from .model_utils import Model, SiameseAccuracy
import tensorflow as tf
from tensorflow.keras import layers, models, metrics, losses
from tensorflow.keras.callbacks import LambdaCallback
import tensorflow_addons as tfa
class UpscalingLayer(layers.Layer):
def __init__(self, out_shape):
super(UpscalingLayer, self).__init__()
self.out_shape = out_shape
def call(self, inputs, **kwargs):
output = tf.repeat(inputs, 2, axis=1)
return tf.ensure_shape(output, self.out_shape)
def get_config(self):
config = super().get_config().copy()
config.update({
'out_shape': self.out_shape,
})
return config
"""
A simple siamese model with convolutional layers, followed by a dense layer and a difference layer.
Model architecture based on https://github.com/ssloxford/auto-phy-fingerprint/blob/main/code/fingerprinting/adsb-siamese/train.py
"""
class SimpleSiameseConvModel(Model):
def build_model(self, input_len, feature_count, conv_layers, dense_size, learning_rate):
"""
Build the siamese model.
Args:
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
Returns:
tf.keras.Model: The siamese model.
"""
input_left = layers.Input(shape=(input_len, feature_count), name="input_left")
input_right = layers.Input(shape=(input_len, feature_count), name="input_right")
input_dummy = layers.Input(shape=(input_len, feature_count), name="input_dummy")
input_norm = layers.BatchNormalization()(input_dummy)
input_padded = layers.ZeroPadding1D(padding=2)(input_norm)
conv = [input_padded]
for i, (filters, kernel_size) in enumerate(conv_layers):
conv.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_{}".format(i))(conv[-1]))
conv.append(layers.MaxPooling1D()(conv[-1]))
flat = layers.Flatten()(conv[-1])
dense = layers.Dense(dense_size, activation='sigmoid', name="dense")(flat)
extractor = models.Model(inputs=[input_dummy], outputs=[dense], name="extractor")
encoded_left = models.Model(inputs=[input_left], outputs=[extractor.call(input_left)], name="encoded_left")
encoded_right = models.Model(inputs=[input_right], outputs=[extractor.call(input_right)], name="encoded_right")
difference = layers.Subtract(name="difference")([encoded_left.output, encoded_right.output])
difference_abs = layers.Lambda(lambda x: tf.abs(x), name="difference_abs")(difference)
prediction = layers.Dense(1, activation='sigmoid', name="output")(difference_abs)
model = models.Model(inputs=[input_left, input_right], outputs=[prediction], name="siamese")
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss='binary_crossentropy',
metrics=['accuracy']
)
return model
def __init__(self, name, input_len, feature_count, conv_layers, dense_size, learning_rate, save_dir=None):
"""
Initialize the siamese model.
Args:
name (str): Name of the model.
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
save_dir (str): Directory to save the model to.
"""
self.model = self.build_model(input_len, feature_count, conv_layers, dense_size, learning_rate)
super().__init__(self.model, name, save_dir=save_dir)
"""
A simple siamese model with convolutional layers (split into separate IQ layers), followed by a dense layer and a difference layer.
Model architecture based on https://github.com/ssloxford/auto-phy-fingerprint/blob/main/code/fingerprinting/adsb-siamese/train.py
"""
class SimpleSiameseSplitConvModel(Model):
def build_model(self, input_len, feature_count, conv_layers, dense_size, learning_rate):
"""
Build the siamese model.
Args:
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
Returns:
tf.keras.Model: The siamese model.
