# Copyright 2019 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """The V2 implementation of Normalization layers.""" import warnings import tensorflow.compat.v2 as tf from tf_keras.src import backend from tf_keras.src import constraints from tf_keras.src import initializers from tf_keras.src import regularizers from tf_keras.src.dtensor import utils from tf_keras.src.engine.base_layer import Layer from tf_keras.src.engine.input_spec import InputSpec from tf_keras.src.utils import control_flow_util from tf_keras.src.utils import tf_utils # isort: off from tensorflow.python.ops.control_flow_ops import ( get_enclosing_xla_context, ) from tensorflow.python.platform import tf_logging as logging from tensorflow.python.util import deprecation from tensorflow.python.util.tf_export import keras_export class BatchNormalizationBase(Layer): r"""Layer that normalizes its inputs. Batch normalization applies a transformation that maintains the mean output close to 0 and the output standard deviation close to 1. Importantly, batch normalization works differently during training and during inference. **During training** (i.e. when using `fit()` or when calling the layer/model with the argument `training=True`), the layer normalizes its output using the mean and standard deviation of the current batch of inputs. That is to say, for each channel being normalized, the layer returns `gamma * (batch - mean(batch)) / sqrt(var(batch) + epsilon) + beta`, where: - `epsilon` is small constant (configurable as part of the constructor arguments) - `gamma` is a learned scaling factor (initialized as 1), which can be disabled by passing `scale=False` to the constructor. - `beta` is a learned offset factor (initialized as 0), which can be disabled by passing `center=False` to the constructor. **During inference** (i.e. when using `evaluate()` or `predict()`) or when calling the layer/model with the argument `training=False` (which is the default), the layer normalizes its output using a moving average of the mean and standard deviation of the batches it has seen during training. That is to say, it returns `gamma * (batch - self.moving_mean) / sqrt(self.moving_var+epsilon) + beta`. `self.moving_mean` and `self.moving_var` are non-trainable variables that are updated each time the layer in called in training mode, as such: - `moving_mean = moving_mean * momentum + mean(batch) * (1 - momentum)` - `moving_var = moving_var * momentum + var(batch) * (1 - momentum)` As such, the layer will only normalize its inputs during inference *after having been trained on data that has similar statistics as the inference data*. Args: axis: Integer or a list of integers, the axis that should be normalized (typically the features axis). For instance, after a `Conv2D` layer with `data_format="channels_first"`, set `axis=1` in `BatchNormalization`. momentum: Momentum for the moving average. epsilon: Small float added to variance to avoid dividing by zero. center: If True, add offset of `beta` to normalized tensor. If False, `beta` is ignored. scale: If True, multiply by `gamma`. If False, `gamma` is not used. When the next layer is linear (also e.g. `nn.relu`), this can be disabled since the scaling will be done by the next layer. beta_initializer: Initializer for the beta weight. gamma_initializer: Initializer for the gamma weight. moving_mean_initializer: Initializer for the moving mean. moving_variance_initializer: Initializer for the moving variance. beta_regularizer: Optional regularizer for the beta weight. gamma_regularizer: Optional regularizer for the gamma weight. beta_constraint: Optional constraint for the beta weight. gamma_constraint: Optional constraint for the gamma weight. renorm: Whether to use [Batch Renormalization]( https://arxiv.org/abs/1702.03275). This adds extra variables during training. The inference is the same for either value of this parameter. renorm_clipping: A dictionary that may map keys 'rmax', 'rmin', 'dmax' to scalar `Tensors` used to clip the renorm correction. The correction `(r, d)` is used as `corrected_value = normalized_value * r + d`, with `r` clipped to [rmin, rmax], and `d` to [-dmax, dmax]. Missing rmax, rmin, dmax are set to inf, 0, inf, respectively. renorm_momentum: Momentum used to update the moving means and standard deviations with renorm. Unlike `momentum`, this affects training and should be neither too small (which would add noise) nor too large (which would give stale estimates). Note that `momentum` is still applied to get the means and variances for inference. fused: if `True`, use a faster, fused implementation, or raise a ValueError if the fused implementation cannot be used. If `None`, use the faster implementation if possible. If False, do not used the fused implementation. Note that in TensorFlow 1.x, the meaning of `fused=True` is different: if `False`, the layer uses the system-recommended implementation. You cannot use `fused=True` if a mask is passed in the `call()` method. trainable: Boolean, if `True` the variables will be marked as trainable. virtual_batch_size: An `int`. By default, `virtual_batch_size` is `None`, which means batch normalization is performed across the whole batch. When `virtual_batch_size` is not `None`, instead perform "Ghost Batch Normalization", which creates virtual sub-batches which are each normalized separately (with shared gamma, beta, and moving statistics). Must divide the actual batch size during execution. adjustment: A function taking the `Tensor` containing the (dynamic) shape of the input tensor and returning a pair (scale, bias) to apply to the normalized values (before gamma and beta), only during training. For example, if `axis=-1`, `adjustment = lambda shape: ( tf.random.uniform(shape[-1:], 0.93, 1.07), tf.random.uniform(shape[-1:], -0.1, 0.1))` will scale the normalized value by up to 7% up or down, then shift the result by up to 0.1 (with independent scaling and bias for each feature but shared across all examples), and finally apply gamma and/or beta. If `None`, no adjustment is applied. Cannot be specified if virtual_batch_size is specified. synchronized: If True, synchronizes the global batch statistics (mean and variance) for the layer across all devices at each training step in a distributed training strategy. If False, each replica uses its own local batch statistics. Only relevant when used inside a `tf.distribute` strategy. Call