src/algos/only_episodic_counts.py [152:198]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    model.share_memory()

    initial_agent_state_buffers = []
    for _ in range(flags.num_buffers):
        state = model.initial_state(batch_size=1)
        for t in state:
            t.share_memory_()
        initial_agent_state_buffers.append(state)
    
    actor_processes = []
    ctx = mp.get_context('fork')
    free_queue = ctx.SimpleQueue()
    full_queue = ctx.SimpleQueue()

    episode_state_count_dict = dict()
    train_state_count_dict = dict()
    for i in range(flags.num_actors):
        actor = ctx.Process(
            target=act,
            args=(i, free_queue, full_queue, model, buffers, 
                episode_state_count_dict, train_state_count_dict, 
                initial_agent_state_buffers, flags))
        actor.start()
        actor_processes.append(actor)

    if 'MiniGrid' in flags.env: 
        learner_model = MinigridPolicyNet(env.observation_space.shape, env.action_space.n)\
            .to(device=flags.device)
    else:
        learner_model = MarioDoomPolicyNet(env.observation_space.shape, env.action_space.n)\
            .to(device=flags.device)
                        
    optimizer = torch.optim.RMSprop(
        learner_model.parameters(),
        lr=flags.learning_rate,
        momentum=flags.momentum,
        eps=flags.epsilon,
        alpha=flags.alpha)


    def lr_lambda(epoch):
        return 1 - min(epoch * T * B, flags.total_frames) / flags.total_frames

    scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)

    logger = logging.getLogger('logfile')
    stat_keys = [
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



src/algos/torchbeast.py [132:178]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    model.share_memory()

    initial_agent_state_buffers = []
    for _ in range(flags.num_buffers):
        state = model.initial_state(batch_size=1)
        for t in state:
            t.share_memory_()
        initial_agent_state_buffers.append(state)
    
    actor_processes = []
    ctx = mp.get_context('fork')
    free_queue = ctx.SimpleQueue()
    full_queue = ctx.SimpleQueue()

    episode_state_count_dict = dict()
    train_state_count_dict = dict()
    for i in range(flags.num_actors):
        actor = ctx.Process(
            target=act,
            args=(i, free_queue, full_queue, model, buffers, 
                episode_state_count_dict, train_state_count_dict, 
                initial_agent_state_buffers, flags))
        actor.start()
        actor_processes.append(actor)

    if 'MiniGrid' in flags.env: 
        learner_model = MinigridPolicyNet(env.observation_space.shape, env.action_space.n)\
            .to(device=flags.device)
    else:
        learner_model = MarioDoomPolicyNet(env.observation_space.shape, env.action_space.n)\
            .to(device=flags.device)

    optimizer = torch.optim.RMSprop(
        learner_model.parameters(),
        lr=flags.learning_rate,
        momentum=flags.momentum,
        eps=flags.epsilon,
        alpha=flags.alpha)


    def lr_lambda(epoch):
        return 1 - min(epoch * T * B, flags.total_frames) / flags.total_frames

    scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)

    logger = logging.getLogger('logfile')
    stat_keys = [
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -