def build_generator(x, nb_rows, nb_cols):
""" Generator sub-component for the CaloGAN
Args:
-----
x: a tensorflow.keras Input with shape (None, latent_dim)
nb_rows: int, number of desired output rows
nb_cols: int, number of desired output cols
Returns:
--------
a tensorflow.keras tensor with the transformation applied
"""
x = Dense((nb_rows + 2) * (nb_cols + 2) * 36)(x)
x = Reshape((nb_rows + 2, nb_cols + 2, 36))(x)
x = Conv2D(16, (2, 2), padding='same', kernel_initializer='he_uniform')(x)
x = LeakyReLU()(x)
x = BatchNormalization()(x)
x = LocallyConnected2D(6, (2, 2), kernel_initializer='he_uniform')(x)
x = LeakyReLU()(x)
x = LocallyConnected2D(1, (2, 2),
use_bias=False,
kernel_initializer='glorot_normal')(x)
return x
def use_deepface(self):
"""
Define DeepFace model.
----------
- Returns
face: Functional
DeepFace model structure.
"""
Input_1 = Input(shape=(152, 152, 3), name='face_Input')
Conv_1 = Conv2D(32, (11, 11), activation='relu',
name='face_Conv_1')(Input_1)
MaxPool_1 = MaxPooling2D(pool_size=(3, 3),
strides=(2, 2),
padding='same',
name='face_MaxPool_1')(Conv_1)
Conv_2 = Conv2D(16, (9, 9), activation='relu',
name='face_Conv_2')(MaxPool_1)
LC_1 = LocallyConnected2D(16, (9, 9),
activation='relu',
name='face_LC_1')(Conv_2)
LC_2 = LocallyConnected2D(16, (7, 7),
strides=(2, 2),
activation='relu',
name='face_LC_2')(LC_1)
LC_3 = LocallyConnected2D(16, (5, 5),
activation='relu',
name='face_LC_3')(LC_2) # output 21, 21, 16
Flatten_1 = Flatten(name='face_Flatten')(LC_3) # output 7056
FC_1 = Dense(4096, activation='relu')(Flatten_1)
Drop_1 = Dropout(0.5)(FC_1)
FC_2 = Dense(8631, activation='softmax')(Drop_1)
face = Model(inputs=Input_1, outputs=FC_2)
return face
def make_lcn():
in2 = Input(shape=(
36,
19,
30,
), name="in2", dtype="float32")
F = LocallyConnected2D(name="layerF",
filters=30,
kernel_size=(3, 2),
padding='valid',
input_shape=(36, 19, 30),
data_format='channels_last',
activation='relu',
use_bias=True)(in2)
G = LocallyConnected2D(name="layerG",
filters=25,
kernel_size=(3, 2),
padding='valid',
input_shape=(34, 18, 30),
data_format='channels_last',
activation='relu',
use_bias=True)(F)
H = LocallyConnected2D(name="layerH",
filters=20,
kernel_size=(3, 2),
padding='valid',
input_shape=(32, 17, 25),
data_format='channels_last',
activation='relu',
use_bias=True)(G)
I = LocallyConnected2D(name="layerI",
filters=15,
kernel_size=(3, 2),
padding='valid',
input_shape=(30, 16, 20),
data_format='channels_last',
activation='relu',
use_bias=True)(H)
J = LocallyConnected2D(name="layerJ",
filters=10,
kernel_size=(3, 2),
padding='valid',
input_shape=(28, 15, 15),
data_format='channels_last',
activation='relu',
use_bias=True)(I)
model = Model(in2, outputs=J)
metrics_to_output = ['accuracy']
model.compile(loss='mean_squared_error',
optimizer='adam',
metrics=metrics_to_output)
return model
def loadModel(
url='https://github.com/swghosh/DeepFace/releases/download/weights-vggface2-2d-aligned/VGGFace2_DeepFace_weights_val-0.9034.h5.zip'
):
base_model = Sequential()
base_model.add(
Convolution2D(32, (11, 11),
activation='relu',
name='C1',
input_shape=(152, 152, 3)))
base_model.add(
MaxPooling2D(pool_size=3, strides=2, padding='same', name='M2'))
base_model.add(Convolution2D(16, (9, 9), activation='relu', name='C3'))
base_model.add(LocallyConnected2D(16, (9, 9), activation='relu',
name='L4'))
base_model.add(
LocallyConnected2D(16, (7, 7), strides=2, activation='relu',
name='L5'))
base_model.add(LocallyConnected2D(16, (5, 5), activation='relu',
name='L6'))
base_model.add(Flatten(name='F0'))
base_model.add(Dense(4096, activation='relu', name='F7'))
base_model.add(Dropout(rate=0.5, name='D0'))
base_model.add(Dense(8631, activation='softmax', name='F8'))
#---------------------------------
home = str(Path.home())
if os.path.isfile(
home + '/.deepface/weights/VGGFace2_DeepFace_weights_val-0.9034.h5'
) != True:
print("VGGFace2_DeepFace_weights_val-0.9034.h5 will be downloaded...")
