"num_gpus": 1,
"hiddens": [],
"schedule_max_timesteps": 7500000,
"timesteps_per_iteration": 4000,
"exploration_fraction": 0.8,
"exploration_final_eps": 0.02,
"lr": 1e-3,
"model": {
"custom_model": "parametric",
"custom_options": {}, # extra options to pass to your model
}
})
# Attempt to restore from checkpoint if possible.
if os.path.exists(CHECKPOINT_FILE):
checkpoint_path = open(CHECKPOINT_FILE).read()
print("Restoring from checkpoint path", checkpoint_path)
dqn.restore(checkpoint_path)
# run the new command using the given tracer
# make a report, placing output in the current directory
# Serving and training loop
while True:
print(pretty_print(dqn.train()))
checkpoint_path = dqn.save()
print("Last checkpoint", checkpoint_path)
with open(CHECKPOINT_FILE, "w") as f:
f.write(checkpoint_path)
"timesteps_per_iteration": 200,
"env_config": {
"observation_size": args.observation_size,
"action_size": args.action_size,
},
})
elif args.run == "PG":
agent = PGAgent(
env="srv",
config={
"num_workers": 0,
"env_config": {
"observation_size": args.observation_size,
"action_size": args.action_size,
},
})
# Attempt to restore from checkpoint if possible.
if os.path.exists(args.checkpoint_file):
checkpoint_file = open(args.checkpoint_file).read()
print("Restoring from checkpoint path", checkpoint_file)
agent.restore(checkpoint_file)
# Serving and training loop
while True:
print(pretty_print(agent.train()))
checkpoint_file = agent.save()
print("Last checkpoint", checkpoint_file)
with open(args.checkpoint_file, "w") as f:
f.write(checkpoint_file)