def build(self) -> tf.keras.models.Model:
x = self.stem_module(self.section_filters[0], self.image_input)
for block, (layers, filters) in enumerate(
zip(self.section_blocks, self.section_filters)):
for layer in range(layers):
if filters != x.shape[-1]:
x = self.transition_block(x, filters, strides=2)
x = self.residual_block(x)
if self.include_top:
x = tf.keras.layers.Activation("relu")(x)
x = utils.global_pool(x)
x = lq.layers.QuantDense(
self.num_classes,
kernel_initializer="glorot_normal",
)(x)
x = tf.keras.layers.Activation("softmax", dtype="float32")(x)
model = tf.keras.Model(inputs=self.image_input,
outputs=x,
name=self.name)
# Load weights.
if self.weights == "imagenet":
weights_path = (self.imagenet_weights_path if self.include_top else
self.imagenet_no_top_weights_path)
model.load_weights(weights_path)
elif self.weights is not None:
model.load_weights(self.weights)
return model
def _scale_binary_conv_output(self, conv_input: tf.Tensor,
conv_output: tf.Tensor,
name: str) -> tf.Tensor:
"""Data-dependent convolution scaling.
Scales the output of the convolution in the (squeeze-and-excite
style) data-dependent way described in Section 4.3 of Martinez at. al.
"""
in_filters = conv_input.shape[-1]
out_filters = conv_output.shape[-1]
z = utils.global_pool(conv_input, name=f"{name}_scaling_pool")
dim_reduction = tf.keras.layers.Dense(
int(in_filters // self.scaling_r),
activation="relu",
kernel_initializer="he_normal",
kernel_regularizer=self.kernel_regularizer,
name=f"{name}_scaling_dense_reduce",
use_bias=False,
)(z)
dim_expansion = tf.keras.layers.Dense(
out_filters,
activation="sigmoid",
kernel_initializer="he_normal",
kernel_regularizer=self.kernel_regularizer,
name=f"{name}_scaling_dense_expand",
use_bias=False,
)(dim_reduction)
scales = tf.keras.layers.Reshape(
(1, 1, out_filters), name=f"{name}_scaling_reshape")(dim_expansion)
return tf.keras.layers.Multiply(name=f"{name}_scaling_multiplication")(
[conv_output, scales])
def build(self) -> tf.keras.models.Model:
x = stem_module(self.stem_filters, self.image_input)
for block, (layers, filters, use_squeeze_and_excite) in enumerate(
zip(
self.blocks_per_section,
self.section_filters,
self.use_squeeze_and_excite_in_section,
)):
for layer in range(layers):
if filters == x.shape[-1]:
x = self.residual_block(x, use_squeeze_and_excite)
else:
strides = 1 if (block == 0 or layer != 0) else 2
x = self.transition_block(x, filters, strides,
use_squeeze_and_excite)
x = tf.keras.layers.Activation("relu")(x)
if self.include_top:
x = utils.global_pool(x)
x = tf.keras.layers.Dense(self.num_classes,
kernel_initializer="glorot_normal")(x)
x = tf.keras.layers.Activation("softmax", dtype="float32")(x)
model = tf.keras.Model(inputs=self.image_input,
outputs=x,
name=self.name)
return model
def build(self) -> tf.keras.models.Model:
# Layer 1
out = tf.keras.layers.Conv2D(
self.filters,
(7, 7),
strides=2,
kernel_initializer=self.kernel_initializer,
padding="same",
use_bias=False,
)(self.image_input)
out = tf.keras.layers.BatchNormalization(momentum=0.8)(out)
out = tf.keras.layers.MaxPool2D((3, 3), strides=2, padding="same")(out)
# Layer 2
out = self.residual_block(out, filters=self.filters)
# Layer 3 - 5
for _ in range(3):
out = self.residual_block(out)
# Layer 6 - 17
for _ in range(3):
out = self.residual_block(out, double_filters=True)
for _ in range(3):
out = self.residual_block(out)
# Layer 18
if self.include_top:
out = utils.global_pool(out)
out = tf.keras.layers.Dense(self.num_classes)(out)
out = tf.keras.layers.Activation("softmax", dtype="float32")(out)
model = tf.keras.Model(inputs=self.image_input,
outputs=out,
name="birealnet18")
