def create_cnn_model(learning_rate, weight_decay, dropout, hidden_layers,
activation):
g_model = Sequential()
for i in range(len(hidden_layers)):
if (i == 0):
g_model.add(
GraphConv(filters=hidden_layers[i],
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors,
activation=activation,
input_shape=(input_size, features),
kernel_regularizer=regularizers.l2(weight_decay)))
g_model.add(Dropout(dropout))
else:
g_model.add(
GraphConv(filters=hidden_layers[i],
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors,
activation=activation,
kernel_regularizer=regularizers.l2(weight_decay)))
g_model.add(GraphConvPro())
g_model.add(Flatten())
g_model.add(Dense(1, activation='sigmoid'))
optimizer = Adam(lr=learning_rate)
g_model.compile(optimizer=optimizer, loss='mean_squared_error')
return g_model
seed_value = 0
if len(sys.argv) > 1:
max_train = int(sys.argv[1])
epochs = int(sys.argv[2])
seed_value = int(sys.argv[3])
os.environ['PYTHONHASHSEED'] = str(seed_value)
random.seed(seed_value)
np.random.seed(seed_value)
g_model = Sequential()
g_model.add(
GraphConv(filters=10,
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors,
activation='elu',
input_shape=(X_train_2.shape[1], features)))
g_model.add(Dropout(0.2))
g_model.add(
GraphConv(filters=5,
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors,
activation='elu'))
g_model.add(GraphConvPro())
g_model.add(Flatten())
g_model.add(Dense(1, activation='sigmoid'))
g_model.compile(loss='mean_squared_error',
optimizer='Adam',
metrics=[soft_AUC_theano, oe])
### Prepare the Graph Correlation matrix
corr_mat = np.array(normalize(np.abs(np.corrcoef(X_train.transpose())),
norm='l1',
axis=1),
dtype='float64')
graph_mat = np.argsort(corr_mat, 1)[:, -nb_neighbors:]
# %%
### Single layer of Graph Convolution
g_model = Sequential()
g_model.add(
GraphConv(nb_filter=nb_filter,
neighbors_ix_mat=graph_mat,
nb_neighbors=nb_neighbors,
activation='relu',
input_shape=(
X_train.shape[1],
1,
)))
g_model.add(Dropout(0.2))
g_model.add(Flatten())
g_model.add(Dense(10, activation='softmax'))
g_model.summary()
g_model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(),
metrics=['accuracy'])
results['g'] = g_model.fit(X_train.reshape(X_train.shape[0], X_train.shape[1],
1),
epochs = np.arange(epoch)
results_train = np.ones(epoch) * 20
results_test = np.ones(epoch) * 20
batch_size = 30
num_hidden = 32
filters_1 = 20
filters_2 = 22
filters_3 = 22
filters_4 = 16
filters_5 = 16
model = Sequential()
model.add(
GraphConv(filters=filters_1,
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors,
input_shape=(X_train.shape[1], 1)))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(
GraphConv(filters=filters_2,
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors))
model.add(Activation('relu'))
model.add(
GraphConv(filters=filters_3,
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(
corr_mat = np.array(normalize(np.abs(np.corrcoef(X_train.transpose())),
norm='l1',
axis=1),
dtype='float64')
graph_mat = np.argsort(corr_mat, 1)[:, -num_neighbors:]
print('graph_mat shape:(%d,%d)' % (graph_mat.shape))
# %%
### 1 layer graph CNN
g_model = Sequential()
g_model.add(
GraphConv(filters=filters_1,
neighbors_ix_mat=graph_mat,
num_neighbors=num_neighbors,
activation='relu',
input_shape=(X_train.shape[1], 1)))
g_model.add(Dropout(0.25))
g_model.add(Flatten())
g_model.add(Dense(1, kernel_regularizer=l2(0.01)))
g_model.add(Dropout(0.1))
g_model.summary()
g_model.compile(loss='mean_squared_error', optimizer='adam')
results['g'] = []
for i in range(epochs):
g_model.fit(
X_train.reshape(X_train.shape[0], X_train.shape[1], 1),
Y_test = np_utils.to_categorical(y_test, nb_classes)
# %% Generate the transition matrix Q
q_mat_layer1 = generate_Q(28, 3)
q_mat_layer2 = generate_Q(14, 3)
q_mat_layer1 = np.argsort(q_mat_layer1, 1)[:, -nb_neighbors:]
q_mat_layer2 = np.argsort(q_mat_layer2, 1)[:, -nb_neighbors:]
# %% Standard LeNet structure with Graph Convolution
model = Sequential()
model.add(
GraphConv(nb_filter=nb_layer_1,
neighbors_ix_mat=q_mat_layer1,
nb_neighbors=nb_neighbors,
activation='relu',
input_shape=(
X_train.shape[1],
1,
)))
model.add(Reshape((28, 28, nb_layer_1)))
model.add(MaxPooling2D(pool_size=pool_size, dim_ordering='tf'))
model.add(Dropout(0.25))
model.add(Reshape((196, nb_layer_1)))
model.add(
GraphConv(nb_filter=nb_layer_2,
neighbors_ix_mat=q_mat_layer2,
nb_neighbors=nb_neighbors,
activation='relu',
bias=True))
model.add(Reshape((14, 14, nb_layer_2)))
model.add(MaxPooling2D(pool_size=pool_size, dim_ordering='tf'))
Y_test = np_utils.to_categorical(y_test, num_classes)
# %% Generate the transition matrix Q
q_mat_layer1 = generate_Q(28, 3)
q_mat_layer2 = generate_Q(14, 3)
q_mat_layer1 = np.argsort(q_mat_layer1, 1)[:, -num_neighbors:]
q_mat_layer2 = np.argsort(q_mat_layer2, 1)[:, -num_neighbors:]
# %% Standard LeNet structure with Graph Convolution
model = Sequential()
model.add(
GraphConv(filters=num_layer_1,
neighbors_ix_mat=q_mat_layer1,
num_neighbors=num_neighbors,
activation='relu',
input_shape=(
X_train.shape[1],
1,
)))
model.add(Reshape((28, 28, num_layer_1)))
model.add(MaxPooling2D(pool_size=pool_size, data_format='channels_last'))
model.add(Dropout(0.25))
model.add(Reshape((196, num_layer_1)))
model.add(
GraphConv(filters=num_layer_2,
neighbors_ix_mat=q_mat_layer2,
num_neighbors=num_neighbors,
activation='relu'))
model.add(Reshape((14, 14, num_layer_2)))
model.add(MaxPooling2D(pool_size=pool_size, data_format='channels_last'))