output = home + '/.deepface/weights/VGGFace2_DeepFace_weights_val-0.9034.h5.zip'
gdown.download(url, output, quiet=False)
#unzip VGGFace2_DeepFace_weights_val-0.9034.h5.zip
with zipfile.ZipFile(output, 'r') as zip_ref:
zip_ref.extractall(home + '/.deepface/weights/')
base_model.load_weights(
home + '/.deepface/weights/VGGFace2_DeepFace_weights_val-0.9034.h5')
#drop F8 and D0. F7 is the representation layer.
deepface_model = Model(inputs=base_model.layers[0].input,
outputs=base_model.layers[-3].output)
return deepface_model
def build_decoder(size,
encoded_shape,
in_channels=1,
latent_dim=8,
dump=False):
"""Create decoder from latent space back to grid."""
latent_inputs = Input(shape=(latent_dim, ), name='z_sampling')
x = Dense(encoded_shape[1] * encoded_shape[2] * encoded_shape[3],
activation=relu)(latent_inputs)
x = Reshape((encoded_shape[1], encoded_shape[2], encoded_shape[3]))(x)
x = LocallyConnected2D(32, (3, 3))(x)
x = ZeroPadding2D(padding=(1, 1))(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(32, (3, 3), activation=relu, padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, (3, 3), activation=relu, padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(16, (3, 3), activation=relu, padding='same')(x)
x = UpSampling2D((2, 2))(x)
x = Conv2D(8, (3, 3), activation=relu, padding='same')(x)
x = UpSampling2D((2, 2))(x)
# can't use Activation(sigmoid)(x) because we need a conv layer to reduce 4-channels to 1-channel
decoded_layer = Conv2D(1, (3, 3), activation=relu, padding='same')(x)
decoder = Model(latent_inputs, decoded_layer, name='vae_decoder')
if dump:
decoder.summary()
plot_model(decoder, to_file='dicom_decoder.png', show_shapes=True)
return decoder
def deepface(input_shape=None):
input_data = Input(shape=input_shape)
conv1 = Conv2D(32, (11, 11), activation='relu', name='C1')(input_data)
maxpool1 = MaxPooling2D(pool_size=3, strides=2, padding='same',
name='M2')(conv1)
conv2 = Conv2D(16, (9, 9), activation='relu', name='C3')(maxpool1)
lconv1 = LocallyConnected2D(16, (9, 9), activation='relu',
name='L4')(conv2)
lconv2 = LocallyConnected2D(16, (7, 7),
strides=2,
activation='relu',
name='L5')(lconv1)
lconv3 = LocallyConnected2D(16, (5, 5), activation='relu',
name='L6')(lconv2)
flat = Flatten(name='F0')(lconv3)
fc1 = Dense(4096, activation='relu', name='F7')(flat)
drop = Dropout(rate=0.5, name='D0')(fc1)
fc2 = Dense(8631, activation='softmax', name='F8')(drop)
model = Model(inputs=input_data, outputs=fc2, name="DeepFace")
return model
def attention_map_model(backbone):
"""
Add attention map branch to CNN
use DenseNet121, MobileNet as backbone
Example:
backbone = DenseNet121(weights='imagenet', include_top=False)
backbone = MobileNet(weights='imagenet', include_top=False)
"""
in_layer = Input((80,80,2)) # ???? dimesnions question ????
