def __init__(
self,
inputs,
blocks,
block,
include_top=True,
classes=1000,
freeze_bn=True,
numerical_names=None,
*args,
**kwargs
):
if tf.keras.backend.image_data_format() == "channels_last":
axis = 3
else:
axis = 1
if numerical_names is None:
numerical_names = [True] * len(blocks)
x = tf.keras.layers.ZeroPadding1D(padding=3, name="padding_conv1")(inputs)
x = tf.keras.layers.Conv1D(64, (7, 7), strides=(2, 2), use_bias=False, name="conv1")(x)
x = layers.BatchNormalization(axis=axis, epsilon=1e-5, freeze=freeze_bn, name="bn_conv1")(x)
x = tf.keras.layers.Activation("relu", name="conv1_relu")(x)
x = tf.keras.layers.MaxPooling1D((3, 3), strides=(2, 2), padding="same", name="pool1")(x)
features = 64
outputs = []
for stage_id, iterations in enumerate(blocks):
for block_id in range(iterations):
x = block(
features,
stage_id,
block_id,
numerical_name=(block_id > 0 and numerical_names[stage_id]),
freeze_bn=freeze_bn
)(x)
features *= 2
outputs.append(x)
if include_top:
assert classes > 0
x = tf.keras.layers.GlobalAveragePooling1D(name="pool5")(x)
x = tf.keras.layers.Dense(classes, activation="softmax", name="fc1000")(x)
super(ResNet1D, self).__init__(inputs=inputs, outputs=x, *args, **kwargs)
else:
# Else output each stages features
super(ResNet1D, self).__init__(inputs=inputs, outputs=outputs, *args, **kwargs)
def f(x):
y = tf.keras.layers.TimeDistributed(
tf.keras.layers.ZeroPadding2D(padding=1),
name="padding{}{}_branch2a".format(stage_char, block_char))(x)
y = tf.keras.layers.TimeDistributed(
tf.keras.layers.Conv2D(filters,
kernel_size,
strides=stride,
use_bias=False,
**parameters),
name="res{}{}_branch2a".format(stage_char, block_char))(y)
y = tf.keras.layers.TimeDistributed(
layers.BatchNormalization(axis=axis,
epsilon=1e-5,
freeze=freeze_bn),
name="bn{}{}_branch2a".format(stage_char, block_char))(y)
y = tf.keras.layers.TimeDistributed(
tf.keras.layers.Activation("relu"),
name="res{}{}_branch2a_relu".format(stage_char, block_char))(y)
y = tf.keras.layers.TimeDistributed(
tf.keras.layers.ZeroPadding2D(padding=1),
name="padding{}{}_branch2b".format(stage_char, block_char))(y)
y = tf.keras.layers.TimeDistributed(
tf.keras.layers.Conv2D(filters,
kernel_size,
use_bias=False,
**parameters),
name="res{}{}_branch2b".format(stage_char, block_char))(y)
y = tf.keras.layers.TimeDistributed(
layers.BatchNormalization(axis=axis,
epsilon=1e-5,
freeze=freeze_bn),
name="bn{}{}_branch2b".format(stage_char, block_char))(y)
if block == 0:
shortcut = tf.keras.layers.TimeDistributed(
tf.keras.layers.Conv2D(filters, (1, 1),
strides=stride,
use_bias=False,
**parameters),
name="res{}{}_branch1".format(stage_char, block_char))(x)
shortcut = tf.keras.layers.TimeDistributed(
layers.BatchNormalization(axis=axis,
epsilon=1e-5,
freeze=freeze_bn),
name="bn{}{}_branch1".format(stage_char, block_char))(shortcut)
else:
shortcut = x
y = tf.keras.layers.Add(name="res{}{}".format(stage_char, block_char))(
[y, shortcut])
y = tf.keras.layers.TimeDistributed(tf.keras.layers.Activation("relu"),
name="res{}{}_relu".format(
stage_char, block_char))(y)
return y
def __init__(self,
inputs,
blocks,
block,
freeze_bn=True,
numerical_names=None,
*args,
**kwargs):
if tf.keras.backend.image_data_format() == "channels_last":
axis = 3
else:
axis = 1
if numerical_names is None:
numerical_names = [True] * len(blocks)
x = tf.keras.layers.Conv2D(64, (7, 7),
strides=(2, 2),
use_bias=False,
name="conv1",
padding="same")(inputs)
x = layers.BatchNormalization(axis=axis,
epsilon=1e-5,
freeze=freeze_bn,
name="bn_conv1")(x)
x = tf.keras.layers.Activation("relu", name="conv1_relu")(x)
x = tf.keras.layers.MaxPooling2D((3, 3),
strides=(2, 2),
padding="same",
name="pool1")(x)
features = 64
outputs = []
for stage_id, iterations in enumerate(blocks):
for block_id in range(iterations):
x = block(features,
stage_id,
block_id,
numerical_name=(block_id > 0
and numerical_names[stage_id]),
freeze_bn=freeze_bn)(x)
features *= 2
outputs.append(x)
c2, c3, c4, c5 = outputs
pyramid_5 = tf.keras.layers.Conv2D(filters=256,
kernel_size=1,
strides=1,
padding="same",
name="c5_reduced")(c5)
upsampled_p5 = tf.keras.layers.UpSampling2D(interpolation="bilinear",
