'GATConv' object is not callable

[slhilyfe]

My spektral is version 1.3.0 and my tensorflow is version 2.10.0. I want to make a model of the integration of space and time. The use of convolution layer in the new version of spektral has changed a lot, but there are a few questions I would like to consult with experts.

1. I think there is no way to call it using the GraphConv layer. I customized it, and I don't know if it is correct:
   class GraphConvLayer(Layer):
   def init(self, **kwargs):
   super(GraphConvLayer, self).init(**kwargs)
   def call(self, inputs, **kwargs):
   A, X = inputs
   return tf.einsum('bji,bjk->bik', A, X) # Change the positions of the adjacency matrix and the feature matrix
   def compute_output_shape(self, input_shape):
   return input_shape[1]
2. The "# Create a GATConv layer gat_conv = GATConv (64, activation='relu') ([graph_conv_layer2, graph_input_layer]) " section of my code always shows" 'GATConv' object is not callable ". I've tried a lot of methods but haven't solved it. This is my code for defining the structure of the model:

Build adjacency matrix

sigma = 10.0 # This is a hyperparameter, you might need to tune it.
A = build_adjacency_matrix(stations_info, sigma)
A = A.squeeze() # Remove unnecessary dimensions

Normalize adjacency matrix

A = spektral.utils.normalized_adjacency(A, symmetric=True)
print(A.shape)

Convert the normalized adjacency matrix into a tensor

tensor_A = tf.convert_to_tensor(A, dtype=tf.float32)
print(tensor_A.shape)

copy the tensor_A to match the training set and test set, because each time step requires a tensor_A

tensor_A_train = np.repeat(tensor_A[np.newaxis, ...], X_train.shape[0], axis=0)
tensor_A_test = np.repeat(tensor_A[np.newaxis, ...], X_test.shape[0], axis=0)
print('X_train shape:', X_train.shape)
print('y_train shape:', y_train.shape)
print('tensor_A_train shape:', tensor_A_train.shape)

Define the model

input_layer = Input(shape=X_train.shape[1:])
graph_input_layer = Input(shape=tensor_A_train.shape[1:], name="graph_input")
print('graph_input_layer:', graph_input_layer.shape)

spatial information

graph_conv_layer = GraphConvLayer()([input_layer, graph_input_layer])
graph_conv_layer = tf.transpose(graph_conv_layer, perm=[0, 2, 1])
print('graph_conv_layer:', graph_conv_layer.shape)

Add a second graph convolution layer to try deeper spatial feature extraction

graph_conv_layer2 = GraphConvLayer()([graph_conv_layer, graph_input_layer])
graph_conv_layer2 = tf.transpose(graph_conv_layer2, perm=[0, 2, 1])
print('graph_conv_layer2:', graph_conv_layer2.shape)

Create a GATConv layer

gat_conv = GATConv(64, activation='relu')([graph_conv_layer2, graph_input_layer])
print(type(gat_conv))

temporal information

multihead_attn = MultiHeadAttention(num_heads=8, key_dim=64)
attn_output = multihead_attn(input_layer, input_layer)
attn_output = LayerNormalization(epsilon=1e-6)(attn_output)

add a second long attention layer to try deeper time feature extraction

multihead_attn2 = MultiHeadAttention(num_heads=8, key_dim=64)
attn_output2 = multihead_attn2(attn_output, attn_output)
attn_output2 = LayerNormalization(epsilon=1e-6)(attn_output2)
lstm = LSTM(64, return_sequences=False, input_shape=(num_timesteps, num_features + 2))(attn_output2)

lstm = Dense(64, activation='relu')(lstm) # New layer to adjust LSTM output shape

Combine spatial and temporal information

lstm_repeated = tf.keras.layers.RepeatVector(131)(lstm) # Shape: (None, 131, 64)
combined_features = Concatenate()([gat_conv, lstm_repeated]) # Shape: (None, 131, 128)

Decoding layer

dense = Dense(64, activation='relu')(combined_features)
output_layer = Dense(1, activation='linear')(dense)

Create and compile the model

model = Model(inputs=[input_layer, graph_input_layer], outputs=output_layer)
model.compile(optimizer=adam_v2.Adam(learning_rate=0.001), loss='mse')

early_stop = EarlyStopping(monitor='val_loss', patience=10)

train the model

model.fit([X_train, tensor_A_train], y_train, epochs=100, batch_size=32, validation_data=([X_test, tensor_A_test], y_test))

Thank you very much for your help!