def projection_block(self, x, strides=(2, 2), **metaparameters):
""" Construct a ResNeXT block with projection shortcut
x : input to the block
strides : whether entry convolution is strided (i.e., (2, 2) vs (1, 1))
filters_in : number of filters (channels) at the input convolution
filters_out: number of filters (channels) at the output convolution
cardinality: width of group convolution
"""
filters_in = metaparameters['filters_in']
filters_out = metaparameters['filters_out']
if 'cardinality' in metaparameters:
cardinality = metaparameters['cardinality']
else:
cardinality = self.cardinality
# Construct the projection shortcut
# Increase filters by 2X to match shape when added to output of block
shortcut = self.Conv2D(x,
filters_out, (1, 1),
strides=strides,
padding='same',
**metaparameters)
shortcut = self.BatchNormalization(shortcut)
# Dimensionality Reduction
x = self.Conv2D(x,
filters_in, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
**metaparameters)
x = self.BatchNormalization(x)
x = self.ReLU(x)
# Cardinality (Wide) Layer (split-transform)
filters_card = filters_in // cardinality
groups = []
for i in range(cardinality):
group = Lambda(lambda z: z[:, :, :, i * filters_card:i *
filters_card + filters_card])(x)
groups.append(
self.Conv2D(group,
filters_card, (3, 3),
strides=strides,
padding='same',
use_bias=False,
**metaparameters))
# Concatenate the outputs of the cardinality layer together (merge)
x = Concatenate()(groups)
x = self.BatchNormalization(x)
x = self.ReLU(x)
# Dimensionality restoration
x = self.Conv2D(x,
filters_out, (1, 1),
strides=(1, 1),
padding='same',
use_bias=False,
**metaparameters)
x = self.BatchNormalization(x)
# Identity Link: Add the shortcut (input) to the output of the block
x = Add()([shortcut, x])
x = self.ReLU(x)
return x
def khop_model_distribute10(): # input/output = num of sensors
sensor_matrix1 = Input(shape=(num_sensors+1, num_sensors+1))
sensor_matrix2 = Input(shape=(num_sensors+1, num_sensors+1))
#sensor_matrix3 = Input(shape=(num_sensors, num_sensors))
s_input1 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input2 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input3 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input4 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input5 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input6 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input7 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input8 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input9 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input10 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_cnn = sensor_cnn(input_shape, repetitions = [2,2,2,2])
extract_cnn1 = s_cnn(s_input1)
extract_cnn2 = s_cnn(s_input2)
extract_cnn3 = s_cnn(s_input3)
extract_cnn4 = s_cnn(s_input4)
extract_cnn5 = s_cnn(s_input5)
extract_cnn6 = s_cnn(s_input6)
extract_cnn7 = s_cnn(s_input7)
extract_cnn8 = s_cnn(s_input8)
extract_cnn9 = s_cnn(s_input9)
extract_cnn10 = s_cnn(s_input10)
extract_cnn = Concatenate(axis=1)([extract_cnn1, extract_cnn2, extract_cnn3,
extract_cnn4, extract_cnn5, extract_cnn6,
extract_cnn7, extract_cnn8, extract_cnn9, extract_cnn10])
#extract_cnn = np.reshape(extract_cnn, (-1,))
G_h1 = GraphConv(256, 'relu')([extract_cnn, sensor_matrix1])
G_h2 = GraphConv(256, 'relu')([extract_cnn, sensor_matrix2])
G_1 = Concatenate(axis=-1)([G_h1, G_h2])
G_2h1 = GraphConv(256, 'relu')([G_1, sensor_matrix1])
G_2h2 = GraphConv(256, 'relu')([G_1, sensor_matrix2])
G_2 = Concatenate(axis=-1)([G_2h1, G_2h2])
gnn_output = tf.split(G_2, num_sensors+1, 1)
mlp_layer = mlp_model()
output1 = mlp_layer(Flatten()(gnn_output[0]))
output2 = mlp_layer(Flatten()(gnn_output[1]))
output3 = mlp_layer(Flatten()(gnn_output[2]))
output4 = mlp_layer(Flatten()(gnn_output[3]))
output5 = mlp_layer(Flatten()(gnn_output[4]))
output6 = mlp_layer(Flatten()(gnn_output[5]))
output7 = mlp_layer(Flatten()(gnn_output[6]))
output8 = mlp_layer(Flatten()(gnn_output[7]))
output9 = mlp_layer(Flatten()(gnn_output[8]))
output10 = mlp_layer(Flatten()(gnn_output[9]))
model = Model(inputs=[s_input1, s_input2, s_input3, s_input4,
s_input5, s_input6, s_input7, s_input8, s_input9,s_input10,
sensor_matrix1, sensor_matrix2],
outputs= [output1,output2,output3,output4,
output5,output6,output7,output8,output9,output10])
return model
def predict_model(input_size, epochs=200, lr=1e-3):
inputs = Input(shape=input_size, name='inputs')
# Embedding input
wday_input = Input(shape=(1,), name='wday')
month_input = Input(shape=(1,), name='month')
year_input = Input(shape=(1,), name='year')
mday_input = Input(shape=(1,), name='mday')
quarter_input = Input(shape=(1,), name='quarter')
event_name_1_input = Input(shape=(1,), name='event_name_1')
event_type_1_input = Input(shape=(1,), name='event_type_1')
event_name_2_input = Input(shape=(1,), name='event_name_2')
event_type_2_input = Input(shape=(1,), name='event_type_2')
item_id_input = Input(shape=(1,), name='item_id')
dept_id_input = Input(shape=(1,), name='dept_id')
store_id_input = Input(shape=(1,), name='store_id')
cat_id_input = Input(shape=(1,), name='cat_id')
state_id_input = Input(shape=(1,), name='state_id')
snap_CA_input = Input(shape=(1,), name='snap_CA')
snap_TX_input = Input(shape=(1,), name='snap_TX')
snap_WI_input = Input(shape=(1,), name='snap_WI')
wday_emb = Flatten()(Embedding(7, 1)(wday_input))
month_emb = Flatten()(Embedding(12, 2)(month_input))
year_emb = Flatten()(Embedding(6, 1)(year_input))
mday_emb = Flatten()(Embedding(31, 2)(mday_input))
quarter_emb = Flatten()(Embedding(4, 1)(quarter_input))
event_name_1_emb = Flatten()(Embedding(31, 2)(event_name_1_input))
event_type_1_emb = Flatten()(Embedding(5, 1)(event_type_1_input))
event_name_2_emb = Flatten()(Embedding(5, 1)(event_name_2_input))
event_type_2_emb = Flatten()(Embedding(5, 1)(event_type_2_input))
item_id_emb = Flatten()(Embedding(3049, 4)(item_id_input))
dept_id_emb = Flatten()(Embedding(7, 1)(dept_id_input))
store_id_emb = Flatten()(Embedding(10, 1)(store_id_input))
cat_id_emb = Flatten()(Embedding(6, 1)(cat_id_input))
state_id_emb = Flatten()(Embedding(3, 1)(state_id_input))
x = Concatenate(-1)([inputs, wday_emb, month_emb, month_emb, year_emb, mday_emb, \
quarter_emb, event_name_1_emb, event_type_1_emb, event_name_2_emb, \
event_type_2_emb, item_id_emb, dept_id_emb, store_id_emb, cat_id_emb, \
state_id_emb])
x = Dense(1024, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(512, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(256, activation='relu')(x)
x = BatchNormalization()(x)
x = Dense(128, activation='relu')(x)
x_deep = Dense(64, activation='relu')(x)
x_deep = BatchNormalization()(x_deep)
x_deep = Dense(128, activation='relu')(x_deep)
x_deep = BatchNormalization()(x_deep)
x_deep = Dense(256, activation='relu')(x_deep)
x = Concatenate(-1)([inputs, x])
x_res = Dense(64, activation='relu')(x)
x_res = BatchNormalization()(x_res)
x_res = Dense(128, activation='relu')(x_res)
x_res = BatchNormalization()(x_res)
x_res = Dense(256, activation='relu')(x_res)
x = Concatenate(-1)([x_deep, x_res])
outputs = Dense(1, activation='sigmoid')(x)
# outputs = resnet_v2(x)
# optimizer = Adam(lr=lr)#Adam(lr=lr)
input_dic = {
'inputs': inputs, 'wday': wday_input, 'month': month_input, 'year': year_input,
'mday': mday_input, 'quarter': quarter_input, 'event_name_1': event_name_1_input,
'event_type_1': event_type_1_input, 'event_name_2': event_name_2_input,
'event_type_2': event_type_2_input, 'item_id': item_id_input, 'dept_id': dept_id_input,
'store_id': store_id_input, 'cat_id': cat_id_input, 'state_id': state_id_input,
}
model = Model(input_dic, outputs)#, name='predict_model')
return model
def __build_model(self, task, num_classes, nets, categorical_columns,
continuous_columns, var_len_categorical_columns, config):
logger.info(f'Building model...')
self.model_desc = ModelDesc()
categorical_inputs, continuous_inputs, var_len_categorical_inputs = \
self.__build_inputs(categorical_columns, continuous_columns, var_len_categorical_columns)
embeddings = self.__build_embeddings(categorical_columns,
categorical_inputs,
var_len_categorical_columns,
var_len_categorical_inputs,
config.embedding_dropout)
dense_layer = self.__build_denses(continuous_columns,
continuous_inputs,
config.dense_dropout)
flatten_emb_layer = None
if len(embeddings) > 0:
if len(embeddings) == 1:
flatten_emb_layer = Flatten(name='flatten_embeddings')(
embeddings[0])
else:
flatten_emb_layer = Flatten(name='flatten_embeddings')(
Concatenate(name='concat_embeddings_axis_0')(embeddings))
self.model_desc.nets = nets
self.model_desc.stacking = config.stacking_op
concat_emb_dense = self.__concat_emb_dense(flatten_emb_layer,
dense_layer)
# concat_emb_dense = flatten_emb_layer
outs = {}
for net in nets:
logit = deepnets.get(net)
out = logit(embeddings, flatten_emb_layer, dense_layer,
concat_emb_dense, self.config, self.model_desc)
if out is not None:
outs[net] = out
if len(outs) > 1:
logits = []
for name, out in outs.items():
if len(out.shape) > 2:
out = Flatten(name=f'flatten_{name}_out')(out)
if out.shape[-1] > 1:
logit = Dense(1,
use_bias=False,
activation=None,
name=f'dense_logit_{name}')(out)
else:
logit = out
logits.append(logit)
if config.stacking_op == consts.STACKING_OP_ADD:
x = Add(name='add_logits')(logits)
elif config.stacking_op == consts.STACKING_OP_CONCAT:
x = Concatenate(name='concat_logits')(logits)
else:
raise ValueError(
f'Unsupported stacking_op:{config.stacking_op}.')
elif (len(outs) == 1):
name, out = outs.popitem()
# out = list(outs.values())[0]
if len(out.shape) > 2:
out = Flatten(name=f'flatten_{name}_out')(out)
x = out
else:
raise ValueError(f'Unexcepted logit output.{outs}')
all_inputs = list(categorical_inputs.values()) + list(var_len_categorical_inputs.values()) + \
list(continuous_inputs.values())
output = self.__output_layer(x,
task,
num_classes,
use_bias=self.config.output_use_bias)
model = Model(inputs=all_inputs, outputs=output)
model = self.__compile_model(model, task, num_classes,
config.optimizer, config.loss,
config.metrics)
print(self.model_desc)
return model
model = MaxPooling2D(pool_size=(3, 3), strides=2, padding='same')(model) model = Conv2D(64, (1, 1), activation='relu')(model) model = Conv2D(192, (3, 3), padding='same', activation='relu')(model) model = BatchNormalization()(model) model = MaxPooling2D((3, 3), strides=2, padding='same')(model) # INCEPTION MODULE tower_1 = Conv2D(64, (1, 1), activation='relu')(model) tower_2 = Conv2D(96, (1, 1), activation='relu')(model) tower_2 = Conv2D(128, (3, 3), padding='same', activation='relu')(tower_2) tower_3 = Conv2D(16, (1, 1), activation='relu')(model) tower_3 = Conv2D(32, (5, 5), padding='same', activation='relu')(tower_3) tower_4 = MaxPooling2D((3, 3), strides=1, padding='same')(model) tower_4 = Conv2D(32, (1, 1), activation='relu')(tower_4) model = Concatenate(axis=-1)([tower_1, tower_2, tower_3, tower_4]) # INCEPTION MODULE tower_1 = Conv2D(128, (1, 1), activation='relu')(model) tower_2 = Conv2D(128, (1, 1), activation='relu')(model) tower_2 = Conv2D(192, (3, 3), padding='same', activation='relu')(tower_2) tower_3 = Conv2D(32, (1, 1), activation='relu')(model) tower_3 = Conv2D(96, (5, 5), padding='same', activation='relu')(tower_3) tower_4 = MaxPooling2D((3, 3), strides=1, padding='same')(model) tower_4 = Conv2D(64, (1, 1), activation='relu')(tower_4) model = Concatenate(axis=-1)([tower_1, tower_2, tower_3, tower_4]) model = MaxPooling2D((3, 3), strides=2, padding='same')(model) # INCEPTION MODULE tower_1 = Conv2D(192, (1, 1), activation='relu')(model) tower_2 = Conv2D(96, (1, 1), activation='relu')(model) tower_2 = Conv2D(208, (3, 3), padding='same', activation='relu')(tower_2)
def InceptionResNetV2():
inputs = Input(shape=(160, 160, 3))
x = Conv2D(32, 3, strides=2, padding='valid', use_bias=False, name= 'Conv2d_1a_3x3') (inputs)
x = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Conv2d_1a_3x3_BatchNorm')(x)
x = Activation('relu', name='Conv2d_1a_3x3_Activation')(x)
x = Conv2D(32, 3, strides=1, padding='valid', use_bias=False, name= 'Conv2d_2a_3x3') (x)
x = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Conv2d_2a_3x3_BatchNorm')(x)
x = Activation('relu', name='Conv2d_2a_3x3_Activation')(x)
x = Conv2D(64, 3, strides=1, padding='same', use_bias=False, name= 'Conv2d_2b_3x3') (x)
x = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Conv2d_2b_3x3_BatchNorm')(x)
x = Activation('relu', name='Conv2d_2b_3x3_Activation')(x)
x = MaxPooling2D(3, strides=2, name='MaxPool_3a_3x3')(x)
x = Conv2D(80, 1, strides=1, padding='valid', use_bias=False, name= 'Conv2d_3b_1x1') (x)
x = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Conv2d_3b_1x1_BatchNorm')(x)
x = Activation('relu', name='Conv2d_3b_1x1_Activation')(x)
x = Conv2D(192, 3, strides=1, padding='valid', use_bias=False, name= 'Conv2d_4a_3x3') (x)
x = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Conv2d_4a_3x3_BatchNorm')(x)
x = Activation('relu', name='Conv2d_4a_3x3_Activation')(x)
x = Conv2D(256, 3, strides=2, padding='valid', use_bias=False, name= 'Conv2d_4b_3x3') (x)
x = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Conv2d_4b_3x3_BatchNorm')(x)
x = Activation('relu', name='Conv2d_4b_3x3_Activation')(x)
# 5x Block35 (Inception-ResNet-A block):
branch_0 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_1_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_1_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block35_1_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_1_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_1_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_1_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_1_Branch_1_Conv2d_0b_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_1_Branch_1_Conv2d_0b_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_1_Branch_1_Conv2d_0b_3x3_Activation')(branch_1)
branch_2 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_1_Branch_2_Conv2d_0a_1x1') (x)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_1_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_1_Branch_2_Conv2d_0a_1x1_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_1_Branch_2_Conv2d_0b_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_1_Branch_2_Conv2d_0b_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_1_Branch_2_Conv2d_0b_3x3_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_1_Branch_2_Conv2d_0c_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_1_Branch_2_Conv2d_0c_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_1_Branch_2_Conv2d_0c_3x3_Activation')(branch_2)
branches = [branch_0, branch_1, branch_2]
mixed = Concatenate(axis=3, name='Block35_1_Concatenate')(branches)
up = Conv2D(256, 1, strides=1, padding='same', use_bias=True, name= 'Block35_1_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.17})(up)
x = add([x, up])
x = Activation('relu', name='Block35_1_Activation')(x)
branch_0 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_2_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_2_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block35_2_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_2_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_2_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_2_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_2_Branch_1_Conv2d_0b_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_2_Branch_1_Conv2d_0b_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_2_Branch_1_Conv2d_0b_3x3_Activation')(branch_1)
branch_2 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_2_Branch_2_Conv2d_0a_1x1') (x)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_2_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_2_Branch_2_Conv2d_0a_1x1_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_2_Branch_2_Conv2d_0b_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_2_Branch_2_Conv2d_0b_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_2_Branch_2_Conv2d_0b_3x3_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_2_Branch_2_Conv2d_0c_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_2_Branch_2_Conv2d_0c_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_2_Branch_2_Conv2d_0c_3x3_Activation')(branch_2)
branches = [branch_0, branch_1, branch_2]
mixed = Concatenate(axis=3, name='Block35_2_Concatenate')(branches)
up = Conv2D(256, 1, strides=1, padding='same', use_bias=True, name= 'Block35_2_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.17})(up)
x = add([x, up])
x = Activation('relu', name='Block35_2_Activation')(x)
branch_0 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_3_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_3_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block35_3_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_3_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_3_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_3_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_3_Branch_1_Conv2d_0b_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_3_Branch_1_Conv2d_0b_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_3_Branch_1_Conv2d_0b_3x3_Activation')(branch_1)
branch_2 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_3_Branch_2_Conv2d_0a_1x1') (x)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_3_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_3_Branch_2_Conv2d_0a_1x1_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_3_Branch_2_Conv2d_0b_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_3_Branch_2_Conv2d_0b_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_3_Branch_2_Conv2d_0b_3x3_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_3_Branch_2_Conv2d_0c_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_3_Branch_2_Conv2d_0c_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_3_Branch_2_Conv2d_0c_3x3_Activation')(branch_2)
branches = [branch_0, branch_1, branch_2]
mixed = Concatenate(axis=3, name='Block35_3_Concatenate')(branches)
up = Conv2D(256, 1, strides=1, padding='same', use_bias=True, name= 'Block35_3_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.17})(up)
x = add([x, up])
x = Activation('relu', name='Block35_3_Activation')(x)
branch_0 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_4_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_4_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block35_4_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_4_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_4_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_4_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_4_Branch_1_Conv2d_0b_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_4_Branch_1_Conv2d_0b_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_4_Branch_1_Conv2d_0b_3x3_Activation')(branch_1)
branch_2 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_4_Branch_2_Conv2d_0a_1x1') (x)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_4_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_4_Branch_2_Conv2d_0a_1x1_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_4_Branch_2_Conv2d_0b_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_4_Branch_2_Conv2d_0b_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_4_Branch_2_Conv2d_0b_3x3_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_4_Branch_2_Conv2d_0c_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_4_Branch_2_Conv2d_0c_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_4_Branch_2_Conv2d_0c_3x3_Activation')(branch_2)
branches = [branch_0, branch_1, branch_2]
mixed = Concatenate(axis=3, name='Block35_4_Concatenate')(branches)
up = Conv2D(256, 1, strides=1, padding='same', use_bias=True, name= 'Block35_4_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.17})(up)
x = add([x, up])
x = Activation('relu', name='Block35_4_Activation')(x)
branch_0 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_5_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_5_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block35_5_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_5_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_5_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_5_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_5_Branch_1_Conv2d_0b_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_5_Branch_1_Conv2d_0b_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block35_5_Branch_1_Conv2d_0b_3x3_Activation')(branch_1)
branch_2 = Conv2D(32, 1, strides=1, padding='same', use_bias=False, name= 'Block35_5_Branch_2_Conv2d_0a_1x1') (x)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_5_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_5_Branch_2_Conv2d_0a_1x1_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_5_Branch_2_Conv2d_0b_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_5_Branch_2_Conv2d_0b_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_5_Branch_2_Conv2d_0b_3x3_Activation')(branch_2)
branch_2 = Conv2D(32, 3, strides=1, padding='same', use_bias=False, name= 'Block35_5_Branch_2_Conv2d_0c_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block35_5_Branch_2_Conv2d_0c_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Block35_5_Branch_2_Conv2d_0c_3x3_Activation')(branch_2)
branches = [branch_0, branch_1, branch_2]
mixed = Concatenate(axis=3, name='Block35_5_Concatenate')(branches)
up = Conv2D(256, 1, strides=1, padding='same', use_bias=True, name= 'Block35_5_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.17})(up)
x = add([x, up])
x = Activation('relu', name='Block35_5_Activation')(x)