reshape_layer = Conv2D(3, (3,3), activation='relu', padding='same', input_shape=(80,80,2))
for layer in backbone.layers:
layer.trainable = True
reshape_ = reshape_layer(in_layer)
pt_depth = backbone.get_output_shape_at(0)[-1]
pt_features = backbone(reshape_)
bn_features = BatchNormalization()(pt_features)
# attention mech to turn pixels in GAP on/ off
attn_layer = Conv2D(64, (1,1), padding='same', activation='relu')(bn_features)
attn_layer = Conv2D(64, (1,1), padding='same', activation='relu')(attn_layer)
attn_layer = LocallyConnected2D(1, (1,1), padding='valid', activation='sigmoid')(attn_layer)
# insert it to backbone branch
# initialize weights
up_c2_w = np.ones((1, 1, 1, pt_depth))
up_c2 = Conv2D(pt_depth, (1,1), padding='same', activation='linear', use_bias=False, weights=[up_c2_w])
up_c2.trainable = False
attn_layer = up_c2(attn_layer)
# get together attn_layer and bn_features branches
mask_features = multiply([attn_layer, bn_features])
gap_features = GlobalAveragePooling2D()(mask_features)
gap_mask = GlobalAveragePooling2D()(attn_layer)
# account for missing values from attention model
gap = Lambda(lambda x: x[0]/x[1], name='RescaleGAP')([gap_features, gap_mask])
gap_dr = Dropout(0.5)(gap)
dr_steps = Dropout(0.25)(Dense(1024, activation='elu')(gap_dr))
# linear 16 bit
out_layer = Dense(1, activation='sigmoid')(dr_steps)
attn_model = models.Model(inputs=[in_layer], outputs=[out_layer])
return attn_model
def inpainting_attention(primary, carryover, constant=-10):
def _initialize_bias(const=-5):
def _(shape, dtype=None):
assert len(shape) == 3, 'must be a 3D shape'
x = np.zeros(shape)
x[:, :, -1] = const
return x
return _
x = concatenate([primary, carryover], axis=-1)
h = ZeroPadding2D((1, 1))(x)
lcn = LocallyConnected2D(filters=2,
kernel_size=(3, 3),
bias_initializer=_initialize_bias(constant))
h = lcn(h)
weights = Lambda(channel_softmax)(h)
channel_sum = Lambda(K.sum, arguments={'axis': -1, 'keepdims': True})
return channel_sum(multiply([x, weights]))
def build_encoded_layer(input_img, l1_l2=(0.0e-4, 0.0e-4), use_dropout=True):
"""Create encoded layer, prior to projection to latent space."""