"""
input_left = layers.Input(shape=(input_len, feature_count), name="input_left")
input_right = layers.Input(shape=(input_len, feature_count), name="input_right")
input_dummy = layers.Input(shape=(input_len, feature_count), name="input_dummy")
input_norm = layers.BatchNormalization()(input_dummy)
input_padded = layers.ZeroPadding1D(padding=2)(input_norm)
input_i, input_q = tf.split(input_padded, feature_count, axis=2)
conv_i = [input_i]
conv_q = [input_q]
for i, (filters, kernel_size) in enumerate(conv_layers):
conv_i.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_i_{}".format(i))(conv_i[-1]))
conv_i.append(layers.MaxPooling1D()(conv_i[-1]))
conv_q.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_{}".format(i))(conv_q[-1]))
conv_q.append(layers.MaxPooling1D()(conv_q[-1]))
conv_out = layers.Concatenate(axis=2, name="conv_out")([conv_i[-1], conv_q[-1]])
flat = layers.Flatten()(conv_out)
dense = layers.Dense(dense_size, activation='sigmoid', name="dense")(flat)
extractor = models.Model(inputs=[input_dummy], outputs=[dense], name="extractor")
encoded_left = models.Model(inputs=[input_left], outputs=[extractor.call(input_left)], name="encoded_left")
encoded_right = models.Model(inputs=[input_right], outputs=[extractor.call(input_right)], name="encoded_right")
difference = layers.Subtract(name="difference")([encoded_left.output, encoded_right.output])
difference_abs = layers.Lambda(lambda x: tf.abs(x), name="difference_abs")(difference)
prediction = layers.Dense(1, activation='sigmoid', name="output")(difference_abs)
model = models.Model(inputs=[input_left, input_right], outputs=[prediction], name="siamese")
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss='binary_crossentropy',
metrics=['accuracy']
)
return model
def __init__(self, name, input_len, feature_count, conv_layers, dense_size, learning_rate, save_dir=None):
"""
Initialize the siamese model.
Args:
name (str): Name of the model.
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
save_dir (str): Directory to save the model to.
"""
if feature_count != 2:
raise ValueError("This model currently only supports 2D input vectors (IQ samples).")
self.model = self.build_model(input_len, feature_count, conv_layers, dense_size, learning_rate)
super().__init__(self.model, name, save_dir=save_dir)
"""
A simple siamese model with convolutional layers (split into separate IQ layers), followed by a dense layer.
Triplet loss is used.
Model architecture based on https://github.com/ssloxford/auto-phy-fingerprint/blob/main/code/fingerprinting/adsb-siamese/train.py
"""
class SimpleTripletSplitConvModel(Model):
def siamese_accuracy(self, y_true, y_pred):
"""
Calculate the accuracy of the siamese model.
Args:
y_true (tf.Tensor): True labels.
y_pred (tf.Tensor): Embeddings.
"""
label_equality = tf.equal(y_true[:, None], y_true[None, :])
distances = tf.linalg.norm(y_pred[:, None, :] - y_pred[None, :, :], axis=-1)
predicted_matches = distances < self.accuracy_threshold
correct_matches = tf.logical_and(label_equality, predicted_matches)
accuracy = tf.reduce_mean(tf.cast(correct_matches, tf.float32))
return accuracy
def siamese_tpr(self, y_true, y_pred):
"""
Calculate the TPR of the siamese model.
Args:
y_true (tf.Tensor): True labels.
y_pred (tf.Tensor): Embeddings.
"""
label_equality = tf.equal(y_true[:, None], y_true[None, :])
distances = tf.linalg.norm(y_pred[:, None, :] - y_pred[None, :, :], axis=-1)
predicted_matches = distances < self.accuracy_threshold
#true_positives = tf.logical_and(label_equality, predicted_matches)
#tpr = tf.reduce_sum(tf.cast(true_positives, tf.float32)) / tf.reduce_sum(tf.cast(label_equality, tf.float32))
true_positives = tf.logical_and(label_equality, predicted_matches)
tp = tf.reduce_sum(tf.cast(true_positives, tf.float32))
fn = tf.reduce_sum(tf.cast(tf.logical_and(tf.logical_not(label_equality), predicted_matches), tf.float32))
tpr = tp / (tp + fn)
return tpr
def siamese_fpr(self, y_true, y_pred):
"""
Calculate the FPR of the siamese model.
Args:
y_true (tf.Tensor): True labels.
y_pred (tf.Tensor): Embeddings.
