Skip to content

GitLab

Commit 89e4b78b authored by Amir MOHAMMADI's avatar Amir MOHAMMADI
Browse files

remove extra models and improve comments

parent 36c3f4b9
Hide whitespace changes
Inline Side-by-side
bob/learn/tensorflow/models/autoencoder_face.py
View file @ 89e4b78b
"""Face Autoencoder used in:
IMPROVING CROSS-DATASET PERFORMANCE OF FACE PRESENTATION ATTACK DETECTION SYSTEMS USING FACE RECOGNITION DATASETS,
Mohammadi, Amir and Bhattacharjee, Sushil and Marcel, Sebastien, ICASSP 2020
"""
import tensorflow as tf
from bob.learn.tensorflow.models.densenet import densenet161
......@@ -9,63 +14,6 @@ def _get_l2_kw(weight_decay):
return l2_kw
# class ConvDecoder(tf.keras.Sequential):
# """The decoder similar to the one in
# https://github.com/google/compare_gan/blob/master/compare_gan/architectures/sndcgan.py
# """
# def __init__(
# self,
# z_dim,
# decoder_layers=(
# (512, 7, 7, 0),
# (256, 4, 2, 1),
# (128, 4, 2, 1),
# (64, 4, 2, 1),
# (32, 4, 2, 1),
# (16, 4, 2, 1),
# (3, 1, 1, 0),
# ),
# weight_decay=1e-5,
# name="Decoder",
# **kwargs,
# ):
# self.z_dim = z_dim
# self.data_format = data_format = "channels_last"
# l2_kw = _get_l2_kw(weight_decay)
# layers = [
# tf.keras.layers.Reshape((1, 1, z_dim), input_shape=(z_dim,), name="reshape")
# ]
# for i, (filters, kernel_size, strides, cropping) in enumerate(decoder_layers):
# dconv = tf.keras.layers.Conv2DTranspose(
# filters,
# kernel_size,
# strides=strides,
# use_bias=i == len(decoder_layers) - 1,
# data_format=data_format,
# name=f"dconv_{i}",
# **l2_kw,
# )
# crop = tf.keras.layers.Cropping2D(
# cropping=cropping, data_format=data_format, name=f"crop_{i}"
# )
# if i == len(decoder_layers) - 1:
# act = tf.keras.layers.Activation("tanh", name=f"tanh_{i}")
# bn = None
# else:
# act = tf.keras.layers.Activation("relu", name=f"relu_{i}")
# bn = tf.keras.layers.BatchNormalization(
# scale=False, fused=False, name=f"bn_{i}"
# )
# if bn is not None:
# layers.extend([dconv, crop, bn, act])
# else:
# layers.extend([dconv, crop, act])
# with tf.name_scope(name):
# super().__init__(layers=layers, name=name, **kwargs)
def ConvDecoder(
z_dim,
decoder_layers=(
......@@ -149,7 +97,7 @@ class Autoencoder(tf.keras.Model):
return z, x_hat
def autoencoder_face(z_dim=256, weight_decay=1e-9, decoder_last_act="tanh"):
def autoencoder_face(z_dim=256, weight_decay=1e-10, decoder_last_act="tanh"):
encoder = densenet161(
output_classes=z_dim, weight_decay=weight_decay, weights=None, name="DenseNet"
)
......
