"""
import torch
import ptan
from lib import data, utils, model
from tensorboardX import SummaryWriter
if __name__=='__main__':
message = '*'*10 + ' A2C on Atari ' +'*'*10
args = utils.argpars_dqn(message)
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,
default='pong',
choices=GAMES,
help='name of the game: invaders(default), pong, breakout')
parser.add_argument('--cuda',
action='store_true',
help='Train on GPU when available')
args = parser.parse_args()
device = 'cuda' if (args.cuda and torch.cuda.is_available()) else 'cpu'
frames = mp.Value('i', 0)
episodes = mp.Value('i', 0)
params = data.params[args.env]
env = utils.createEnvs(params)[0]
shape = env.observation_space.shape
actions = env.action_space.n
net = model.DDQN(shape, actions).to(device)
net.share_memory()
print(net)
tgt_net = ptan.agent.TargetNet(net)
selector = ptan.actions.EpsilonGreedyActionSelector()
agent = ptan.agent.DQNAgent(net, selector, device=device)
buffer = ptan.experience.ExperienceReplayBuffer(None, params.buffer_size)
optimizer = torch.optim.Adam(net.parameters(), lr=params.lr)
exp_queue = mp.Queue(THREADS)
proc_list = []
for n in range(THREADS):
# mp.set_start_method('spawn')
os.environ['OMP_NUM_THREADS'] = '1'
os.environ['MKL_THREADING_LAYER'] = 'GNU'
message = '*' * 10 + ' A3C on Atari ' + '*' * 10
args = utils.argpars_dqn(message)
params = data.params[args.env]
utils.update_params(params, args)
# For A2C/A3C to converge we need a high number of environments
params.n_envs = max(params.n_envs, 8)
device = 'cuda' if args.cuda else 'cpu'
env = utils.createEnvs(
params,
stack_frames=params.frame_stack)[0] # we can get rid of this one
shape = env.observation_space.shape
actions = env.action_space.n
net = model.A2CNet(shape, actions)
net.to(device)
net.share_memory()
optimizer = torch.optim.Adam(net.parameters(), lr=params.lr)
mean_monitor = utils.MeanRewardsMonitor(env, net, ALGORITHM,
params.solve_rewards)
writer = SummaryWriter(logdir=mean_monitor.runs_dir,
comment=params.frame_stack)
parser.add_argument('-s', '--steps', default=4, type=int,
help='steps to skip while training')
parser.add_argument('-n', '--envs', default=3, type=int,
help='Number of environments to run simultaneously')
parser.add_argument('-g', '--game', default='invaders', choices=GAMES,
help='OpenAI gym environment name')
parser.add_argument('--play', action='store_true',
help='Play a game when the environment is solved')
args = parser.parse_args()
device = 'cuda' if args.cuda else 'cpu'
params = data.params[args.game]
envs = utils.createEnvs(params)
shape = envs[0].observation_space.shape
actions = envs[0].action_space.n
net = model.DDQN(shape, actions).to(device)
tgt_net = ptan.agent.TargetNet(net)
selector = ptan.actions.EpsilonGreedyActionSelector(
epsilon=params.eps_start)
eps_tracker = ptan.actions.EpsilonTracker(
selector, params.eps_start, params.eps_final, params.eps_frames*args.envs)
agent = ptan.agent.DQNAgent(net, selector, device=device)
exp_source = ptan.experience.ExperienceSourceFirstLast(
envs, agent, params.gamma, steps_count=args.steps)
mp.set_start_method('spawn')
os.environ['OMP_NUM_THREADS'] = '1'
os.environ['MKL_THREADING_LAYER'] = 'GNU'
message = '*' * 10 + ' A3C GRU on Atari ' + '*' * 10
args = utils.argpars_dqn(message)
params = data.params[args.env]
utils.update_params(params, args)
# For A2C/A3C to converge we need a lot of environments to draw observations from
# This will ensure sample i.i.d (somehow!)
params.n_envs = max(params.n_envs, 8)
device = 'cuda' if args.cuda else 'cpu'
env = utils.createEnvs(params, stack_frames=params.frame_stack)[0]
shape = env.observation_space.shape
actions = env.action_space.n
net = model.A2Cgru(shape, actions)
net.to(device)
net.share_memory()
optimizer = torch.optim.Adam(net.parameters(), lr=params.lr)
mean_monitor = utils.MeanRewardsMonitor(env, net, ALGORITHM,
params.solve_rewards)
writer = SummaryWriter(logdir=mean_monitor.runs_dir,
comment=params.frame_stack)
lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
action='store_true',
help='Activate GPU in training')
parser.add_argument(
'--stack',
default=4,
type=int,
help=
'stack N frames in each observation: this will change the input of the nn'
)
args = parser.parse_args()
params = data.params[args.env]
torch.manual_seed(params.seed)
device = 'cuda' if (args.cuda and torch.cuda.is_available()) else 'cpu'
envs = utils.createEnvs(params, stack_frames=args.stack)
shape = envs[0].observation_space.shape
actions = envs[0].action_space.n
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(
envs, agent, params.gamma, steps_count=params.steps)
type=int,
help='Frames to skip when stacking. Must specifiy when selecting --lw')
parser.add_argument('--model',
type=str,
help='Path to the trained state dict model')
parser.add_argument(
'--record',
action='store_true',
help='record a video of the game and store it in ~/Videos')
args = parser.parse_args()
params = data.params[args.env]
params.max_steps = None
env = utils.createEnvs(params,
stack_frames=args.stack,
episodic_life=False,
reward_clipping=False,
skip=args.skip)[0]
# recording the game
if args.record: env = gym.wrappers.Monitor(env, "Videos", force=True)
shape, actions = env.observation_space.shape, env.action_space.n
net = model.A2Cgru(shape, actions)
print(net)
if args.model: net.load_state_dict(load(args.model, map_location='cpu'))
selector = ProbabilityActionSelector()
agent = ActorCriticAgent(net, selector, apply_softmax=True)
utils.play(env, agent, wait=args.wait, render=args.render)