arguments: inputs: Input tensor (of any rank). training: Python boolean indicating whether the layer should behave in training mode or in inference mode. - `training=True`: The layer will normalize its inputs using the mean and variance of the current batch of inputs. - `training=False`: The layer will normalize its inputs using the mean and variance of its moving statistics, learned during training. mask: Binary tensor of shape broadcastable to `inputs` tensor, indicating the positions for which the mean and variance should be computed. Input shape: Arbitrary. Use the keyword argument `input_shape` (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model. Output shape: Same shape as input. Reference: - [Ioffe and Szegedy, 2015](https://arxiv.org/abs/1502.03167). """ # By default, the base class uses V2 behavior. The BatchNormalization V1 # subclass sets this to False to use the V1 behavior. _USE_V2_BEHAVIOR = True def __init__( self, axis=-1, momentum=0.99, epsilon=1e-3, center=True, scale=True, beta_initializer="zeros", gamma_initializer="ones", moving_mean_initializer="zeros", moving_variance_initializer="ones", beta_regularizer=None, gamma_regularizer=None, beta_constraint=None, gamma_constraint=None, renorm=False, renorm_clipping=None, renorm_momentum=0.99, fused=None, trainable=True, virtual_batch_size=None, adjustment=None, name=None, synchronized=False, **kwargs, ): super().__init__(name=name, **kwargs) if isinstance(axis, (list, tuple)): self.axis = axis[:] elif isinstance(axis, int): self.axis = axis else: raise TypeError( "Expected an int or a list/tuple of ints for the " "argument 'axis', but received: %r" % axis ) if synchronized and fused: raise ValueError( "`fused=True` is not supported when `synchronized=True`." ) self.synchronized = synchronized if self.synchronized: fused = False self.momentum = momentum self.epsilon = epsilon self.center = center self.scale = scale self.beta_initializer = initializers.get(beta_initializer) self.gamma_initializer = initializers.get(gamma_initializer) self.moving_mean_initializer = initializers.get(moving_mean_initializer) self.moving_variance_initializer = initializers.get( moving_variance_initializer ) self.beta_regularizer = regularizers.get(beta_regularizer) self.gamma_regularizer = regularizers.get(gamma_regularizer) self.beta_constraint = constraints.get(beta_constraint) self.gamma_constraint = constraints.get(gamma_constraint) self.renorm = renorm self.virtual_batch_size = virtual_batch_size self.adjustment = adjustment if self._USE_V2_BEHAVIOR: if fused: self._raise_if_fused_cannot_be_used() # We leave fused as None if self._fused_can_be_used()==True, since # we still may set it to False in self.build() if the input rank is # not 4. elif fused is None and not self._fused_can_be_used(): fused = False elif fused is None: fused = True self.supports_masking = True self.fused = fused self._bessels_correction_test_only = True self.trainable = trainable if renorm: renorm_clipping = renorm_clipping or {} keys = ["rmax", "rmin", "dmax"] if set(renorm_clipping) - set(keys): raise ValueError( "Received invalid keys for `renorm_clipping` argument: " f"{renorm_clipping}. Supported values: {keys}." ) self.renorm_clipping = renorm_clipping self.renorm_momentum = renorm_momentum def _raise_if_fused_cannot_be_used(self): """Raises a ValueError if fused implementation cannot be used. In addition to the checks done in this function, the input tensors rank must be 4 or 5. The input rank check can only be done once the input shape is known. """ # Note the ValueErrors in this function are caught and not reraised in # _fused_can_be_used(). No other exception besides ValueError should be # raised here. # Currently fused batch norm doesn't support renorm. It also only # supports a channel dimension on axis 1 or 3 (rank=4) / 1 or 4 (rank5), # when no virtual batch size or adjustment is used. if self.renorm: raise ValueError( "Passing both `fused=True` and `renorm=True` is not supported" ) axis = [self.axis] if isinstance(self.axis, int) else self.axis # Axis -3 is equivalent to 1, and axis -1 is equivalent to 3, when the # input rank is 4. Similarly, the valid axis is -4, -1, 1, 4 when the # rank is 5. The combination of ranks and axes will be checked later. if len(axis) > 1 or axis[0] not in (-4, -3, -1, 1, 3, 4): raise ValueError( "Passing `fused=True` is only supported when axis is 1 " "or 3 for input rank = 4 or 1 or 4 for input rank = 5. " "Got axis %s" % (axis,) ) if self.virtual_batch_size is not None: raise ValueError( "Passing `fused=True` is not supported when " "`virtual_batch_size` is specified." ) if self.adjustment is not None: raise ValueError( "Passing `fused=True` is not supported when " "`adjustment` is specified." ) # TODO(reedwm): Support fp64 in FusedBatchNorm then remove this check. if self._compute_dtype not in ("float16", "bfloat16", "float32", None): raise ValueError( "Passing `fused=True` is only supported when the compute " "dtype is float16, bfloat16, or float32. Got dtype: %s" % (self._compute_dtype,) ) def _fused_can_be_used(self): try: self._raise_if_fused_cannot_be_used() return True except ValueError: return False @property def trainable(self): return self._trainable @trainable.setter def trainable(self, value): self._trainable = value @property def _param_dtype(self): # Raise parameters of fp16 batch norm to fp32 if self.dtype == tf.float16 or self.dtype == tf.bfloat16: return tf.float32 else: return self.dtype or tf.float32 def build(self, input_shape): self.axis = tf_utils.validate_axis(self.axis, input_shape) input_shape = tf.TensorShape(input_shape) rank = input_shape.rank if self.virtual_batch_size is not None: if self.virtual_batch_size <= 0: raise ValueError( "`virtual_batch_size` must be a positive integer that " "divides the true batch size of the input tensor. " f"Received: virtual_batch_size={self.virtual_batch_size}" ) # If using virtual batches, the first dimension must be the batch # dimension and cannot be the batch norm axis if 0 in self.axis: raise ValueError( "When using `virtual_batch_size`, the batch dimension " "must be 0 and thus axis cannot include 0. " f"Received axis={self.axis}" ) if self.adjustment is not None: raise ValueError( "When using `virtual_batch_size`, adjustment cannot " "be specified" ) if self.fused in (None, True): # TODO(yaozhang): if input is not 4D, reshape it to 4D and reshape # the output back to its original shape accordingly. if self._USE_V2_BEHAVIOR: if self.fused is None: self.fused = rank in (4, 5) elif self.fused and rank not in (4, 5): raise ValueError( "Batch normalization layers with `fused=True` only " "support 4D or 5D input tensors. " f"Received tensor with shape: {tuple(input_shape)}" ) else: assert self.fused is not None self.fused = rank in (4, 5) and self._fused_can_be_used() # TODO(chrisying): fused batch norm is currently not supported for # multi-axis batch norm and by extension virtual batches. In some # cases, it might be possible to use fused batch norm but would # require reshaping the Tensor to 4D with the axis in 1 or 3 # (preferred 1) which is particularly tricky. A compromise might be # to just support the most common use case (turning 5D w/ virtual # batch to NCHW) if self.fused: if self.axis == [1] and rank == 4: self._data_format = "NCHW" elif self.axis == [1] and rank == 5: self._data_format = "NCDHW" elif self.axis == [3] and rank == 4: self._data_format = "NHWC" elif self.axis == [4] and rank == 5: self._data_format = "NDHWC" elif rank == 5: # 5D tensors that can be passed in but should not use fused # batch norm due to unsupported axis. self.fused = False else: if rank == 4: raise ValueError( "Unsupported axis. The use of `fused=True` is only " "possible with `axis=1` or `axis=3` for 4D input " f"tensors. Received: axis={tuple(self.axis)}" ) else: raise ValueError( "Unsupported axis. The use of `fused=True` is only " "possible with `axis=1` or `axis=4` for 5D input " f"tensors. Received: axis={tuple(self.axis)}" ) axis_to_dim = {x: input_shape.dims[x].value for x in self.axis} for x in axis_to_dim: if axis_to_dim[x] is None: raise ValueError( "Input has undefined `axis` dimension. Received input " f"with shape {tuple(input_shape)} " f"and axis={tuple(self.axis)}" ) self.input_spec = InputSpec(ndim=rank, axes=axis_to_dim) if len(axis_to_dim) == 1 and self.virtual_batch_size is None: # Single axis batch norm (most common/default use-case) param_shape = (list(axis_to_dim.values())[0],) else: # Parameter shape is the original shape but with 1 in all non-axis # dims param_shape = [ axis_to_dim[i] if i in axis_to_dim else 1 for i in range(rank) ] if self.virtual_batch_size is not None: # When using virtual batches, add an extra dim at index 1 param_shape.insert(1, 1) for idx, x in enumerate(self.axis): self.axis[idx] = x + 1 # Account for added dimension self._param_shape = param_shape if self.scale: self.gamma = self.add_weight( name="gamma", shape=param_shape, dtype=self._param_dtype, initializer=self.gamma_initializer, regularizer=self.gamma_regularizer, constraint=self.gamma_constraint, trainable=True, experimental_autocast=False, ) else: self.gamma = None if self.center: self.beta = self.add_weight( name="beta", shape=param_shape, dtype=self._param_dtype, initializer=self.beta_initializer, regularizer=self.beta_regularizer, constraint=self.beta_constraint, trainable=True, experimental_autocast=False, ) else: self.beta = None try: # Disable variable partitioning when creating the moving mean and # variance if hasattr(self, "_scope") and self._scope: partitioner = self._scope.partitioner self._scope.set_partitioner(None) else: partitioner = None self.moving_mean = self.add_weight( name="moving_mean", shape=param_shape, dtype=self._param_dtype, initializer=self.moving_mean_initializer, synchronization=tf.VariableSynchronization.ON_READ, trainable=False, aggregation=tf.VariableAggregation.MEAN, experimental_autocast=False, ) self.moving_variance = self.add_weight( name="moving_variance", shape=param_shape, dtype=self._param_dtype, initializer=self.moving_variance_initializer, synchronization=tf.VariableSynchronization.ON_READ, trainable=False, aggregation=tf.VariableAggregation.MEAN, experimental_autocast=False, ) if self.renorm: # In batch renormalization we track the inference moving stddev # instead of the moving variance to more closely align with the # paper. def moving_stddev_initializer(*args, **kwargs): return tf.sqrt( self.moving_variance_initializer(*args, **kwargs) ) with tf.distribute.get_strategy().extended.colocate_vars_with( self.moving_variance ): self.moving_stddev = self.add_weight( name="moving_stddev", shape=param_shape, dtype=self._param_dtype, initializer=moving_stddev_initializer, synchronization=tf.VariableSynchronization.ON_READ, trainable=False, aggregation=tf.VariableAggregation.MEAN, experimental_autocast=False, ) # Create variables to maintain the moving mean and standard # deviation. These are used in training and thus are different # from the moving averages above. The renorm variables are # colocated with moving_mean and moving_stddev. # NOTE: below, the outer `with device` block causes the current # device stack to be cleared. The nested ones use a `lambda` to # set the desired device and ignore any devices that may be set # by the custom getter. def _renorm_variable(name, shape, initializer="zeros"): """Create a renorm variable.""" var = self.add_weight( name=name, shape=shape, dtype=self._param_dtype, initializer=initializer, synchronization=tf.VariableSynchronization.ON_READ, trainable=False, aggregation=tf.VariableAggregation.MEAN, experimental_autocast=False, ) return var with tf.distribute.get_strategy().extended.colocate_vars_with( self.moving_mean ): self.renorm_mean = _renorm_variable( "renorm_mean", param_shape, self.moving_mean_initializer ) with tf.distribute.get_strategy().extended.colocate_vars_with( self.moving_stddev ): self.renorm_stddev = _renorm_variable( "renorm_stddev", param_shape, moving_stddev_initializer ) finally: if partitioner: self._scope.set_partitioner(partitioner) super().build(input_shape) def call(self, inputs, training=None, mask=None): inputs = tf.cast(inputs, self.compute_dtype) training = self._get_training_value(training) # Determine a boolean value for `training`: could be True, False, or # None. training_value = control_flow_util.constant_value(training) _raise_for_non_sync_bn_with_renorm_and_dtensor_strategy( synchronized=self.synchronized, training=training, renorm=self.renorm, ) if self.virtual_batch_size is not None: # Virtual batches (aka ghost batches) can be simulated by reshaping # the Tensor and reusing the existing batch norm implementation original_shape = tf.shape(inputs) original_shape = tf.concat( [tf.constant([-1]), original_shape[1:]], axis=0 ) if tf.