bob/learn/tensorflow/models/autoencoder_yz.py deleted 100644 → 0
View file @ 36c3f4b9
import tensorflow as tf
from .densenet import densenet161, ConvBlock
def _get_l2_kw(weight_decay):
l2_kw = {}
if weight_decay is not None:
l2_kw = {"kernel_regularizer": tf.keras.regularizers.l2(weight_decay)}
return l2_kw
class ConvEncoder(tf.keras.Model):
"""The encoder part"""
def __init__(
self,
encoder_layers,
data_format="channels_last",
weight_decay=1e-5,
name="Encoder",
**kwargs,
):
super().__init__(name=name, **kwargs)
self.data_format = data_format
l2_kw = _get_l2_kw(weight_decay)
layers = []
for i, (filters, kernel_size, strides, padding) in enumerate(encoder_layers):
bn = tf.keras.layers.BatchNormalization(
scale=False, fused=False, name=f"bn_{i}"
)
if i == 0:
act = tf.keras.layers.Activation("linear", name=f"linear_{i}")
else:
act = tf.keras.layers.Activation("relu", name=f"relu_{i}")
pad = tf.keras.layers.ZeroPadding2D(
padding=padding, data_format=data_format, name=f"pad_{i}"
)
conv = tf.keras.layers.Conv2D(
filters,
kernel_size,
strides=strides,
use_bias=(i == len(encoder_layers) - 1),
data_format=data_format,
name=f"conv_{i}",
**l2_kw,
)
if i == len(encoder_layers) - 1:
pool = tf.keras.layers.AvgPool2D(
data_format=data_format, name=f"pool_{i}"
)
else:
pool = tf.keras.layers.MaxPooling2D(
data_format=data_format, name=f"pool_{i}"
)
layers.extend([bn, act, pad, conv, pool])
self.sequential_layers = layers
def call(self, x, training=None):
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
class ConvDecoder(tf.keras.Model):
"""The encoder part"""
def __init__(
self, decoder_layers, y_dim, weight_decay=1e-5, name="Decoder", **kwargs
):
super().__init__(name=name, **kwargs)
self.data_format = data_format = "channels_last"
self.y_dim = y_dim
l2_kw = _get_l2_kw(weight_decay)
layers = []
for i, (filters, kernel_size, strides, cropping) in enumerate(decoder_layers):
dconv = tf.keras.layers.Conv2DTranspose(
filters,
kernel_size,
strides=strides,
use_bias=False,
data_format=data_format,
name=f"dconv_{i}",
**l2_kw,
)
crop = tf.keras.layers.Cropping2D(
cropping=cropping, data_format=data_format, name=f"crop_{i}"
)
bn = tf.keras.layers.BatchNormalization(
scale=(i == len(decoder_layers) - 1), fused=False, name=f"bn_{i}"
)
if i == len(decoder_layers) - 1:
act = tf.keras.layers.Activation("tanh", name=f"tanh_{i}")
else:
act = tf.keras.layers.Activation("relu", name=f"relu_{i}")
layers.extend([dconv, crop, bn, act])
self.sequential_layers = layers
def call(self, x, y, training=None):
y = tf.reshape(tf.cast(y, x.dtype), (-1, 1, 1, self.y_dim))
x = tf.concat([x, y], axis=-1)
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
class Autoencoder(tf.keras.Model):
"""
A class defining a simple convolutional autoencoder.
Attributes
----------
data_format : str
channels_last is only supported
decoder : object
The encoder part
encoder : object
The decoder part
"""
def __init__(
self, encoder, decoder, z_dim, weight_decay=1e-5, name="Autoencoder", **kwargs
):
super().__init__(name=name, **kwargs)
data_format = "channels_last"
self.data_format = data_format
self.weight_decay = weight_decay
self.encoder = encoder
self.decoder = decoder
self.z_dim = z_dim
def call(self, x, y, training=None):
self.encoder_output = tf.reshape(
self.encoder(x, training=training), (-1, 1, 1, self.z_dim)
)
self.decoder_output = self.decoder(self.encoder_output, y, training=training)
return self.decoder_output
def densenet161_autoencoder(z_dim=256, y_dim=3, weight_decay=1e-10):
encoder = densenet161(output_classes=z_dim, weight_decay=weight_decay, weights=None)
decoder_layers = (
(128, 7, 7, 0),
(64, 4, 2, 1),
(32, 4, 2, 1),
(16, 4, 2, 1),
(8, 4, 2, 1),
(4, 4, 2, 1),
(3, 1, 1, 0),
)
decoder = ConvDecoder(
decoder_layers, y_dim=y_dim, weight_decay=weight_decay, name="Decoder"
)
autoencoder = Autoencoder(encoder, decoder, z_dim=z_dim, weight_decay=weight_decay)
return autoencoder
class ConvDecoderSupervised(tf.keras.Model):
"""The encoder part"""
def __init__(
self,
decoder_layers,
weight_decay=1e-5,
data_format="channels_last",
name="Decoder",
y_dim=None,
**kwargs,
):
super().__init__(name=name, **kwargs)
self.data_format = data_format
self.y_dim = y_dim
l2_kw = _get_l2_kw(weight_decay)
layers = []
for i, (filters, kernel_size, strides, cropping) in enumerate(decoder_layers):
dconv = tf.keras.layers.Conv2DTranspose(
filters,
kernel_size,
strides=strides,
use_bias=False,
data_format=data_format,
name=f"dconv_{i}",
**l2_kw,
)
crop = tf.keras.layers.Cropping2D(
