net = model.DDQN(shape, actions).to(device)
tgt_net = ptan.agent.TargetNet(net)
selector = ptan.actions.EpsilonGreedyActionSelector()
eps_tracker = ptan.actions.EpsilonTracker(selector, params.eps_start,
params.eps_final, params.eps_frames)
agent = ptan.agent.DQNAgent(net, selector, device=device)
exp_src = ptan.experience.ExperienceSourceFirstLast(
env, agent, params.gamma, steps_count=1)
buffer = ptan.experience.ExperienceReplayBuffer(
exp_src, params.buffer_size)
mean_monitor = utils.MeanRewardsMonitor(
env, net, ALGORITHM, params.solve_rewards)
writer = SummaryWriter(logdir=mean_monitor.runs_dir,
comment=params.frame_stack)
optimizer = torch.optim.Adam(net.parameters(), lr=params.lr)
lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,
mode='max', patience=10_000, verbose=True,
factor=0.75, min_lr=params.min_lr, cooldown=10_000)
print(net)
print('*'*10, ' Start Training ',
env.game, ' {} '.format(device), '*'*10)
frame = 0
episode = 0
params = data.params[args.env]
utils.update_params(params, args)
params.n_envs = max(params.n_envs, 8)
device = 'cuda' if args.cuda else 'cpu'
envs = utils.createEnvs(params, stack_frames=2)
shape = envs[0].observation_space.shape
actions = envs[0].action_space.n
net = model.A2CNet(shape, actions)
net.to(device)
agent = ptan.agent.ActorCriticAgent(net, device=device, apply_softmax=True)
exp_src = ptan.experience.ExperienceSourceFirstLast(envs, agent, params.gamma,steps_count=params.steps)
generator = utils.BatchGenerator(exp_src, params)
mean_monitor = utils.MeanRewardsMonitor(envs[0], net, 'A2C', params.solve_rewards)
writer = SummaryWriter(logdir=mean_monitor.runs_dir,comment=params.frame_stack)
optimizer = torch.optim.Adam(net.parameters(), lr=params.lr)
# lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', factor=0.75, patience=20000,
# cooldown=20000, verbose=True, min_lr=params.min_lr)
print('# Parameters: ', utils.count_parameters(net))
print(net)
print('*'*10, ' Start Training ',
envs[0].game, ' {} '.format(device), '*'*10)