x = Conv2D(8, (3, 3), activation=relu, padding='same')(input_img)
x = MaxPooling2D((2, 2), padding='same')(x)
if use_dropout:
x = SpatialDropout2D(0.1)(x)
x = Conv2D(16, (3, 3), activation=relu, padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(16, (3, 3), activation=relu, padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = Conv2D(32, (3, 3), activation=relu, padding='same')(x)
x = MaxPooling2D((2, 2), padding='same')(x)
x = LocallyConnected2D(32, (3, 3))(x)
x = ZeroPadding2D(padding=(1, 1))(x)
x = MaxPooling2D((2, 2), padding='same')(x)
l1, l2 = l1_l2
encoded_layer = ActivityRegularization(l1=l1, l2=l2)(x)
return encoded_layer
# Phase 1: in2 -> ABCDE pre = Reshape( target_shape=(36, 19, 27,) )( input2 )#ALWAYS DO WIDTH, HEIGHT, CHANNELS A = Conv2D( name="layerA", filters=30, kernel_size=(1,1), padding='valid', input_shape=(36, 19, 27), data_format='channels_last', activation='relu', use_bias=True )( pre ) B = Conv2D( name="layerB", filters=30, kernel_size=(1,1), padding='valid', input_shape=A.shape, data_format='channels_last', activation='relu', use_bias=True )( A ) C = Conv2D( name="layerC", filters=30, kernel_size=(1,1), padding='valid', input_shape=B.shape, data_format='channels_last', activation='relu', use_bias=True )( B ) D = Conv2D( name="layerD", filters=30, kernel_size=(1,1), padding='valid', input_shape=C.shape, data_format='channels_last', activation='relu', use_bias=True )( C ) E = Conv2D( name="layerE", filters=30, kernel_size=(1,1), padding='valid', input_shape=D.shape, data_format='channels_last', activation='relu', use_bias=True )( D ) # Phase 2: FGHIJ E2 = tensorflow.keras.layers.concatenate( [E, E], name="merge_E_E", axis=1 ) #print( E2.shape ) #exit( 0 ) F = LocallyConnected2D( name="layerF", filters=20, kernel_size=(4,2), strides=(2,1), padding='valid', input_shape=E2.shape, data_format='channels_last', activation='relu', use_bias=True )( E2 ) G = LocallyConnected2D( name="layerG", filters=14, kernel_size=(3,2), strides=(1,1), padding='valid', input_shape=F.shape, data_format='channels_last', activation='relu', use_bias=True )( F ) H = LocallyConnected2D( name="layerH", filters=13, kernel_size=(3,2), strides=(1,1), padding='valid', input_shape=G.shape, data_format='channels_last', activation='relu', use_bias=True )( G ) I = LocallyConnected2D( name="layerI", filters=12, kernel_size=(3,2), strides=(1,1), padding='valid', input_shape=H.shape, data_format='channels_last', activation='relu', use_bias=True )( H ) J = LocallyConnected2D( name="layerJ", filters=10, kernel_size=(3,2), strides=(1,1), padding='valid', input_shape=I.shape, data_format='channels_last', activation='relu', use_bias=True )( I ) #print( Ia.shape ) #exit( 0 ) # Phase 3: flatJ, merge, KLMN, output flatJ = Flatten( name="flatJ", data_format='channels_last' )( J ) print( flatJ.shape )
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(pre)
B1 = Conv2D(name="layerB1",
filters=10,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(A)
B2 = LocallyConnected2D(name="layerB2",
filters=5,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(B1)
B3 = LocallyConnected2D(name="layerB3",
filters=5,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(B2)
merge = tensorflow.keras.layers.concatenate([B3, in1up], name="merge", axis=-1)
# Phase 2: FGHIJ
C = LocallyConnected2D(name="layerC",
filters=3,
def buildNet(inputShape, numUniqueClasses):
"""
:param inputShape: The shape of the network input. The first element is
always set to None to allow for arbitrary batch sizes.
:param numUniqueClasses: Number of different classes.
"""
model_input = Input(shape=inputShape, name='model_input')
layers = ResidualLayer(model_input,
name='res1',
num_filters=8,
filter_size=(3, 1))
# First spatial reduction
layers = ResidualLayer(layers,
name='res2',
num_filters=8,
filter_size=(3, 1),
stride=(5, 1))
layers = ResidualLayer(layers,
name='res3',
num_filters=8,
filter_size=(3, 1))
# Second spatial reduction
layers = ResidualLayer(layers,
name='res4',
num_filters=1,
filter_size=(3, 1),
stride=(3, 1))
layers = Activation(activation='relu', name='act1')(layers)
layers = Dropout(rate=0.3, name='dropout1')(layers)
layers = ZeroPadding2D((2, 0))(layers)
