load-weights

[vkaragounis]

Is it possible to provide further details for the parameter 'load-weights'. I would like to test an example after the training completion.
What is the weight file's name? Is it "episode_reward.csv"? Can you please provide an example? Thank you.

[Jmiceli17]

Hi!
I apologize, the load-weights parameter is currently unused. I have a ticket to remove deprecated config parameters but just haven't gotten to it yet.

After training, you should be seeing logs generated in a nn/<experiment_name> directory in the form of .pt files. These are the weights of your model. For example, if you're using ac-indepnedent-SUMO.py with parameter sharing, you will see a .pt file that was saved every number of max-cycles.

Those files can be loaded using the load_state_dict() function in torch. An example of this can be seen in sumo_analysis.py

Note though that it matters if you used parameter sharing or not during training. If you DID NOT, then the .pt files will have their agent index assigned to them (i.e. training round 39500 will produce logs of 39500-0.pt, 39500-1.pt, 39500-2.pt, etc.). Example of loading these types of weights here here

[Jmiceli17]

I've opened the discussions page so I'm moving this thread there.

Feel free to provide more context with what you're trying to do and I can try to help out!

[vkaragounis]

Thanks for your help. I managed to run those commands below

python3 dqn-independent-ps.py -c experiments/sumo-2x2-ac-independent-ps.config
python3 sumo_analysis.py -c experiments/sumo-2x2-ac-independent-ps.config

with success and also watched the trained simulation working perfect. However when i tried some
other methods (ex. dqn-independent-empathetic.py, dqn-independent-greedy.py) i got an error:

File "/marl-experiments-main/dqn-independent-empathetic.py", line 372, in
rb[agent].put((obses[agent], actions[agent], rewards[agent], next_obses[agent], dones[agent]))
TypeError: tuple indices must be integers or slices, not str

Also i realized that 'global_obs' should always be TRUE because 'else' is missing. With 'True'
i got the error:

File "/marl-experiments-main/dqn-independent-empathetic.py", line 319, in
global_obs = np.hstack(list(obses.values()))
AttributeError: 'tuple' object has no attribute 'values'

Maybe the error occurs in more methods. I tested only 'dqn-independent-empathetic.py' and 'fdqn-independent-greedy.py'
Thanks again for your time.

[Jmiceli17]

Glad to hear you got those working!

I'll have to look into dqn-independent-empathetic.py I'm not sure why you're seeing that error.

global_obs is set to False by default in all the experiment files (see here for example)

I have not really tested global observations very much but I can take a look at it.