Gcn_lstm layer: Stateful lstm layers

[ajikmr]

My time series prediction requires Stateful lstm layers. In gcn_lstm layer, currently it is not so.

I have tried to add 'stateful=True' in the gcn_lstm.py file. But this option requires that I also add batch input shape information. What would be appropriate way to add this parameter.

It will be great to also have a demo file for gcn_lstm with stateful lstm layers.

To make code work for stateful LSTM layers, I have created a class GCN_LSTM_STATEFUL which is a copy of GCN_LSTM class, except lstm layers are defined as below. I have hard-coded batch_input_shape for now to see it can run. Function call 'in_out_tensors()' throws error as shown further below.

---

    self._lstm_layers = [
        LSTM(layer_size, activation=activation, batch_input_shape=(1,6,5), stateful=True, return_sequences=True)
        for layer_size, activation in zip(
            self.lstm_layer_sizes[:-1], self.lstm_activations
        )
    ]
    self._lstm_layers.append(
        LSTM(
            self.lstm_layer_sizes[-1],
            activation=self.lstm_activations[-1],
            batch_input_shape=(1,6,200),
            stateful=True,
            return_sequences=False,
        )
    )

---

ERROR

---

ValueError Traceback (most recent call last)
/tmp/ipykernel_9509/572892708.py in
----> 1 x_input, x_output = gcn_lstm.in_out_tensors()

~/install/anaconda3/lib/python3.9/site-packages/stellargraph/layer/gcn_lstm.py in in_out_tensors(self)
636
637 x_inp = [x_t, out_indices_t]
--> 638 x_out = self(x_inp)
639
640 return x_inp[0], x_out

~/install/anaconda3/lib/python3.9/site-packages/stellargraph/layer/gcn_lstm.py in call(self, x)
604
605 for layer in self._lstm_layers:
--> 606 h_layer = layer(h_layer)
607
608 h_layer = Dropout(self.dropout)(h_layer)

~/install/anaconda3/lib/python3.9/site-packages/keras/layers/rnn/base_rnn.py in reset_states(self, states)
874 batch_size = spec_shape[1] if self.time_major else spec_shape[0]
875 if not batch_size:
--> 876 raise ValueError(
877 "If a RNN is stateful, it needs to know "
878 "its batch size. Specify the batch size "

ValueError: If a RNN is stateful, it needs to know its batch size. Specify the batch size of your input tensors:

• If using a Sequential model, specify the batch size by passing a batch_input_shape argument to your first layer.
• If using the functional API, specify the batch size by passing a batch_shape argument to your Input layer.