def build_simple_gcn(num_filters, graph_conv_filters):
model = Sequential()
model.add(
GraphCNN(Y.shape[1],
num_filters,
graph_conv_filters,
input_shape=(X.shape[1], ),
activation='elu',
kernel_regularizer=l2(5e-4)))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.01),
metrics=['acc'])
model.summary()
return model
Y = to_categorical(
Y_o_dim
) # labels, whatever we wanna classify things into... in categorical form.
train_on_weight = np.array([1, 1, 0])
print("Now we won't do any fancy preprocessing, just basic training.")
NUM_FILTERS = 1
graph_conv_filters = A # you may try np.eye(3)
graph_conv_filters = K.constant(graph_conv_filters)
model = Sequential()
model.add(
GraphCNN(Y.shape[1],
NUM_FILTERS,
graph_conv_filters,
input_shape=(X.shape[1], ),
activation='elu',
kernel_regularizer=l2(5e-4)))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.01),
metrics=['acc'])
model.summary()
model.fit(X,
Y,
batch_size=A.shape[0],
sample_weight=train_on_weight,
epochs=200,
shuffle=False,
verbose=0)
# print(X.shape, Y_train.shape)
# Build Graph Convolution filters
SYM_NORM = True
A_norm = preprocess_adj_numpy(A, SYM_NORM)
num_filters = 2
graph_conv_filters = np.concatenate([A_norm, np.matmul(A_norm, A_norm)],
axis=0)
graph_conv_filters = K.constant(graph_conv_filters)
# Build Model
input_tensor = Input(shape=(X.shape[1], ))
x = GraphCNN(16,
num_filters,
graph_conv_filters,
input_shape=(X.shape[1], ),
activation='elu',
kernel_regularizer=l2(5e-4))(input_tensor)
x = Dropout(0.2)(x)
x = GraphCNN(16,
num_filters,
graph_conv_filters,
activation='elu',
kernel_regularizer=l2(5e-4))(x)
print(x.shape)
output_tensor = Dense(Y.shape[1], activation='softmax')(x)
model = Model(input_tensor, output_tensor)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.01),
metrics=['acc'])
def load_gnn_model():
X, Y = get_data()
A = get_adj_matrix(X)
_, Y_val, _, train_idx, val_idx, test_idx, train_mask = get_splits(Y)
train_idx = np.array(train_idx)
val_idx = np.array(val_idx)
test_idx = np.array(test_idx)
labels = np.argmax(Y, axis=1) + 1
Y_train = np.zeros(Y.shape)
labels_train = np.zeros(labels.shape)
Y_train[train_idx] = Y[train_idx]
labels_train[train_idx] = labels[train_idx]
Y_test = np.zeros(Y.shape)
labels_test = np.zeros(labels.shape)
Y_test[test_idx] = Y[test_idx]
labels_test[test_idx] = labels[test_idx]
# Build Graph Convolution filters
SYM_NORM = True
# A_norm = preprocess_adj_numpy(A, SYM_NORM)
A_norm = A
num_filters = 2
graph_conv_filters = np.concatenate(
[A_norm, np.matmul(A_norm, A_norm)], axis=0)
graph_conv_filters = K.constant(graph_conv_filters)
# build GAT model
input_tensor = Input(shape=(X.shape[1], ))
x = Dropout(0.6, input_shape=(X.shape[1], ))(x)
model.add(
GraphAttentionCNN(8,
A,
num_filters,
graph_conv_filters,
num_attention_heads=8,
attention_combine='concat',
attention_dropout=0.6,
activation='elu',
kernel_regularizer=l2(5e-4)))
model.add(Dropout(0.6))
model.add(
GraphAttentionCNN(Y.shape[1],
A,
num_filters,
graph_conv_filters,
num_attention_heads=1,
attention_combine='average',
attention_dropout=0.6,
activation='elu',
kernel_regularizer=l2(5e-4)))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.005),
metrics=['accuracy'])
# Build Model
input_tensor = Input(shape=(X.shape[1], ))
x = GraphCNN(16,
num_filters,
graph_conv_filters,
input_shape=(X.shape[1], ),
activation='elu',
kernel_regularizer=l2(5e-4))(input_tensor)
x = Dropout(0.2)(x)
x = GraphCNN(16,
num_filters,
graph_conv_filters,
activation='elu',
kernel_regularizer=l2(5e-4),
name='g_out')(x)
output_tensor = Dense(Y.shape[1], activation='softmax', name='sub_out')(x)
model = Model(input_tensor, output_tensor)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.01),
metrics=['acc'])
model.summary()
# load
model.load_weights('gnn.h5')
# # Y_pred = model.predict(X, batch_size=A.shape[0])
sub_model = Model(inputs=model.input,
outputs=model.get_layer('g_out').output)
# result = sub_model.predict(X, batch_size=A.shape[0])
# print(result)
return sub_model
# ## The model itself
# + {"outputId": "a3142dd2-4ff0-4bb6-a833-a7046f4e0596", "colab_type": "code", "id": "3lOBizVa4rVt", "colab": {"base_uri": "https://localhost:8080/", "height": 1295}}
A_norm = normalize_adj_numpy(A, True)
num_filters = 2
graph_conv_filters = np.concatenate([A_norm, np.matmul(A_norm, A_norm)],
axis=0)
print graph_conv_filters.shape
graph_conv_filters = K.constant(graph_conv_filters)
# Build Model
inp = Input(shape=(X.shape[1], ))
x = GraphCNN(16,
num_filters,
graph_conv_filters,
activation='elu',
kernel_regularizer=l2(5e-4))(inp)
x = Dropout(0.2)(x)
x = GraphCNN(Y_train.shape[1],
num_filters,
graph_conv_filters,
activation='elu',
kernel_regularizer=l2(5e-4))(x)
x = Activation('softmax')(x)
model = Model(inputs=inp, outputs=x)
model.compile(loss='categorical_crossentropy',
optimizer=Adam(lr=0.01),
metrics=['acc'])
model.summary()
Y_test = np.zeros(Y.shape)
labels_test = np.zeros(labels.shape)
Y_test[test_idx] = Y[test_idx]
labels_test[test_idx] = labels[test_idx]
# Build Graph Convolution filters
SYM_NORM = True
A_norm = preprocess_adj_numpy(A, SYM_NORM)
num_filters = 2
graph_conv_filters = np.concatenate([A_norm, np.matmul(A_norm, A_norm)], axis=0)
graph_conv_filters = K.constant(graph_conv_filters)
# Build Model
model = Sequential()
model.add(GraphCNN(16, num_filters, graph_conv_filters, input_shape=(X.shape[1],), activation='elu', kernel_regularizer=l2(5e-4)))
model.add(Dropout(0.2))
model.add(GraphCNN(Y.shape[1], num_filters, graph_conv_filters, activation='elu', kernel_regularizer=l2(5e-4)))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy', optimizer=Adam(lr=0.01), metrics=['acc'])
nb_epochs = 500
start_time = time.time() # START: Training Time Tracker
for epoch in range(nb_epochs):
model.fit(X, Y_train, sample_weight=train_mask, batch_size=A.shape[0], epochs=1, shuffle=False, verbose=0)
Y_pred = model.predict(X, batch_size=A.shape[0])
Y_pred_proba = model.predict_proba(X, batch_size=A.shape[0])
_, train_acc = evaluate_preds(Y_pred, [Y_train], [train_idx])
_, test_acc = evaluate_preds(Y_pred, [Y_test], [test_idx])
print("Epoch: {:04d}".format(epoch), "train_acc= {:.4f}".format(train_acc[0]), "test_acc= {:.4f}".format(test_acc[0]))