cloud integration

cleanrl/utils/submit_exp.py

@@ -12,7 +12,7 @@
 
 parser = argparse.ArgumentParser(description='CleanRL Experiment Submission')
 # experiment generation
-parser.add_argument('--exp-script', type=str, default="exp.sh",
+parser.add_argument('--exp-script', type=str, default="debug.sh",
                     help='the file name of this experiment')
 parser.add_argument('--algo', type=str, default="ppo.py",
                     help='the algorithm that will be used')
@@ -22,15 +22,13 @@
                     help='total timesteps of the experiments')
 parser.add_argument('--other-args', type=str, default="",
                     help="the entity (team) of wandb's project")
-parser.add_argument('--wandb-project-name', type=str, default="cleanRL",
-                    help="the wandb's project name")
 
 # experiment submission
 parser.add_argument('--job-queue', type=str, default="cleanrl",
                    help='the name of the job queue')
 parser.add_argument('--wandb-key', type=str, default="",
                    help='the wandb key. If not provided, the script will try to read the env variable `WANDB_KEY`')
-parser.add_argument('--docker-repo', type=str, default="vwxyzjn/gym-microrts:latest",
+parser.add_argument('--docker-repo', type=str, default="vwxyzjn/cleanrl:latest",
                    help='the name of the job queue')
 parser.add_argument('--job-definition', type=str, default="cleanrl",
                    help='the name of the job definition')
@@ -53,29 +51,21 @@
 
 args = parser.parse_args()
 
-template = '''
-for seed in {{1..2}}
+final_str = f'''
+for seed in {{1..{args.num_seed}}}
 do
-    (sleep 0.3 && nohup xvfb-run -a python {} \\
-    --gym-id {} \\
-    --total-timesteps {} \\
-    --wandb-project-name {} \\
+    (sleep 0.3 && nohup xvfb-run -a python {args.algo} \\
+    {args.other_args} \\
+    --seed $seed
     --prod-mode \\
-    {} \\
     --capture-video \\
-    --seed $seed
     ) >& /dev/null &
 done
 '''
 
-final_str = ""
-for env in args.gym_ids:
-    final_str += template.format(args.algo, env, args.total_timesteps, args.wandb_project_name, args.other_args)
-
 with open(f"{args.exp_script}", "w+") as f:
     f.write(final_str)
 
-
 # get env variable values
 if not args.wandb_key:
     args.wandb_key = os.environ['WANDB_KEY']
@@ -96,7 +86,7 @@
             final_run_cmds[-1] = ['curl', '-O', link, ';'] + final_run_cmds[-1]
 
 
-run_ids = [wandb.util.generate_id() for _ in range (len(args.gym_ids)*args.num_seed)]
+run_ids = [wandb.util.generate_id() for _ in range (args.num_seed)]
 
 final_str = ""
 cores = 40

cloud/README.md

@@ -8,56 +8,79 @@ easy to manage and reproducible, we use Terraform to spin up services.
 
 ```bash
 wandb login
+pip install cleanrl --upgrade
 git clone https://github.com/vwxyzjn/cleanrl
 cd cleanrl/cloud
 terraform init
 terraform apply
 pip install awscli
 python -m awscli authenticate
 
-# submit a job using AWS's compute-optimized spot instances 
-python -m cleanrl.utils.submit_exp --exp-script ppo.sh \
-    --algo ppo.py \
+# dry run to inspect the generated docker command
+python -m cleanrl.utils.submit_exp --algo ppo.py \
     --other-args "--gym-id CartPole-v0 --wandb-project-name cleanrl --total-timesteps 100000 --cuda True" \
     --job-queue cpu_spot \
     --job-definition cleanrl \
     --num-seed 1 \
     --num-vcpu 1 \
     --num-memory 2000 \
     --num-hours 48.0
+```
+The generated docker command should look like
+```
+docker run -d --cpuset-cpus="0" -e WANDB=xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx -e WANDB_RESUME=allow -e WANDB_RUN_ID=2avt9i7l vwxyzjn/cleanrl:latest /bin/bash -c "python ppo.py --gym-id CartPole-v0 --wandb-project-name cleanrl --total-timesteps 100000 --cuda True --seed 1"
+```
 
-# submit a job using AWS's accelerated-computing spot instances 
-python -m cleanrl.utils.submit_exp --exp-script ppo_atari_visual.sh \
-    --algo ppo_atari_visual.py \
+Submit a job using AWS's compute-optimized spot instances 
+```
+python -m cleanrl.utils.submit_exp --algo ppo.py \
+    --other-args "--gym-id CartPole-v0 --wandb-project-name cleanrl --total-timesteps 100000 --cuda True" \
+    --job-queue cpu_spot \
+    --job-definition cleanrl \
+    --num-seed 1 \
+    --num-vcpu 1 \
+    --num-memory 2000 \
+    --num-hours 48.0 \
+    --submit-aws
+```
+
+Submit a job using AWS's accelerated-computing spot instances 
+```
+python -m cleanrl.utils.submit_exp --algo ppo_atari_visual.py \
     --other-args "--gym-id BreakoutNoFrameskip-v4 --wandb-project-name cleanrl --cuda True" \
     --job-queue gpu_spot \
     --job-definition cleanrl \
     --num-seed 1 \
     --num-vcpu 1 \
     --num-gpu 1 \
     --num-memory 4000 \
-    --num-hours 48.0
+    --num-hours 48.0 \
+    --submit-aws
+```
 
-# submit a job using AWS's compute-optimized on-demand instances 
-python -m cleanrl.utils.submit_exp --exp-script ppo.sh \
-    --algo ppo.py \
+Submit a job using AWS's compute-optimized on-demand instances 
+```
+python -m cleanrl.utils.submit_exp --algo ppo.py \
     --other-args "--gym-id CartPole-v0 --wandb-project-name cleanrl --total-timesteps 100000 --cuda True" \
     --job-queue cpu_on_demand \
     --job-definition cleanrl \
     --num-seed 1 \
     --num-vcpu 1 \
     --num-memory 2000 \
-    --num-hours 48.0
+    --num-hours 48.0 \
+    --submit-aws
+```
 
-# submit a job using AWS's accelerated-computing on-demand instances 
-python -m cleanrl.utils.submit_exp --exp-script ppo_atari_visual.sh \
-    --algo ppo_atari_visual.py \
+Submit a job using AWS's accelerated-computing on-demand instances 
+```
+python -m cleanrl.utils.submit_exp --algo ppo_atari_visual.py \
     --other-args "--gym-id BreakoutNoFrameskip-v4 --wandb-project-name cleanrl --cuda True" \
     --job-queue gpu_on_demand \
     --job-definition cleanrl \
     --num-seed 1 \
     --num-vcpu 1 \
     --num-gpu 1 \
     --num-memory 4000 \
-    --num-hours 48.0
+    --num-hours 48.0 \
+    --submit-aws
 ```