# Copyright 2022 The TF-Keras 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. # ============================================================================== """Group normalization layer""" 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.layers import InputSpec from tf_keras.src.layers import Layer from tf_keras.src.utils import tf_utils # isort: off from tensorflow.python.util.tf_export import keras_export @keras_export("keras.layers.GroupNormalization", v1=[]) class GroupNormalization(Layer): """Group normalization layer. Group Normalization divides the channels into groups and computes within each group the mean and variance for normalization. Empirically, its accuracy is more stable than batch norm in a wide range of small batch sizes, if learning rate is adjusted linearly with batch sizes. Relation to Layer Normalization: If the number of groups is set to 1, then this operation becomes nearly identical to Layer Normalization (see Layer Normalization docs for details). Relation to Instance Normalization: If the number of groups is set to the input dimension (number of groups is equal to number of channels), then this operation becomes identical to Instance Normalization. Args: groups: Integer, the number of groups for Group Normalization. Can be in the range [1, N] where N is the input dimension. The input dimension must be divisible by the number of groups. Defaults to `32`. axis: Integer or List/Tuple. The axis or axes to normalize across. Typically, this is the features axis/axes. The left-out axes are typically the batch axis/axes. `-1` is the last dimension in the input. Defaults to `-1`. epsilon: Small float added to variance to avoid dividing by zero. Defaults to 1e-3 center: If True, add offset of `beta` to normalized tensor. If False, `beta` is ignored. Defaults to `True`. 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. Defaults to `True`. beta_initializer: Initializer for the beta weight. Defaults to zeros. gamma_initializer: Initializer for the gamma weight. Defaults to ones. beta_regularizer: Optional regularizer for the beta weight. None by default. gamma_regularizer: Optional regularizer for the gamma weight. None by default. beta_constraint: Optional constraint for the beta weight. None by default. gamma_constraint: Optional constraint for the gamma weight. None by default. 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. Call arguments: inputs: Input tensor (of any rank). mask: The mask parameter is a tensor that indicates the weight for each position in the input tensor when computing the mean and variance. Reference: - [Yuxin Wu & Kaiming He, 2018](https://arxiv.org/abs/1803.08494) """ def __init__( self, groups=32, axis=-1, epsilon=1e-3, center=True, scale=True, beta_initializer="zeros", gamma_initializer="ones", beta_regularizer=None, gamma_regularizer=None, beta_constraint=None, gamma_constraint=None, **kwargs, ): super().__init__(**kwargs) self.supports_masking = True self.groups = groups self.axis = axis self.epsilon = epsilon self.center = center self.scale = scale self.beta_initializer = initializers.get(beta_initializer) self.gamma_initializer = initializers.get(gamma_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) def build(self, input_shape): tf_utils.validate_axis(self.axis, input_shape) dim = input_shape[self.axis] if dim is None: raise ValueError( f"Axis {self.axis} of input tensor should have a defined " "dimension but the layer received an input with shape " f"{input_shape}." ) if self.groups == -1: self.groups = dim if dim < self.groups: raise ValueError( f"Number of groups ({self.groups}) cannot be more than the " f"number of channels ({dim})." ) if dim % self.groups != 0: raise ValueError( f"Number of groups ({self.groups}) must be a multiple " f"of the number of channels ({dim})." ) self.input_spec = InputSpec( ndim=len(input_shape), axes={self.axis: dim} ) if self.scale: self.gamma = self.add_weight( shape=(dim,), name="gamma", initializer=self.gamma_initializer, regularizer=self.gamma_regularizer, constraint=self.gamma_constraint, ) else: self.gamma = None if self.center: self.beta = self.add_weight( shape=(dim,), name="beta", initializer=self.beta_initializer, regularizer=self.beta_regularizer, constraint=self.beta_constraint, ) else: self.beta = None super().build(input_shape) def call(self, inputs, mask=None): input_shape = tf.shape(inputs) if mask is None: mask = tf.ones_like(inputs) else: # We broadcast before we group in case the mask does not have the # same shape as the input. mask = tf.broadcast_to(mask, input_shape) reshaped_inputs = self._reshape_into_groups(inputs) reshaped_mask = self._reshape_into_groups(mask) normalized_inputs = self._apply_normalization( reshaped_inputs=reshaped_inputs, input_shape=input_shape, reshaped_mask=reshaped_mask, ) return tf.reshape(normalized_inputs, input_shape) def _reshape_into_groups(self, inputs): input_shape = tf.shape(inputs) group_shape = [input_shape[i] for i in range(inputs.shape.rank)] group_shape[self.axis] = input_shape[self.axis] // self.groups group_shape.insert(self.axis, self.groups) group_shape = tf.stack(group_shape) reshaped_inputs = tf.reshape(inputs, group_shape) return reshaped_inputs def _apply_normalization( self, *, reshaped_inputs, reshaped_mask, input_shape, ): group_reduction_axes = list(range(1, reshaped_inputs.shape.rank)) axis = self.axis - 1 group_reduction_axes.pop(axis) mask_weights = tf.cast(reshaped_mask, reshaped_inputs.dtype) mean, variance = tf.nn.weighted_moments( reshaped_inputs, axes=group_reduction_axes, frequency_weights=mask_weights, keepdims=True, ) gamma, beta = self._get_reshaped_weights(input_shape) normalized_inputs = tf.nn.batch_normalization( reshaped_inputs, mean=mean, variance=variance, scale=gamma, offset=beta, variance_epsilon=self.epsilon, ) return normalized_inputs def _get_reshaped_weights(self, input_shape): broadcast_shape = self._create_broadcast_shape(input_shape) gamma = None beta = None if self.scale: gamma = tf.reshape(self.gamma, broadcast_shape) if self.center: beta = tf.reshape(self.beta, broadcast_shape) return gamma, beta def _create_broadcast_shape(self, input_shape): broadcast_shape = [1] * backend.int_shape(input_shape)[0] broadcast_shape[self.axis] = input_shape[self.axis] // self.groups broadcast_shape.insert(self.axis, self.groups) return broadcast_shape def compute_output_shape(self, input_shape): return input_shape def get_config(self): config = { "groups": self.groups, "axis": self.axis, "epsilon": self.epsilon, "center": self.center, "scale": self.scale, "beta_initializer": initializers.serialize(self.beta_initializer), "gamma_initializer": initializers.serialize(self.gamma_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), } base_config = super().get_config() return {**base_config, **config}