"""
label_equality = tf.equal(y_true[:, None], y_true[None, :])
distances = tf.linalg.norm(y_pred[:, None, :] - y_pred[None, :, :], axis=-1)
predicted_matches = distances < self.accuracy_threshold
#false_positives = tf.logical_and(tf.logical_not(label_equality), predicted_matches)
#fpr = tf.reduce_sum(tf.cast(false_positives, tf.float32)) / tf.reduce_sum(tf.cast(tf.logical_not(label_equality), tf.float32))
false_positives = tf.logical_and(tf.logical_not(label_equality), predicted_matches)
fp = tf.reduce_sum(tf.cast(false_positives, tf.float32))
tn = tf.reduce_sum(tf.cast(tf.logical_and(tf.logical_not(label_equality), tf.logical_not(predicted_matches)), tf.float32))
fpr = fp / (fp + tn)
return fpr
def siamese_fscore(self, y_true, y_pred):
"""
Calculate the F-score of the siamese model.
Args:
y_true (tf.Tensor): True labels.
y_pred (tf.Tensor): Embeddings.
"""
label_equality = tf.equal(y_true[:, None], y_true[None, :])
distances = tf.linalg.norm(y_pred[:, None, :] - y_pred[None, :, :], axis=-1)
predicted_matches = distances < self.accuracy_threshold
true_positives = tf.logical_and(label_equality, predicted_matches)
false_positives = tf.logical_and(tf.logical_not(label_equality), predicted_matches)
false_negatives = tf.logical_and(label_equality, tf.logical_not(predicted_matches))
tp = tf.reduce_sum(tf.cast(true_positives, tf.float32))
fp = tf.reduce_sum(tf.cast(false_positives, tf.float32))
fn = tf.reduce_sum(tf.cast(false_negatives, tf.float32))
precision = tp / (tp + fp)
recall = tp / (tp + fn)
f_score = 2 * precision * recall / (precision + recall)
return f_score
def build_model(self, input_len, feature_count, conv_layers, dense_size, learning_rate, triplet_margin, triplet_distance_metric, normalization):
"""
Build the siamese model.
Args:
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
triplet_margin (float): Margin for the triplet loss.
triplet_distance_metric (str): Distance metric to use for the triplet loss. One of 'L2', 'squared-L2', 'angular', or a callable.
normalization (str): Normalization to use for the embeddings. One of 'L1', 'L2', None.
Returns:
tf.keras.Model: The siamese model.
"""
input = layers.Input(shape=(input_len, feature_count), name="input")
input_norm = layers.BatchNormalization()(input)
input_padded = layers.ZeroPadding1D(padding=2)(input_norm)
input_i, input_q = tf.split(input_padded, feature_count, axis=2)
conv_i = [input_i]
conv_q = [input_q]
for i, (filters, kernel_size) in enumerate(conv_layers):
conv_i.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_i_{}".format(i))(conv_i[-1]))
conv_i.append(layers.MaxPooling1D()(conv_i[-1]))
conv_q.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_{}".format(i))(conv_q[-1]))
conv_q.append(layers.MaxPooling1D()(conv_q[-1]))
conv_out = layers.Concatenate(axis=2, name="conv_out")([conv_i[-1], conv_q[-1]])
flat = layers.Flatten()(conv_out)
dense = layers.Dense(dense_size, activation=None, name="dense")(flat)
norm = None
if normalization == 'l1':
norm = layers.Lambda(lambda x: tf.linalg.normalize(x, ord=1, axis=1)[0], name="l1_norm")(dense)
elif normalization == 'l2':
norm = layers.Lambda(lambda x: tf.math.l2_normalize(x, axis=1), name="l2_norm")(dense)
elif normalization is None:
norm = dense
model = models.Model(inputs=[input], outputs=[norm], name="model")
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss=tfa.losses.TripletSemiHardLoss(
margin=triplet_margin,
distance_metric=triplet_distance_metric,
),
metrics=[
#self.siamese_accuracy,
#self.siamese_tpr,
#self.siamese_fpr,
#self.siamese_fscore,
]
)
return model
def __init__(self, name, input_len, feature_count, conv_layers, dense_size, learning_rate, triplet_margin=1.0, triplet_distance_metric='L2', normalization='L2', accuracy_threshold=0.5, save_dir=None):
"""
Initialize the siamese model.
Args:
name (str): Name of the model.
