def _resample_feature_map(self,
inputs,
input_level,
target_level,
target_num_filters=256):
x = inputs
_, _, _, input_num_filters = x.get_shape().as_list()
if input_num_filters != target_num_filters:
x = self._conv_op(
filters=target_num_filters,
kernel_size=1,
padding='same',
**self._conv_kwargs)(x)
x = self._norm_op(**self._norm_kwargs)(x)
if input_level < target_level:
stride = int(2 ** (target_level - input_level))
return tf.keras.layers.MaxPool2D(
pool_size=stride, strides=stride, padding='same')(x)
if input_level > target_level:
scale = int(2 ** (input_level - target_level))
return spatial_transform_ops.nearest_upsampling(x, scale=scale)
# Force output x to be the same dtype as mixed precision policy. This avoids
# dtype mismatch when one input (by default float32 dtype) does not meet all
# the above conditions and is output unchanged, while other inputs are
# processed to have different dtype, e.g., using bfloat16 on TPU.
compute_dtype = tf.keras.layers.Layer().dtype_policy.compute_dtype
if (compute_dtype is not None) and (x.dtype != compute_dtype):
return tf.cast(x, dtype=compute_dtype)
else:
return x
def call(self, inputs: Tuple[Union[tf.Tensor, Mapping[str, tf.Tensor]],
Union[tf.Tensor, Mapping[str, tf.Tensor]]]):
"""Forward pass of the segmentation head.
It supports both a tuple of 2 tensors or 2 dictionaries. The first is
backbone endpoints, and the second is decoder endpoints. When inputs are
tensors, they are from a single level of feature maps. When inputs are
dictionaries, they contain multiple levels of feature maps, where the key
is the index of feature map.
Args:
inputs: A tuple of 2 feature map tensors of shape
[batch, height_l, width_l, channels] or 2 dictionaries of tensors:
- key: A `str` of the level of the multilevel features.
- values: A `tf.Tensor` of the feature map tensors, whose shape is
[batch, height_l, width_l, channels].
The first is backbone endpoints, and the second is decoder endpoints.
Returns:
segmentation prediction mask: A `tf.Tensor` of the segmentation mask
scores predicted from input features.
"""
backbone_output = inputs[0]
decoder_output = inputs[1]
if self._config_dict['feature_fusion'] == 'deeplabv3plus':
# deeplabv3+ feature fusion
x = decoder_output[str(self._config_dict['level'])] if isinstance(
decoder_output, dict) else decoder_output
y = backbone_output[str(
self._config_dict['low_level'])] if isinstance(
backbone_output, dict) else backbone_output
y = self._dlv3p_norm(self._dlv3p_conv(y))
y = self._activation(y)
x = tf.image.resize(x,
tf.shape(y)[1:3],
method=tf.image.ResizeMethod.BILINEAR)
x = tf.cast(x, dtype=y.dtype)
x = tf.concat([x, y], axis=self._bn_axis)
elif self._config_dict['feature_fusion'] == 'pyramid_fusion':
if not isinstance(decoder_output, dict):
raise ValueError('Only support dictionary decoder_output.')
x = nn_layers.pyramid_feature_fusion(decoder_output,
self._config_dict['level'])
elif self._config_dict['feature_fusion'] == 'panoptic_fpn_fusion':
x = self._panoptic_fpn_fusion(decoder_output)
else:
x = decoder_output[str(self._config_dict['level'])] if isinstance(
decoder_output, dict) else decoder_output
for conv, norm in zip(self._convs, self._norms):
x = conv(x)
x = norm(x)
x = self._activation(x)
if self._config_dict['upsample_factor'] > 1:
x = spatial_transform_ops.nearest_upsampling(
x, scale=self._config_dict['upsample_factor'])
return self._classifier(x)
def _resample_with_sepconv(self, inputs, input_width, target_width,
target_num_filters):
"""Matches resolution and feature dimension."""