# Mixed 6a (Reduction-A block):
branch_0 = Conv2D(384, 3, strides=2, padding='valid', use_bias=False, name= 'Mixed_6a_Branch_0_Conv2d_1a_3x3') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_6a_Branch_0_Conv2d_1a_3x3_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Mixed_6a_Branch_0_Conv2d_1a_3x3_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Mixed_6a_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_6a_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Mixed_6a_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, 3, strides=1, padding='same', use_bias=False, name= 'Mixed_6a_Branch_1_Conv2d_0b_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_6a_Branch_1_Conv2d_0b_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Mixed_6a_Branch_1_Conv2d_0b_3x3_Activation')(branch_1)
branch_1 = Conv2D(256, 3, strides=2, padding='valid', use_bias=False, name= 'Mixed_6a_Branch_1_Conv2d_1a_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_6a_Branch_1_Conv2d_1a_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Mixed_6a_Branch_1_Conv2d_1a_3x3_Activation')(branch_1)
branch_pool = MaxPooling2D(3, strides=2, padding='valid', name='Mixed_6a_Branch_2_MaxPool_1a_3x3')(x)
branches = [branch_0, branch_1, branch_pool]
x = Concatenate(axis=3, name='Mixed_6a')(branches)
# 10x Block17 (Inception-ResNet-B block):
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_1_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_1_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_1_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_1_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_1_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_1_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_1_Branch_1_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_1_Branch_1_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_1_Branch_1_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_1_Branch_1_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_1_Branch_1_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_1_Branch_1_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_1_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_1_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_1_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_2_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_2_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_2_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_2_Branch_2_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_2_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_2_Branch_2_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_2_Branch_2_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_2_Branch_2_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_2_Branch_2_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_2_Branch_2_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_2_Branch_2_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_2_Branch_2_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_2_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_2_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_2_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_3_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_3_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_3_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_3_Branch_3_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_3_Branch_3_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_3_Branch_3_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_3_Branch_3_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_3_Branch_3_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_3_Branch_3_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_3_Branch_3_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_3_Branch_3_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_3_Branch_3_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_3_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_3_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_3_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_4_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_4_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_4_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_4_Branch_4_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_4_Branch_4_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_4_Branch_4_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_4_Branch_4_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_4_Branch_4_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_4_Branch_4_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_4_Branch_4_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_4_Branch_4_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_4_Branch_4_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_4_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_4_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_4_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_5_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_5_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_5_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_5_Branch_5_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_5_Branch_5_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_5_Branch_5_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_5_Branch_5_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_5_Branch_5_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_5_Branch_5_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_5_Branch_5_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_5_Branch_5_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_5_Branch_5_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_5_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_5_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_5_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_6_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_6_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_6_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_6_Branch_6_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_6_Branch_6_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_6_Branch_6_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_6_Branch_6_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_6_Branch_6_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_6_Branch_6_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_6_Branch_6_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_6_Branch_6_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_6_Branch_6_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_6_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_6_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_6_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_7_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_7_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_7_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_7_Branch_7_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_7_Branch_7_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_7_Branch_7_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_7_Branch_7_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_7_Branch_7_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_7_Branch_7_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_7_Branch_7_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_7_Branch_7_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_7_Branch_7_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_7_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_7_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_7_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_8_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_8_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_8_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_8_Branch_8_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_8_Branch_8_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_8_Branch_8_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_8_Branch_8_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_8_Branch_8_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_8_Branch_8_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_8_Branch_8_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_8_Branch_8_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_8_Branch_8_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_8_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_8_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_8_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_9_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_9_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_9_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_9_Branch_9_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_9_Branch_9_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_9_Branch_9_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_9_Branch_9_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_9_Branch_9_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_9_Branch_9_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_9_Branch_9_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_9_Branch_9_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_9_Branch_9_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_9_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_9_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_9_Activation')(x)
branch_0 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_10_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_10_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block17_10_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(128, 1, strides=1, padding='same', use_bias=False, name= 'Block17_10_Branch_10_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_10_Branch_10_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_10_Branch_10_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(128, [1, 7], strides=1, padding='same', use_bias=False, name= 'Block17_10_Branch_10_Conv2d_0b_1x7') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_10_Branch_10_Conv2d_0b_1x7_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_10_Branch_10_Conv2d_0b_1x7_Activation')(branch_1)
branch_1 = Conv2D(128, [7, 1], strides=1, padding='same', use_bias=False, name= 'Block17_10_Branch_10_Conv2d_0c_7x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block17_10_Branch_10_Conv2d_0c_7x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block17_10_Branch_10_Conv2d_0c_7x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block17_10_Concatenate')(branches)
up = Conv2D(896, 1, strides=1, padding='same', use_bias=True, name= 'Block17_10_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.1})(up)
x = add([x, up])
x = Activation('relu', name='Block17_10_Activation')(x)
# Mixed 7a (Reduction-B block): 8 x 8 x 2080
branch_0 = Conv2D(256, 1, strides=1, padding='same', use_bias=False, name= 'Mixed_7a_Branch_0_Conv2d_0a_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_0_Conv2d_0a_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Mixed_7a_Branch_0_Conv2d_0a_1x1_Activation')(branch_0)
branch_0 = Conv2D(384, 3, strides=2, padding='valid', use_bias=False, name= 'Mixed_7a_Branch_0_Conv2d_1a_3x3') (branch_0)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_0_Conv2d_1a_3x3_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Mixed_7a_Branch_0_Conv2d_1a_3x3_Activation')(branch_0)
branch_1 = Conv2D(256, 1, strides=1, padding='same', use_bias=False, name= 'Mixed_7a_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Mixed_7a_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(256, 3, strides=2, padding='valid', use_bias=False, name= 'Mixed_7a_Branch_1_Conv2d_1a_3x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_1_Conv2d_1a_3x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Mixed_7a_Branch_1_Conv2d_1a_3x3_Activation')(branch_1)
branch_2 = Conv2D(256, 1, strides=1, padding='same', use_bias=False, name= 'Mixed_7a_Branch_2_Conv2d_0a_1x1') (x)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Mixed_7a_Branch_2_Conv2d_0a_1x1_Activation')(branch_2)
branch_2 = Conv2D(256, 3, strides=1, padding='same', use_bias=False, name= 'Mixed_7a_Branch_2_Conv2d_0b_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_2_Conv2d_0b_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Mixed_7a_Branch_2_Conv2d_0b_3x3_Activation')(branch_2)
branch_2 = Conv2D(256, 3, strides=2, padding='valid', use_bias=False, name= 'Mixed_7a_Branch_2_Conv2d_1a_3x3') (branch_2)
branch_2 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Mixed_7a_Branch_2_Conv2d_1a_3x3_BatchNorm')(branch_2)
branch_2 = Activation('relu', name='Mixed_7a_Branch_2_Conv2d_1a_3x3_Activation')(branch_2)
branch_pool = MaxPooling2D(3, strides=2, padding='valid', name='Mixed_7a_Branch_3_MaxPool_1a_3x3')(x)
branches = [branch_0, branch_1, branch_2, branch_pool]
x = Concatenate(axis=3, name='Mixed_7a')(branches)
# 5x Block8 (Inception-ResNet-C block):
branch_0 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_1_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_1_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block8_1_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_1_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_1_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_1_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, [1, 3], strides=1, padding='same', use_bias=False, name= 'Block8_1_Branch_1_Conv2d_0b_1x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_1_Branch_1_Conv2d_0b_1x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_1_Branch_1_Conv2d_0b_1x3_Activation')(branch_1)
branch_1 = Conv2D(192, [3, 1], strides=1, padding='same', use_bias=False, name= 'Block8_1_Branch_1_Conv2d_0c_3x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_1_Branch_1_Conv2d_0c_3x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_1_Branch_1_Conv2d_0c_3x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block8_1_Concatenate')(branches)
up = Conv2D(1792, 1, strides=1, padding='same', use_bias=True, name= 'Block8_1_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.2})(up)
x = add([x, up])
x = Activation('relu', name='Block8_1_Activation')(x)
branch_0 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_2_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_2_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block8_2_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_2_Branch_2_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_2_Branch_2_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_2_Branch_2_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, [1, 3], strides=1, padding='same', use_bias=False, name= 'Block8_2_Branch_2_Conv2d_0b_1x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_2_Branch_2_Conv2d_0b_1x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_2_Branch_2_Conv2d_0b_1x3_Activation')(branch_1)
branch_1 = Conv2D(192, [3, 1], strides=1, padding='same', use_bias=False, name= 'Block8_2_Branch_2_Conv2d_0c_3x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_2_Branch_2_Conv2d_0c_3x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_2_Branch_2_Conv2d_0c_3x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block8_2_Concatenate')(branches)
up = Conv2D(1792, 1, strides=1, padding='same', use_bias=True, name= 'Block8_2_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.2})(up)
x = add([x, up])
x = Activation('relu', name='Block8_2_Activation')(x)
branch_0 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_3_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_3_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block8_3_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_3_Branch_3_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_3_Branch_3_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_3_Branch_3_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, [1, 3], strides=1, padding='same', use_bias=False, name= 'Block8_3_Branch_3_Conv2d_0b_1x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_3_Branch_3_Conv2d_0b_1x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_3_Branch_3_Conv2d_0b_1x3_Activation')(branch_1)
branch_1 = Conv2D(192, [3, 1], strides=1, padding='same', use_bias=False, name= 'Block8_3_Branch_3_Conv2d_0c_3x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_3_Branch_3_Conv2d_0c_3x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_3_Branch_3_Conv2d_0c_3x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block8_3_Concatenate')(branches)
up = Conv2D(1792, 1, strides=1, padding='same', use_bias=True, name= 'Block8_3_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.2})(up)
x = add([x, up])
x = Activation('relu', name='Block8_3_Activation')(x)
branch_0 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_4_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_4_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block8_4_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_4_Branch_4_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_4_Branch_4_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_4_Branch_4_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, [1, 3], strides=1, padding='same', use_bias=False, name= 'Block8_4_Branch_4_Conv2d_0b_1x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_4_Branch_4_Conv2d_0b_1x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_4_Branch_4_Conv2d_0b_1x3_Activation')(branch_1)
branch_1 = Conv2D(192, [3, 1], strides=1, padding='same', use_bias=False, name= 'Block8_4_Branch_4_Conv2d_0c_3x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_4_Branch_4_Conv2d_0c_3x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_4_Branch_4_Conv2d_0c_3x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block8_4_Concatenate')(branches)
up = Conv2D(1792, 1, strides=1, padding='same', use_bias=True, name= 'Block8_4_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.2})(up)
x = add([x, up])
x = Activation('relu', name='Block8_4_Activation')(x)
branch_0 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_5_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_5_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block8_5_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_5_Branch_5_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_5_Branch_5_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_5_Branch_5_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, [1, 3], strides=1, padding='same', use_bias=False, name= 'Block8_5_Branch_5_Conv2d_0b_1x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_5_Branch_5_Conv2d_0b_1x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_5_Branch_5_Conv2d_0b_1x3_Activation')(branch_1)
branch_1 = Conv2D(192, [3, 1], strides=1, padding='same', use_bias=False, name= 'Block8_5_Branch_5_Conv2d_0c_3x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_5_Branch_5_Conv2d_0c_3x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_5_Branch_5_Conv2d_0c_3x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block8_5_Concatenate')(branches)
up = Conv2D(1792, 1, strides=1, padding='same', use_bias=True, name= 'Block8_5_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 0.2})(up)
x = add([x, up])
x = Activation('relu', name='Block8_5_Activation')(x)
branch_0 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_6_Branch_0_Conv2d_1x1') (x)
branch_0 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_6_Branch_0_Conv2d_1x1_BatchNorm')(branch_0)
branch_0 = Activation('relu', name='Block8_6_Branch_0_Conv2d_1x1_Activation')(branch_0)
branch_1 = Conv2D(192, 1, strides=1, padding='same', use_bias=False, name= 'Block8_6_Branch_1_Conv2d_0a_1x1') (x)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_6_Branch_1_Conv2d_0a_1x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_6_Branch_1_Conv2d_0a_1x1_Activation')(branch_1)
branch_1 = Conv2D(192, [1, 3], strides=1, padding='same', use_bias=False, name= 'Block8_6_Branch_1_Conv2d_0b_1x3') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_6_Branch_1_Conv2d_0b_1x3_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_6_Branch_1_Conv2d_0b_1x3_Activation')(branch_1)
branch_1 = Conv2D(192, [3, 1], strides=1, padding='same', use_bias=False, name= 'Block8_6_Branch_1_Conv2d_0c_3x1') (branch_1)
branch_1 = BatchNormalization(axis=3, momentum=0.995, epsilon=0.001, scale=False, name='Block8_6_Branch_1_Conv2d_0c_3x1_BatchNorm')(branch_1)
branch_1 = Activation('relu', name='Block8_6_Branch_1_Conv2d_0c_3x1_Activation')(branch_1)
branches = [branch_0, branch_1]
mixed = Concatenate(axis=3, name='Block8_6_Concatenate')(branches)
up = Conv2D(1792, 1, strides=1, padding='same', use_bias=True, name= 'Block8_6_Conv2d_1x1') (mixed)
up = Lambda(scaling, output_shape=K.int_shape(up)[1:], arguments={'scale': 1})(up)
x = add([x, up])
# Classification block
x = GlobalAveragePooling2D(name='AvgPool')(x)
x = Dropout(1.0 - 0.8, name='Dropout')(x)
# Bottleneck
x = Dense(128, use_bias=False, name='Bottleneck')(x)
x = BatchNormalization(momentum=0.995, epsilon=0.001, scale=False, name='Bottleneck_BatchNorm')(x)
# Create model
model = Model(inputs, x, name='inception_resnet_v1')
return model
def tiny_yolo4lite_mobilenet_body(inputs,
num_anchors,
num_classes,
alpha=1.0,
use_spp=True):
'''Create Tiny YOLO_v3 Lite MobileNet model CNN body in keras.'''