cropping=cropping, data_format=data_format, name=f"crop_{i}"
)
bn = tf.keras.layers.BatchNormalization(
scale=(i == len(decoder_layers) - 1), fused=False, name=f"bn_{i}"
)
if i == len(decoder_layers) - 1:
act = tf.keras.layers.Activation("tanh", name=f"tanh_{i}")
else:
act = tf.keras.layers.Activation("relu", name=f"relu_{i}")
layers.extend([dconv, crop, bn, act])
self.sequential_layers = layers
def call(self, x, training=None):
x = tf.reshape(x, (-1, 1, 1, x.get_shape().as_list()[-1]))
if self.y_dim is not None:
y_fixed = tf.one_hot([[[0]]], self.y_dim, dtype=x.dtype)
y_fixed = tf.tile(y_fixed, multiples=[tf.shape(x)[0], 1, 1, 1])
x = tf.concat([x, y_fixed], axis=-1)
x = tf.keras.Input(tensor=x)
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
def densenet161_autoencoder_supervised(
x,
training,
weight_decay=1e-10,
z_dim=256,
y_dim=1,
deeppixbis_add_one_more_layer=False,
start_from_face_autoencoder=False,
):
data_format = "channels_last"
with tf.name_scope("Autoencoder"):
densenet = densenet161(
output_classes=z_dim,
weight_decay=weight_decay,
weights=None,
data_format=data_format,
)
z = densenet(x, training=training)
transition = tf.keras.Input(tensor=densenet.transition_blocks[1].output)
layers = [
tf.keras.layers.Conv2D(
filters=1,
kernel_size=1,
kernel_initializer="he_normal",
kernel_regularizer=tf.keras.regularizers.l2(weight_decay),
data_format=data_format,
name="dec",
),
tf.keras.layers.Flatten(
data_format=data_format, name="Pixel_Logits_Flatten"
),
]
if deeppixbis_add_one_more_layer:
layers.insert(
0,
ConvBlock(
num_filters=32,
data_format=data_format,
bottleneck=True,
weight_decay=weight_decay,
name="prelogits",
),
)
y = transition
with tf.name_scope("DeepPixBiS"):
for l in layers:
try:
y = l(y, training=training)
except TypeError:
y = l(y)
deep_pix_bis_final_layers = tf.keras.Model(
inputs=transition, outputs=y, name="DeepPixBiS"
)
encoder = tf.keras.Model(inputs=[x, transition], outputs=[y, z], name="Encoder")
encoder.densenet = densenet
if deeppixbis_add_one_more_layer:
encoder.prelogits = deep_pix_bis_final_layers.layers[-3].output
else:
encoder.prelogits = transition
encoder.deep_pix_bis = deep_pix_bis_final_layers
decoder_layers = (
(128, 7, 7, 0),
(64, 4, 2, 1),
(32, 4, 2, 1),
(16, 4, 2, 1),
(8, 4, 2, 1),
(4, 4, 2, 1),
(3, 1, 1, 0),
)
decoder = ConvDecoderSupervised(
decoder_layers,
weight_decay=weight_decay,
name="Decoder",
data_format=data_format,
y_dim=3 if start_from_face_autoencoder else None,
)
x_hat = decoder(z, training=training)
autoencoder = tf.keras.Model(
inputs=[x, transition], outputs=[y, z, x_hat], name="Autoencoder"
)
autoencoder.encoder = encoder
autoencoder.decoder = decoder
return autoencoder, y, z, x_hat
bob/learn/tensorflow/models/densenet.py
View file @ 89e4b78b
"""Densely Connected Convolutional Networks.
Reference [
Densely Connected Convolutional Networks](https://arxiv.org/abs/1608.06993)
Reference [Densely Connected Convolutional Networks](https://arxiv.org/abs/1608.06993)
"""
import tensorflow as tf
......
bob/learn/tensorflow/models/discriminator.py
View file @ 89e4b78b
"""Discriminator used in training of autoencoder in:
IMPROVING CROSS-DATASET PERFORMANCE OF FACE PRESENTATION ATTACK DETECTION SYSTEMS USING FACE RECOGNITION DATASETS,
Mohammadi, Amir and Bhattacharjee, Sushil and Marcel, Sebastien, ICASSP 2020
"""
import tensorflow as tf
from ..gan.spectral_normalization import spectral_norm_regularizer
from ..utils import gram_matrix
class ConvDiscriminator(tf.keras.Model):
"""A discriminator that can sit on top of DenseNet 161's transition 1 block.
The output of that block given 224x224x3 inputs is 14x14x384."""
def __init__(self, data_format="channels_last", n_classes=1, **kwargs):
super().__init__(**kwargs)
self.data_format = data_format
self.n_classes = n_classes
act = "sigmoid" if n_classes == 1 else "softmax"
self.sequential_layers = [
tf.keras.layers.Conv2D(200, 1, data_format=data_format),
tf.keras.layers.Activation("relu"),
tf.keras.layers.AveragePooling2D(3, 2, data_format=data_format),
tf.keras.layers.Conv2D(100, 1, data_format=data_format),
tf.keras.layers.Activation("relu"),
tf.keras.layers.AveragePooling2D(3, 2, data_format=data_format),
tf.keras.layers.Flatten(data_format=data_format),
tf.keras.layers.Dense(n_classes),
tf.keras.layers.Activation(act),
]
def call(self, x, training=None):
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
class ConvDiscriminator2(tf.keras.Model):
"""A discriminator that can sit on top of DenseNet 161's transition 1 block.
