New hpset for dataparallel training

Author: cswinter

hyper_params.py

@@ -57,8 +57,7 @@ def __init__(self):
 
         # Data parallel
         self.rank = 0
-        self.parallelism = 1
-        self.discovery_port = 29500
+        self.parallelism = 1           # Number of data parallel processes. Must be set explicitly when using schedule.py, otherwise runner.py will just spawn a single process.
 
         # Observations
         self.obs_allies = 10            # Max number of allied drones returned by the env
@@ -228,6 +227,20 @@ def standard():
 
         return hps
 
+    @staticmethod
+    def standard_dataparallel():
+        hps = HyperParams.standard()
+
+        hps.steps = 300e6
+
+        hps.adr_hstepsize = 2e-06
+        hps.batches_per_update = 16
+        hps.num_envs = 64
+        hps.num_self_play = 32
+        hps.lr = 0.0003
+        hps.final_lr = 0.00003
+
+        return hps
 
     @staticmethod
     def arena():

main.py

@@ -1003,6 +1003,8 @@ def main():
         hps = HyperParams.arena()
     elif args.hpset == 'standard':
         hps = HyperParams.standard()
+    elif args.hpset == 'standard_dataparallel':
+        hps = HyperParams.standard_dataparallel()
     elif args.hpset == 'micro_practice':
         hps = HyperParams.micro_practice()
     elif args.hpset == 'scout':
@@ -1025,8 +1027,6 @@ def main():
         hps.objective = envs.Objective(hps.objective)
 
     if hps.parallelism > 1:
-        os.environ['MASTER_ADDR'] = 'localhost'
-        os.environ['MASTER_PORT'] = str(hps.discovery_port)
         dist.init_process_group(backend='gloo', rank=hps.rank, world_size=hps.parallelism)
 
     if hps.rank == 0:

runner.py

@@ -129,7 +129,12 @@ def git(args, workdir=dir):
                 with open(logpath, "w+") as outfile:
                     retcode = subprocess.call(
                         ["python3", "main.py", "--out-dir", out_dir] + args,
-                        env=dict(os.environ, CUDA_VISIBLE_DEVICES=str(job.device)),
+                        env=dict(
+                            os.environ,
+                            CUDA_VISIBLE_DEVICES=str(job.device),
+                            MASTER_ADDR='localhost',
+                            MASTER_PORT=str(job.discovery_port),
+                        ),
                         stdout=outfile, stderr=outfile, cwd=dir
                     )
                 if retcode != 0:
@@ -197,13 +202,14 @@ def __init__(self, repo_path, revision, params, handle, parallelism):
         self.parallelism = parallelism
         self.descriptor = "-".join([revision[:6]] + [f'{k}{v}' for k, v in params.items()])
         self.rank = 0
+        self.discovery_port = None
 
     def set_device(self, device, rank, discovery_port):
         self.device = device
         self.rank = rank
+        self.discovery_port = discovery_port
         self.params['device'] = device
         self.params['rank'] = rank
-        self.params['discovery_port'] = discovery_port
 
 
 @click.command()