mobilenet = MobileNet(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# conv_pw_13_relu :13 x 13 x (1024*alpha)
# conv_pw_11_relu :26 x 26 x (512*alpha)
# conv_pw_5_relu : 52 x 52 x (256*alpha)
# f1 :13 x 13 x (1024*alpha) for 416 input
f1 = mobilenet.get_layer('conv_pw_13_relu').output
# f2: 26 x 26 x (512*alpha) for 416 input
f2 = mobilenet.get_layer('conv_pw_11_relu').output
f1_channel_num = int(1024 * alpha)
f2_channel_num = int(512 * alpha)
#feature map 1 head (13 x 13 x (512*alpha) for 416 input)
x1 = DarknetConv2D_BN_Leaky(f1_channel_num // 2, (1, 1))(f1)
if use_spp:
x1 = Spp_Conv2D_BN_Leaky(x1, f1_channel_num // 2)
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x1)
x2 = compose(
Concatenate(),
#DarknetConv2D_BN_Leaky(f2_channel_num, (3,3)),
Depthwise_Separable_Conv2D_BN_Leaky(filters=f2_channel_num,
kernel_size=(3, 3),
block_id_str='15'))(
[x1_upsample, f2])
#feature map 2 output (26 x 26 x (512*alpha) for 416 input)
y2 = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
#DarknetConv2D_BN_Leaky(f1_channel_num//2, (3,3), strides=(2,2)),
Darknet_Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num // 2,
(3, 3),
strides=(2, 2),
block_id_str='16'))(x2)
x1 = compose(
Concatenate(),
#DarknetConv2D_BN_Leaky(f1_channel_num, (3,3)),
Depthwise_Separable_Conv2D_BN_Leaky(filters=f1_channel_num,
kernel_size=(3, 3),
block_id_str='17'))(
[x2_downsample, x1])
#feature map 1 output (13 x 13 x (1024*alpha) for 416 input)
y1 = DarknetConv2D(num_anchors * (num_classes + 5), (1, 1))(x1)
return Model(inputs, [y1, y2])
def yoloNano(anchors,input_size=416,num_classes = 1,expension = .75,decay=0.0005):
#f**k tensorflow 2.x
#backbone
input_0 = Input(shape=(input_size,input_size,3))
input_gt = [Input(shape=(input_size//{0:32, 1:16, 2:8}[l], input_size//{0:32, 1:16, 2:8}[l],len(anchors)//3, num_classes+5)) for l in range(3)]
x = Conv2D(filters=12,strides=(1,1),kernel_size=(3,3),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(input_0)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=24,strides=(2,2),kernel_size=(3,3),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_0 = LeakyReLU(alpha = 0.1)(x)
#PEP(7)(208x208x24)
x = Conv2D(filters=7,strides=(1,1),kernel_size=(1,1),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x_0)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(24 * expension),strides=(1,1),kernel_size=(1,1),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1,1),kernel_size=(3,3),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=24,strides=(1,1),kernel_size=(1,1),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = Add()([x_0 ,x])
#EP(104x104x70)
x = Conv2D(filters=math.ceil(70 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(2, 2), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=70, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x_1 = BatchNormalization()(x)
#PEP(25)(104x104x70)
x = Conv2D(filters=25, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_1)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(70 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=70, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_2 = Add()([x_1, x])
# PEP(24)(104x104x70)
x = Conv2D(filters=24, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_2)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(70 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=70, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = Add()([x_2, x])
# EP(52x52x150)
x = Conv2D(filters=math.ceil(150 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(2, 2), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=150, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x_3 = BatchNormalization()(x)
# PEP(56)(52x52x150)
x = Conv2D(filters=56, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_3)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(150 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=150, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = Add()([x_3, x])
#Conv1x1
x = Conv2D(filters=150,kernel_size=(1,1),strides=(1,1),use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_4 = LeakyReLU(alpha = 0.1)(x)
#FCA(8)
x = AvgPool2D(pool_size=(52,52))(x_4)
x = Dense(units=150 // 8,activation='relu',use_bias=False,kernel_regularizer=l2(l=decay))(x)
x = Dense(units=150, activation='sigmoid', use_bias=False,kernel_regularizer=l2(l=decay))(x)
x_5 = Multiply()([x_4,x])
#PEP(73)(52x52x150)
x = Conv2D(filters=73, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_5)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(150 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=150, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_6 = Add()([x_5, x])
# PEP(71)(52x52x150)
x = Conv2D(filters=71, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_6)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(150 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=150, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_7 = Add()([x_6, x])
# PEP(75)(52x52x150)
x = Conv2D(filters=75, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_7)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(150 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=150, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_8 = Add()([x_7, x]) #output 52x52x150
#EP(26x26x325)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_8)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(2, 2), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x_9 = BatchNormalization()(x)
# PEP(132)(26x26x325)
x = Conv2D(filters=132, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_9)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_10 = Add()([x_9, x])
# PEP(124)(26x26x325)
x = Conv2D(filters=124, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_10)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_11 = Add()([x_10, x])
# PEP(141)(26x26x325)
x = Conv2D(filters=141, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_11)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_12 = Add()([x_11, x])
# PEP(140)(26x26x325)
x = Conv2D(filters=140, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_12)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_13 = Add()([x_12, x])
# PEP(137)(26x26x325)
x = Conv2D(filters=137, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_13)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_14 = Add()([x_13, x])
# PEP(135)(26x26x325)
x = Conv2D(filters=135, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_14)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_15 = Add()([x_14, x])
# PEP(133)(26x26x325)
x = Conv2D(filters=133, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_15)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_16 = Add()([x_15, x])
# PEP(140)(26x26x325)
x = Conv2D(filters=140, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_16)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_17 = Add()([x_16, x]) #output 26x26x325
# EP(13x13x545)
x = Conv2D(filters=math.ceil(545 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_17)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(2, 2), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=545, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x_18 = BatchNormalization()(x)
# PEP(276)(13x13x545)
x = Conv2D(filters=276, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_18)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(545 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=545, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_19 = Add()([x_18, x])
#Conv1x1
x = Conv2D(filters=230, kernel_size=(1, 1), strides=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_19)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
# EP(13x13x489)
x = Conv2D(filters=math.ceil(489 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=489, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# PEP(213)(13x13x469)
x = Conv2D(filters=213, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(469 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=469, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# Conv1x1
x = Conv2D(filters=189, kernel_size=(1, 1), strides=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_20 = LeakyReLU(alpha = 0.1)(x) #output 13x13x189
# EP(13x13x462)
x = Conv2D(filters=math.ceil(462 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_20)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=462, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# feature 13x13x[(num_classes+5)x3]
feature_13x13 = Conv2D(filters=3 * (num_classes + 5), kernel_size=(1, 1), strides=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
# Conv1x1
x = Conv2D(filters=105, kernel_size=(1, 1), strides=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_20)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
# upsampling 26x26x105
x = UpSampling2D()(x)
# concatenate
x = Concatenate()([x,x_17])
# PEP(113)(26x26x325)
x = Conv2D(filters=113, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(325 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=325, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# PEP(99)(26x26x207)
x = Conv2D(filters=99, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(207 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=207, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# Conv1x1
x = Conv2D(filters=98, kernel_size=(1, 1), strides=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x_21 = LeakyReLU(alpha = 0.1)(x)
# EP(13x13x183)
x = Conv2D(filters=math.ceil(183 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_21)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=183, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# feature 26x26x[(num_classes+5)x3]
feature_26x26 = Conv2D(filters=3 * (num_classes + 5), kernel_size=(1, 1), strides=(1, 1), use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
# Conv1x1
x = Conv2D(filters=47, kernel_size=(1, 1), strides=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x_21)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
#upsampling
x = UpSampling2D()(x)
#concatenate
x = Concatenate()([x,x_8])
# PEP(58)(52x52x132)
x = Conv2D(filters=58, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(132 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=132, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# PEP(52)(52x52x87)
x = Conv2D(filters=52, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(87 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=87, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
# PEP(47)(52x52x93)
x = Conv2D(filters=47, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=math.ceil(93 * expension), strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = DepthwiseConv2D(strides=(1, 1), kernel_size=(3, 3), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
x = LeakyReLU(alpha = 0.1)(x)
x = Conv2D(filters=93, strides=(1, 1), kernel_size=(1, 1), use_bias=False, padding='same',kernel_regularizer=l2(l=decay))(x)
x = BatchNormalization()(x)
feature_52x52 = Conv2D(filters=3 * (num_classes + 5), kernel_size=(1, 1), strides=(1, 1), use_bias=False,padding='same',kernel_regularizer=l2(l=decay))(x)
#loss layer
loss = Lambda(yolo_loss, output_shape=(1,), name='yolo_loss',arguments={'anchors': anchors, 'num_classes': num_classes, 'ignore_thresh': 0.5})([feature_13x13,feature_26x26,feature_52x52, *input_gt])
debug_model = tf.keras.Model(inputs=input_0,outputs=[feature_13x13,feature_26x26,feature_52x52])
train_model = tf.keras.Model(inputs=[input_0,*input_gt],outputs=loss)
return train_model,debug_model
# import numpy as np
# anchors = np.array([[6.,9.],[8.,13.],[11.,16.],[14.,22.],[17.,37.],[21.,26.],[29.,38.],[39.,62.],[79.,99.]],dtype='float32')
# model,_ = yoloNano(anchors,input_size=416,num_classes=1)
# model.summary()
def build(self):
'''
1. Build Code Representation Model
'''
logger.debug('Building Code Representation Model')
methname = Input(shape=(self.data_params['methname_len'], ),
dtype='int32',
name='methname')
apiseq = Input(shape=(self.data_params['apiseq_len'], ),
dtype='int32',
name='apiseq')
tokens = Input(shape=(self.data_params['tokens_len'], ),
dtype='int32',
name='tokens')
## method name representation ##
#1.embedding
init_emb_weights = np.load(
self.model_params['init_embed_weights_methname']
) if self.model_params[
'init_embed_weights_methname'] is not None else None
if init_emb_weights is not None: init_emb_weights = [init_emb_weights]
embedding = Embedding(
input_dim=self.data_params['n_words'],
output_dim=self.model_params.get('n_embed_dims', 100),
weights=init_emb_weights,
mask_zero=
False, #Whether 0 in the input is a special "padding" value that should be masked out.
#If True, all subsequent layers in the model must support masking, otherwise an exception will be raised.
name='embedding_methname')
methname_embedding = embedding(methname)
dropout = Dropout(0.25, name='dropout_methname_embed')
methname_dropout = dropout(methname_embedding)
#2.rnn
f_rnn = LSTM(self.model_params.get('n_lstm_dims', 128),
recurrent_dropout=0.2,
return_sequences=True,
name='lstm_methname_f')
b_rnn = LSTM(self.model_params.get('n_lstm_dims', 128),
return_sequences=True,
recurrent_dropout=0.2,
name='lstm_methname_b',
go_backwards=True)
methname_f_rnn = f_rnn(methname_dropout)
methname_b_rnn = b_rnn(methname_dropout)
dropout = Dropout(0.25, name='dropout_methname_rnn')
methname_f_dropout = dropout(methname_f_rnn)
methname_b_dropout = dropout(methname_b_rnn)
#3.maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]),
name='maxpool_methname')
methname_pool = Concatenate(name='concat_methname_lstms')(
[maxpool(methname_f_dropout),
maxpool(methname_b_dropout)])
activation = Activation('tanh', name='active_methname')
methname_repr = activation(methname_pool)
## API Sequence Representation ##
#1.embedding
embedding = Embedding(
input_dim=self.data_params['n_words'],
output_dim=self.model_params.get('n_embed_dims', 100),
#weights=weights,
mask_zero=
False, #Whether 0 in the input is a special "padding" value that should be masked out.
#If True, all subsequent layers must support masking, otherwise an exception will be raised.
name='embedding_apiseq')
apiseq_embedding = embedding(apiseq)
dropout = Dropout(0.25, name='dropout_apiseq_embed')
apiseq_dropout = dropout(apiseq_embedding)
#2.rnn
f_rnn = LSTM(self.model_params.get('n_lstm_dims', 100),
return_sequences=True,
recurrent_dropout=0.2,
name='lstm_apiseq_f')
b_rnn = LSTM(self.model_params.get('n_lstm_dims', 100),
return_sequences=True,
recurrent_dropout=0.2,
name='lstm_apiseq_b',
go_backwards=True)
apiseq_f_rnn = f_rnn(apiseq_dropout)
apiseq_b_rnn = b_rnn(apiseq_dropout)
dropout = Dropout(0.25, name='dropout_apiseq_rnn')
apiseq_f_dropout = dropout(apiseq_f_rnn)
apiseq_b_dropout = dropout(apiseq_b_rnn)
#3.maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]),
name='maxpool_apiseq')
apiseq_pool = Concatenate(name='concat_apiseq_lstms')(
[maxpool(apiseq_f_dropout),
maxpool(apiseq_b_dropout)])
activation = Activation('tanh', name='active_apiseq')
apiseq_repr = activation(apiseq_pool)
## Tokens Representation ##
#1.embedding
init_emb_weights = np.load(
self.model_params['init_embed_weights_tokens']
) if self.model_params[
'init_embed_weights_tokens'] is not None else None
if init_emb_weights is not None: init_emb_weights = [init_emb_weights]
embedding = Embedding(
input_dim=self.data_params['n_words'],
output_dim=self.model_params.get('n_embed_dims', 100),
weights=init_emb_weights,
#mask_zero=True,#Whether 0 in the input is a special "padding" value that should be masked out.
#If True, all subsequent layers must support masking, otherwise an exception will be raised.
name='embedding_tokens')
tokens_embedding = embedding(tokens)
dropout = Dropout(0.25, name='dropout_tokens_embed')
tokens_dropout = dropout(tokens_embedding)
#4.maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]),
name='maxpool_tokens')
tokens_pool = maxpool(tokens_dropout)
activation = Activation('tanh', name='active_tokens')
tokens_repr = activation(tokens_pool)
## concatenate the representation of code ##
merged_methname_api = Concatenate(name='merge_methname_api')(
[methname_repr, apiseq_repr])
merged_code_repr = Concatenate(name='merge_coderepr')(
[merged_methname_api, tokens_repr])
code_repr = Dense(self.model_params.get('n_hidden', 400),
activation='tanh',
name='dense_coderepr')(merged_code_repr)
self._code_repr_model = Model(inputs=[methname, apiseq, tokens],
outputs=[code_repr],
name='code_repr_model')
'''
2. Build Desc Representation Model
'''
## Desc Representation ##
logger.debug('Building Desc Representation Model')
desc = Input(shape=(self.data_params['desc_len'], ),
dtype='int32',
name='desc')
#1.embedding
init_emb_weights = np.load(
self.model_params['init_embed_weights_desc']
) if self.model_params['init_embed_weights_desc'] is not None else None
if init_emb_weights is not None: init_emb_weights = [init_emb_weights]
embedding = Embedding(
input_dim=self.data_params['n_words'],
output_dim=self.model_params.get('n_embed_dims', 100),
weights=init_emb_weights,
mask_zero=
True, #Whether 0 in the input is a special "padding" value that should be masked out.