__internal__.tf2.enabled(): expanded_shape = ( [self.virtual_batch_size, -1] if training_value else [-1, 1] ) expanded_shape = tf.concat( [ tf.constant(expanded_shape), original_shape[1:], ], axis=0, ) else: # Preserve incorrect legacy behavior for backwards compatibility expanded_shape = tf.concat( [ tf.constant([self.virtual_batch_size, -1]), original_shape[1:], ], axis=0, ) # Will cause errors if virtual_batch_size does not divide the batch # size inputs = tf.reshape(inputs, expanded_shape) def undo_virtual_batching(outputs): outputs = tf.reshape(outputs, original_shape) return outputs if self.fused: outputs = self._fused_batch_norm( inputs, mask=mask, training=training ) if self.virtual_batch_size is not None: # Currently never reaches here since fused_batch_norm does not # support virtual batching outputs = undo_virtual_batching(outputs) return outputs inputs_dtype = inputs.dtype.base_dtype if inputs_dtype in (tf.float16, tf.bfloat16): # Do all math in float32 if given 16-bit inputs for numeric # stability. In particular, it's very easy for variance to overflow # in float16 and for safety we also choose to cast bfloat16 to # float32. inputs = tf.cast(inputs, tf.float32) # Compute the axes along which to reduce the mean / variance input_shape = inputs.shape ndims = len(input_shape) reduction_axes = [i for i in range(ndims) if i not in self.axis] if self.virtual_batch_size is not None: del reduction_axes[1] # Do not reduce along virtual batch dim # Broadcasting only necessary for single-axis batch norm where the axis # is not the last dimension broadcast_shape = [1] * ndims broadcast_shape[self.axis[0]] = input_shape.dims[self.axis[0]].value def _broadcast(v): if ( v is not None and len(v.shape) != ndims and reduction_axes != list(range(ndims - 1)) ): return tf.reshape(v, broadcast_shape) return v scale, offset = _broadcast(self.gamma), _broadcast(self.beta) def _compose_transforms(scale, offset, then_scale, then_offset): if then_scale is not None: scale *= then_scale offset *= then_scale if then_offset is not None: offset += then_offset return (scale, offset) if training_value == False: # noqa: E712 mean, variance = self.moving_mean, self.moving_variance else: # The following long block are handling mean/variance update during # the training stage in various of different settings. if self.adjustment: adj_scale, adj_bias = self.adjustment(tf.shape(inputs)) # Adjust only during training. adj_scale = control_flow_util.smart_cond( training, lambda: adj_scale, lambda: tf.ones_like(adj_scale) ) adj_bias = control_flow_util.smart_cond( training, lambda: adj_bias, lambda: tf.zeros_like(adj_bias) ) scale, offset = _compose_transforms( adj_scale, adj_bias, scale, offset ) # Some of the computations here are not necessary when # training==False but not a constant. However, this makes the code # simpler. keep_dims = ( self.virtual_batch_size is not None or len(self.axis) > 1 ) mean, variance = self._moments( tf.cast(inputs, self._param_dtype), reduction_axes, keep_dims=keep_dims, mask=mask, ) moving_mean = self.moving_mean moving_variance = self.moving_variance mean = control_flow_util.smart_cond( training, lambda: mean, lambda: tf.convert_to_tensor(moving_mean), ) variance = control_flow_util.smart_cond( training, lambda: variance, lambda: tf.convert_to_tensor(moving_variance), ) if self.virtual_batch_size is not None: # This isn't strictly correct since in ghost batch norm, you are # supposed to sequentially update the moving_mean and # moving_variance with each sub-batch. However, since the moving # statistics are only used during evaluation, it is more # efficient to just update in one step and should not make a # significant difference in the result. new_mean = tf.reduce_mean(mean, axis=1, keepdims=True) new_variance = tf.reduce_mean(variance, axis=1, keepdims=True) else: if ( utils.running_with_dtensor_strategy() and not self.synchronized ): new_mean = tf.math.reduce_mean(mean, axis=reduction_axes) new_variance = tf.math.reduce_mean( variance, axis=reduction_axes ) else: new_mean, new_variance = mean, variance if self._support_zero_size_input(): # TF-Keras assumes that batch dimension is the first dimension # for Batch Normalization. input_batch_size = tf.shape(inputs)[0] else: input_batch_size = None if self.renorm: ( r, d, new_mean, new_variance, ) = self._renorm_correction_and_moments( new_mean, new_variance, training, input_batch_size ) # When training, the normalized values (say, x) will be # transformed as x * gamma + beta without renorm, and (x * r + # d) * gamma + beta = x * (r * gamma) + (d * gamma + beta) with # renorm. r = _broadcast(tf.stop_gradient(r, name="renorm_r")) d = _broadcast(tf.stop_gradient(d, name="renorm_d")) scale, offset = _compose_transforms(r, d, scale, offset) def _do_update(var, value): """Compute the updates for mean and variance.""" return self._assign_moving_average( var, value, self.momentum, input_batch_size ) def mean_update(): true_branch = lambda: _do_update(self.moving_mean, new_mean) false_branch = lambda: self.moving_mean return control_flow_util.smart_cond( training, true_branch, false_branch ) def variance_update(): """Update the moving variance.""" def true_branch_renorm(): # We apply epsilon as part of the moving_stddev to mirror # the training code path. moving_stddev = _do_update( self.moving_stddev, tf.sqrt(new_variance + self.epsilon) ) return self._assign_new_value( self.moving_variance, # Apply relu in case floating point rounding causes it # to go negative. backend.relu( moving_stddev * moving_stddev - self.epsilon ), ) if self.renorm: true_branch = true_branch_renorm else: true_branch = lambda: _do_update( self.moving_variance, new_variance ) false_branch = lambda: self.moving_variance return control_flow_util.smart_cond( training, true_branch, false_branch ) self.add_update(mean_update) self.add_update(variance_update) # End of handling mean/variance calculation and update. mean = tf.cast(mean, inputs.dtype) variance = tf.cast(variance, inputs.dtype) if offset is not None: offset = tf.cast(offset, inputs.dtype) if scale is not None: scale = tf.cast(scale, inputs.dtype) outputs = tf.nn.batch_normalization( inputs, _broadcast(mean), _broadcast(variance), offset, scale, self.epsilon, ) if inputs_dtype in (tf.float16, tf.bfloat16): outputs = tf.cast(outputs, inputs_dtype) # If some components of the shape got lost due to adjustments, fix that. outputs.set_shape(input_shape) if self.virtual_batch_size is not None: outputs = undo_virtual_batching(outputs) return outputs def compute_output_shape(self, input_shape): return input_shape def get_config(self): config = { "axis": self.axis, "momentum": self.momentum, "epsilon": self.epsilon, "center": self.center, "scale": self.scale, "beta_initializer": initializers.serialize(self.beta_initializer), "gamma_initializer": initializers.serialize(self.gamma_initializer), "moving_mean_initializer": initializers.serialize( self.moving_mean_initializer ), "moving_variance_initializer": initializers.serialize( self.moving_variance_initializer ), "beta_regularizer": regularizers.serialize(self.beta_regularizer), "gamma_regularizer": regularizers.serialize(self.gamma_regularizer), "beta_constraint": constraints.serialize(self.beta_constraint), "gamma_constraint": constraints.serialize(self.gamma_constraint), } # Only add TensorFlow-specific parameters if they are set, so as to # preserve model compatibility with external TF-Keras. if self.renorm: config["renorm"] = True config["renorm_clipping"] = self.renorm_clipping config["renorm_momentum"] = self.renorm_momentum if self.virtual_batch_size is not None: config["virtual_batch_size"] = self.virtual_batch_size # Note: adjustment is not serializable. if self.adjustment is not None: logging.warning( "The `adjustment` function of this `BatchNormalization` " "layer cannot be serialized and has been omitted from " "the layer config. It will not be included when " "re-creating the layer from the saved config." ) base_config = super().get_config() return dict(list(base_config.items()) + list(config.items())) ######################## Start of private methods ########################## def _support_zero_size_input(self): if not tf.distribute.has_strategy(): return False strategy = tf.distribute.get_strategy() # TODO(b/195085185): remove experimental_enable_get_next_as_optional # after migrating all users. return getattr( strategy.extended, "enable_partial_batch_handling", getattr( strategy.extended, "experimental_enable_get_next_as_optional", False, ), ) def _assign_moving_average(self, variable, value, momentum, inputs_size): def calculate_update_delta(): decay = tf.convert_to_tensor(1.0 - momentum, name="decay") if decay.dtype != variable.dtype.base_dtype: decay = tf.cast(decay, variable.dtype.base_dtype) update_delta = (variable - tf.cast(value, variable.dtype)) * decay if inputs_size is not None: update_delta = tf.where( inputs_size > 0, update_delta, backend.zeros_like(update_delta), ) return update_delta with backend.name_scope("AssignMovingAvg") as scope: if tf.compat.v1.executing_eagerly_outside_functions(): return variable.assign_sub(calculate_update_delta(), name=scope) else: with tf.compat.v1.colocate_with(variable): return tf.compat.v1.assign_sub( variable, calculate_update_delta(), name=scope ) def _assign_new_value(self, variable, value): with backend.name_scope("AssignNewValue") as scope: if tf.compat.v1.executing_eagerly_outside_functions(): return variable.assign(value, name=scope) else: with tf.compat.v1.colocate_with(variable): return tf.compat.v1.assign(variable, value, name=scope) def _fused_batch_norm(self, inputs, mask, training): """Returns the output of fused batch norm.""" if mask is not None: warnings.warn( "Masking is not supported with `fused=True`. " "You should either turn off fusing " "(`fused=False`) or you should not pass a `mask` " "argument when calling the layer. " "For the moment `mask` will be ignored for the " "normalization." ) if self.center: beta = self.beta else: beta = backend.constant( 0.0, dtype=self._param_dtype, shape=self._param_shape ) if self.scale: gamma = self.gamma else: gamma = backend.constant( 1.0, dtype=self._param_dtype, shape=self._param_shape ) # TODO(b/129279393): Support zero batch input in non # DistributionStrategy code as well. if self._support_zero_size_input(): # TF-Keras assumes that batch dimension is the first dimension for # Batch Normalization. input_batch_size = tf.shape(inputs)[0] else: input_batch_size = None # TODO(rmlarsen): Support using fused avg updates for non-eager # execution after fixing graph pattern matching and enabling # fused_batch_norm to take exponential_avg_factor as a tensor input. use_fused_avg_updates = ( tf.compat.v1.executing_eagerly_outside_functions() and isinstance(self.momentum, (float, int)) and get_enclosing_xla_context() is None ) if use_fused_avg_updates: exponential_avg_factor = 1.0 - self.momentum else: exponential_avg_factor = None def _maybe_add_or_remove_bessels_correction(variance, remove=True): r"""Add or remove Bessel's correction.""" # Removes Bessel's correction if remove == True, adds it otherwise. # This is to be consistent with non-fused batch norm. Note that the # variance computed by fused batch norm is with Bessel's correction. # This is only used in legacy V1 batch norm tests. if self._bessels_correction_test_only: return variance sample_size = tf.cast( tf.size(inputs) / tf.size(variance), variance.dtype ) if remove: factor = ( sample_size - tf.cast(1.0, variance.dtype) ) / sample_size else: factor = sample_size / ( sample_size - tf.cast(1.0, variance.dtype) ) return variance * factor def _fused_batch_norm_training(): return tf.compat.v1.nn.fused_batch_norm( inputs, gamma, beta, mean=self.moving_mean, variance=_maybe_add_or_remove_bessels_correction( self.moving_variance, remove=False ), epsilon=self.epsilon, is_training=True, data_format=self._data_format, exponential_avg_factor=exponential_avg_factor, ) def _fused_batch_norm_inference(): return tf.compat.v1.nn.fused_batch_norm( inputs, gamma, beta, mean=self.moving_mean, variance=self.moving_variance, epsilon=self.epsilon, is_training=False, data_format=self._data_format, ) output, mean, variance = control_flow_util.smart_cond( training, _fused_batch_norm_training, _fused_batch_norm_inference ) variance = _maybe_add_or_remove_bessels_correction( variance, remove=True ) training_value = control_flow_util.constant_value(training) if training_value or training_value is None: if not use_fused_avg_updates: if training_value is None: momentum = control_flow_util.smart_cond( training, lambda: self.momentum, lambda: 1.0 ) else: momentum = tf.convert_to_tensor(self.momentum) def mean_update(): """Update self.moving_mean with the most recent data point.""" if use_fused_avg_updates: if input_batch_size is not