The output of that block given 224x224 inputs is 14x14x384. Here we want to output
15x15x128 features which is going to match the output of encoder in mcae.py given
these layers::
ENCODER_LAYERS = (
(32, 5, 1, 2),
(64, 5, 1, 2),
(128, 3, 1, 2),
(128, 3, 1, 2)
)
DECODER_LAYERS = (
(64, 3, 2, 1),
(32, 3, 2, 1),
(16, 5, 2, 2),
(8, 5, 2, 2),
(3, 2, 1, 1),
)
"""
def __init__(self, data_format="channels_last", **kwargs):
super().__init__(**kwargs)
self.data_format = data_format
self.sequential_layers = [
tf.keras.layers.ZeroPadding2D(
padding=((1, 0), (1, 0)), data_format=data_format
),
tf.keras.layers.Conv2D(256, 5, data_format=data_format, padding="same"),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Conv2D(128, 5, data_format=data_format, padding="same"),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Conv2D(128, 1, data_format=data_format, padding="same"),
tf.keras.layers.Activation("relu"),
]
def call(self, x, training=None):
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
class ConvDiscriminator3(tf.keras.Model):
"""A discriminator that takes images and tries its best.
Be careful, this one returns logits."""
def __init__(self, data_format="channels_last", n_classes=1, **kwargs):
super().__init__(**kwargs)
self.data_format = data_format
self.n_classes = n_classes
spectral_norm = spectral_norm_regularizer(scale=1.0)
conv2d_kw = {"kernel_regularizer": spectral_norm, "data_format": data_format}
self.sequential_layers = [
tf.keras.layers.Conv2D(64, 3, strides=1, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.Conv2D(64, 4, strides=2, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.Conv2D(128, 3, strides=1, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.Conv2D(128, 4, strides=2, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.Conv2D(256, 3, strides=1, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.Conv2D(256, 4, strides=2, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.Conv2D(512, 3, strides=1, **conv2d_kw),
tf.keras.layers.LeakyReLU(0.1),
tf.keras.layers.GlobalAveragePooling2D(data_format=data_format),
tf.keras.layers.Dense(n_classes),
]
def call(self, x, training=None):
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
class DenseDiscriminator(tf.keras.Model):
def DenseDiscriminator(n_classes=1, name="DenseDiscriminator", **kwargs):
"""A discriminator that takes vectors as input and tries its best.
Be careful, this one returns logits."""
def __init__(self, n_classes=1, **kwargs):
super().__init__(**kwargs)
self.n_classes = n_classes
self.sequential_layers = [
return tf.keras.Sequential(
[
tf.keras.layers.Dense(1000),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Dense(1000),
tf.keras.layers.Activation("relu"),
tf.keras.layers.Dense(n_classes),
]
def call(self, x, training=None):
for l in self.sequential_layers:
try:
x = l(x, training=training)
except TypeError:
x = l(x)
return x
class GramComparer1(tf.keras.Model):
"""A model to compare images based on their gram matrices."""
def __init__(self, **kwargs):
super().__init__(**kwargs)
self.batchnorm = tf.keras.layers.BatchNormalization()
self.conv2d = tf.keras.layers.Conv2D(128, 7)
def call(self, x_1_2, training=None):
def _call(x):
x = self.batchnorm(x, training=training)
x = self.conv2d(x)
return gram_matrix(x)
gram1 = _call(x_1_2[..., :3])
gram2 = _call(x_1_2[..., 3:])
return -tf.reduce_mean((gram1 - gram2) ** 2, axis=[1, 2])[:, None]
],
name=name,
**kwargs
)
bob/learn/tensorflow/models/msu_patch.py
View file @ 89e4b78b
"""Patch-based CNN used for face PAD in:
Y. Atoum, Y. Liu, A. Jourabloo, and X. Liu, “Face anti-spoofing using patch and
depth-based CNNs,” in 2017 IEEE International Joint Conference on Biometrics (IJCB),
Denver, CO, 2017, pp. 319–328.
"""
import tensorflow as tf
......
bob/learn/tensorflow/models/simple_cnn.py
View file @ 89e4b78b
"""
The network using keras (same as new_architecture function below)::
from tensorflow.python.keras import *
from tensorflow.python.keras.layers import *
simplecnn = Sequential([
Conv2D(32,(3,3),padding='same',use_bias=False, input_shape=(28,28,3)),
BatchNormalization(scale=False),
Activation('relu'),
MaxPool2D(padding='same'),