#If True, all subsequent layers must support masking, otherwise an exception will be raised.
name='embedding_desc')
desc_embedding = embedding(desc)
dropout = Dropout(0.25, name='dropout_desc_embed')
desc_dropout = dropout(desc_embedding)
#2. rnn
f_rnn = LSTM(self.model_params.get('n_lstm_dims', 100),
return_sequences=True,
recurrent_dropout=0.2,
name='lstm_desc_f')
b_rnn = LSTM(self.model_params.get('n_lstm_dims', 100),
return_sequences=True,
recurrent_dropout=0.2,
name='lstm_desc_b',
go_backwards=True)
desc_f_rnn = f_rnn(desc_dropout)
desc_b_rnn = b_rnn(desc_dropout)
dropout = Dropout(0.25, name='dropout_desc_rnn')
desc_f_dropout = dropout(desc_f_rnn)
desc_b_dropout = dropout(desc_b_rnn)
#3. maxpooling
maxpool = Lambda(lambda x: K.max(x, axis=1, keepdims=False),
output_shape=lambda x: (x[0], x[2]),
name='maxpool_desc')
desc_pool = Concatenate(name='concat_desc_rnns')(
[maxpool(desc_f_dropout),
maxpool(desc_b_dropout)])
activation = Activation('tanh', name='active_desc')
desc_repr = activation(desc_pool)
self._desc_repr_model = Model(inputs=[desc],
outputs=[desc_repr],
name='desc_repr_model')
"""
3: calculate the cosine similarity between code and desc
"""
logger.debug('Building similarity model')
code_repr = self._code_repr_model([methname, apiseq, tokens])
desc_repr = self._desc_repr_model([desc])
cos_sim = Dot(axes=1, normalize=True,
name='cos_sim')([code_repr, desc_repr])
sim_model = Model(inputs=[methname, apiseq, tokens, desc],
outputs=[cos_sim],
name='sim_model')
self._sim_model = sim_model #for model evaluation
'''
4:Build training model
'''
good_sim = sim_model(
[self.methname, self.apiseq, self.tokens,
self.desc_good]) # similarity of good output
bad_sim = sim_model(
[self.methname, self.apiseq, self.tokens,
self.desc_bad]) #similarity of bad output
loss = Lambda(lambda x: K.maximum(
1e-6, self.model_params['margin'] - x[0] + x[1]),
output_shape=lambda x: x[0],
name='loss')([good_sim, bad_sim])
logger.debug('Building training model')
self._training_model = Model(inputs=[
self.methname, self.apiseq, self.tokens, self.desc_good,
self.desc_bad
],
outputs=[loss],
name='training_model')
def model_unet(inp):
"""
Define the unet model. See relation for deeper explanation of this part of the code.
:param inp (tf.keras.layers.Layer): Input of the NN
:return: output (tf.keras.layers.Layer): last layer of the network (see relation)
"""
padding = 'same'
strides = (2, 2)
kernel_size = (3, 3)
conv1 = Conv2D(32, kernel_size, padding=padding)(inp)
conv1 = LeakyReLU(alpha=0.2)(conv1)
conv1 = Conv2D(32, kernel_size, padding=padding)(conv1)
conv1 = LeakyReLU(alpha=0.2)(conv1)
pool1 = MaxPooling2D(pool_size=(3, 3), strides=strides, padding=padding)(conv1)
conv2 = Conv2D(64, kernel_size, padding=padding)(pool1)
conv2 = LeakyReLU(alpha=0.2)(conv2)
conv2 = Conv2D(64, kernel_size, padding=padding)(conv2)
conv2 = LeakyReLU(alpha=0.2)(conv2)
pool2 = MaxPooling2D(pool_size=(3, 3), strides=strides, padding=padding)(conv2)
conv3 = Conv2D(128, kernel_size, padding=padding)(pool2)
conv3 = LeakyReLU(alpha=0.2)(conv3)
conv3 = Conv2D(128, kernel_size, padding=padding)(conv3)
conv3 = LeakyReLU(alpha=0.2)(conv3)
pool3 = MaxPooling2D(pool_size=kernel_size, strides=strides, padding=padding)(conv3)
conv4 = Conv2D(256, kernel_size, padding=padding)(pool3)
conv4 = LeakyReLU(alpha=0.2)(conv4)
conv4 = Conv2D(256, kernel_size, padding=padding)(conv4)
conv4 = LeakyReLU(alpha=0.2)(conv4)
pool4 = MaxPooling2D(pool_size=(3, 3), strides=strides, padding=padding)(conv4)
conv5 = Conv2D(512, kernel_size, padding=padding)(pool4)
conv5 = LeakyReLU(alpha=0.2)(conv5)
conv5 = Conv2D(512, kernel_size, padding=padding)(conv5)
conv5 = LeakyReLU(alpha=0.2)(conv5)
up6 = Conv2DTranspose(256, kernel_size, strides=strides, padding=padding)(
conv5)
up6 = Concatenate()([conv4, up6])
conv6 = Conv2D(256, kernel_size, padding=padding)(up6)
conv6 = LeakyReLU(alpha=0.2)(conv6)
conv6 = Conv2D(256, kernel_size, padding=padding)(conv6)
conv6 = LeakyReLU(alpha=0.2)(conv6)
up7 = Conv2DTranspose(128, kernel_size, strides=strides, padding=padding)(
conv6)
up7 = Concatenate()([conv3, up7])
conv7 = Conv2D(128, kernel_size, padding=padding)(up7)
conv7 = LeakyReLU(alpha=0.2)(conv7)
conv7 = Conv2D(128, kernel_size, padding=padding)(conv7)
conv7 = LeakyReLU(alpha=0.2)(conv7)
up8 = Conv2DTranspose(64, kernel_size, strides=strides, padding=padding)(
conv7)
up8 = Concatenate()([conv2, up8])
conv8 = Conv2D(64, kernel_size, padding=padding)(up8)
conv8 = LeakyReLU(alpha=0.2)(conv8)
conv8 = Conv2D(64, kernel_size, padding=padding)(conv8)
conv8 = LeakyReLU(alpha=0.2)(conv8)
up9 = Conv2DTranspose(32, kernel_size, strides=strides, padding=padding)(
conv8)
up9 = Concatenate()([conv1, up9])
conv9 = Conv2D(32, kernel_size, padding=padding)(up9)
conv9 = LeakyReLU(alpha=0.2)(conv9)
conv9 = Conv2D(32, kernel_size, padding=padding)(conv9)
conv9 = LeakyReLU(alpha=0.2)(conv9)
conv10 = Conv2D(12, kernel_size, padding=padding)(conv9)
conv10 = LeakyReLU(alpha=0.2)(conv10)
drop = Dropout(0.3)(conv10)
output = Conv2D(3, kernel_size, padding=padding, activation='sigmoid')(drop)
return output
def create_model(l2, num_classes, num_ctrl_classes):
##############
# BRANCH MODEL
##############
regul = regularizers.l2(l2)
# optim = Adam(lr=lr)
# kwargs = {'kernel_regularizer': regul}
base_model = efn.EfficientNetB2(input_shape=(network_shape[0],
network_shape[1], 3),
weights='imagenet',
include_top=False)
input_tensor = Input(shape=network_shape, dtype=K.floatx())
conv1 = Conv2D(32,
kernel_size=(3, 3),
strides=(2, 2),
activation='relu',
use_bias=False,
padding='same',
input_shape=(input_shape[0] + 6, input_shape[1] + 6,
input_shape[2]))
layers = []
layers.append(input_tensor)
layers.append(conv1)
layers[2:] = base_model.layers[2:]
new_model = copy_model_graph(layers, base_model, input_tensor)
weights = base_model.layers[1].get_weights()
weight0 = weights[0]
w = np.concatenate((weight0, weight0), axis=2)
w = w / 2.0
weights[0] = w
# weights.append(np.zeros((64),dtype='float32'))
new_model.layers[1].set_weights(weights)
inp = Input(shape=input_shape, dtype='uint8') # 384x384x6
x = Lambda(augment)(inp)
for layer in new_model.layers:
if type(layer) is Conv2D:
layer.kernel_regularizer = regul
x = new_model(x)
x = GlobalMaxPooling2D()(x)
x = BatchNormalization()(x)
x = Dropout(rate=0.5)(x)
x = Flatten()(x)
x = Dense(512, use_bias=False, kernel_initializer='he_normal')(x)
x = BatchNormalization()(x)
encoder_model = Model(inputs=inp, outputs=x)
# softmax model for training encoder
output_softmax = Dense(num_classes, use_bias=False,
activation='softmax')(x)
softmax_model = Model(inputs=inp, outputs=output_softmax)
#################
# COMPARE MODEL #
#################
mid = 32
xa_inp = Input(shape=encoder_model.output_shape[1:])
xb_inp = Input(shape=encoder_model.output_shape[1:])
x1 = Lambda(lambda x: x[0] * x[1])([xa_inp, xb_inp])
x2 = Lambda(lambda x: x[0] + x[1])([xa_inp, xb_inp])
x3 = Lambda(lambda x: x[0] - x[1])([xa_inp, xb_inp])
x4 = Lambda(lambda x: K.square(x))(x3)
head = Concatenate()([x1, x2, x3, x4])
head = Reshape((4, encoder_model.output_shape[1], 1),
name='reshape1')(head)
# Per feature NN with shared weight is implemented using CONV2D with appropriate stride.
head = Conv2D(mid, (4, 1), activation='relu', padding='valid')(head)
head = Reshape((encoder_model.output_shape[1], mid, 1))(head)
head = Conv2D(1, (1, mid), activation='linear', padding='valid')(head)
head = Flatten()(head)
compare_model = Model([xa_inp, xb_inp], head)
# process encoding from control
# compare the current features to all controls
features_controls = Input(
shape=[num_ctrl_classes, encoder_model.output_shape[1]])
fs = Lambda(lambda x: tf.unstack(x, axis=1))(features_controls)
# def create_mask(features_controls):
# # Use a function with a Keras Lambda layer wrapper to resolve a tensorflow issue.
# # https://stackoverflow.com/questions/50715928/valueerror-output-tensors-to-a-model-must-be-the-output-of-a-tensorflow-layer
# max_abs_features = K.max(K.abs(features_controls), axis=2)
# mask = tf.greater(max_abs_features, K.epsilon())
# mask = tf.expand_dims(tf.expand_dims(tf.dtypes.cast(mask, K.floatx()), axis=-1), axis=-1)
# return mask
# mask = Lambda(create_mask)(features_controls)
comps = []
for f in fs:
comp = compare_model([x, f])
comps.append(comp)
c = Concatenate()(comps)
c = Reshape((num_ctrl_classes, encoder_model.output_shape[1], 1))(c)
# c = Lambda(lambda x: tf.math.multiply(x[0], x[1]))([c, mask])
# compare = Lambda(compare_features)([x, features_controls])
# Per feature NN with shared weight is implemented using CONV2D with appropriate stride.
compare = Conv2D(mid, (num_ctrl_classes, 1),
activation='relu',
padding='valid')(c)
compare = Reshape((encoder_model.output_shape[1], mid, 1))(compare)
compare = Conv2D(1, (1, mid), activation='linear',
padding='valid')(compare)
compare = Flatten(name='flatten2')(compare)
feature_model = Model(inputs=[inp, features_controls], outputs=compare)
label = Input(shape=(num_classes, ))
output_arcface = ArcFace(num_classes, regularizer=regul)([compare, label])
arcface_model = Model([inp, features_controls, label], output_arcface)
output_cosface = CosFace(num_classes, regularizer=regul)([compare, label])
cosface_model = Model([inp, features_controls, label], output_cosface)
return encoder_model, softmax_model, feature_model, arcface_model, cosface_model
features_num = 10
# Users embedding features.
user_input = Input(shape=[1])
user_embedding = Embedding(len(dataset.user_id.unique()) + 1,
features_num)(user_input)
user_vector = Flatten()(user_embedding)
# Books embedding features.
book_input = Input(shape=[1])
book_embedding = Embedding(len(dataset.book_id.unique()) + 1,
features_num)(book_input)
book_vector = Flatten()(book_embedding)
# Concatenate features.
concatenated = Concatenate()([book_vector, user_vector])
# Create model.
layer1 = Dense(128, activation='relu')(concatenated)
dropout1 = Dropout(0.1)(layer1)
layer2 = Dense(64, activation='relu')(dropout1)
dropout2 = Dropout(0.1)(layer2)
output = Dense(1)(dropout2)
model = Model([user_input, book_input], output)
model.compile('adam', 'mean_squared_error')
# Train.
model.fit([train.user_id, train.book_id],
train.rating,
epochs=15,
batch_size=256)
def __init__(self):
super(MedModel, self).__init__()
# Optimizer
self.optimizer = tf.keras.optimizers.SGD(
learning_rate=hp.learning_rate, momentum=hp.momentum)
# Load instance of small model .h5 (get_appropriate layer and those weight)
small_model = SmallModel()
small_model(tf.keras.Input(shape=(8, 8, 3)))
print(os.getcwd())
small_model.load_weights("./models/small_weights.h5")
self.small_model = small_model
initializer = tf.keras.initializers.Ones()
# Define Model Layers
# First Conv Block
self.med_conv1 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv1")
self.upsamp_small_filters_conv1 = Conv2D(
filters=64,
kernel_size=3,
kernel_initializer=self.small_conv1_init,
padding='SAME',
name='upsamp_small_filters_conv1',
trainable=False)
self.comb_tensors1 = Concatenate(axis=3, name="med_concat1")
self.med_bn1 = BatchNormalization(name="med_bn1")
self.med_relu1 = ReLU(name="med_relu1")
# Second Conv Block
self.med_conv2 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv2")
self.down_med_relu1 = Conv2D(filters=64,
kernel_size=1,
padding='SAME',
activation=None,
kernel_initializer=initializer,
name="reduce_filters",
trainable=False)
self.upsamp_small_filters_conv2 = Conv2D(
filters=64,
kernel_size=3,
kernel_initializer=self.small_conv2_init,
padding='SAME',
name='upsamp_small_filters_conv2',
trainable=False)
self.comb_tensors2 = Concatenate(axis=3, name="med_concat2")
self.med_bn2 = BatchNormalization(name="med_bn2")
self.med_relu2 = ReLU(name="med_relu2")
# Third Conv Block
self.med_conv3 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv3")
self.med_bn3 = BatchNormalization(name="med_bn3")
self.med_relu3 = ReLU(name="med_relu3")
# Fourth Conv Block
self.med_conv4 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv4")
self.med_bn4 = BatchNormalization(name="med_bn4")
self.med_relu4 = ReLU(name="med_relu4")
# Classification Part
self.med_class_conv1 = Conv2D(filters=128,
kernel_size=3,
strides=2,
padding='same',
name="med_class_conv1")
self.med_class_conv2 = Conv2D(filters=128,
kernel_size=3,
strides=2,
padding='same',
name="med_class_conv2")
self.med_class_flatten = Flatten(name="med_class_flatten")
self.med_class_dense = Dense(units=10, activation='softmax')
class MedModel(tf.keras.Model):
def __init__(self):
super(MedModel, self).__init__()
# Optimizer
self.optimizer = tf.keras.optimizers.SGD(
learning_rate=hp.learning_rate, momentum=hp.momentum)
# Load instance of small model .h5 (get_appropriate layer and those weight)
small_model = SmallModel()
small_model(tf.keras.Input(shape=(8, 8, 3)))
print(os.getcwd())
small_model.load_weights("./models/small_weights.h5")
self.small_model = small_model
initializer = tf.keras.initializers.Ones()
# Define Model Layers
# First Conv Block
self.med_conv1 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv1")
self.upsamp_small_filters_conv1 = Conv2D(
filters=64,
kernel_size=3,
kernel_initializer=self.small_conv1_init,
padding='SAME',
name='upsamp_small_filters_conv1',
trainable=False)
self.comb_tensors1 = Concatenate(axis=3, name="med_concat1")
self.med_bn1 = BatchNormalization(name="med_bn1")
self.med_relu1 = ReLU(name="med_relu1")
# Second Conv Block
self.med_conv2 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv2")
self.down_med_relu1 = Conv2D(filters=64,
kernel_size=1,
padding='SAME',
activation=None,
kernel_initializer=initializer,
name="reduce_filters",
trainable=False)
self.upsamp_small_filters_conv2 = Conv2D(
filters=64,
kernel_size=3,
kernel_initializer=self.small_conv2_init,
padding='SAME',
name='upsamp_small_filters_conv2',
trainable=False)
self.comb_tensors2 = Concatenate(axis=3, name="med_concat2")
self.med_bn2 = BatchNormalization(name="med_bn2")
self.med_relu2 = ReLU(name="med_relu2")
# Third Conv Block
self.med_conv3 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv3")
self.med_bn3 = BatchNormalization(name="med_bn3")
self.med_relu3 = ReLU(name="med_relu3")
# Fourth Conv Block
self.med_conv4 = Conv2D(filters=64,
kernel_size=3,
strides=1,
padding='SAME',
activation=None,
name="med_conv4")
self.med_bn4 = BatchNormalization(name="med_bn4")
self.med_relu4 = ReLU(name="med_relu4")
# Classification Part
self.med_class_conv1 = Conv2D(filters=128,
kernel_size=3,
strides=2,
padding='same',
name="med_class_conv1")
self.med_class_conv2 = Conv2D(filters=128,
kernel_size=3,
strides=2,
padding='same',
name="med_class_conv2")
self.med_class_flatten = Flatten(name="med_class_flatten")
self.med_class_dense = Dense(units=10, activation='softmax')
# This function returns small_conv1_filters
def small_conv1_init(self, shape, dtype=None):
small_conv1_filters, biases = self.small_model.get_layer(
"small_conv1").get_weights()
return small_conv1_filters
# This function returns small_conv2_filters
def small_conv2_init(self, shape, dtype=None):
small_conv2_filters, biases = self.small_model.get_layer(
"small_conv2").get_weights()
return small_conv2_filters
def call(self, inputs, training=False):
"""
The call function is inherited from Model. It defines the behaviour of the model.
