I just want to say your trained model has no effect about rl_a3c_pytorch

[dgriff777]

what os you running this on? I just ran it works fine for me

[2017-08-01 11:30:30,111] Making new env: Pong-v0
[2017-08-01 11:30:30,336] Clearing 6 monitor files from previous run (because force=True was provided)
[2017-08-01 11:30:30,345] Starting new video recorder writing to /Users/dgriffis/rl_a3c_pytorch/Pong-v0_monitor/openaigym.video.0.38559.video000000.mp4
2017-08-01 11:30:42,785 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:30:42,785] reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:30:42,804] Starting new video recorder writing to /Users/dgriffis/rl_a3c_pytorch/Pong-v0_monitor/openaigym.video.0.38559.video000001.mp4
2017-08-01 11:30:55,304 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:30:55,304] reward sum: 21.0, reward mean: 21.0000
2017-08-01 11:31:07,255 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:31:07,255] reward sum: 21.0, reward mean: 21.0000
2017-08-01 11:31:19,209 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:31:19,209] reward sum: 21.0, reward mean: 21.0000
2017-08-01 11:31:31,044 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:31:31,044] reward sum: 21.0, reward mean: 21.0000
2017-08-01 11:31:43,474 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:31:43,474] reward sum: 21.0, reward mean: 21.0000
2017-08-01 11:31:55,597 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:31:55,597] reward sum: 21.0, reward mean: 21.0000
2017-08-01 11:32:07,620 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:32:07,620] reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:32:07,628] Starting new video recorder writing to /Users/dgriffis/rl_a3c_pytorch/Pong-v0_monitor/openaigym.video.0.38559.video000008.mp4
2017-08-01 11:32:20,379 : reward sum: 21.0, reward mean: 21.0000
[2017-08-01 11:32:20,379] reward sum: 21.0, reward mean: 21.0000

[lucasjinreal]

I am running macOS, why am I just got -21 all the time. Which command are you using?

[lucasjinreal]

➜  rl_a3c_pytorch git:(master) ✗ python gym_eval.py --env Pong-v0 --num-episodes 100
[2017-08-02 09:06:09,852] Making new env: Pong-v0
[2017-08-02 09:06:10,107] Clearing 6 monitor files from previous run (because force=True was provided)
[2017-08-02 09:06:10,145] Starting new video recorder writing to /Volumes/xs/CodeSpace/AISpace/rl_space/rl_a3c_pytorch/Pong-v0_monitor/openaigym.video.0.35879.video000000.mp4
2017-08-02 09:06:20,499 : reward sum: -21.0, reward mean: -21.0000
[2017-08-02 09:06:20,499] reward sum: -21.0, reward mean: -21.0000
[2017-08-02 09:06:20,529] Starting new video recorder writing to /Volumes/xs/CodeSpace/AISpace/rl_space/rl_a3c_pytorch/Pong-v0_monitor/openaigym.video.0.35879.video000001.mp4
2017-08-02 09:06:30,942 : reward sum: -21.0, reward mean: -21.0000
[2017-08-02 09:06:30,942] reward sum: -21.0, reward mean: -21.0000

I trained whole night but when I cut it, nothing saved, can not find any model saved.....

[dgriff777]

Well first I would update repo cause I tinkered a lot with it past couple days but I know it's working fine now.. are you seeing the models in the trained_models folder?

[dgriff777]

Oh this your trained model? Are you seeing a saved model in the folder or and models? Should be a Pong-v0.dat file

[lucasjinreal]

Yeah, I seen it, but seems this model is you have trained already in your repo, cause besides Pong there are other moels. Anyway, how should I exactly call my model and the render env at mean time to see AI play?

[dgriff777]

Well I have set up up so models save in trained models folder and load there. If you want to watch gym_eval you have to do

Python gym_eval.py --env Pong-v0 --num-episodes 100 --render True

[lucasjinreal]

Well, got this old-fasioned black-white screen, and the result still -21, is the model didn't update?

[dgriff777]

That looks like you having dependencies issues with gym

Can u go in terminal: start python
Type:

Import gym
Import cv2

env=gym.make('Pong-v0')
frame=env.reset()
cv2.imshow('tt', frame)
cv2.waitKey(0)

Let me know what you see from that..

[lucasjinreal]

Well, weired, I also have gym on python3, it work totally fine. on Python2.7 it shows like this, no matter using cv2 or just env.render(), should I update this code to python3?

[lucasjinreal]

It has problem in save model, I update main.py default saved dir to trained_models_me, when I cut it, there has no my dir created.

[dgriff777]

You have to create directory first if not using folder trained_models. I did not set up to create saved folder directories

[dgriff777]

Yeah try that same code in python3 and see if pic of Atari screen comes up

[lucasjinreal]

Thanks dgriff, you are a master in reinforcement learning.

[dgriff777]

It's working now?! You welcome. Happy to help😄

[lucasjinreal]

Yeah, really thanks your help Pal.

[dgriff777]

Awesome! Have fun 👍

[lucasjinreal]

Hi, dgriff, sorry for the bother but I have one last question, in train.py I can't find codes to save model, I am new to pytorch, is there a way to store weights into specific dir and load it when run again?

[dgriff777]

When running training command you can do:

python main.py --env Pong-v0 --workers 32 --save-dir 'example_folder/'

And to load from specific folder:

python main.py --env Pong-v0 --workers 32  --load-dir 'example_folder/' --load True

Can also specify both in command

[dgriff777]

Loading code in training is in main.py

    if args.load:
        saved_state = torch.load(
            '{0}{1}.dat'.format(args.load_model_dir, args.env))

Saving model code is in test.py

            if reward_sum > args.save_score_level:
                player.model.load_state_dict(shared_model.state_dict())
                state_to_save = player.model.state_dict()
                torch.save(state_to_save, '{0}{1}.dat'.format(
                    args.save_model_dir, args.env))

And load model code in gym_eval.py

saved_state = torch.load(
    '{0}{1}.dat'.format(args.load_model_dir, args.env),
    map_location=lambda storage, loc: storage)