# Note: deleted the bias and moved the order of layer-BN-NL to match
# the original lasagne version, which was done with the batch_norm method.
layers = LocallyConnected2D(filters=1,
kernel_size=(5, 1),
kernel_initializer='he_uniform',
activation='linear',
use_bias=False,
name='loc_con1')(layers)
layers = BatchNormalization(name='bn1')(layers)
layers = Activation(activation='relu', name='act_after_loc')(layers)
layers = Flatten(name='flatten1')(layers)
layers = Dense(units=numUniqueClasses,
activation='linear',
name='model_logits')(layers)
model_output = Activation(activation='softmax',
name='model_output')(layers)
model = Model(inputs=model_input, outputs=model_output)
return model
input_shape=(36, 19, 30),
data_format='channels_last',
activation='relu',
use_bias=True)(D)
# Phase 2: FGHIJ
#Ea = UpSampling2D( name="layerEa", size=(2, 1), data_format='channels_last' )( E )
Ea = tensorflow.keras.layers.concatenate([E, E], name="merge_E_E", axis=1)
#print( Ea.shape )
#exit( 0 )
F = LocallyConnected2D(name="layerF",
filters=30,
strides=(3, 1),
kernel_size=(4, 2),
padding='valid',
input_shape=(72, 19, 30),
data_format='channels_last',
activation='relu',
use_bias=True)(Ea)
Fa = tensorflow.keras.layers.concatenate([F, F], name="merge_F_F", axis=1)
G = LocallyConnected2D(name="layerG",
filters=25,
strides=(3, 1),
kernel_size=(4, 2),
padding='valid',
input_shape=(46, 18, 30),
data_format='channels_last',
activation='relu',
use_bias=True)(Fa)
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(pre)
A = LeakyReLU()(A)
B1 = Conv2D(name="layerB1",
filters=8,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(A)
B1 = LeakyReLU()(B1)
B4 = LocallyConnected2D(name="layerB4",
filters=4,
kernel_size=(2, 3),
strides=(2, 2),
padding='valid',
data_format='channels_last',
activation=None,
use_bias=True)(B1)
X = Conv2D(name="layerX",
filters=8,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(pre)
X = LeakyReLU()(X)
Y = LocallyConnected2D(name="layerY",
filters=6,
kernel_size=(1, 1),
padding='valid',
def main():
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
###################### LOADING DATA #############################
print("\n--------------- MY DATA -----------------------\n")
# Load training data into an array
train_array = np.loadtxt("zip.train")
# extracting the levels from the training array
train_levels_data = train_array[0:7291, 0:1]
# convert the class vector to binary class matrix
# as many columns as there are classes
train_levels = to_categorical(train_levels_data)
# extracting the features from the training array
train_features = train_array[0:7291, 1:257]
# Get total number of training data
total_train = len(train_features)
# Load testing data into an array
test_array = np.loadtxt("zip.test")
# extract test levels from test_array, for a sample of data
test_levels = test_array[0:2007, 0:1]
# extract teat features from test_array, for a sample of data
test_features = test_array[0:2007, 1:257]
# Get total number of test data
total_test = len(test_features)
# total size of all data
total_val = total_test + total_train
# View the data shape
# print(test_features.shape)
# Output is (2007, 257)
# this size is not given in the book
batch_size = 256
# figure had 30 training epochs
start_epoch = 0
num_epochs = 30
interval = 1
# 16 x 16 greyscale images
IMG_HEIGHT = 16