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
triplet_margin (float): Margin for the triplet loss.
triplet_distance_metric (str): Distance metric to use for the triplet loss. One of 'L2', 'squared-L2', 'angular', or a callable.
normalization (str): Normalization to use for the embeddings. One of 'L1', 'L2', None.
accuracy_threshold (float): Accuracy threshold for the model's SiameseAccuracy metric.
save_dir (str): Directory to save the model to.
"""
self.accuracy_threshold = accuracy_threshold
if feature_count != 2:
raise ValueError("This model currently only supports 2D input vectors (IQ samples).")
if normalization is not None:
normalization = normalization.lower()
if normalization not in ['l1', 'l2']:
raise ValueError("Invalid normalization: {}".format(normalization))
self.model = self.build_model(input_len, feature_count, conv_layers, dense_size, learning_rate, triplet_margin, triplet_distance_metric, normalization)
super().__init__(self.model, name, save_dir=save_dir)
"""
A simple siamese model with convolutional layers (split into separate IQ layers), followed by a dense layer.
Triplet loss is used.
"""
class AETripletSplitConvModel(Model):
def build_model(self, input_len, feature_count, conv_layers, dense_size, learning_rate, triplet_margin, triplet_distance_metric, normalization):
"""
Build the siamese model.
Args:
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
triplet_margin (float): Margin for the triplet loss.
triplet_distance_metric (str): Distance metric to use for the triplet loss. One of 'L2', 'squared-L2', 'angular', or a callable.
normalization (str): Normalization to use for the embeddings. One of 'L1', 'L2', None.
Returns:
tf.keras.Model: The siamese model.
"""
input = layers.Input(shape=(input_len, feature_count), name="input")
input_norm = layers.BatchNormalization()(input)
input_padded = layers.ZeroPadding1D(padding=2)(input_norm)
input_i, input_q = tf.split(input_padded, feature_count, axis=2)
conv_shapes = []
conv_i = [input_i]
conv_q = [input_q]
for i, (filters, kernel_size) in enumerate(conv_layers):
conv_shapes.append(conv_i[-1].get_shape().as_list()[1])
conv_i.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_i_{}".format(i))(conv_i[-1]))
conv_i.append(layers.MaxPooling1D()(conv_i[-1]))
conv_q.append(layers.Conv1D(filters, kernel_size, activation='relu', name="conv_q_{}".format(i))(conv_q[-1]))
conv_q.append(layers.MaxPooling1D()(conv_q[-1]))
conv_out = layers.Concatenate(axis=2, name="conv_out")([conv_i[-1], conv_q[-1]])
flat = layers.Flatten()(conv_out)
dense = layers.Dense(dense_size, activation=None, name="dense")(flat)
norm = None
if normalization == 'l1':
norm = layers.Lambda(lambda x: tf.linalg.normalize(x, ord=1, axis=1)[0], name="embedding")(dense)
elif normalization == 'l2':
norm = layers.Lambda(lambda x: tf.math.l2_normalize(x, axis=1), name="embedding")(dense)
elif normalization is None:
norm = layers.Identity(name="embedding")(dense)
# Build the decoder
decoder_dense = layers.Dense(flat.get_shape().as_list()[1], activation='relu', name="decoder_dense")(norm)
deconv_in = layers.Reshape(conv_out.get_shape().as_list()[1:])(decoder_dense)
deconv_i, deconv_q = tf.split(deconv_in, feature_count, axis=2)
deconv_i = [deconv_i]
deconv_q = [deconv_q]
for i, (filters, kernel_size) in enumerate(reversed(conv_layers)):
upscaling_shape = deconv_i[-1].get_shape().as_list()
upscaling_shape[1] *= 2
#deconv_i.append(layers.UpSampling1D()(deconv_i[-1]))
#deconv_i.append(layers.Conv1DTranspose(filters, kernel_size, activation='relu', name="deconv_i_{}".format(i))(deconv_i[-1]))
deconv_i.append(UpscalingLayer(upscaling_shape)(deconv_i[-1]))
deconv_i.append(layers.Conv1D(filters, kernel_size, activation='relu', name="deconv_i_{}".format(i))(deconv_i[-1]))