None: new_mean = control_flow_util.smart_cond( input_batch_size > 0, lambda: mean, lambda: self.moving_mean, ) else: new_mean = mean return self._assign_new_value(self.moving_mean, new_mean) else: return self._assign_moving_average( self.moving_mean, mean, momentum, input_batch_size ) def variance_update(): """Update self.moving_variance with the most recent data point.""" if use_fused_avg_updates: if input_batch_size is not None: new_variance = control_flow_util.smart_cond( input_batch_size > 0, lambda: variance, lambda: self.moving_variance, ) else: new_variance = variance return self._assign_new_value( self.moving_variance, new_variance ) else: return self._assign_moving_average( self.moving_variance, variance, momentum, input_batch_size, ) self.add_update(mean_update) self.add_update(variance_update) return output def _renorm_correction_and_moments( self, mean, variance, training, inputs_size ): """Returns the correction and update values for renorm.""" stddev = tf.sqrt(variance + self.epsilon) # Compute the average mean and standard deviation, as if they were # initialized with this batch's moments. renorm_mean = self.renorm_mean # Avoid divide by zero early on in training. renorm_stddev = tf.maximum(self.renorm_stddev, tf.sqrt(self.epsilon)) # Compute the corrections for batch renorm. r = stddev / renorm_stddev d = (mean - renorm_mean) / renorm_stddev # Ensure the corrections use pre-update moving averages. with tf.control_dependencies([r, d]): mean = tf.identity(mean) stddev = tf.identity(stddev) rmin, rmax, dmax = [ self.renorm_clipping.get(key) for key in ["rmin", "rmax", "dmax"] ] if rmin is not None: r = tf.maximum(r, rmin) if rmax is not None: r = tf.minimum(r, rmax) if dmax is not None: d = tf.maximum(d, -dmax) d = tf.minimum(d, dmax) # When not training, use r=1, d=0. r = control_flow_util.smart_cond( training, lambda: r, lambda: tf.ones_like(r) ) d = control_flow_util.smart_cond( training, lambda: d, lambda: tf.zeros_like(d) ) def _update_renorm_variable(var, value, inputs_size): """Updates a moving average and weight, returns the unbiased value.""" value = tf.identity(value) def _do_update(): """Updates the var, returns the updated value.""" new_var = self._assign_moving_average( var, value, self.renorm_momentum, inputs_size ) return new_var def _fake_update(): return tf.identity(var) return control_flow_util.smart_cond( training, _do_update, _fake_update ) # TODO(yuefengz): colocate the operations update_new_mean = _update_renorm_variable( self.renorm_mean, mean, inputs_size ) update_new_stddev = _update_renorm_variable( self.renorm_stddev, stddev, inputs_size ) # Update the inference mode moving averages with the batch value. with tf.control_dependencies([update_new_mean, update_new_stddev]): out_mean = tf.identity(mean) out_variance = tf.identity(variance) return (r, d, out_mean, out_variance) def _calculate_mean_and_var( self, inputs, reduction_axes, keep_dims, mask=None ): if self.synchronized: return self._sync_calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask=mask ) return self._no_sync_calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask=mask ) def _no_sync_calculate_mean_and_var( self, inputs, reduction_axes, keep_dims, mask=None ): if mask is None: return tf.nn.moments(inputs, reduction_axes, keepdims=keep_dims) else: mask_weights = tf.cast( mask, self.compute_dtype, name="mask_weights" ) mask_weights = tf.expand_dims( mask_weights, axis=-1, name="mask_weights_broadcasted" ) return tf.nn.weighted_moments( inputs, axes=reduction_axes, frequency_weights=mask_weights, keepdims=keep_dims, ) def _sync_calculate_mean_and_var( self, x, reduction_axes, keep_dims, mask=None ): with backend.name_scope("moments"): # The dynamic range of fp16 is too limited to support the collection # of sufficient statistics. As a workaround we simply perform the # operations on 32-bit floats before converting the mean and # variance back to fp16 y = tf.cast(x, tf.float32) if x.dtype == tf.float16 else x replica_ctx = tf.distribute.get_replica_context() if not replica_ctx: return self._no_sync_calculate_mean_and_var( x, reduction_axes, keep_dims, mask=mask ) if mask is not None: mask_weights = tf.cast(mask, y.dtype, name="mask_weights") mask_weights = tf.expand_dims( mask_weights, axis=-1, name="mask_weights_broadcasted" ) y *= mask_weights local_count = tf.broadcast_to( mask_weights, tf.shape(y), name="count" ) else: local_count = tf.ones_like(y, name="count") local_sum = tf.reduce_sum(y, axis=reduction_axes, keepdims=True) local_squared_sum = tf.reduce_sum( tf.square(y), axis=reduction_axes, keepdims=True ) local_count = tf.reduce_sum( local_count, axis=reduction_axes, keepdims=True ) # TODO(b/163099951): batch the all-reduces once we sort out the # ordering issue for NCCL. We don't have a mechanism to launch # NCCL in the same order in each replica nowadays, so we limit # NCCL to batch all-reduces. y_sum = replica_ctx.all_reduce( tf.distribute.ReduceOp.SUM, local_sum ) y_squared_sum = replica_ctx.all_reduce( tf.distribute.ReduceOp.SUM, local_squared_sum ) count_sum = replica_ctx.all_reduce( tf.distribute.ReduceOp.SUM, local_count ) mean = tf.math.divide_no_nan(y_sum, count_sum) y_squared_mean = tf.math.divide_no_nan(y_squared_sum, count_sum) # var = E(x^2) - E(x)^2 # The substraction operation does not guarantee a non-negative # result given float precision operations. variance = tf.maximum(y_squared_mean - tf.square(mean), 0.0) if not keep_dims: mean = tf.squeeze(mean, reduction_axes) variance = tf.squeeze(variance, reduction_axes) if x.dtype == tf.float16: return ( tf.cast(mean, tf.float16), tf.cast(variance, tf.float16), ) else: return (mean, variance) def _dtensor_calculate_mean_and_var( self, inputs, reduction_axes, keep_dims, mask=None ): if self.synchronized: return self._dtensor_sync_calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask=mask ) return self._dtensor_no_sync_calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask=mask ) def _dtensor_no_sync_calculate_mean_and_var( self, inputs, reduction_axes, keep_dims, mask=None ): replica_tensor = _expand_tensor_with_local_replica_group(inputs) local_batch_size = tf.shape(replica_tensor)[1] # Since we added a new axis in the beginning, all the value in # reduction_axes need to be incremented by 1. updated_reduction_axes = [n + 1 for n in reduction_axes] if mask is None: mean, var = tf.nn.moments( replica_tensor, updated_reduction_axes, keepdims=keep_dims ) else: mask_weights = tf.cast( mask, self.compute_dtype, name="mask_weights" ) mask_weights = tf.expand_dims( mask_weights, axis=-1, name="mask_weights_broadcasted" ) mask_weights = _expand_tensor_with_local_replica_group(mask_weights) mean, var = tf.nn.weighted_moments( replica_tensor, axes=updated_reduction_axes, frequency_weights=mask_weights, keepdims=keep_dims, ) # Also note that the mean/var we have here will have an extra dim in # axis 0, which is represented for num local replica. Down the # stream, the mean/var will be used to update the moving_mean/var # and also normalize the inputs. To make the shape match, we will # expand the tensor shape from [num_replica, x, y] to # [batch_size, x, y] so that it can be properly used for # normalization. When it reaches the mean/var update, a separate # logic will be there to reduce_mean the value based on the batch # dim. mean = tf.repeat(mean, local_batch_size, axis=0) var = tf.repeat(var, local_batch_size, axis=0) if not keep_dims: # We need to fill the reduced dims so that the mean/var can be # properly broadcast to the input shapes. In the example above, # the original reduction_axes is [0, 1]. We ignore the first 0 # (batch dim) here since we already expand and use it as num_replica for dim in reduction_axes[1:]: mean = tf.expand_dims(mean, axis=dim) var = tf.expand_dims(var, axis=dim) return mean, var def _dtensor_sync_calculate_mean_and_var( self, inputs, reduction_axes, keep_dims, mask=None ): # In the DTensor sync BN, since the input tensor is already in global # context, we just need to use the normal moments/weighted_moments # to calculate mean/var, which is same as the non-sync BN in the normal # mode. return self._no_sync_calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask ) def _moments(self, inputs, reduction_axes, keep_dims, mask=None): if utils.running_with_dtensor_strategy(): mean, variance = self._dtensor_calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask=mask ) else: mean, variance = self._calculate_mean_and_var( inputs, reduction_axes, keep_dims, mask=mask ) # TODO(b/129279393): Support zero batch input in non # DistributionStrategy code as well. if self._support_zero_size_input(): input_batch_size = tf.shape(inputs)[0] mean = tf.where( input_batch_size > 0, mean, backend.zeros_like(mean) ) variance = tf.where( input_batch_size > 0, variance, backend.zeros_like(variance) ) return mean, variance def _get_training_value(self, training=None): if training is None: training = backend.learning_phase() if self._USE_V2_BEHAVIOR: if isinstance(training, int): training = bool(training) if not self.trainable: # When the layer is not trainable, it overrides the value passed # from model. training = False return training @keras_export("keras.layers.BatchNormalization", v1=[]) class BatchNormalization(BatchNormalizationBase): """Layer that normalizes its inputs. Batch normalization applies a transformation that maintains the mean output close to 0 and the output standard deviation close to 1. Importantly, batch normalization works differently during training and during inference. **During training** (i.e. when using `fit()` or when calling the layer/model with the argument `training=True`), the layer normalizes its output using the mean and standard deviation of the current batch of inputs. That is to say, for each channel being normalized, the layer returns `gamma * (batch - mean(batch)) / sqrt(var(batch) + epsilon) + beta`, where: - `epsilon` is small constant (configurable as part of the constructor arguments) - `gamma` is a learned scaling factor (initialized as 1), which can be disabled by passing `scale=False` to the constructor. - `beta` is a learned offset factor (initialized as 0), which can be disabled by passing `center=False` to the constructor. **During inference** (i.e. when using `evaluate()` or `predict()` or when calling the layer/model with the argument `training=False` (which is the default), the layer normalizes its output using a moving average of the mean and standard deviation of the batches it has seen during training. That is to say, it returns `gamma * (batch - self.moving_mean) / sqrt(self.moving_var+epsilon) + beta`. `self.moving_mean` and `self.moving_var` are non-trainable variables that are updated each time the layer in called in training mode, as such: - `moving_mean = moving_mean * momentum + mean(batch) * (1 - momentum)` - `moving_var = moving_var * momentum + var(batch) * (1 - momentum)` As such, the layer will only normalize its inputs during inference *after having been trained on data that has similar statistics as the inference data*. When `synchronized=True` is set and if this layer is used within a `tf.distribute` strategy, there will be an `allreduce` call to aggregate batch statistics across all replicas at every training step. Setting `synchronized` has no impact when the model is trained without specifying any distribution strategy. Example usage: ```python strategy = tf.distribute.MirroredStrategy() with strategy.scope(): model = tf.keras.Sequential() model.add(tf.keras.layers.Dense(16)) model.add(tf.keras.layers.BatchNormalization(synchronized=True)) ``` Args: axis: Integer, the axis that should be normalized (typically the features axis). For instance, after a `Conv2D` layer with `data_format="channels_first"`, set `axis=1` in `BatchNormalization`. momentum: Momentum for the moving average. epsilon: Small float added to variance to avoid dividing by zero. center: If True, add offset of `beta` to normalized tensor. If False, `beta` is ignored. scale: If True, multiply by `gamma`. If False, `gamma` is not used. When the next layer is linear (also e.g. `nn.relu`), this can be disabled since the scaling will be done by the next layer. beta_initializer: Initializer for the beta weight. gamma_initializer: Initializer for the gamma weight. moving_mean_initializer: Initializer for the moving mean. moving_variance_initializer: Initializer for the moving variance. beta_regularizer: Optional regularizer for the beta weight. gamma_regularizer: Optional regularizer for the gamma weight. beta_constraint: Optional constraint for the beta weight. gamma_constraint: Optional constraint for the gamma weight. synchronized: If True, synchronizes the global batch statistics (mean and variance) for the layer across all devices at each training step in a