IMG_WIDTH = 16
# Net-2 Archetecture (two layers dense)
model_two = Sequential([
Dense(12, input_dim=256, activation='sigmoid'),
Dense(10, input_dim=256, activation='sigmoid')
])
# Net-3 Archetecture (two local and one Dense)
model_three = Sequential([
LocallyConnected2D(64, (3, 3),
input_shape=(16, 16, 1),
activation='sigmoid'),
LocallyConnected2D(16, (5, 5), activation='sigmoid'),
Flatten(),
Dense(10, activation='sigmoid')
])
# Net-4 Archetecture (conv2d locally connected, and dense last layer)
model_four = Sequential([
Conv2D(128, (3, 3), input_shape=(16, 16, 1), activation='sigmoid'),
LocallyConnected2D(16, (5, 5), activation='sigmoid'),
Flatten(),
Dense(10, activation='sigmoid')
])
# Net-5 Archetecture (conv2d, dense layer)
model_five = Sequential([
Conv2D(128, (3, 3), input_shape=(16, 16, 1), activation='sigmoid'),
LocallyConnected2D(16, (5, 5), activation='sigmoid'),
Flatten(),
Dense(10, activation='sigmoid')
])
# currently only trying toget the network one correct percent on test data
# archetecture seems tight but isnt giving vallues shown in the figure
net1_correct = net_one(train_levels, train_features, test_levels,
test_features, num_epochs, batch_size)
graph_properties = np.arange(start_epoch, num_epochs, interval)
plt.figure(figsize=(8, 8))
plt.subplot(2, 1, 1)
plt.plot(graph_properties, net1_correct, label='Net-1')
plt.xlabel("Training Epochs")
plt.ylabel("% Correct on Test Data")
plt.legend(loc='lower right')
plt.show()
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(pre)
A = LeakyReLU()(A)
B1 = Conv2D(name="layerB1",
filters=10,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(A)
B1 = LeakyReLU()(B1)
B4 = LocallyConnected2D(name="layerB4",
filters=5,
kernel_size=(4, 3),
strides=(4, 2),
padding='valid',
data_format='channels_last',
activation=None,
use_bias=True)(B1)
C = Add()([B4, in1merge])
C = LeakyReLU()(C)
E = LocallyConnected2D(name="layerE",
filters=4,
kernel_size=(3, 3),
strides=(3, 3),
padding='valid',
data_format='channels_last',
use_bias=True)(C)
E = LeakyReLU()(E)
'''
batch_size_i_array = np.logspace(0, 4, 10)
make_x_particles = False
loops = 100
''' Create the network. '''
latent_size = 200
loc = Sequential([
Dense(128 * 7 * 7, input_dim=latent_size),
Reshape((7, 7, 128)),
Conv2D(64, (5, 5), padding='same', kernel_initializer='he_uniform'),
LeakyReLU(),
BatchNormalization(),
UpSampling2D(size=(2, 2), interpolation='bilinear'),
ZeroPadding2D((2, 2)),
LocallyConnected2D(6, (5, 5), kernel_initializer='he_uniform'),
LeakyReLU(),
BatchNormalization(),
UpSampling2D(size=(2, 2), interpolation='bilinear'),
LocallyConnected2D(6, (3, 3), kernel_initializer='he_uniform'),
LeakyReLU(),
LocallyConnected2D(1, (2, 2),
use_bias=False,
kernel_initializer='glorot_normal'),
Activation('relu')
])
latent = Input(shape=(latent_size, ))
image_class = Input(shape=(1, ), dtype='int32')
emb = Flatten()(Embedding(2,
def build_discriminator(image, mbd=False, sparsity=False, sparsity_mbd=False):
""" Generator sub-component for the CaloGAN
Args:
-----
image: tensorflow.keras tensor of 4 dimensions (i.e. the output of one calo layer)
mdb: bool, perform feature level minibatch discrimination
sparsiry: bool, whether or not to calculate and include sparsity
sparsity_mdb: bool, perform minibatch discrimination on the sparsity
values in a batch
Returns:
--------
a tensorflow.keras tensor of features
"""
x = Conv2D(64, (2, 2), padding='same')(image)
x = LeakyReLU()(x)
x = ZeroPadding2D((1, 1))(x)
x = LocallyConnected2D(16, (3, 3), padding='valid', strides=(1, 2))(x)
x = LeakyReLU()(x)
x = BatchNormalization()(x)
x = ZeroPadding2D((1, 1))(x)
x = LocallyConnected2D(8, (2, 2), padding='valid')(x)
x = LeakyReLU()(x)
x = BatchNormalization()(x)
x = ZeroPadding2D((1, 1))(x)
x = LocallyConnected2D(8, (2, 2), padding='valid', strides=(1, 2))(x)
x = LeakyReLU()(x)
x = BatchNormalization()(x)
x = Flatten()(x)
if mbd or sparsity or sparsity_mbd:
minibatch_featurizer = Lambda(minibatch_discriminator,
output_shape=minibatch_output_shape)
features = [x]
nb_features = 10
vspace_dim = 10
# creates the kernel space for the minibatch discrimination
if mbd:
K_x = Dense3D(nb_features, vspace_dim)(x)
features.append(Activation('tanh')(minibatch_featurizer(K_x)))
if sparsity or sparsity_mbd:
sparsity_detector = Lambda(sparsity_level, sparsity_output_shape)
empirical_sparsity = sparsity_detector(image)
if sparsity:
features.append(empirical_sparsity)
if sparsity_mbd:
K_sparsity = Dense3D(nb_features,
vspace_dim)(empirical_sparsity)
features.append(
Activation('tanh')(minibatch_featurizer(K_sparsity)))
return concatenate(features)
else:
return x
activation='relu',
use_bias=True)(B1)
B3 = Conv2D(name="layerB3",
filters=15,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(B2)
merge = tensorflow.keras.layers.concatenate([B3, in1up], name="merge", axis=-1)
# Phase 2: FGHIJ
C = LocallyConnected2D(name="layerC",
filters=12,
kernel_size=(1, 1),
strides=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(merge)
D = LocallyConnected2D(name="layerD",
filters=10,
kernel_size=(1, 1),
strides=(1, 1),
padding='valid',
data_format='channels_last',
activation='relu',
use_bias=True)(C)
E = LocallyConnected2D(name="layerE",
filters=10,
kernel_size=(1, 1),
strides=(1, 1),
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(pre)
A = LeakyReLU()(A)
B1 = Conv2D(name="layerB1",
filters=8,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(A)
B1 = LeakyReLU()(B1)
B4 = LocallyConnected2D(name="layerB4",
filters=merge_depth,
kernel_size=(2, 3),
strides=(2, 2),
padding='valid',
data_format='channels_last',
activation=None,
use_bias=True)(B1)
X1a = Conv2D(name="layerX1a",
filters=4,
kernel_size=(1, 1),
padding='valid',
data_format='channels_last',
use_bias=True)(pre)
X1a = LeakyReLU()(X1a)
Y1a = LocallyConnected2D(name="layerY1a",
filters=4,
kernel_size=(1, 1),
padding='valid',
num_input_dimensions2 = 18494 - num_input_dimensions1
input1 = Input(shape=(num_input_dimensions1, ), name="in1", dtype="float32")
#input2 = Input(shape=(36, 19, 27,), name="in2", dtype="float32" )
input2 = Input(shape=(18468, ), name="in2", dtype="float32")
pre = Reshape(target_shape=(
36,
19,
27,
))(input2) #ALWAYS DO WIDTH, HEIGHT, CHANNELS
l1 = LocallyConnected2D(name="layer1",
filters=30,
kernel_size=(3, 2),
padding='valid',
input_shape=(36, 19, 27),
data_format='channels_last',
activation='relu',
use_bias=True)(pre)
l2 = LocallyConnected2D(name="layer2",
filters=25,
kernel_size=(3, 2),
padding='valid',
input_shape=(34, 18, 30),
data_format='channels_last',
activation='relu',
use_bias=True)(l1)
l3 = LocallyConnected2D(name="layer3",
filters=20,
kernel_size=(3, 2),
padding='valid',