x = inputs
# Spatial resampling.
if input_width > target_width:
while input_width > target_width:
x = layers.DepthwiseConv2D(
kernel_size=3,
strides=2,
padding='SAME',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
x)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
x = tf_utils.get_activation(
self._activation, use_keras_layer=True)(x)
input_width /= 2
elif input_width < target_width:
scale = target_width // input_width
x = spatial_transform_ops.nearest_upsampling(
x, scale=scale, use_keras_layer=self._use_keras_upsampling_2d)
# Last 1x1 conv to match filter size.
x = layers.Conv2D(
filters=target_num_filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
x)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
return x
def call(self, inputs: Tuple[Union[tf.Tensor, Mapping[str, tf.Tensor]],
Union[tf.Tensor, Mapping[str, tf.Tensor]]],
training=None):
"""Forward pass of the head.
It supports both a tuple of 2 tensors or 2 dictionaries. The first is
backbone endpoints, and the second is decoder endpoints. When inputs are
tensors, they are from a single level of feature maps. When inputs are
dictionaries, they contain multiple levels of feature maps, where the key
is the index of feature map.
Args:
inputs: A tuple of 2 feature map tensors of shape
[batch, height_l, width_l, channels] or 2 dictionaries of tensors:
- key: A `str` of the level of the multilevel features.
- values: A `tf.Tensor` of the feature map tensors, whose shape is
[batch, height_l, width_l, channels].
training: A bool, runs the model in training/eval mode.
Returns:
A `tf.Tensor` of the fused backbone and decoder features.
"""
if training is None:
training = tf.keras.backend.learning_phase()
x = self._panoptic_deeplab_fusion(inputs, training=training)
for conv, norm in zip(self._convs, self._norms):
x = conv(x)
x = norm(x, training=training)
x = self._activation(x)
if self._config_dict['upsample_factor'] > 1:
x = spatial_transform_ops.nearest_upsampling(
x, scale=self._config_dict['upsample_factor'])
return x
def __init__(self,
input_specs: Mapping[str, tf.TensorShape],
min_level: int = 3,
max_level: int = 7,
num_filters: int = 256,
fusion_type: str = 'sum',
use_separable_conv: bool = False,
use_keras_layer: bool = False,
activation: str = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
kernel_initializer: str = 'VarianceScaling',
kernel_regularizer: Optional[
tf.keras.regularizers.Regularizer] = None,
bias_regularizer: Optional[
tf.keras.regularizers.Regularizer] = None,
**kwargs):
"""Initializes a Feature Pyramid Network (FPN).
Args:
input_specs: A `dict` of input specifications. A dictionary consists of
{level: TensorShape} from a backbone.
min_level: An `int` of minimum level in FPN output feature maps.
max_level: An `int` of maximum level in FPN output feature maps.
num_filters: An `int` number of filters in FPN layers.
fusion_type: A `str` of `sum` or `concat`. Whether performing sum or
concat for feature fusion.
use_separable_conv: A `bool`. If True use separable convolution for
convolution in FPN layers.
use_keras_layer: A `bool`. If Ture use keras layers as many as possible.
activation: A `str` name of the activation function.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
kernel_initializer: A `str` name of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
**kwargs: Additional keyword arguments to be passed.
"""
self._config_dict = {
'input_specs': input_specs,
'min_level': min_level,
'max_level': max_level,
'num_filters': num_filters,
'fusion_type': fusion_type,
'use_separable_conv': use_separable_conv,
'use_keras_layer': use_keras_layer,
'activation': activation,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
'kernel_initializer': kernel_initializer,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
}
if use_separable_conv:
conv2d = tf.keras.layers.SeparableConv2D
else:
conv2d = tf.keras.layers.Conv2D
if use_sync_bn:
norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
norm = tf.keras.layers.BatchNormalization
activation_fn = tf_utils.get_activation(activation,
use_keras_layer=True)
