How can I apply input data with `ndim=1` to LSTM network?

General Discussion
,
#1

Straight to the question is “How can I apply input data with ndim=1 to LSTM network?”

I have 12 input data for LSTM network with two layers of 128 units each, and I’m trying to use it for PPO TF-Agent. I defined the observation_spec of the environment as

self._observation_spec = array_spec.BoundedArraySpec(shape=(12,), dtype=np.float32, name='observation')

and networks as

def create_networks(tf_env):
    actor_net = ActorDistributionRnnNetwork(
        tf_env.observation_spec(),
        tf_env.action_spec(),
        input_fc_layer_params=None,
        lstm_size=(128,128),
        output_fc_layer_params=None,
        activation_fn=None)
    value_net = ValueRnnNetwork(
        tf_env.observation_spec(),
        input_fc_layer_params=None,
        lstm_size=(128,128),
        output_fc_layer_params=None,
        activation_fn=None)

But I’m facing an error saying

ValueError: Input 0 of layer bias_layer is incompatible with the layer: : expected min_ndim=2, found ndim=1. Full shape received: [0]

I see that the observation_spec I defined has only one array dimension [vel_y, vel_z, right_hip_theta, …]
But I can’t think of any other way to define the observation_spec with more than two array dimensions.

How can I apply input data with ndim=1 to LSTM network?

Do I need to replicate the data like
[ [vel_y, vel_z, right_hip_theta, … ] ,
[vel_y, vel_z, right_hip_theta, … ] , ]
to make ndim=2?

Or am I missing something else?

I’ve been struggling with this for more than two weeks now, and I hope I could make some progress…

#3

Does any of this code help? นโยบาย  |  TensorFlow Agents

In particular, if you run through the example make liberal use of print and .shape to understand what it is doing, specifically around the shape of the data fed into the network. And then also do the same thing with your code.