distributed training strategy. If False, each replica uses its own local batch statistics. Only relevant when used inside a `tf.distribute` strategy. Call arguments: inputs: Input tensor (of any rank). training: Python boolean indicating whether the layer should behave in training mode or in inference mode. - `training=True`: The layer will normalize its inputs using the mean and variance of the current batch of inputs. - `training=False`: The layer will normalize its inputs using the mean and variance of its moving statistics, learned during training. Input shape: Arbitrary. Use the keyword argument `input_shape` (tuple of integers, does not include the samples axis) when using this layer as the first layer in a model. Output shape: Same shape as input. Reference: - [Ioffe and Szegedy, 2015](https://arxiv.org/abs/1502.03167). **About setting `layer.trainable = False` on a `BatchNormalization` layer:** The meaning of setting `layer.trainable = False` is to freeze the layer, i.e. its internal state will not change during training: its trainable weights will not be updated during `fit()` or `train_on_batch()`, and its state updates will not be run. Usually, this does not necessarily mean that the layer is run in inference mode (which is normally controlled by the `training` argument that can be passed when calling a layer). "Frozen state" and "inference mode" are two separate concepts. However, in the case of the `BatchNormalization` layer, **setting `trainable = False` on the layer means that the layer will be subsequently run in inference mode** (meaning that it will use the moving mean and the moving variance to normalize the current batch, rather than using the mean and variance of the current batch). This behavior has been introduced in TensorFlow 2.0, in order to enable `layer.trainable = False` to produce the most commonly expected behavior in the convnet fine-tuning use case. Note that: - Setting `trainable` on an model containing other layers will recursively set the `trainable` value of all inner layers. - If the value of the `trainable` attribute is changed after calling `compile()` on a model, the new value doesn't take effect for this model until `compile()` is called again. """ _USE_V2_BEHAVIOR = True @utils.allow_initializer_layout def __init__( self, axis=-1, momentum=0.99, epsilon=1e-3, center=True, scale=True, beta_initializer="zeros", gamma_initializer="ones", moving_mean_initializer="zeros", moving_variance_initializer="ones", beta_regularizer=None, gamma_regularizer=None, beta_constraint=None, gamma_constraint=None, synchronized=False, **kwargs, ): # Currently we only support aggregating over the global batch size. super().__init__( axis=axis, momentum=momentum, epsilon=epsilon, center=center, scale=scale, beta_initializer=beta_initializer, gamma_initializer=gamma_initializer, moving_mean_initializer=moving_mean_initializer, moving_variance_initializer=moving_variance_initializer, beta_regularizer=beta_regularizer, gamma_regularizer=gamma_regularizer, beta_constraint=beta_constraint, gamma_constraint=gamma_constraint, synchronized=synchronized, **kwargs, ) @keras_export("keras.layers.experimental.SyncBatchNormalization", v1=[]) @deprecation.deprecated_endpoints( "keras.layers.experimental.SyncBatchNormalization" ) class SyncBatchNormalization(BatchNormalizationBase): """Deprecated. Please use `tf.keras.layers.BatchNormalization` instead. Caution: `tf.keras.layers.experimental.SyncBatchNormalization` endpoint is deprecated and will be removed in a future release. Please use `tf.keras.layers.BatchNormalization` with parameter `synchronized` set to True """ def __init__( self, axis=-1, momentum=0.99, epsilon=1e-3, center=True, scale=True, beta_initializer="zeros", gamma_initializer="ones", moving_mean_initializer="zeros", moving_variance_initializer="ones", beta_regularizer=None, gamma_regularizer=None, beta_constraint=None, gamma_constraint=None, **kwargs, ): warning = ( "`tf.keras.layers.experimental.SyncBatchNormalization` endpoint is " "deprecated and will be removed in a future release. Please use " "`tf.keras.layers.BatchNormalization` with parameter " "`synchronized` set to True." ) logging.log_first_n(logging.WARN, warning, 1) super().__init__( axis=axis, momentum=momentum, epsilon=epsilon, center=center, scale=scale, beta_initializer=beta_initializer, gamma_initializer=gamma_initializer, moving_mean_initializer=moving_mean_initializer, moving_variance_initializer=moving_variance_initializer, beta_regularizer=beta_regularizer, gamma_regularizer=gamma_regularizer, beta_constraint=beta_constraint, gamma_constraint=gamma_constraint, synchronized=True, **kwargs, ) def _expand_tensor_with_local_replica_group(inputs): """Reshape the input tensor to have an extra dimension of replica group. Under the DTensor usage, the normal batch norm still need to perform on a local batch size, which mean we can't directly do mean/var on a global tensor. In order to do a local mean/var, we have to add a new dimention to the tensor, so that the ops will not cross the replica boundary. E.g, a global tensor with shape [8, x, y] and has 2 local replica, the output of this will be [2, 4, x, y], where the first dim is for num of replica, and the second dim is for the local batch size. The follow ops can do reduces among the local batch dimension. Note that this function should only be used under DTensor based strategy, and it will use the current strategy in the context to get the number of replica. Args: inputs: Tensor with shape [global_batch_size, ...] Returns: Tensor with shape [num_replica, local_batch_size, ...] """ # TODO(b/272382109): Implement this an an Op. input_shape = tf.shape(inputs) global_batch_size = input_shape[0] num_replica = tf.distribute.get_strategy().num_replicas_in_sync local_batch_size = global_batch_size // num_replica replica_shape = tf.stack([num_replica, local_batch_size]) replica_shape = tf.concat([replica_shape, input_shape[1:]], axis=0) return tf.reshape(inputs, replica_shape) def _raise_for_non_sync_bn_with_renorm_and_dtensor_strategy( synchronized, training, renorm ): if ( utils.running_with_dtensor_strategy() and not synchronized and training == True and renorm ): raise NotImplementedError( "Renorm for BatchNormalization under DTensor based distribution " "strategy is not supported at the moment. Please file a feature " "request if this is blocking your adoption." )