batch_end) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs) progbar.on_batch_end(batch_index, batch_logs) if callbacks.model.stop_training: break aggregator.finalize() results = aggregator.results epoch_logs = cbks.make_logs(model, epoch_logs, results, mode) if len(results) == 1: results = results[0] # Run the test loop every `validation_freq` epochs during training. if (do_validation and training_utils.should_run_validation( validation_freq, epoch) and not callbacks.model.stop_training): val_results = model_iteration( model, val_inputs, targets=val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps_per_epoch=validation_steps, callbacks=callbacks, verbose=0, mode=ModeKeys.TEST, validation_in_fit=True, prepared_feed_values_from_dataset=(val_iterator is not None), steps_name='validation_steps') if not isinstance(val_results, list): val_results = [val_results] epoch_logs = cbks.make_logs(model, epoch_logs, val_results, mode, prefix='val_') if val_iterator and epoch < epochs - 1: _reinitialize_iterator(val_iterator, model._distribution_strategy) if mode == ModeKeys.TRAIN: # Epochs only apply to `fit`. callbacks.on_epoch_end(epoch, epoch_logs) progbar.on_epoch_end(epoch, epoch_logs) # Reinitialize dataset iterator for the next epoch. if reset_dataset_after_each_epoch and epoch < epochs - 1: _reinitialize_iterator(input_iterator, model._distribution_strategy) callbacks._call_end_hook(mode) if model._distribution_strategy: if model._compile_distribution and not validation_in_fit: # TODO(priyag, psv): Copy back metrics to the original model as well? distributed_training_utils._copy_weights_to_original_model( model, model._distributed_model, mode) scope.__exit__(None, None, None) if mode == ModeKeys.TRAIN: return model.history return results
def model_iteration(model, inputs, targets=None, sample_weights=None, batch_size=None, epochs=1, verbose=1, callbacks=None, val_inputs=None, val_targets=None, val_sample_weights=None, shuffle=True, initial_epoch=0, steps_per_epoch=None, validation_steps=None, mode=ModeKeys.TRAIN, validation_in_fit=False, **kwargs): """Loop function for arrays of data with modes TRAIN/TEST/PREDICT. Arguments: model: Keras Model instance. inputs: Either a list of arrays or a dictionary. targets: List of target arrays. sample_weights: Optional list of sample weight arrays. batch_size: Integer batch size or None if unknown. epochs: Number of times to iterate over the data verbose: Verbosity mode, 0, 1 or 2 callbacks: List of callbacks to be called during training val_inputs: List of input arrays. val_targets: List of target arrays. val_sample_weights: Optional list of sample weight arrays. shuffle: Whether to shuffle the data at the beginning of each epoch concatenation of list the display names of the outputs of `f` and the list of display names of the outputs of `f_val`. initial_epoch: Epoch at which to start training (useful for resuming a previous training run) steps_per_epoch: Total number of steps (batches of samples) before declaring one epoch finished and starting the next epoch. Ignored with the default value of `None`. validation_steps: Number of steps to run validation for (only if doing validation from data tensors). Ignored with the default value of `None`. mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT. validation_in_fit: DEPRECATED: if true, then this method is invoked from within training iteration (for validation). In this case, do not copy weights when using a tf.distribute.Strategy. The input is deprecated as it is not required if the user creates a distributed model under the distribution strategy scope rather than passing it to compile. **kwargs: Additional arguments for backwards compatibility. Returns: - In TRAIN mode: `History` object. - In TEST mode: Evaluation metrics. - In PREDICT mode: Outputs of the Model called on inputs. Raises: ValueError: in case of invalid arguments. """ # Backwards compatibility. if 'steps' in kwargs: steps_per_epoch = kwargs['steps'] _validate_arguments(steps_per_epoch, validation_steps, kwargs) if mode == ModeKeys.TRAIN: _print_train_info(inputs, val_inputs, steps_per_epoch, verbose) # Enter DistributionStrategy scope. if model._distribution_strategy: scope = model._distribution_strategy.scope() scope.__enter__() # Get step function and loop type. f = _make_execution_function(model, mode) use_steps = steps_per_epoch is not None do_validation = val_inputs is not None # Convert Eager Tensors to NumPy arrays to support batching/shuffling. inputs, targets, sample_weights = training_utils. \ convert_eager_tensors_to_numpy((inputs, targets, sample_weights)) # Prepare input data. ins = _prepare_feed_values(model, inputs, targets, sample_weights, mode) num_samples_or_steps = _get_num_samples_or_steps(ins, batch_size, steps_per_epoch) # Configure callbacks. count_mode = 'steps' if use_steps else 'samples' callbacks = cbks.configure_callbacks( callbacks, model, do_validation=do_validation, batch_size=batch_size, epochs=epochs, steps_per_epoch=steps_per_epoch, samples=num_samples_or_steps, verbose=0, # Handle ProgBarLogger separately in this loop. mode=mode) # TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready. progbar = training_utils.get_progbar(model, count_mode) progbar.params = callbacks.params progbar.params['verbose'] = verbose # Find beforehand arrays that need sparse-to-dense conversion. if issparse is not None and not use_steps: indices_for_conversion_to_dense = [] feed = _get_model_feed(model, mode) for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)): if issparse(input_data) and not K.is_sparse(feed_tensor): indices_for_conversion_to_dense.append(i) # Select aggregation method. if mode == ModeKeys.PREDICT: aggregator = training_utils.OutputsAggregator(use_steps, num_samples_or_steps) else: aggregator = training_utils.MetricsAggregator(use_steps, num_samples_or_steps) if model._compile_distribution and not validation_in_fit: distributed_training_utils._copy_weights_to_distributed_model( model, model._distributed_model) callbacks.model.stop_training = False callbacks._call_begin_hook(mode) progbar.on_train_begin() for epoch in range(initial_epoch, epochs): if callbacks.model.stop_training: break # Setup work for each epoch epoch_logs = {} model.reset_metrics() if mode == ModeKeys.TRAIN: callbacks.on_epoch_begin(epoch, epoch_logs) progbar.on_epoch_begin(epoch, epoch_logs) if use_steps: # Step-wise loop. for step in range(steps_per_epoch): batch_logs = {'batch': step, 'size': 1} callbacks._call_batch_hook(mode, 'begin', step, batch_logs) progbar.on_batch_begin(step, batch_logs) # Get outputs. try: # `ins` can be callable in DistributionStrategy + eager case. actual_inputs = ins() if callable(ins) else ins batch_outs = f(actual_inputs) except errors.OutOfRangeError: logging.warning('Your dataset iterator ran out of data; ' 'interrupting training. Make sure that your dataset ' 'can generate at least `steps_per_epoch * epochs` ' 'batches (in this case, %d batches). You may need to' 'use the repeat() function when building your ' 'dataset.' % steps_per_epoch * epochs) break if not isinstance(batch_outs, list): batch_outs = [batch_outs] if model._distribution_strategy: batch_outs = distributed_training_utils._per_device_aggregate_batch( batch_outs, model, mode) # Aggregate results. if step == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', step, batch_logs) progbar.on_batch_end(step, batch_logs) if callbacks.model.stop_training: break else: # Sample-wise loop. index_array = np.arange(num_samples_or_steps) if shuffle == 'batch': index_array = training_utils.batch_shuffle(index_array, batch_size) elif shuffle: np.random.shuffle(index_array) batches = make_batches(num_samples_or_steps, batch_size) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] # Slice into a batch. try: if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) except TypeError: raise TypeError('TypeError while preparing batch. ' 'If using HDF5 input data, ' 'pass shuffle="batch".') # Sparse to dense conversion. if issparse is not None: for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() # Callbacks batch_begin. batch_logs = {'batch': batch_index, 'size': len(batch_ids)} callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs) progbar.on_batch_begin(batch_index, batch_logs) # Get outputs. batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] # Aggregate results. if batch_index == 0: aggregator.create(batch_outs) aggregator.aggregate(batch_outs, batch_start, batch_end) # Callbacks batch end. batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode) callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs) progbar.on_batch_end(batch_index, batch_logs) if callbacks.model.stop_training: break aggregator.finalize() results = aggregator.results epoch_logs = cbks.make_logs(model, epoch_logs, results, mode) if len(results) == 1: results = results[0] # Run the test loop every epoch during training. if do_validation and not callbacks.model.stop_training: val_results = model_iteration( model, val_inputs, targets=val_targets, sample_weights=val_sample_weights, batch_size=batch_size, steps_per_epoch=validation_steps, callbacks=callbacks, verbose=0, mode=ModeKeys.TEST, validation_in_fit=True) if not isinstance(val_results, list): val_results = [val_results] epoch_logs = cbks.make_logs( model, epoch_logs, val_results, mode, prefix='val_') if mode == ModeKeys.TRAIN: # Epochs only apply to `fit`. callbacks.on_epoch_end(epoch, epoch_logs) progbar.on_epoch_end(epoch, epoch_logs) callbacks._call_end_hook(mode) if model._distribution_strategy: if model._compile_distribution and not validation_in_fit: # TODO(priyag, psv): Copy back metrics to the original model as well? distributed_training_utils._copy_weights_to_original_model( model, model._distributed_model, mode) scope.__exit__(None, None, None) if mode == ModeKeys.TRAIN: return model.history return results
if len(unconcatenated_outs) == 1: return np.concatenate(unconcatenated_outs[0], axis=0) return [ np.concatenate(unconcatenated_outs[i], axis=0) for i in range(len(unconcatenated_outs)) ] else: # Sample-based predictions. outs = [] batches = make_batches(num_samples, batch_size) index_array = np.arange(num_samples) for batch_index, (batch_start, batch_end) in enumerate(batches): batch_ids = index_array[batch_start:batch_end] if ins and isinstance(ins[-1], int): # Do not slice the training phase flag. ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]] else: ins_batch = slice_arrays(ins, batch_ids) for i in indices_for_conversion_to_dense: ins_batch[i] = ins_batch[i].toarray() batch_outs = f(ins_batch) if not isinstance(batch_outs, list): batch_outs = [batch_outs] if batch_index == 0: # Pre-allocate the results arrays. for batch_out in batch_outs: shape = (num_samples,) + batch_out.shape[1:] outs.append(np.zeros(shape, dtype=batch_out.dtype)) for i, batch_out in enumerate(batch_outs): outs[i][batch_start:batch_end] = batch_out