#deconv_q.append(layers.UpSampling1D()(deconv_q[-1]))
#deconv_q.append(layers.Conv1DTranspose(filters, kernel_size, activation='relu', name="deconv_{}".format(i))(deconv_q[-1]))
deconv_q.append(UpscalingLayer(upscaling_shape)(deconv_q[-1]))
deconv_q.append(layers.Conv1D(filters, kernel_size, activation='relu', name="deconv_q_{}".format(i))(deconv_q[-1]))
conv_shape = list(reversed(conv_shapes))[i]
if deconv_i[-1].get_shape().as_list()[1] != conv_shape:
deconv_i.append(layers.ZeroPadding1D(padding=(0, conv_shape - deconv_i[-1].get_shape().as_list()[1]))(deconv_i[-1]))
deconv_q.append(layers.ZeroPadding1D(padding=(0, conv_shape - deconv_q[-1].get_shape().as_list()[1]))(deconv_q[-1]))
output_i = layers.Conv1DTranspose(1, 5, activation='linear', padding='valid', name="output_i")(deconv_i[-1])
output_q = layers.Conv1DTranspose(1, 5, activation='linear', padding='valid', name="output_q")(deconv_q[-1])
output_concatenate = layers.Concatenate(axis=2, name="output_concatenate")([output_i, output_q])
cropping = (output_concatenate.get_shape().as_list()[1] - input_len) // 2
output = layers.Cropping1D(cropping=cropping, name="output")(output_concatenate)
model = models.Model(inputs=[input], outputs=[norm, output], name="model")
#model.add_loss(tfa.losses.TripletSemiHardLoss(
# margin=triplet_margin,
# distance_metric=triplet_distance_metric,
#)(norm), name="embedding_loss")
mse = tf.keras.losses.MeanSquaredError()(input, output)
model.add_loss(mse)
model.add_metric(mse, name="reconstruction_loss")
losses = {
"embedding": tfa.losses.TripletSemiHardLoss(
margin=triplet_margin,
distance_metric=triplet_distance_metric,
),
# MSE between input and output
#"output": tf.keras.losses.MeanSquaredError()(input, output),
}
loss_weights = {
"embedding": 1.0,
#"output": 1.0,
}
model.compile(
optimizer=tf.keras.optimizers.Adam(learning_rate=learning_rate),
loss=losses,
loss_weights=loss_weights,
#loss=None,
metrics=[
]
)
return model
def __init__(self, name, input_len, feature_count, conv_layers, dense_size, learning_rate, triplet_margin=1.0, triplet_distance_metric='L2', normalization='L2', accuracy_threshold=0.5, save_dir=None):
"""
Initialize the siamese model.
Args:
name (str): Name of the model.
input_len (int): Length of the input vector.
feature_count (int): Number of features in the input vector.
conv_layers (list): List of tuples (filters, kernel_size) for each convolutional layer.
dense_size (int): Size of the dense layer.
learning_rate (float): Learning rate for the optimizer.
triplet_margin (float): Margin for the triplet loss.
triplet_distance_metric (str): Distance metric to use for the triplet loss. One of 'L2', 'squared-L2', 'angular', or a callable.
normalization (str): Normalization to use for the embeddings. One of 'L1', 'L2', None.
accuracy_threshold (float): Accuracy threshold for the model's SiameseAccuracy metric.
save_dir (str): Directory to save the model to.
"""
self.accuracy_threshold = accuracy_threshold
if feature_count != 2:
raise ValueError("This model currently only supports 2D input vectors (IQ samples).")
if normalization is not None:
normalization = normalization.lower()
if normalization not in ['l1', 'l2']:
raise ValueError("Invalid normalization: {}".format(normalization))
self.model = self.build_model(input_len, feature_count, conv_layers, dense_size, learning_rate, triplet_margin, triplet_distance_metric, normalization)
super().__init__(self.model, name, save_dir=save_dir)
这是模型的脚本
卷积与反卷积层的正确使用顺序
本文探讨了卷积层(ConvLayers)、激活层(ActivationLayers)及反卷积层(DeconvLayers)在神经网络中的正确使用顺序。强调了在上采样操作中DeconvLayer的重要性，并提供了Caffe框架下具体的参数设置，同时指出了连接损失层(loss layer)时的注意事项。
扫一扫
2120
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