In this function we will connect the layers we defined in __init__ together.
Please review Connection Scheme and observe naming conventions.
:param inputs: these are the images that are passed in shape (batches, height, width, channels)
:param training: BOOL this is a MODEL param that indicates if we are training or testing... I'm still trying to figure this out...
:return: stuff (softmax class probabilities in this case)
"""
# Connect First Med Conv Block
med_conv1 = self.med_conv1.apply(inputs)
upsamp_small_filters_conv1 = self.upsamp_small_filters_conv1.apply(
inputs)
comb_tensors1 = self.comb_tensors1.apply(
[med_conv1, upsamp_small_filters_conv1])
med_bn1 = self.med_bn1.apply(comb_tensors1)
med_relu1 = self.med_relu1.apply(med_bn1)
# Connect Second Med Conv Block
med_conv2 = self.med_conv2.apply(med_relu1)
down_samp_relu1 = self.down_med_relu1.apply(med_relu1)
upsamp_small_filters_conv2 = self.upsamp_small_filters_conv2.apply(
down_samp_relu1)
comb_tensors2 = self.comb_tensors2.apply(
[med_conv2, upsamp_small_filters_conv2])
med_bn2 = self.med_bn2.apply(comb_tensors2)
med_relu2 = self.med_relu2.apply(med_bn2)
# Connect Third Med Conv Block
med_conv3 = self.med_conv3.apply(med_relu2)
med_bn3 = self.med_bn3.apply(med_conv3)
med_relu3 = self.med_relu3.apply(med_bn3)
# Connect Fourth Med Conv Block
med_conv4 = self.med_conv4.apply(med_relu3)
med_bn4 = self.med_bn4.apply(med_conv4)
med_relu4 = self.med_relu4.apply(med_bn4)
# Connect Small Class Block
med_class_conv1 = self.med_class_conv1.apply(med_relu4)
med_class_conv2 = self.med_class_conv2.apply(med_class_conv1)
med_class_flatten = self.med_class_flatten.apply(med_class_conv2)
med_class_dense = self.med_class_dense.apply(med_class_flatten)
# if training:
# output = med_class_dense
# else:
# #pred = np.argmax(med_class_dense)
# #conf = np.max(med_class_dense)
# #output = [pred, conf]
return med_class_dense
@staticmethod
def loss_fn(labels, predictions):
""" Loss function for model. """
return tf.keras.losses.sparse_categorical_crossentropy(
labels, predictions, from_logits=False)
def identity_block(x,
filters_in,
filters_out,
cardinality=32,
ratio=16,
init_weights=None):
""" Construct a ResNeXT block with identity link
x : input to block
filters_in : number of filters (channels) at the input convolution
filters_out: number of filters (channels) at the output convolution
cardinality: width of cardinality layer
ratio : amount of filter reduction during squeeze
"""
if init_weights is None:
init_weights = SEResNeXt.init_weights
# Remember the input
shortcut = x
# Dimensionality Reduction
x = Conv2D(filters_in,
kernel_size=(1, 1),
strides=(1, 1),
padding='same',
kernel_initializer=init_weights)(shortcut)
x = BatchNormalization()(x)
x = ReLU()(x)
# Cardinality (Wide) Layer (split-transform)
filters_card = filters_in // cardinality
groups = []
for i in range(cardinality):
group = Lambda(lambda z: z[:, :, :, i * filters_card:i *
filters_card + filters_card])(x)
groups.append(
Conv2D(filters_card,
kernel_size=(3, 3),
strides=(1, 1),
padding='same',
kernel_initializer=init_weights)(group))
# Concatenate the outputs of the cardinality layer together (merge)
x = Concatenate()(groups)
x = BatchNormalization()(x)
x = ReLU()(x)
# Dimensionality restoration
x = Conv2D(filters_out,
kernel_size=(1, 1),
strides=(1, 1),
padding='same',
kernel_initializer=init_weights)(x)
x = BatchNormalization()(x)
# Pass the output through the squeeze and excitation block
x = SEResNeXt.squeeze_excite_block(x, ratio, init_weights)
# Identity Link: Add the shortcut (input) to the output of the block
x = Add()([shortcut, x])
x = ReLU()(x)
return x
# Decoder Layer
decoder_layer = GRU(hidden_size,
return_sequences=True,
dropout=dropout,
recurrent_dropout=dropout / 2,
name='decode_layer')
decoder_outputs = decoder_layer(decoder_embedding,
initial_state=encoder_states)
# Attention Layer
attention_layer = Attention(name='attention_layer')
attention_outputs = attention_layer([decoder_outputs, encoder_outputs])
# Concatenate the Result of Attention and the Hidden States of Decoder
decoder_concat_inputs = Concatenate(
axis=-1, name='concatenate_layer')([decoder_outputs, attention_outputs])
# Output Layer
output_layer = Dense(num_of_morphemes, activation=softmax, name='output_layer')
outputs = output_layer(decoder_concat_inputs)
# Define Model
model = Model(inputs=[encoder_inputs, decoder_inputs],
outputs=outputs,
name='training_model')
# Compile
model.compile(optimizer=Adam(learning_rate=learning_late),
loss=sparse_categorical_crossentropy)
# Display Model Summary
def projection_block(x,
filters_in,
filters_out,
cardinality=32,
strides=1,
ratio=16,
init_weights=None):
""" Construct a ResNeXT block with projection shortcut
x : input to the block
filters_in : number of filters (channels) at the input convolution
filters_out: number of filters (channels) at the output convolution
cardinality: width of cardinality layer
strides : whether entry convolution is strided (i.e., (2, 2) vs (1, 1))
ratio : amount of filter reduction during squeeze
"""
if init_weights is None:
init_weights = SEResNeXt.init_weights
# Construct the projection shortcut
# Increase filters by 2X to match shape when added to output of block
shortcut = Conv2D(filters_out,
kernel_size=(1, 1),
strides=strides,
padding='same',
kernel_initializer=init_weights)(x)
shortcut = BatchNormalization()(shortcut)
# Dimensionality Reduction
x = Conv2D(filters_in,
kernel_size=(1, 1),
strides=(1, 1),
padding='same',
kernel_initializer=init_weights)(x)
x = BatchNormalization()(x)
x = ReLU()(x)
# Cardinality (Wide) Layer (split-transform)
filters_card = filters_in // cardinality
groups = []
for i in range(cardinality):
group = Lambda(lambda z: z[:, :, :, i * filters_card:i *
filters_card + filters_card])(x)
groups.append(
Conv2D(filters_card,
kernel_size=(3, 3),
strides=strides,
padding='same',
kernel_initializer=init_weights)(group))
# Concatenate the outputs of the cardinality layer together (merge)
x = Concatenate()(groups)
x = BatchNormalization()(x)
x = ReLU()(x)
# Dimensionality restoration
x = Conv2D(filters_out,
kernel_size=(1, 1),
strides=(1, 1),
padding='same',
kernel_initializer=init_weights)(x)
x = BatchNormalization()(x)
# Pass the output through the squeeze and excitation block
x = SEResNeXt.squeeze_excite_block(x, ratio, init_weights)
# Add the projection shortcut (input) to the output of the block
x = Add()([shortcut, x])
x = ReLU()(x)
return x
# 모델2 input2 = Input(shape=(3, )) dense2 = Dense(10, activation='relu')(input2) dense2 = Dense(5, activation='relu')(dense2) dense2 = Dense(5, activation='relu')(dense2) dense2 = Dense(5, activation='relu')(dense2) # output2 = Dense(3)(dense2) # 모델 병합 / concatenate from tensorflow.keras.layers import Concatenate # from keras.layers.merge import concatenate, Concatenate # from keras.layers import concatenate, Concatenate # merge = 합치다 merge1 = Concatenate()([dense1, dense2]) # 제일 끝의 dense 변수명 넣기 middle1 = Dense(30)(merge1) middle1 = Dense(10)(middle1) middle1 = Dense(10)(middle1) # 모델 분기1 output1 = Dense(30)(middle1) output1 = Dense(7)(output1) output1 = Dense(3)(output1) # 모델 분기2 output2 = Dense(15)(middle1) output2 = Dense(7)(output2) output2 = Dense(7)(output2) output2 = Dense(3)(output2)
def bayes_unet_model_3d_hybrid(self,
filter_root,
depth,
categorical_kccs,
voxel_dim=64,
deviation_channels=3,
output_heads=2):
"""Build the 3D Model using the specified loss function, the inputs are parsed from the assemblyconfig_<case_study_name>.py file
:param voxel_dim: The voxel dimension of the input, required to build input to the 3D CNN model
:type voxel_dim: int (required)
:param voxel_channels: The number of voxel channels in the input structure, required to build input to the 3D CNN model
:type voxel_channels: int (required)
"""
import tensorflow as tf
import tensorflow_probability as tfp
tfd = tfp.distributions
import numpy as np
from tensorflow.keras.models import Model
import tensorflow.keras.backend as K
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Conv3D, MaxPooling3D, Add, BatchNormalization, Input, Activation, Lambda, Concatenate, Flatten, Dense, UpSampling3D, GlobalAveragePooling3D
from tensorflow.keras.utils import plot_model
from tensorflow.keras.layers import add, multiply
from tensorflow.keras.layers import Input
from tensorflow.keras.utils import plot_model
#Testing
#reg_kccs=self.output_dimension-categorical_kccs
#Testing
reg_kccs = 148 - categorical_kccs
# Probabalistic Losses
negloglik = lambda y, rv_y: -rv_y.log_prob(y)
bin_crossentropy = tf.keras.losses.BinaryCrossentropy()
mse_basic = tf.keras.losses.MeanSquaredError()
#Can also Try
#scale_factor=4000/64 #number of samples/batchsize
#kl = sum(model.losses) / scale_factor
#Annealing of KL divergence
#bin_crossentropy=tf.keras.losses.BinaryCrossentropy()+kl* K.get_value(kl_alpha)
overall_loss_dict = {
"regression_outputs": negloglik,
"classification_outputs": bin_crossentropy,
"shape_error_outputs": mse_basic
}
overall_loss_weights = {
"regression_outputs": 2.0,
"classification_outputs": 2.0,
"shape_error_outputs": 1.0
}
overall_metrics_dict = {
"regression_outputs": [tf.keras.metrics.MeanAbsoluteError()],
"classification_outputs": [tf.keras.metrics.CategoricalAccuracy()],
"shape_error_outputs": [tf.keras.metrics.MeanAbsoluteError()]
}
#constant aleatoric uncertainty
aleatoric_std = 0.001
aleatoric_std_cop = 0.001
aleatoric_tensor = [aleatoric_std] * reg_kccs
tfd = tfp.distributions
kl_divergence_function = (
lambda q, p, _: tfd.kl_divergence(q, p) /
(tf.cast(4000, dtype=tf.float32) / tf.cast(64, dtype=tf.float32)))
c = np.log(np.expm1(1.))
long_connection_store = {}
Conv = Conv3D
MaxPooling = MaxPooling3D
UpSampling = UpSampling3D
activation = 'relu'
final_activation = 'linear'
def attention_block(x, g, inter_channel):
theta_x = Conv(inter_channel, [1, 1, 1], strides=[1, 1, 1])(x)
phi_g = Conv(inter_channel, [1, 1, 1], strides=[1, 1, 1])(g)
f = Activation('relu')(add([theta_x, phi_g]))
psi_f = Conv(1, [1, 1, 1], strides=[1, 1, 1])(f)
rate = Activation('sigmoid')(psi_f)
att_x = multiply([x, rate])
return att_x
input_size = (voxel_dim, voxel_dim, voxel_dim, deviation_channels)
inputs = Input(input_size)
x = inputs
# Down sampling
for i in range(depth):
out_channel = 2**i * filter_root
# Residual/Skip connection
res = tfp.layers.Convolution3DFlipout(
out_channel,
kernel_size=1,
kernel_divergence_fn=kl_divergence_function,
padding='same',
name="Identity{}_1".format(i))(x)
# First Conv Block with Conv, BN and activation
conv1 = tfp.layers.Convolution3DFlipout(
out_channel,
kernel_size=3,
kernel_divergence_fn=kl_divergence_function,
padding='same',
name="Conv{}_1".format(i))(x)
#if batch_norm:
#conv1 = BatchNormalization(name="BN{}_1".format(i))(conv1)
act1 = Activation(activation, name="Act{}_1".format(i))(conv1)
# Second Conv block with Conv and BN only
conv2 = tfp.layers.Convolution3DFlipout(
out_channel,
kernel_size=3,
padding='same',
kernel_divergence_fn=kl_divergence_function,
name="Conv{}_2".format(i))(act1)
#if batch_norm:
#conv2 = BatchNormalization(name="BN{}_2".format(i))(conv2)
resconnection = Add(name="Add{}_1".format(i))([res, conv2])
act2 = Activation(activation,
name="Act{}_2".format(i))(resconnection)
# Max pooling
if i < depth - 1:
long_connection_store[str(i)] = act2
x = MaxPooling(padding='same',
name="MaxPooling{}_1".format(i))(act2)
else:
x = act2
feature_vector_reg = Conv(reg_kccs,
1,
padding='same',
activation=final_activation,
name='Process_Parameter_Reg_output')(x)
process_parameter_reg = GlobalAveragePooling3D()(feature_vector_reg)
#process_parameter_reg=tfp.layers.DenseFlipout(64,kernel_divergence_fn=kl_divergence_function)(process_parameter_reg)
#process_parameter_reg=tfp.layers.DenseFlipout(123,kernel_divergence_fn=kl_divergence_function)(process_parameter_reg)
feature_vector_cla = Conv(categorical_kccs,
1,
padding='same',
activation=final_activation,
name='Process_Parameter_Cla_output')(x)
process_parameter_cla = GlobalAveragePooling3D()(feature_vector_cla)
#process_parameter_cla=tfp.layers.DenseFlipout(64,kernel_divergence_fn=kl_divergence_function)(process_parameter_cla)
#process_parameter_cla=tfp.layers.DenseFlipout(25,kernel_divergence_fn=kl_divergence_function)(process_parameter_cla)
#feature_categorical=Flatten()(feature_vector)
#reg_output=tfp.layers.DenseFlipout(output_dimension,kernel_divergence_fn=kl_divergence_function)(process_parameter)
#Process Parameter Outputs
reg_distrbution = tfp.layers.DistributionLambda(
lambda t: tfd.MultivariateNormalDiag(loc=t[..., :reg_kccs],
scale_diag=aleatoric_tensor),
name="regression_outputs")(process_parameter_reg)
cla_distrbution = Activation(
'sigmoid', name="classification_outputs")(process_parameter_cla)
#cla_distrbution=tfp.layers.DenseFlipout(categorical_kccs, kernel_divergence_fn=kl_divergence_function,activation=tf.nn.sigmoid,name="classification_outputs")(process_parameter_cla)
# Upsampling
for i in range(depth - 2, -1, -1):
out_channel = 2**(i) * filter_root
# long connection from down sampling path.
long_connection = long_connection_store[str(i)]
up1 = UpSampling(name="UpSampling{}_1".format(i))(x)
up_conv1 = Conv(out_channel,
2,
activation='relu',
padding='same',
name="upConvSam{}_1".format(i))(up1)
attention_layer = attention_block(x=long_connection,
g=up_conv1,
inter_channel=out_channel // 4)
