def create_model(shape):
"""Implementation of a modified LeNet-5.
Modified Architecture -- ConvNet --> Pool --> ConvNet --> Pool --> (
Flatten) --> FullyConnected --> FullyConnected --> Softmax
:param shape: -- shape of the images of the dataset
:return: result -- a Model() instance in Keras
"""
result = Sequential()
# Layer 1
result.add(
Conv2D(filters=6,
kernel_size=5,
strides=1,
activation='relu',
input_shape=shape,
name='convolution_1'))
result.add(MaxPooling2D(pool_size=2, strides=2, name='max_pool_1'))
# Layer 2
result.add(
Conv2D(filters=16,
kernel_size=5,
strides=1,
activation='relu',
name='convolution_2'))
result.add(MaxPooling2D(pool_size=2, strides=2, name='max_pool_2'))
# Layer 3
result.add(Flatten(name='flatten'))
result.add(Dense(units=120, activation='relu', name='fully_connected_1'))
# Layer 4
result.add(Dense(units=84, activation='relu', name='fully_connected_2'))
# Output
result.add(Dense(units=10, activation='softmax', name='output'))
result._name = 'lenet5'
return result
def OptmizedCNNv1(input_shape=(28, 28, 1), classes=10):
# Implementation of optimized CNN:
# Modified Architecture -- ConvNet --> MaxPool --> ConvNet --> MaxPool --> (Flatten) --> FullyConnected --> FullyConnected --> Softmax
model = Sequential([
# Layer 1
Conv2D(
filters=6,
kernel_size=3,
strides=1,
activation="relu",
input_shape=(28, 28, 1),
name="convolution1",
),
MaxPooling2D(pool_size=2, strides=2, name="maxpool1"),
# Layer 2
Conv2D(
filters=16,
kernel_size=3,
strides=1,
activation="relu",
name="convolution2",
),
MaxPooling2D(pool_size=2, strides=2, name="maxpool2"),
# Layer 3
Flatten(name="flatten"),
Dense(units=120, activation="relu", name="fullyconnected1"),
# Layer 4
Dense(units=84, activation="relu", name="fullyconnected2"),
# Output
Dense(units=10, activation="softmax", name="output"),
])
model._name = "OptimizedCNNv1"
return model
def init_convlrs_model(conv_size, lrs_size, lrs_levels, conv_time=False, lrs_diff=True, lrs_time=True, recursive_tensors=True, name_only=False, input_shape=None, num_classes=None, lrs_norm='BN'):
conv_name = 'TConv' if conv_time else 'Conv'
lrs_time_tag = 'T' if lrs_time else ''
lrs_diff_tag = 'D' if lrs_diff else ''
lrs_recurrent_tag = 'R' if recursive_tensors else ''
lrs_name = '{}{}{}LRS{}'.format(lrs_time_tag, lrs_diff_tag, lrs_recurrent_tag, lrs_norm)
model_name = '{}{}_H{}_W{}'.format(conv_name, lrs_name, conv_size, lrs_size)
if name_only: return model_name
num_sig_layers = 3
model = Sequential()
model.add(InputLayer(input_shape=input_shape))
if conv_time:
model.add(Time())
model.add(Conv1D(conv_size, 8, padding='same', kernel_initializer='he_uniform'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
if conv_time:
model.add(Time())
model.add(Conv1D(conv_size, 5, padding='same' , kernel_initializer='he_uniform'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
if conv_time:
model.add(Time())
model.add(Conv1D(conv_size, 3, padding='same', kernel_initializer='he_uniform'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
for i in range(num_sig_layers-1):
# model.add(Conv1D(conv_size, 3, padding='same', kernel_initializer='he_uniform'))
# model.add(BatchNormalization(axis=-1))
if lrs_time:
model.add(Time())
if lrs_diff:
model.add(Difference())
model.add(LRS(lrs_size, lrs_levels, return_sequences=True, recursive_tensors=recursive_tensors))
model.add(Reshape((input_shape[0]-i-1, lrs_levels, lrs_size,)))
if lrs_norm == 'BN':
model.add(BatchNormalization(axis=[-2]))
elif lrs_norm == 'TBN':
model.add(BatchNormalization(axis=[-2, -1]))
elif lrs_norm == 'LN':
model.add(LayerNormalization(axis=[-3, -1]))
# model.add(LayerNormalization(axis=[1, 3]))
#model.add(LayerNormalization(axis=[1, 3]))
model.add(Reshape((input_shape[0]-i-1, lrs_levels * lrs_size,)))
# model.add(Activation('relu'))
# if renorm:
# model.add(BatchNormalization(axis=[1, 2], renorm=True, center=False, scale=False))
# else:
#
# model.add(Conv1D(conv_size, 3, padding='same', kernel_initializer='he_uniform'))
# model.add(BatchNormalization(axis=-1))
if lrs_time:
model.add(Time())
if lrs_diff:
model.add(Difference())
model.add(LRS(lrs_size, lrs_levels, return_sequences=False, recursive_tensors=recursive_tensors))
# model.add(BatchNormalization(axis=1, center=False, scale=False))
model.add(Reshape((lrs_levels, lrs_size,)))
# model.add(LayerNormalization(axis=[2]))
model.add(BatchNormalization(axis=1))
model.add(Reshape((lrs_levels * lrs_size,)))
# model.add(BatchNormalization(axis=-1, renorm=renorm, center=False, scale=False))
# model.add(tf.keras.layers.Dropout(0.5))
# if gap:
# model.add(GlobalAveragePooling1D())
# else:
# model.add(Lambda(lambda X: X[:, -1]))
model.add(Dense(num_classes, activation='softmax'))
