batching dynamic number of nodes

[encorechow]

The background is I was running a reinforcement learning task which has a dynamic number of agents. I was trying to build these agents as a graph. In the training phase, I found it seems impossible to batching each sample together since the number of agents was dynamically changing. for example, one sample may has [1(batch size), 3(number of nodes/agents), 1024(feature number)], the other may has [1(batch size), 2(number of nodes/agents), 1024(feature number)]. How could I batching them together for training?
I will be very appreciate if anyone can help me with this.

[rusty1s]

This section briefly explains the mini-batching technique used in PyG that is well able to operate on a dynamic number of nodes. The trick is that you concatenate your "agents" in the first dimension (and drop the batch dimension). Let me know if you need any more guidance.