# Concatenate.
#up_conc = Concatenate(axis=-1, name="upConcatenate{}_1".format(i))([up_conv1, long_connection])
up_conc = Concatenate(axis=-1, name="upConcatenate{}_1".format(i))(
[up_conv1, attention_layer])
# Convolutions
up_conv2 = Conv(out_channel,
3,
padding='same',
name="upConv{}_1".format(i))(up_conc)
up_act1 = Activation(activation,
name="upAct{}_1".format(i))(up_conv2)
up_conv2 = Conv(out_channel,
3,
padding='same',
name="upConv{}_2".format(i))(up_act1)
# Residual/Skip connection
res = Conv(out_channel,
kernel_size=1,
padding='same',
use_bias=False,
name="upIdentity{}_1".format(i))(up_conc)
resconnection = Add(name="upAdd{}_1".format(i))([res, up_conv2])
x = Activation(activation,
name="upAct{}_2".format(i))(resconnection)
output_list = []
output_list.append(reg_distrbution)
output_list.append(cla_distrbution)
output = Conv(deviation_channels * output_heads,
1,
padding='same',
activation=final_activation,
name='shape_error_outputs')(x)
output_list.append(output)
model = Model(inputs, outputs=output_list, name='Hybrid_Unet_Model')
#Loss Dictonary Created
#model_losses=[]
#model_losses.append(negloglik)
#model_losses.append(bin_crossentropy)
model.compile(optimizer=tf.keras.optimizers.Adam(lr=0.001),
experimental_run_tf_function=False,
loss=overall_loss_dict,
metrics=overall_metrics_dict,
loss_weights=overall_loss_weights)
print("3D CNN model successfully compiled")
print(model.summary())
#plot_model(model,to_file='Bayes_OSER2.png',show_shapes=True, show_layer_names=True)
return model
def u_net_module_hp(hp):
############################## Special Fun ###################################################
# embedding = tf.Variable(initial_value = np.load('../Embedding/embedd_w_5_ogt.npy').reshape(1,21,5))
############################## Transformation module #################################################
p = {
'n_fil_1': hp.Choice('number_filters_conv1', [8, 16, 32, 64]),
's_fil_1': hp.Choice('size_filters_conv1', [3, 5, 7, 10]),
'stride_1': hp.Choice('stride_length_sampling1', [2, 4]),
'dOut_1': hp.Choice('Dropout_module1', [0.1, 0.2, 0.3, 0.4]),
'n_fil_2': hp.Choice('number_filters_conv2', [32, 64, 128]),
's_fil_2': hp.Choice('size_filters_conv2', [3, 5, 7, 10]),
'stride_2': hp.Choice('stride_length_sampling2', [2, 4]),
'dOut_2': hp.Choice('Dropout_module2', [0.1, 0.2, 0.3, 0.4]),
'n_fil_3': hp.Choice('number_filters_conv3', [64, 128, 256]),
's_fil_3': hp.Choice('size_filters_conv3', [3, 5, 7, 10]),
'stride_3': hp.Choice('stride_length_sampling3', [2, 4]),
'dOut_3': hp.Choice('Dropout_module3', [0.1, 0.2, 0.3, 0.4]),
'n_fil_4': hp.Choice('number_filters_conv4', [128, 256, 512]),
's_fil_4': hp.Choice('size_filters_conv4', [3, 5, 7, 10]),
'dOut_4': hp.Choice('Dropout_module4', [0.1, 0.2, 0.3, 0.4]),
'dOut_5': hp.Choice('Dropout_module5', [0.05, 0.1, 0.2, 0.3]),
's_fil_5': hp.Choice('size_filters_conv5', [3, 5, 7, 10])
}
# Layers of stage 0 contraction
inp = Input(shape=(1024, 21))
#tct0_in = Dot(axes=(2,1))([inp,embedding])
tct0_bn1 = BatchNormalization()(inp) #tct0_in)
#tf.keras.regularizers.L1L2(l1=0.0, l2=0.0)
tct0_conv1 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Ct0_0')(tct0_bn1)
tct0_bn2 = BatchNormalization()(tct0_conv1)
tct0_conv2 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Ct0_1')(tct0_bn2)
tct0_bn3 = BatchNormalization()(tct0_conv2)
tct0_conv3 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
strides=int(p['stride_1']),
padding='same',
name='Convolution_Ct0_2')(tct0_bn3)
tct0_bn4 = BatchNormalization()(tct0_conv3)
#tct0_max = MaxPool1D(pool_size=2, strides=2)(tct0_bn2)
tct0_dp = Dropout(p['dOut_1'])(tct0_bn4)
# Layers of stage 1 contraction
tct1_conv1 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
padding='same',
name='Convolution_Ct1_0')(tct0_dp)
tct1_bn1 = BatchNormalization()(tct1_conv1)
tct1_conv2 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct1_1')(tct1_bn1)
tct1_bn2 = BatchNormalization()(tct1_conv2)
tct1_conv3 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
strides=int(p['stride_2']),
padding='same',
name='Convolution_Ct1_2')(tct1_bn2)
tct1_bn3 = BatchNormalization()(tct1_conv3)
#tct1_max = MaxPool1D(pool_size=2, strides=2)(tct1_bn2)
tct1_dp = Dropout(p['dOut_2'])(tct1_bn3)
# Layers of stage 2 contraction
tct2_conv1 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
padding='same',
name='Convolution_Ct2_0')(tct1_dp)
tct2_bn1 = BatchNormalization()(tct2_conv1)
tct2_conv2 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct2_1')(tct2_bn1)
tct2_bn2 = BatchNormalization()(tct2_conv2)
tct2_conv3 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
strides=int(p['stride_3']),
padding='same',
name='Convolution_Ct2_2')(tct2_bn2)
tct2_bn3 = BatchNormalization()(tct2_conv3)
#tct2_max = MaxPool1D(pool_size=2, strides=2)(tct2_bn2)
tct2_dp = Dropout(p['dOut_3'])(tct2_bn3)
# Layers of stage 3 contraction
tct3_conv1 = Conv1D(int(p['n_fil_4']),
int(p['s_fil_4']),
activation='relu',
padding='same',
name='Convolution_Ce3_0')(tct2_dp)
tct3_bn1 = BatchNormalization()(tct3_conv1)
tct3_conv2 = Conv1D(int(p['n_fil_4']),
int(p['s_fil_4']),
activation='relu',
padding='same',
name='Convolution_Ce3_1')(tct3_bn1)
tct3_bn2 = BatchNormalization()(tct3_conv2)
tct3_dp = Dropout(p['dOut_4'])(tct3_bn2)
# Layers of stage 1 expansion
tet1_Tconv = Conv1DTranspose(int(p['n_fil_3']),
int(p['s_fil_3']),
strides=int(p['stride_3']),
activation='relu',
padding='same',
name='TransConv_Et1')(tct3_dp)
tet1_Concat = Concatenate(axis=2)([tet1_Tconv, tct2_conv1])
tet1_bn1 = BatchNormalization()(tet1_Concat)
tet1_conv1 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
padding='same',
name='Convolution_Et1_0')(tet1_bn1)
tet1_bn2 = BatchNormalization()(tet1_conv1)
tet1_conv2 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
padding='same',
name='Convolution_Et1_1')(tet1_bn2)
tet1_bn3 = BatchNormalization()(tet1_conv2)
tet1_dp = Dropout(p['dOut_3'])(tet1_bn3)
#Layers of stage 2 expansion
tet2_Tconv = Conv1DTranspose(int(p['n_fil_2']),
int(p['s_fil_2']),
strides=int(p['stride_2']),
activation='relu',
padding='same',
name='TransConv_Et2')(tet1_dp)
tet2_Concat = Concatenate(axis=2)([tet2_Tconv, tct1_conv1])
tet2_bn1 = BatchNormalization()(tet2_Concat)
tet2_conv1 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
padding='same',
name='Convolution_Et2_0')(tet2_bn1)
tet2_bn2 = BatchNormalization()(tet2_conv1)
tet2_conv2 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
padding='same',
name='Convolution_Et2_1')(tet2_bn2)
tet2_bn3 = BatchNormalization()(tet2_conv2)
tet2_dp = Dropout(p['dOut_2'])(tet2_bn3)
#Layers of stage 3 expansion
tet3_Tconv = Conv1DTranspose(int(p['n_fil_1']),
int(p['s_fil_1']),
strides=int(p['stride_1']),
activation='relu',
padding='same',
name='TransConv_Et3')(tet2_dp)
tet3_Concat = Concatenate(axis=2)([tet3_Tconv, tct0_conv1])
tet3_bn1 = BatchNormalization()(tet3_Concat)
tet3_conv1 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Et3_1')(tet3_bn1)
tet3_bn2 = BatchNormalization()(tet3_conv1)
tet3_conv2 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Et3_2')(tet3_bn2)
tet3_bn3 = BatchNormalization()(tet3_conv2)
tet3_dp = Dropout(p['dOut_5'])(tet3_bn3)
tet3_conv3 = Conv1D(3,
int(p['s_fil_5']),
activation='softmax',
padding='same',
name='Convolution_Et3_3')(tet3_dp)
model = Model(inputs=inp, outputs=tet3_conv3)
cce = tf.keras.losses.CategoricalCrossentropy(
from_logits=True) #, sample_weight = )
add = tf.keras.optimizers.Adam(learning_rate=hp.Choice(
'learning_rate', [1e-2, 1e-3, 1e-4]),
name='Adam')
model.compile(optimizer=add, loss=cce, metrics=['accuracy'])
return model
def yolo4lite_mobilenet_body(inputs, num_anchors, num_classes, alpha=1.0):
'''Create YOLO_v4 Lite MobileNet model CNN body in keras.'''
mobilenet = MobileNet(input_tensor=inputs,
weights='imagenet',
include_top=False,
alpha=alpha)
# input: 416 x 416 x 3
# conv_pw_13_relu :13 x 13 x (1024*alpha)
# conv_pw_11_relu :26 x 26 x (512*alpha)
# conv_pw_5_relu : 52 x 52 x (256*alpha)
# f1: 13 x 13 x (1024*alpha) for 416 input
f1 = mobilenet.get_layer('conv_pw_13_relu').output
# f2: 26 x 26 x (512*alpha) for 416 input
f2 = mobilenet.get_layer('conv_pw_11_relu').output
# f3: 52 x 52 x (256*alpha) for 416 input
f3 = mobilenet.get_layer('conv_pw_5_relu').output
f1_channel_num = int(1024 * alpha)
f2_channel_num = int(512 * alpha)
f3_channel_num = int(256 * alpha)
#feature map 1 head (13 x 13 x (512*alpha) for 416 input)
x1 = make_yolo_spp_depthwise_separable_head(f1,
f1_channel_num // 2,
block_id_str='14')
#upsample fpn merge for feature map 1 & 2
x1_upsample = compose(DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1)),
UpSampling2D(2))(x1)
x2 = DarknetConv2D_BN_Leaky(f2_channel_num // 2, (1, 1))(f2)
x2 = Concatenate()([x2, x1_upsample])
#feature map 2 head (26 x 26 x (256*alpha) for 416 input)
x2 = make_yolo_depthwise_separable_head(x2,
f2_channel_num // 2,
block_id_str='15')
#upsample fpn merge for feature map 2 & 3
x2_upsample = compose(DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1)),
UpSampling2D(2))(x2)
x3 = DarknetConv2D_BN_Leaky(f3_channel_num // 2, (1, 1))(f3)
x3 = Concatenate()([x3, x2_upsample])
#feature map 3 head & output (52 x 52 x (256*alpha) for 416 input)
#x3, y3 = make_depthwise_separable_last_layers(x3, f3_channel_num//2, num_anchors*(num_classes+5), block_id_str='16')
x3 = make_yolo_depthwise_separable_head(x3,
f3_channel_num // 2,
block_id_str='16')
y3 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(f3_channel_num, (3, 3),
block_id_str='16_3'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x3)
#downsample fpn merge for feature map 3 & 2
x3_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
Darknet_Depthwise_Separable_Conv2D_BN_Leaky(f2_channel_num // 2,
(3, 3),
strides=(2, 2),
block_id_str='16_4'))(x3)
x2 = Concatenate()([x3_downsample, x2])
#feature map 2 output (26 x 26 x (512*alpha) for 416 input)
#x2, y2 = make_depthwise_separable_last_layers(x2, f2_channel_num//2, num_anchors*(num_classes+5), block_id_str='17')
x2 = make_yolo_depthwise_separable_head(x2,
f2_channel_num // 2,
block_id_str='17')
y2 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(f2_channel_num, (3, 3),
block_id_str='17_3'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x2)
#downsample fpn merge for feature map 2 & 1
x2_downsample = compose(
ZeroPadding2D(((1, 0), (1, 0))),
Darknet_Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num // 2,
(3, 3),
strides=(2, 2),
block_id_str='17_4'))(x2)
x1 = Concatenate()([x2_downsample, x1])
#feature map 1 output (13 x 13 x (1024*alpha) for 416 input)
#x1, y1 = make_depthwise_separable_last_layers(x1, f1_channel_num//2, num_anchors*(num_classes+5), block_id_str='18')
x1 = make_yolo_depthwise_separable_head(x1,
f1_channel_num // 2,
block_id_str='18')
y1 = compose(
Depthwise_Separable_Conv2D_BN_Leaky(f1_channel_num, (3, 3),
block_id_str='18_3'),
DarknetConv2D(num_anchors * (num_classes + 5), (1, 1)))(x1)
return Model(inputs, [y1, y2, y3])
def u_net_module(p):
############################## Special Fun ###################################################
# embedding = tf.Variable(initial_value = np.load('../Embedding/embedd_w_5_ogt.npy').reshape(1,21,5))
############################## Transformation module #################################################
# Layers of stage 0 contraction
inp = Input(shape=(1024, ))
#w = Input(shape=(1024,))
#tct0_in = Dot(axes=(2,1))([inp,embedding])
#tct0_bn1 = BatchNormalization()(inp)#tct0_in)
#tf.keras.regularizers.L1L2(l1=0.0, l2=0.0)
x = Embedding(22, 5)(inp)
x = BatchNormalization()(x)
tct0_conv1 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Ct0_0')(x)
tct0_bn2 = BatchNormalization()(tct0_conv1)
tct0_conv2 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Ct0_1')(tct0_bn2)
tct0_bn3 = BatchNormalization()(tct0_conv2)
tct0_conv3 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
strides=int(p['stride_1']),
padding='same',
name='Convolution_Ct0_2')(tct0_bn3)
tct0_bn4 = BatchNormalization()(tct0_conv3)
#tct0_max = MaxPool1D(pool_size=2, strides=2)(tct0_bn2)
tct0_dp = Dropout(p['dOut_1'])(tct0_bn4)
# Layers of stage 1 contraction
tct1_conv1 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
padding='same',
name='Convolution_Ct1_0')(tct0_dp)
tct1_bn1 = BatchNormalization()(tct1_conv1)
tct1_conv2 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct1_1')(tct1_bn1)
tct1_bn2 = BatchNormalization()(tct1_conv2)
tct1_conv3 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
strides=int(p['stride_2']),
padding='same',
name='Convolution_Ct1_2')(tct1_bn2)
tct1_bn3 = BatchNormalization()(tct1_conv3)
#tct1_max = MaxPool1D(pool_size=2, strides=2)(tct1_bn2)
tct1_dp = Dropout(p['dOut_2'])(tct1_bn3)
# Layers of stage 2 contraction
tct2_conv1 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
padding='same',
name='Convolution_Ct2_0')(tct1_dp)
tct2_bn1 = BatchNormalization()(tct2_conv1)
tct2_conv2 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct2_1')(tct2_bn1)
tct2_bn2 = BatchNormalization()(tct2_conv2)
tct2_conv3 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
strides=int(p['stride_3']),
padding='same',
name='Convolution_Ct2_2')(tct2_bn2)
tct2_bn3 = BatchNormalization()(tct2_conv3)
#tct2_max = MaxPool1D(pool_size=2, strides=2)(tct2_bn2)
tct2_dp = Dropout(p['dOut_3'])(tct2_bn3)
# Layers of stage 3 contraction
tct3_conv1 = Conv1D(int(p['n_fil_4']),
int(p['s_fil_4']),
activation='relu',
padding='same',
name='Convolution_Ce3_0')(tct2_dp)
tct3_bn1 = BatchNormalization()(tct3_conv1)
tct3_conv2 = Conv1D(int(p['n_fil_4']),
int(p['s_fil_4']),
activation='relu',
padding='same',
name='Convolution_Ce3_1')(tct3_bn1)
tct3_bn2 = BatchNormalization()(tct3_conv2)
tct3_dp = Dropout(p['dOut_4'])(tct3_bn2)
# Layers of stage 1 expansion
tet1_Tconv = Conv1DTranspose(int(p['n_fil_3']),
int(p['s_fil_3']),
strides=int(p['stride_3']),
name='TransConv_Et1')(tct3_dp)
tet1_Concat = Concatenate(axis=2)([tet1_Tconv, tct2_conv1])
tet1_bn1 = BatchNormalization()(tet1_Concat)
tet1_conv1 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
padding='same',
name='Convolution_Et1_0')(tet1_bn1)
tet1_bn2 = BatchNormalization()(tet1_conv1)
tet1_conv2 = Conv1D(int(p['n_fil_3']),
int(p['s_fil_3']),
activation='relu',
padding='same',
name='Convolution_Et1_1')(tet1_bn2)
tet1_bn3 = BatchNormalization()(tet1_conv2)
tet1_dp = Dropout(p['dOut_3'])(tet1_bn3)
#Layers of stage 2 expansion
tet2_Tconv = Conv1DTranspose(int(p['n_fil_2']),
int(p['s_fil_2']),
strides=int(p['stride_2']),
name='TransConv_Et2')(tet1_dp)
tet2_Concat = Concatenate(axis=2)([tet2_Tconv, tct1_conv1])
tet2_bn1 = BatchNormalization()(tet2_Concat)
tet2_conv1 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
padding='same',
name='Convolution_Et2_0')(tet2_bn1)