# Load weights.
if self.weights == "imagenet":
# Download appropriate file
if self.include_top:
weights_path = utils.download_pretrained_model(
model="birealnet",
version="v0.3.0",
file="birealnet_weights.h5",
file_hash=
"6e6efac1584fcd60dd024198c87f42eb53b5ec719a5ca1f527e1fe7e8b997117",
)
else:
weights_path = utils.download_pretrained_model(
model="birealnet",
version="v0.3.0",
file="birealnet_weights_notop.h5",
file_hash=
"5148b61c0c2a1094bdef811f68bf4957d5ba5f83ad26437b7a4a6855441ab46b",
)
model.load_weights(weights_path)
elif self.weights is not None:
model.load_weights(self.weights)
return model
def last_block(self, x: tf.Tensor, name: str = "") -> tf.Tensor:
"""Last block, shared across ResNet, StrongBaselineNet and Real-to-Bin nets."""
x = utils.global_pool(x, name=f"{name}_global_pool")
x = tf.keras.layers.Dense(
self.num_classes,
name=f"{name}_logits",
)(x)
return tf.keras.layers.Softmax(name=f"{name}_probs",
dtype=tf.float32)(x)
def build_model(input_res=(32, 32), data_format="channels_first"):
if data_format == "channels_first":
input_shape = (3, input_res[0], input_res[1])
else:
input_shape = (input_res[0], input_res[1], 3)
inp = tf.keras.Input(shape=input_shape)
x = tf.keras.layers.Conv2D(6, 3, 2, data_format=data_format)(inp)
x = utils.global_pool(x, data_format=data_format)
x = tf.keras.layers.Dense(2)(x)
return tf.keras.Model(inputs=inp, outputs=x)
def test_global_pool(input_res, data_format):
shape = (3,
*input_res) if data_format == "channels_first" else (*input_res,
3)
inp = tf.keras.Input(shape=shape)
x = tf.keras.layers.Conv2D(6, 3, 2, data_format=data_format)(inp)
x = utils.global_pool(x, data_format=data_format)
x = tf.keras.layers.Dense(2)(x)
model = tf.keras.Model(inp, x)
assert model.outputs[0].shape.as_list() == [None, 2]
def squeeze_and_excite(inp: tf.Tensor, filters: int, r: int = 16):
"""Squeeze and Excite as per [Squeeze-and-Excitation Networks](https://arxiv.org/abs/1709.01507).
Use of S&E in BNNs was pioneered in [Training binary neural networks with
real-to-binary convolutions](https://openreview.net/forum?id=BJg4NgBKvH).
"""
out = utils.global_pool(inp)
out = tf.keras.layers.Dense(
inp.shape[-1] // r,
activation="relu",
kernel_initializer="he_normal",
use_bias=False,
kernel_regularizer=tf.keras.regularizers.l2(1e-5),
)(out)
out = tf.keras.layers.Dense(
filters,
activation="sigmoid",
kernel_initializer="he_normal",
use_bias=False,
kernel_regularizer=tf.keras.regularizers.l2(1e-5),
)(out)
return tf.reshape(out, [-1, 1, 1, filters])
def build(self) -> tf.keras.models.Model:
x = self.image_input
x = self.group_stem(x, name="stem")
for i, (n, f) in enumerate(zip(self.num_blocks, self.transition_features)):
for j in range(n):
x = self.block(x, f"section_{i}_block_{j}")
if f:
x = self.transition_block(x, f, f"section_{i}_transition")
x = self.norm(x, "head_bn")
x = self.act(x, "head_relu")
if self.include_top:
x = utils.global_pool(x, name="head_globalpool")
x = tf.keras.layers.Dense(
self.num_classes,
kernel_initializer=self.kernel_initializer,
name="head_dense",
)(x)
x = tf.keras.layers.Activation(
"softmax", dtype="float32", name="head_softmax"
)(x)
model = tf.keras.models.Model(
inputs=self.image_input, outputs=x, name=self.name
)
if self.weights == "imagenet":
model.load_weights(
self.imagenet_weights_path
if self.include_top
else self.imagenet_no_top_weights_path
)
elif self.weights is not None:
model.load_weights(self.weights)
return model
def build(self) -> BinaryDenseNet:
if self.image_input.shape[1] and self.image_input.shape[1] < 50:
x = tf.keras.layers.Conv2D(
self.initial_filters,
kernel_size=3,
padding="same",
kernel_initializer="he_normal",
use_bias=False,
)(self.image_input)
else:
x = tf.keras.layers.Conv2D(
self.initial_filters,
kernel_size=7,
strides=2,
padding="same",
kernel_initializer="he_normal",
use_bias=False,
)(self.image_input)
x = tf.keras.layers.BatchNormalization(momentum=0.9,
epsilon=1e-5)(x)
x = tf.keras.layers.Activation("relu")(x)
x = tf.keras.layers.MaxPool2D(3, strides=2, padding="same")(x)
for block, layers_per_block in enumerate(self.layers):
for _ in range(layers_per_block):
x = self.densely_connected_block(x, self.dilation_rate[block])
if block < len(self.layers) - 1:
x = tf.keras.layers.BatchNormalization(momentum=0.9,
epsilon=1e-5)(x)
if self.dilation_rate[block + 1] == 1:
x = tf.keras.layers.MaxPooling2D(2, strides=2)(x)
x = tf.keras.layers.Activation("relu")(x)
x = tf.keras.layers.Conv2D(
round(x.shape.as_list()[-1] // self.reduction[block] / 32)
* 32,
kernel_size=1,
kernel_initializer="he_normal",
use_bias=False,
)(x)
x = tf.keras.layers.BatchNormalization(momentum=0.9, epsilon=1e-5)(x)
x = tf.keras.layers.Activation("relu")(x)
if self.include_top:
x = utils.global_pool(x)
x = tf.keras.layers.Dense(self.num_classes,
kernel_initializer="he_normal")(x)
x = tf.keras.layers.Activation("softmax", dtype="float32")(x)
model = BinaryDenseNet(inputs=self.image_input,
outputs=x,
name=self.name)
if self.weights == "imagenet":
if self.include_top:
weights_path = self.imagenet_weights_path
else:
weights_path = self.imagenet_no_top_weights_path
model.load_weights(weights_path)
elif self.weights is not None:
model.load_weights(self.weights)
return model
def build(self) -> tf.keras.models.Model:
if self.image_input.shape[1] and self.image_input.shape[1] < 50:
x = tf.keras.layers.Conv2D(
self.initial_filters,
kernel_size=3,
padding="same",
kernel_initializer="he_normal",
use_bias=False,
)(self.image_input)
else:
x = tf.keras.layers.Conv2D(
self.initial_filters,
kernel_size=7,
strides=2,
padding="same",
kernel_initializer="he_normal",
use_bias=False,
)(self.image_input)
x = tf.keras.layers.BatchNormalization(momentum=0.9,
epsilon=1e-5)(x)
x = tf.keras.layers.Activation("relu")(x)
x = tf.keras.layers.MaxPool2D(3, strides=2, padding="same")(x)
x = tf.keras.layers.BatchNormalization(momentum=0.9,
epsilon=1e-5)(x)
for block, (layers, filters) in enumerate(zip(*self.spec)):
# This trick adds shortcut connections between original ResNet
# blocks. We wultiply the number of blocks by two, but add only one
# layer instead of two in each block
for layer in range(layers * 2):
strides = 1 if block == 0 or layer != 0 else 2
x = self.residual_block(x, filters, strides=strides)
x = tf.keras.layers.Activation("relu")(x)
if self.include_top:
x = utils.global_pool(x)
x = tf.keras.layers.Dense(self.num_classes,
kernel_initializer="glorot_normal")(x)
x = tf.keras.layers.Activation("softmax", dtype="float32")(x)
model = tf.keras.Model(
inputs=self.image_input,
outputs=x,
name=f"binary_resnet_e_{self.num_layers}",
)
# Load weights.
if self.weights == "imagenet":
# Download appropriate file
if self.include_top:
weights_path = utils.download_pretrained_model(
model="resnet_e",
version="v0.1.0",
file="resnet_e_18_weights.h5",
file_hash=
"bde4a64d42c164a7b10a28debbe1ad5b287c499bc0247ecb00449e6e89f3bf5b",
)
else:
weights_path = utils.download_pretrained_model(
model="resnet_e",
version="v0.1.0",
file="resnet_e_18_weights_notop.h5",
file_hash=
"14cb037e47d223827a8d09db88ec73d60e4153a4464dca847e5ae1a155e7f525",
)
model.load_weights(weights_path)
elif self.weights is not None:
model.load_weights(self.weights)
return model
