def train_shared_model(agent_params,env_params,training_params,id,actor,critic):
actor.train()
critic.train()
pid = os.getpid()
print(f"Intiantiating process PID {pid}")
env = Poker(env_params)
nS = env.state_space
nO = env.observation_space
nA = env.action_space
nB = env.betsize_space
nC = nA - 2 + nB
print_every = (training_params['epochs']+1) // 5
seed = 154
agent = ParallelAgent(nS,nO,nA,nB,seed,agent_params,actor,critic)
training_data = copy.deepcopy(training_params['training_data'])
for e in range(1,training_params['epochs']+1):
last_state,state,obs,done,mask,betsize_mask = env.reset()
while not done:
if env.game == pdt.GameTypes.HISTORICALKUHN:
actor_outputs = agent(state,mask,betsize_mask) if env.rules.betsize == True else agent(state,mask)
else:
actor_outputs = agent(last_state,mask,betsize_mask) if env.rules.betsize == True else agent(last_state,mask)
last_state,state,obs,done,mask,betsize_mask = env.step(actor_outputs)
ml_inputs = env.ml_inputs()
agent.learn(ml_inputs)
ml_inputs = detach_ml(ml_inputs)
for position in ml_inputs.keys():
training_data[position].append(ml_inputs[position])
if id == 0 and e % print_every == 0:
print(f'PID {pid}, Epoch {e}')
mongo = MongoDB()
mongo.clean_db()
mongo.store_data(training_data,env.db_mapping,training_params['training_round'],env.game,id,training_params['epochs'])
# return_dict = manager.dict()
# online training
seed = 123
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
actor = env_networks['actor'](seed, nS, nA, nB, agent_params)
critic = env_networks['critic'][args.critic](seed, nS, nA, nB,
agent_params)
del env
actor.share_memory() #.to(device)
critic.share_memory() #.to(device)
processes = []
num_processes = mp.cpu_count()
if args.clean:
print('Cleaning db')
mongo = MongoDB()
mongo.clean_db()
del mongo
for i in range(num_processes): # No. of processes
p = mp.Process(target=train_shared_model,
args=(agent_params, env_params, training_params, i,
actor, critic))
p.start()
processes.append(p)
for p in processes:
p.join()
basepath = os.path.abspath(sys.argv[0])
torch.save(
actor.state_dict(),
os.path.join(basepath, 'checkpoints/Historical_kuhn' + '_actor'))
torch.save(
critic.state_dict(),