tet2_bn2 = BatchNormalization()(tet2_conv1)
tet2_conv2 = Conv1D(int(p['n_fil_2']),
int(p['s_fil_2']),
activation='relu',
padding='same',
name='Convolution_Et2_1')(tet2_bn2)
tet2_bn3 = BatchNormalization()(tet2_conv2)
tet2_dp = Dropout(p['dOut_2'])(tet2_bn3)
#Layers of stage 3 expansion
tet3_Tconv = Conv1DTranspose(int(p['n_fil_1']),
int(p['s_fil_1']),
strides=int(p['stride_1']),
name='TransConv_Et3')(tet2_dp)
tet3_Concat = Concatenate(axis=2)([tet3_Tconv, tct0_conv1])
tet3_bn1 = BatchNormalization()(tet3_Concat)
tet3_conv1 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Et3_1')(tet3_bn1)
tet3_bn2 = BatchNormalization()(tet3_conv1)
tet3_conv2 = Conv1D(int(p['n_fil_1']),
int(p['s_fil_1']),
activation='relu',
padding='same',
name='Convolution_Et3_2')(tet3_bn2)
tet3_bn3 = BatchNormalization()(tet3_conv2)
tet3_dp = Dropout(p['dOut_5'])(tet3_bn3)
tet3_conv3 = Conv1D(3,
int(p['s_fil_5']),
activation='softmax',
padding='same',
name='Convolution_Et3_3')(tet3_dp)
cce = tf.keras.losses.CategoricalCrossentropy(
from_logits=True) # , sample_weight = w)
add = tf.keras.optimizers.Adam(learning_rate=1e-2, name='Adam')
model = Model(inputs=inp, outputs=tet3_conv3)
model.compile(optimizer=add,
loss=cce,
metrics=['accuracy'],
sample_weight_mode="temporal")
return model
def tiny_unit(telescope,
image_mode,
image_mask,
input_img_shape,
input_features_shape,
targets,
target_mode,
target_shapes=None,
latent_variables=64,
conv_layers_blocks=2,
dense_layer_blocks=3,
ignore_telescopes=False,
activity_regularizer_l2=None):
"""Build Deterministic CNN Unit Model
Parameters
==========
telescope
image_mode
image_mask
input_img_shape
input_features_shape
output_mode
output_shape
Return
======
keras.Model
"""
# Soport lineal only
if target_mode != 'lineal':
raise ValueError(f"Invalid target_mode: '{target_mode}'")
# Image Encoding Block
## HexConvLayer
input_img = Input(name="image_input", shape=input_img_shape)
if image_mode == "simple-shift":
front = HexConvLayer(filters=32,
kernel_size=(3, 3),
name="encoder_hex_conv_layer")(input_img)
elif image_mode == "simple":
front = Conv2D(name="encoder_conv_layer_0",
filters=32,
kernel_size=(3, 3),
kernel_initializer="he_uniform",
padding="valid",
activation="relu")(input_img)
front = MaxPooling2D(name=f"encoder_max_poolin_layer_0",
pool_size=(2, 2))(front)
else:
raise ValueError(f"Invalid image mode {image_mode}")
## convolutional layers
conv_kernel_sizes = [3] * conv_layers_blocks
filters = 32
for i, kernel_size in enumerate(conv_kernel_sizes, start=1):
front = Conv2D(name=f"encoder_conv_layer_{i}_a",
filters=filters,
kernel_size=kernel_size,
kernel_initializer="he_uniform",
padding="same")(front)
front = Activation(name=f"encoder_ReLU_{i}_a",
activation="relu")(front)
front = BatchNormalization(name=f"encoder_batchnorm_{i}_a")(front)
front = Conv2D(name=f"encoder_conv_layer_{i}_b",
filters=filters,
kernel_size=kernel_size,
kernel_initializer="he_uniform",
padding="same")(front)
front = Activation(name=f"encoder_ReLU_{i}_b",
activation="relu")(front)
front = BatchNormalization(name=f"encoder_batchnorm_{i}_b")(front)
front = MaxPooling2D(name=f"encoder_maxpool_layer_{i}",
pool_size=(2, 2))(front)
filters *= 2
front = Flatten(name="encoder_flatten_to_latent")(front)
# Logic Block
## extra Telescope Features
input_params = Input(name="feature_input", shape=input_features_shape)
if ignore_telescopes:
input_params = Lambda(lambda x: x * 0)(input_params)
front = Concatenate()([input_params, front])
## dense blocks
l2_ = lambda activity_regularizer_l2: None if activity_regularizer_l2 is None else l2(
activity_regularizer_l2)
for dense_i in range(dense_layer_blocks):
front = Dense(name=f"logic_dense_{dense_i}",
units=latent_variables,
kernel_regularizer=l2_(activity_regularizer_l2))(front)
front = Activation(name=f"logic_ReLU_{dense_i}",
activation="relu")(front)
front = BatchNormalization(name=f"logic_batchnorm_{dense_i}")(front)
# Outpout block
output = Dense(len(targets), activation="linear")(front)
model_name = f"Tiny_Unit_{telescope}"
model = Model(name=model_name,
inputs=[input_img, input_params],
outputs=output)
return model
def u_net_module_2(in_):
############################## Special Fun ###################################################
embedding = tf.Variable(
initial_value=np.load('../Embedding/embedd_w_5.npy').reshape(1, 21, 5),
trainable=True)
############################## Transformation module #################################################
# Layers of stage 0 contraction
inp = InputLayer(input_shape=(in_, 21))
tct0_in = Dot(axes=(2, 1))([inp, embedding])
tct0_bn1 = BatchNormalization()(tct0_in)
tct0_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_0')(tct0_bn1)
tct0_bn2 = BatchNormalization()(tct0_conv1)
tct0_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_1')(tct0_bn2)
tct0_bn3 = BatchNormalization()(tct0_conv2)
tct0_conv3 = Conv1D(32,
5,
activation='relu',
strides=2,
padding='same',
name='Convolution_Ct0_2')(tct0_bn3)
tct0_bn4 = BatchNormalization()(tct0_conv3)
#tct0_max = MaxPool1D(pool_size=2, strides=2)(tct0_bn2)
tct0_dp = Dropout(0.2)(tct0_bn4)
# Layers of stage 1 contraction
tct1_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct1_0')(tct0_dp)
tct1_bn1 = BatchNormalization()(tct1_conv1)
tct1_conv2 = Conv1D(32,
5,
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct1_1')(tct1_bn1)
tct1_bn2 = BatchNormalization()(tct1_conv2)
tct1_conv3 = Conv1D(32,
5,
activation='relu',
strides=2,
padding='same',
name='Convolution_Ct1_2')(tct1_bn2)
tct1_bn3 = BatchNormalization()(tct1_conv3)
#tct1_max = MaxPool1D(pool_size=2, strides=2)(tct1_bn2)
tct1_dp = Dropout(0.2)(tct1_bn3)
# Layers of stage 2 contraction
tct2_conv1 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Ct2_0')(tct1_dp)
tct2_bn1 = BatchNormalization()(tct2_conv1)
tct2_conv2 = Conv1D(64,
5,
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct2_1')(tct2_bn1)
tct2_bn2 = BatchNormalization()(tct2_conv2)
tct2_conv3 = Conv1D(64,
5,
activation='relu',
strides=2,
padding='same',
name='Convolution_Ct2_2')(tct2_bn2)
tct2_bn3 = BatchNormalization()(tct2_conv3)
#tct2_max = MaxPool1D(pool_size=2, strides=2)(tct2_bn2)
tct2_dp = Dropout(0.2)(tct2_bn3)
# Layers of stage 3 contraction
tct3_conv1 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_0')(tct2_dp)
tct3_bn1 = BatchNormalization()(tct3_conv1)
tct3_conv2 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_1')(tct3_bn1)
tct3_bn2 = BatchNormalization()(tct3_conv2)
tct3_dp = Dropout(0.2)(tct3_bn2)
# Layers of stage 1 expansion
tet1_Tconv = Conv1DTranspose(64,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et1')(tct3_dp)
tet1_Concat = Concatenate(axis=2)([tet1_Tconv, tct2_conv1])
tet1_bn1 = BatchNormalization()(tet1_Concat)
tet1_conv1 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_0')(tet1_bn1)
tet1_bn2 = BatchNormalization()(tet1_conv1)
tet1_conv2 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_1')(tet1_bn2)
tet1_bn3 = BatchNormalization()(tet1_conv2)
tet1_dp = Dropout(0.2)(tet1_bn3)
#Layers of stage 2 expansion
tet2_Tconv = Conv1DTranspose(32,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et2')(tet1_dp)
tet2_Concat = Concatenate(axis=2)([tet2_Tconv, tct1_conv1])
tet2_bn1 = BatchNormalization()(tet2_Concat)
tet2_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_0')(tet2_bn1)
tet2_bn2 = BatchNormalization()(tet2_conv1)
tet2_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_1')(tet2_bn2)
tet2_bn3 = BatchNormalization()(tet2_conv2)
tet2_dp = Dropout(0.2)(tet2_bn3)
#Layers of stage 3 expansion
tet3_Tconv = Conv1DTranspose(32,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et3')(tet2_dp)
tet3_Concat = Concatenate(axis=2)([tet3_Tconv, tct0_conv1])
tet3_bn1 = BatchNormalization()(tet3_Concat)
tet3_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et3_1')(tet3_bn1)
tet3_bn2 = BatchNormalization()(tet3_conv1)
tet3_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et3_2')(tet3_bn2)
tet3_bn3 = BatchNormalization()(tet3_conv2)
tet3_dp = Dropout(0.1)(tet3_bn3)
tet3_conv3 = Conv1D(3,
5,
activation='softmax',
padding='same',
name='Convolution_Et3_3')(tet3_dp)
################################################ Stage 2 ######################################################
tet3_Concat2 = Concatenate(axis=2)([tct0_in, tet3_conv3])
tct0_bn12 = BatchNormalization()(tet3_Concat2)
tct0_conv12 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_02')(tct0_bn12)
tct0_bn22 = BatchNormalization()(tct0_conv12)
tct0_conv22 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_12')(tct0_bn22)
tct0_bn32 = BatchNormalization()(tct0_conv22)
tct0_conv32 = Conv1D(32,
5,
activation='relu',
strides=2,
padding='same',
name='Convolution_Ct0_22')(tct0_bn32)
tct0_bn42 = BatchNormalization()(tct0_conv32)
#tct0_max = MaxPool1D(pool_size=2, strides=2)(tct0_bn2)
tct0_dp2 = Dropout(0.2)(tct0_bn42)
# Layers of stage 1 contraction
tct1_conv12 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct1_02')(tct0_dp2)
tct1_bn12 = BatchNormalization()(tct1_conv12)
tct1_conv22 = Conv1D(32,
5,
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct1_12')(tct1_bn12)
tct1_bn22 = BatchNormalization()(tct1_conv22)
tct1_conv32 = Conv1D(32,
5,
activation='relu',
strides=2,
padding='same',
name='Convolution_Ct1_22')(tct1_bn22)
tct1_bn32 = BatchNormalization()(tct1_conv32)
#tct1_max = MaxPool1D(pool_size=2, strides=2)(tct1_bn2)
tct1_dp2 = Dropout(0.2)(tct1_bn32)
# Layers of stage 2 contraction
tct2_conv12 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Ct2_02')(tct1_dp2)
tct2_bn12 = BatchNormalization()(tct2_conv12)
tct2_conv22 = Conv1D(64,
5,
activation='relu',
strides=1,
padding='same',
name='Convolution_Ct2_12')(tct2_bn12)
tct2_bn22 = BatchNormalization()(tct2_conv22)
tct2_conv32 = Conv1D(64,
5,
activation='relu',
strides=2,
padding='same',
name='Convolution_Ct2_22')(tct2_bn22)
tct2_bn32 = BatchNormalization()(tct2_conv32)
#tct2_max = MaxPool1D(pool_size=2, strides=2)(tct2_bn2)
tct2_dp2 = Dropout(0.2)(tct2_bn32)
# Layers of stage 3 contraction
tct3_conv12 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_02')(tct2_dp2)
tct3_bn12 = BatchNormalization()(tct3_conv12)
tct3_conv22 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_12')(tct3_bn12)
tct3_bn22 = BatchNormalization()(tct3_conv22)
tct3_dp2 = Dropout(0.2)(tct3_bn22)
# Layers of stage 1 expansion
tet1_Tconv2 = Conv1DTranspose(64,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et12')(tct3_dp2)
tet1_Concat2 = Concatenate(axis=2)([tet1_Tconv2, tct2_conv12])
tet1_bn12 = BatchNormalization()(tet1_Concat2)
tet1_conv12 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_02')(tet1_bn12)
tet1_bn22 = BatchNormalization()(tet1_conv12)
tet1_conv22 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_12')(tet1_bn22)
tet1_bn32 = BatchNormalization()(tet1_conv22)
tet1_dp2 = Dropout(0.2)(tet1_bn32)
#Layers of stage 2 expansion
tet2_Tconv2 = Conv1DTranspose(32,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et22')(tet1_dp2)
tet2_Concat2 = Concatenate(axis=2)([tet2_Tconv2, tct1_conv12])
tet2_bn12 = BatchNormalization()(tet2_Concat2)
tet2_conv12 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_02')(tet2_bn12)
tet2_bn22 = BatchNormalization()(tet2_conv12)
tet2_conv22 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_12')(tet2_bn22)
tet2_bn32 = BatchNormalization()(tet2_conv22)
tet2_dp2 = Dropout(0.2)(tet2_bn32)
#Layers of stage 3 expansion
tet3_Tconv2 = Conv1DTranspose(32,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et32')(tet2_dp2)
tet3_Concat2 = Concatenate(axis=2)([tet3_Tconv2, tct0_conv12])
tet3_bn12 = BatchNormalization()(tet3_Concat2)
tet3_conv12 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et3_12')(tet3_bn12)
tet3_bn22 = BatchNormalization()(tet3_conv12)
tet3_conv22 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et3_22')(tet3_bn22)
tet3_bn32 = BatchNormalization()(tet3_conv22)
tet3_dp2 = Dropout(0.1)(tet3_bn32)
tet3_conv32 = Conv1D(3,
5,
activation='softmax',
padding='same',
name='Convolution_Et3_32')(tet3_dp2)
model = Model(inputs=inp, outputs=[tet3_conv3, tet3_conv32])
return model
x2, y2, y3, train_size=0.6, shuffle=False) m1_input = Input(shape=(3, ), name='model1_input') m1_dense1 = Dense(50, activation='relu', name='m1_d1')(m1_input) m1_dense2 = Dense(45, activation='relu', name='m1_d2')(m1_dense1) m1_dense3 = Dense(40, activation='relu', name='m1_d3')(m1_dense2) m1_output = Dense(35, name='m1_out')(m1_dense3) m2_input = Input(shape=(3, )) m2_dense1 = Dense(50, activation='relu', name='m2_d1')(m2_input) m2_dense2 = Dense(45, activation='relu', name='m2_d2')(m2_dense1) m2_dense3 = Dense(40, activation='relu', name='m2_d3')(m2_dense2) m2_output = Dense(35, name='m2_out')(m2_dense3) #model의 병합 merge1 = Concatenate()([m1_output, m2_output]) # merge1 = Concatenate(axis=1)([m1_output, m2_output]) # merge1 = Concatenate(axis=0)([m1_output, m2_output]) middle1 = Dense(33, name='middle_d1')(merge1) middle2 = Dense(31, name='middle_d2')(middle1) middle3 = Dense(29, name='middle_d3')(middle2) #output 모델 output1 = Dense(25, name='out1_d1')(middle3) output1 = Dense(10, name='out1_d2')(output1) output1 = Dense(3, name='out1_d3')(output1) output2 = Dense(25, name='out2_d1')(middle3) output2 = Dense(10, name='out2_d2')(output2) output2 = Dense(3, name='out2_d3')(output2)
def cnn_cycle_gan():
############################## Transformation module #################################################
# Layers of stage 0 contraction
InputLayer(input_shape=(in_, 21))
tct0_conv1 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_1')(tct0_in)
tct0_bn1 = BatchNormalization()(ct0_conv1)
tct0_conv2 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_2')(tct0_bn1)
tct0_bn2 = BatchNormalization()(tct0_conv2)
tct0_max = MaxPool1D(pool_size=2, strides=2)(tct0_bn2)
tct0_dp = Dropout(0.2)(tct0_max)
# Layers of stage 1 contraction
tct1_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct1_1')(tct0_dp)
tct1_bn1 = BatchNormalization()(tct1_conv1)
tct1_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct1_2')(tct1_bn1)
tct1_bn2 = BatchNormalization()(tct1_conv2)
tct1_max = MaxPool1D(pool_size=2, strides=2)(tct1_bn2)
tct1_dp = Dropout(0.2)(tct1_max)
# Layers of stage 2 contraction
tct2_conv1 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Ct2_1')(tct1_dp)
tct2_bn1 = BatchNormalization()(tct2_conv1)
tct2_conv2 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Ct2_2')(tct2_bn1)
tct2_bn2 = BatchNormalization()(tct2_conv2)
tct2_max = MaxPool1D(pool_size=2, strides=2)(tct2_bn2)
tct2_dp = Dropout(0.2)(tct2_max)
# Layers of stage 3 contraction
tct3_conv1 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_1')(tct2_dp)
tct3_bn1 = BatchNormalization()(tct3_conv1)
tct3_conv2 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_2')(tct3_bn1)