model._name = model_name
return model
def init_convlrs2_model(input_shape,
num_levels,
num_hidden,
num_classes,
difference=True,
add_time=True,
recursive_tensors=True,
reverse=False,
layer_norm=False):
num_sig_layers = 3
num_sig_hidden = int(num_hidden / num_levels)
# num_levels = 3
model = Sequential()
model.add(InputLayer(input_shape=input_shape))
if add_time:
model.add(Time())
model.add(Conv1D(num_hidden, 8, padding='same'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
if add_time:
model.add(Time())
model.add(Conv1D(num_hidden, 5, padding='same'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
if add_time:
model.add(Time())
model.add(Conv1D(num_hidden, 3, padding='same'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
for i in range(num_sig_layers - 1):
if add_time:
model.add(Time())
if difference:
model.add(Difference())
model.add(
LRS(num_sig_hidden,
num_levels,
return_sequences=True,
reverse=reverse,
recursive_tensors=recursive_tensors))
model.add(Reshape((
input_shape[0],
num_levels,
num_sig_hidden,
)))
if layer_norm:
model.add(LayerNormalization(axis=-1, center=False, scale=False))
else:
model.add(
BatchNormalization(axis=[1, 2], center=False, scale=False))
model.add(Reshape((
input_shape[0],
num_sig_hidden * num_levels,
)))
if add_time:
model.add(Time())
if difference:
model.add(Difference())
model.add(
LRS(num_sig_hidden,
num_levels,
reverse=reverse,
recursive_tensors=recursive_tensors))
model.add(Reshape((
num_sig_hidden,
num_levels,
)))
if layer_norm:
model.add(LayerNormalization(axis=-1, center=False, scale=False))
else:
model.add(BatchNormalization(axis=1, center=False, scale=False))
model.add(Reshape((num_sig_hidden * num_levels, )))
model.add(Dense(num_classes, activation='softmax'))
diff_tag = 'D' if difference else ''
norm_tag = 'LN' if layer_norm else 'BN'
model._name = 'Conv{}LRS{}2_M{}_H{}'.format(diff_tag, norm_tag, num_levels,
num_hidden)
return model
def init_ls2t_model(preprocess_size,
ls2t_size,
ls2t_order,
ls2t_depth,
preprocess='conv',
preprocess_time=True,
ls2t_diff=True,
ls2t_time=True,
recursive_tensors=True,
name_only=False,
input_shape=None,
num_classes=None):
preprocess = preprocess.lower()
if preprocess != 'conv' and preprocess != 'dense':
preprocess_name = ''
else:
preprocess_name = 'Conv' if preprocess == 'conv' else 'Dense'
ls2t_recurrent_tag = 'R' if recursive_tensors else ''
ls2t_name = 'LS2T{}'.format(ls2t_recurrent_tag)
model_name = '{}{}_H{}_N{}_M{}_D{}'.format(preprocess_name, ls2t_name,
preprocess_size, ls2t_size,
ls2t_order, ls2t_depth)
if name_only: return model_name
model = Sequential()
model.add(InputLayer(input_shape=input_shape))
model.add(Masking(mask_value=0.))
if preprocess_name != '':
if preprocess_time:
model.add(seq2tens.layers.Time())
if preprocess == 'conv':
model.add(
Conv1D(preprocess_size,
8,
padding='same',
kernel_initializer='he_uniform'))
else:
model.add(
TimeDistributed(
Dense(preprocess_size, kernel_initializer='he_uniform')))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
if preprocess_time:
model.add(seq2tens.layers.Time())
if preprocess == 'conv':
model.add(
Conv1D(preprocess_size,
5,
padding='same',
kernel_initializer='he_uniform'))
else:
model.add(
TimeDistributed(
Dense(preprocess_size, kernel_initializer='he_uniform')))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
if preprocess_time:
model.add(seq2tens.layers.Time())
if preprocess == 'conv':
model.add(
Conv1D(preprocess_size,
3,
padding='same',
kernel_initializer='he_uniform'))
else:
model.add(Dense(preprocess_size, kernel_initializer='he_uniform'))
model.add(BatchNormalization(axis=-1))
model.add(Activation('relu'))
for i in range(ls2t_depth - 1):
if ls2t_time:
model.add(seq2tens.layers.Time())
if ls2t_diff:
model.add(seq2tens.layers.Difference())
model.add(
seq2tens.layers.LS2T(ls2t_size,
ls2t_order,
return_sequences=True,
recursive_weights=recursive_tensors))
model.add(Reshape((
input_shape[0],
ls2t_order,
ls2t_size,
)))
model.add(BatchNormalization(axis=[-2]))
model.add(Reshape((
input_shape[0],
ls2t_order * ls2t_size,
)))
if ls2t_time:
model.add(seq2tens.layers.Time())
if ls2t_diff:
model.add(seq2tens.layers.Difference())
model.add(
seq2tens.layers.LS2T(ls2t_size,
ls2t_order,
return_sequences=False,
recursive_weights=recursive_tensors))
model.add(Reshape((
ls2t_order,
ls2t_size,
)))
model.add(BatchNormalization(axis=1))
model.add(Reshape((ls2t_order * ls2t_size, )))
model.add(Dense(num_classes, activation='softmax'))
model._name = model_name
return model