# Build input feature pyramid.
if tf.keras.backend.image_data_format() == 'channels_last':
bn_axis = -1
else:
bn_axis = 1
# Get input feature pyramid from backbone.
logging.info('FPN input_specs: %s', input_specs)
inputs = self._build_input_pyramid(input_specs, min_level)
backbone_max_level = min(int(max(inputs.keys())), max_level)
# Build lateral connections.
feats_lateral = {}
for level in range(min_level, backbone_max_level + 1):
feats_lateral[str(level)] = conv2d(
filters=num_filters,
kernel_size=1,
padding='same',
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer)(inputs[str(level)])
# Build top-down path.
feats = {
str(backbone_max_level): feats_lateral[str(backbone_max_level)]
}
for level in range(backbone_max_level - 1, min_level - 1, -1):
feat_a = spatial_transform_ops.nearest_upsampling(
feats[str(level + 1)], 2, use_keras_layer=use_keras_layer)
feat_b = feats_lateral[str(level)]
if fusion_type == 'sum':
if use_keras_layer:
feats[str(level)] = tf.keras.layers.Add()([feat_a, feat_b])
else:
feats[str(level)] = feat_a + feat_b
elif fusion_type == 'concat':
if use_keras_layer:
feats[str(level)] = tf.keras.layers.Concatenate(axis=-1)(
[feat_a, feat_b])
else:
feats[str(level)] = tf.concat([feat_a, feat_b], axis=-1)
else:
raise ValueError(
'Fusion type {} not supported.'.format(fusion_type))
# TODO(xianzhi): consider to remove bias in conv2d.
# Build post-hoc 3x3 convolution kernel.
for level in range(min_level, backbone_max_level + 1):
feats[str(level)] = conv2d(filters=num_filters,
strides=1,
kernel_size=3,
padding='same',
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer)(
feats[str(level)])
# TODO(xianzhi): consider to remove bias in conv2d.
# Build coarser FPN levels introduced for RetinaNet.
for level in range(backbone_max_level + 1, max_level + 1):
feats_in = feats[str(level - 1)]
if level > backbone_max_level + 1:
feats_in = activation_fn(feats_in)
feats[str(level)] = conv2d(
filters=num_filters,
strides=2,
kernel_size=3,
padding='same',
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer)(feats_in)
# Apply batch norm layers.
for level in range(min_level, max_level + 1):
feats[str(level)] = norm(axis=bn_axis,
momentum=norm_momentum,
epsilon=norm_epsilon)(feats[str(level)])
self._output_specs = {
str(level): feats[str(level)].get_shape()
for level in range(min_level, max_level + 1)
}
super(FPN, self).__init__(inputs=inputs, outputs=feats, **kwargs)
def _resample_with_alpha(self,
inputs,
input_width,
input_block_fn,
target_width,
target_num_filters,
target_block_fn,
alpha=0.5):
"""Matches resolution and feature dimension."""
_, _, _, input_num_filters = inputs.get_shape().as_list()
if input_block_fn == 'bottleneck':
input_num_filters /= 4
new_num_filters = int(input_num_filters * alpha)
x = layers.Conv2D(
filters=new_num_filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
inputs)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
x = tf_utils.get_activation(self._activation_fn)(x)
# Spatial resampling.
if input_width > target_width:
x = layers.Conv2D(
filters=new_num_filters,
kernel_size=3,
strides=2,
padding='SAME',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
x)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
x = tf_utils.get_activation(self._activation_fn)(x)
input_width /= 2
while input_width > target_width:
x = layers.MaxPool2D(pool_size=3, strides=2, padding='SAME')(x)
input_width /= 2
elif input_width < target_width:
scale = target_width // input_width
x = spatial_transform_ops.nearest_upsampling(x, scale=scale)
# Last 1x1 conv to match filter size.
if target_block_fn == 'bottleneck':
target_num_filters *= 4
x = layers.Conv2D(
filters=target_num_filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
x)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
return x