tct3_bn2 = BatchNormalization()(tct3_conv2)
tct3_max = MaxPool1D(pool_size=2, strides=2)(tct3_bn2)
tct3_dp = Dropout(0.2)(tct3_max)
# Layers of stage 1 expansion
tet1_Tconv = Conv1DTranspose(64,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et1')(tct3_dp)
tet1_Concat = Concatenate(axis=1)([tet1_Tconv, tct3_conv2])
tet1_bn1 = BatchNormalization()(et1_Concat)
tet1_conv1 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_1')(tet1_bn1)
tet1_bn2 = BatchNormalization()(tet1_conv1)
tet1_conv2 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_2')(tet1_bn2)
tet1_dp = Dropout(0.2)(tet1_conv2)
#Layers of stage 2 expansion
tet2_Tconv = Conv1DTranspose(32,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et2')(tet1_dp)
tet2_Concat = Concatenate(axis=1)([tet2_Tconv, tct1_conv2])
tet2_bn1 = BatchNormalization()(tet2_Concat)
tet2_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_1')(tet2_bn1)
tet2_bn2 = BatchNormalization()(tet2_conv1)
tet2_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_2')(tet2_bn2)
tet2_dp = Dropout(0.2)(tet2_conv2)
#Layers of stage 3 expansion
tet3_Tconv = Conv1DTranspose(16,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et3')(tet2_dp)
tet3_Concat = Concatenate(axis=1)([tet3_Tconv, tce0_conv2])
tet3_bn1 = BatchNormalization()(tet3_Concat)
tet3_conv1 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Et3_1')(tet3_bn1)
tet3_bn2 = BatchNormalization()(tet3_conv1)
tet3_conv2 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Et3_2')(tet3_bn2)
tet3_bn3 = BatchNormalization()(tet3_conv2)
tet3_dp = Dropout(0.1)(tet3_bn3)
tet3_conv3 = Conv1D(1,
5,
activation='sigmoid',
padding='same',
name='Convolution_Et3_3')(tet3_dp)
########################################## Reconstroction Module ######################################
# Layers of stage 0 contraction
rct0_conv1 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_1')(tet3_conv3)
rct0_bn1 = BatchNormalization()(rct0_conv1)
rct0_conv2 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Ct0_2')(rct0_bn1)
rct0_bn2 = BatchNormalization()(rct0_conv2)
rct0_max = MaxPool1D(pool_size=2, strides=2)(rct0_bn2)
rct0_dp = Dropout(0.2)(rct0_max)
# Layers of stage 1 contraction
rct1_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct1_1')(rct0_dp)
rct1_bn1 = BatchNormalization()(rct1_conv1)
rct1_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Ct1_2')(rct1_bn1)
rct1_bn2 = BatchNormalization(rct1_conv2)
rct1_max = MaxPool1D(pool_size=2, strides=2)(rct1_bn2)
rct1_dp = Dropout(0.2)(rct1_max)
# Layers of stage 2 contraction
rct2_conv1 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Ct2_1')(rct1_dp)
rct2_bn1 = BatchNormalization()(rct2_conv1)
rct2_conv2 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Ct2_2')(rct2_bn1)
rct2_bn2 = BatchNormalization()(rct2_conv2)
rct2_max = MaxPool1D(pool_size=2, strides=2)(rct2_bn2)
rct2_dp = Dropout(0.2)(rct2_max)
# Layers of stage 3 contraction
rct3_conv1 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_1')(rct2_dp)
rct3_bn1 = BatchNormalization()(rct3_conv1)
rct3_conv2 = Conv1D(128,
5,
activation='relu',
padding='same',
name='Convolution_Ce3_2')(rct3_bn1)
rct3_bn2 = BatchNormalization()(rct3_conv2)
rct3_max = MaxPool1D(pool_size=2, strides=2)(rct3_bn2)
rct3_dp = Dropout(0.2)(rct3_max)
# Layers of stage 1 expansion
ret1_Tconv = Conv1DTranspose(64,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et1')(rct3_dp)
ret1_Concat = Concatenate(axis=1)([ret1_Tconv, rct3_conv2])
ret1_bn1 = BatchNormalization()(ret1_Concat)
ret1_conv1 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_1')(ret1_bn1)
ret1_bn2 = BatchNormalization()(ret1_conv1)
ret1_conv2 = Conv1D(64,
5,
activation='relu',
padding='same',
name='Convolution_Et1_2')(ret1_bn2)
ret1_dp = Dropout(0.2)(ret1_conv2)
#Layers of stage 2 expansion
ret2_Tconv = Conv1DTranspose(32,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et2')(ret1_dp)
ret2_Concat = Concatenate(axis=1)([ret2_Tconv, rct1_conv2])
ret2_bn1 = BatchNormalization()(ret2_Concat)
ret2_conv1 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_1')(ret2_bn1)
ret2_bn2 = BatchNormalization()(ret2_conv1)
ret2_conv2 = Conv1D(32,
5,
activation='relu',
padding='same',
name='Convolution_Et2_2')(ret2_bn2)
ret2_dp = Dropout(0.2)(ret2_conv2)
#Layers of stage 3 expansion
ret3_Tconv = Conv1DTranspose(16,
5,
strides=2,
activation='relu',
padding='same',
name='TransConv_Et3')(ret2_dp)
ret3_Concat = Concatenate(axis=1)([ret3_Tconv, rce0_conv2])
ret3_bn1 = BatchNormalization()(ret3_Concat)
ret3_conv1 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Et3_1')(ret3_bn1)
ret3_bn2 = BatchNormalization()(ret3_conv1)
ret3_conv2 = Conv1D(16,
5,
activation='relu',
padding='same',
name='Convolution_Et3_2')(ret3_bn2)
ret3_bn3 = BatchNormalization()(ret3_conv2)
ret3_dp = Dropout(0.1)(ret3_bn3)
ret3_conv2 = Conv1D(1,
5,
activation='sigmoid',
padding='same',
name='Convolution_Et3_3')(ret3_dp)
######################################## Discriminator Module ###############################################
distrib_Y = Input(shape=(2000, 20))
def khop_model_share(): # input/output = num of sensors
sensor_matrix1 = Input(shape=(num_sensors, num_sensors))
sensor_matrix2 = Input(shape=(num_sensors, num_sensors))
#sensor_matrix3 = Input(shape=(num_sensors, num_sensors))
s_input1 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input2 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input3 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input4 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input5 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input6 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input7 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input8 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_input9 = Input(shape=(input_shape[0], input_shape[1], input_shape[2]))
s_cnn = sensor_cnn(input_shape, repetitions = [2,2,2,2])
extract_cnn1 = s_cnn(s_input1)
extract_cnn2 = s_cnn(s_input2)
extract_cnn3 = s_cnn(s_input3)
extract_cnn4 = s_cnn(s_input4)
extract_cnn5 = s_cnn(s_input5)
extract_cnn6 = s_cnn(s_input6)
extract_cnn7 = s_cnn(s_input7)
extract_cnn8 = s_cnn(s_input8)
extract_cnn9 = s_cnn(s_input9)
extract_cnn = Concatenate(axis=1)([extract_cnn1, extract_cnn2, extract_cnn3,
extract_cnn4, extract_cnn5, extract_cnn6,
extract_cnn7, extract_cnn8, extract_cnn9])
#extract_cnn = np.reshape(extract_cnn, (-1,))
G_h1 = GraphConv(256, 'relu')([extract_cnn, sensor_matrix1])
G_h2 = GraphConv(256, 'relu')([extract_cnn, sensor_matrix2])
G_1 = Concatenate(axis=-1)([G_h1, G_h2])
G_2h1 = GraphConv(256, 'relu')([G_1, sensor_matrix1])
G_2h2 = GraphConv(256, 'relu')([G_1, sensor_matrix2])
G_2 = Concatenate(axis=-1)([G_2h1, G_2h2])
gnn_output = tf.split(G_2, num_sensors, 1)
output1 = Dense(32, activation='relu')(Flatten()(gnn_output[0]))
output1 = Dense(2, activation='linear', name='sensor_1')(output1)
output2 = Dense(32, activation='relu')(Flatten()(gnn_output[1]))
output2 = Dense(2, activation='linear', name='sensor_2')(output2)
output3 = Dense(32, activation='relu')(Flatten()(gnn_output[2]))
output3 = Dense(2, activation='linear', name='sensor_3')(output3)
output4 = Dense(32, activation='relu')(Flatten()(gnn_output[3]))
output4 = Dense(2, activation='linear', name='sensor_4')(output4)
output5 = Dense(32, activation='relu')(Flatten()(gnn_output[4]))
output5 = Dense(2, activation='linear', name='sensor_5')(output5)
output6 = Dense(32, activation='relu')(Flatten()(gnn_output[5]))
output6 = Dense(2, activation='linear', name='sensor_6')(output6)
output7 = Dense(32, activation='relu')(Flatten()(gnn_output[6]))
output7 = Dense(2, activation='linear', name='sensor_7')(output7)
output8 = Dense(32, activation='relu')(Flatten()(gnn_output[7]))
output8 = Dense(2, activation='linear', name='sensor_8')(output8)
output9 = Dense(32, activation='relu')(Flatten()(gnn_output[8]))
output9 = Dense(2, activation='linear', name='sensor_9')(output9)
model = Model(inputs=[s_input1, s_input2, s_input3, s_input4,
s_input5, s_input6, s_input7, s_input8, s_input9,
sensor_matrix1, sensor_matrix2],
outputs= [output1,output2,output3,output4,
output5,output6,output7,output8,output9])
return model
def __init__(self):
super(Concatenate_, self).__init__()
self.concat = Concatenate(axis=-1)
def define_ensemble_model(input_size, lr=1e-3):
try:
df = pd.read_pickle('train_data_df.pkl')
except:
FIRST_DAY = 350# If you want to load all the data set it to '1' --> Great memory overflow risk !
df = create_dt(is_train=True, first_day= FIRST_DAY)
create_fea(df)
df.dropna(inplace = True)
df.to_pickle('train_data_df.pkl')
# del df; gc.collect()
inputs = Input(shape=(input_size, ), name='inputs')
# Embedding input
wday_input = Input(shape=(1,), name='wday')
month_input = Input(shape=(1,), name='month')
year_input = Input(shape=(1,), name='year')
mday_input = Input(shape=(1,), name='mday')
quarter_input = Input(shape=(1,), name='quarter')
event_name_1_input = Input(shape=(1,), name='event_name_1')
event_type_1_input = Input(shape=(1,), name='event_type_1')
event_name_2_input = Input(shape=(1,), name='event_name_2')
event_type_2_input = Input(shape=(1,), name='event_type_2')
item_id_input = Input(shape=(1,), name='item_id')
dept_id_input = Input(shape=(1,), name='dept_id')
store_id_input = Input(shape=(1,), name='store_id')
cat_id_input = Input(shape=(1,), name='cat_id')
state_id_input = Input(shape=(1,), name='state_id')
snap_CA_input = Input(shape=(1,), name='snap_CA')
snap_TX_input = Input(shape=(1,), name='snap_TX')
snap_WI_input = Input(shape=(1,), name='snap_WI')
wday_emb = Flatten()(Embedding(7, 1)(wday_input))
month_emb = Flatten()(Embedding(12, 2)(month_input))
year_emb = Flatten()(Embedding(6, 1)(year_input))
mday_emb = Flatten()(Embedding(31, 2)(mday_input))
quarter_emb = Flatten()(Embedding(4, 1)(quarter_input))
event_name_1_emb = Flatten()(Embedding(31, 2)(event_name_1_input))
event_type_1_emb = Flatten()(Embedding(5, 1)(event_type_1_input))
event_name_2_emb = Flatten()(Embedding(5, 1)(event_name_2_input))
event_type_2_emb = Flatten()(Embedding(5, 1)(event_type_2_input))
item_id_emb = Flatten()(Embedding(3049, 4)(item_id_input))
dept_id_emb = Flatten()(Embedding(7, 1)(dept_id_input))
store_id_emb = Flatten()(Embedding(10, 1)(store_id_input))
cat_id_emb = Flatten()(Embedding(6, 1)(cat_id_input))
state_id_emb = Flatten()(Embedding(3, 1)(state_id_input))
x = Concatenate(-1)([inputs, wday_emb, month_emb, month_emb, year_emb, mday_emb, \
quarter_emb, event_name_1_emb, event_type_1_emb, event_name_2_emb, \
event_type_2_emb, item_id_emb, dept_id_emb, store_id_emb, cat_id_emb, \
state_id_emb])
input_dic = {
'inputs': inputs, 'wday': wday_input, 'month': month_input, 'year': year_input,
'mday': mday_input, 'quarter': quarter_input, 'event_name_1': event_name_1_input,
'event_type_1': event_type_1_input, 'event_name_2': event_name_2_input,
'event_type_2': event_type_2_input, 'item_id': item_id_input, 'dept_id': dept_id_input,
'store_id': store_id_input, 'cat_id': cat_id_input, 'state_id': state_id_input,
}
models = [predict_model(len(input_dense)), predict_model(len(input_dense)), predict_model(len(input_dense))]
models[0].load_weights('./m5_predict5_mse.h5')
models[1].load_weights('./m5_predict5_rmsse.h5')
models[2].load_weights('./m5_predict5_wrmsse.h5')
i = 0
for model in models:
for layer in model.layers:
layer.trainable = False
# layer.name = 'ensemble_' + str(i) + layer.name
i += 1
ensemble_outputs = [model(input_dic) for model in models]
merge = Concatenate(-1)(ensemble_outputs)
merge = Concatenate(-1)([x, merge])
x = Dense(64, activation='relu')(merge)
x = BatchNormalization()(x)
x = Dense(32, activation='relu')(x)
outputs = Dense(1, activation='linear')(x)
model = Model(inputs=input_dic, outputs=outputs)
model.compile(optimizer=Adam(lr=lr), loss=rmse)
return model
def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'):
"""Adds a Inception-ResNet block.
This function builds 3 types of Inception-ResNet blocks mentioned
in the paper, controlled by the `block_type` argument (which is the
block name used in the official TF-slim implementation):
- Inception-ResNet-A: `block_type='block35'`
- Inception-ResNet-B: `block_type='block17'`
- Inception-ResNet-C: `block_type='block8'`
# Arguments
x: input tensor.
scale: scaling factor to scale the residuals (i.e., the output of
passing `x` through an inception module) before adding them
to the shortcut branch. Let `r` be the output from the residual branch,
the output of this block will be `x + scale * r`.
block_type: `'block35'`, `'block17'` or `'block8'`, determines
the network structure in the residual branch.
block_idx: an `int` used for generating layer names. The Inception-ResNet blocks
are repeated many times in this network. We use `block_idx` to identify
each of the repetitions. For example, the first Inception-ResNet-A block
will have `block_type='block35', block_idx=0`, ane the layer names will have
a common prefix `'block35_0'`.
activation: activation function to use at the end of the block
(see [activations](../activations.md)).
When `activation=None`, no activation is applied
(i.e., "linear" activation: `a(x) = x`).
# Returns
Output tensor for the block.
# Raises
ValueError: if `block_type` is not one of `'block35'`,
`'block17'` or `'block8'`.
"""
if block_type == 'block35':
branch_0 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(x, 32, 1)
branch_1 = conv2d_bn(branch_1, 32, 3)
branch_2 = conv2d_bn(x, 32, 1)
branch_2 = conv2d_bn(branch_2, 48, 3)
branch_2 = conv2d_bn(branch_2, 64, 3)
branches = [branch_0, branch_1, branch_2]
elif block_type == 'block17':
branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 128, 1)
branch_1 = conv2d_bn(branch_1, 160, [1, 7])
branch_1 = conv2d_bn(branch_1, 192, [7, 1])
branches = [branch_0, branch_1]
elif block_type == 'block8':
branch_0 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(x, 192, 1)
branch_1 = conv2d_bn(branch_1, 224, [1, 3])
branch_1 = conv2d_bn(branch_1, 256, [3, 1])
branches = [branch_0, branch_1]
else:
raise ValueError('Unknown Inception-ResNet block type. '
'Expects "block35", "block17" or "block8", '
'but got: ' + str(block_type))
block_name = block_type + '_' + str(block_idx)
channel_axis = 1 if K.image_data_format() == 'channels_first' else 3
mixed = Concatenate(axis=channel_axis,
name=block_name + '_mixed')(branches)
up = conv2d_bn(mixed,
K.int_shape(x)[channel_axis],
1,
activation=None,
use_bias=True,
name=block_name + '_conv')
x = Lambda(lambda inputs, scale: inputs[0] + inputs[1] * scale,
output_shape=K.int_shape(x)[1:],
arguments={'scale': scale},
name=block_name)([x, up])
if activation is not None:
x = Activation(activation, name=block_name + '_ac')(x)
return x