name="p5_upsampled",
size=(2, 2))(pyramid_5)
pyramid_4 = tf.keras.layers.Conv2D(filters=256,
kernel_size=1,
strides=1,
padding="same",
name="c4_reduced")(c4)
pyramid_4 = tf.keras.layers.Add(name="p4_merged")(
[upsampled_p5, pyramid_4])
upsampled_p4 = tf.keras.layers.UpSampling2D(interpolation="bilinear",
name="p4_upsampled",
size=(2, 2))(pyramid_4)
pyramid_4 = tf.keras.layers.Conv2D(filters=256,
kernel_size=3,
strides=1,
padding="same",
name="p4")(pyramid_4)
pyramid_3 = tf.keras.layers.Conv2D(filters=256,
kernel_size=1,
strides=1,
padding="same",
name="c3_reduced")(c3)
pyramid_3 = tf.keras.layers.Add(name="p3_merged")(
[upsampled_p4, pyramid_3])
upsampled_p3 = tf.keras.layers.UpSampling2D(interpolation="bilinear",
name="p3_upsampled",
size=(2, 2))(pyramid_3)
pyramid_3 = tf.keras.layers.Conv2D(filters=256,
kernel_size=3,
strides=1,
padding="same",
name="p3")(pyramid_3)
pyramid_2 = tf.keras.layers.Conv2D(filters=256,
kernel_size=1,
strides=1,
padding="same",
name="c2_reduced")(c2)
pyramid_2 = tf.keras.layers.Add(name="p2_merged")(
[upsampled_p3, pyramid_2])
pyramid_2 = tf.keras.layers.Conv2D(filters=256,
kernel_size=3,
strides=1,
padding="same",
name="p2")(pyramid_2)
pyramid_6 = tf.keras.layers.MaxPooling2D(strides=2,
name="p6")(pyramid_5)
outputs = [pyramid_2, pyramid_3, pyramid_4, pyramid_5, pyramid_6]
super(FPN2D, self).__init__(inputs=inputs,
outputs=outputs,
*args,
**kwargs)
def TimeDistributedResNet(inputs,
blocks,
block,
include_top=True,
classes=1000,
freeze_bn=True,
*args,
**kwargs):
"""
Constructs a time distributed `keras.models.Model` object using the given block count.
:param inputs: input tensor (e.g. an instance of `keras.layers.Input`)
:param blocks: the network’s residual architecture
:param block: a time distributed residual block (e.g. an instance of `keras_resnet.blocks.time_distributed_basic_2d`)
:param include_top: if true, includes classification layers
:param classes: number of classes to classify (include_top must be true)
:param freeze_bn: if true, freezes BatchNormalization layers (ie. no updates are done in these layers)
:return model: Time distributed ResNet model with encoding output (if `include_top=False`) or classification output (if `include_top=True`)
Usage:
>>> import keras_resnet.blocks
>>> import keras_resnet.models
>>> shape, classes = (224, 224, 3), 1000
>>> x = keras.layers.Input(shape)
>>> blocks = [2, 2, 2, 2]
>>> blocks = keras_resnet.blocks.time_distributed_basic_2d
>>> y = keras_resnet.models.TimeDistributedResNet(x, classes, blocks, blocks)
>>> y = keras.layers.TimeDistributed(keras.layers.Flatten())(y.output)
>>> y = keras.layers.TimeDistributed(keras.layers.Dense(classes, activation="softmax"))(y)
>>> model = keras.models.Model(x, y)
>>> model.compile("adam", "categorical_crossentropy", ["accuracy"])
"""
if tf.keras.backend.image_data_format() == "channels_last":
axis = 3
else:
axis = 1
x = tf.keras.layers.TimeDistributed(
tf.keras.layers.ZeroPadding2D(padding=3), name="padding_conv1")(inputs)
x = tf.keras.layers.TimeDistributed(tf.keras.layers.Conv2D(64, (7, 7),
strides=(2, 2),
use_bias=False),
name="conv1")(x)
x = tf.keras.layers.TimeDistributed(layers.BatchNormalization(
axis=axis, epsilon=1e-5, freeze=freeze_bn),
name="bn_conv1")(x)
x = tf.keras.layers.TimeDistributed(tf.keras.layers.Activation("relu"),
name="conv1_relu")(x)
x = tf.keras.layers.TimeDistributed(tf.keras.layers.MaxPooling2D(
(3, 3), strides=(2, 2), padding="same"),
name="pool1")(x)
features = 64
outputs = []
for stage_id, iterations in enumerate(blocks):
for block_id in range(iterations):
x = block(features,
stage_id,
block_id,
numerical_name=(blocks[stage_id] > 6),
freeze_bn=freeze_bn)(x)
features *= 2
outputs.append(x)
if include_top:
assert classes > 0
x = tf.keras.layers.TimeDistributed(
tf.keras.layers.GlobalAveragePooling2D(), name="pool5")(x)
x = tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(
classes, activation="softmax"),
name="fc1000")(x)
return tf.keras.models.Model(inputs=inputs, outputs=x, *args, **kwargs)
else:
# Else output each stages features
return tf.keras.models.Model(inputs=inputs,
outputs=outputs,
*args,
**kwargs)