def main():
parser = argparse.ArgumentParser(description=None)
parser.add_argument('-v',
'--verbose',
action='count',
dest='verbosity',
default=0,
help='Set verbosity.')
args = parser.parse_args()
if args.verbosity == 0:
logger.setLevel(logging.INFO)
elif args.verbosity >= 1:
logger.setLevel(logging.DEBUG)
env = gym.make('wob.mini.ClickDialog-v0')
env = jiminy.wrappers.experimental.SoftmaxClickMouse(env)
env = wob_vnc.MiniWoBCropper(env)
wob_vnc.configure(env,
wob_vnc.remotes_url(port_ofs=0,
hostname='0.0.0.0',
count=REMOTES_COUNT)
) # automatically creates a local docker container
observation_n = env.reset()
idx = 0
while True:
# your agent here
#
# Try sending this instead of a random action: ('KeyEvent', 'ArrowUp', True)
action_n = [env.action_space.sample() for ob in observation_n]
observation_n, reward_n, done_n, info = env.step(action_n)
print("idx: {}, reward: {}".format(idx * REMOTES_COUNT, reward_n))
idx += 1
return 0
writer = SummaryWriter(comment="-wob_click_" + name)
saves_path = os.path.join(SAVES_DIR, name)
os.makedirs(saves_path, exist_ok=True)
demo_samples = None
if args.demo:
demo_samples = vnc_demo.load_demo(args.demo, env_name)
if not demo_samples:
demo_samples = None
else:
print("Loaded %d demo samples, will use them during training" % len(demo_samples))
env = gym.make(env_name)
env = universe.wrappers.experimental.SoftmaxClickMouse(env)
env = wob_vnc.MiniWoBCropper(env)
wob_vnc.configure(env, wob_vnc.remotes_url(port_ofs=args.port_ofs, hostname=args.host, count=REMOTES_COUNT))
net = model_vnc.Model(input_shape=wob_vnc.WOB_SHAPE, n_actions=env.action_space.n)
if args.cuda:
net.cuda()
print(net)
optimizer = optim.Adam(net.parameters(), lr=LEARNING_RATE, eps=1e-3)
agent = ptan.agent.PolicyAgent(lambda x: net(x)[0], cuda=args.cuda,
apply_softmax=True)
exp_source = ptan.experience.ExperienceSourceFirstLast(
[env], agent, gamma=GAMMA, steps_count=REWARD_STEPS, vectorized=True)
best_reward = None
with common.RewardTracker(writer) as tracker:
with ptan.common.utils.TBMeanTracker(writer, batch_size=10) as tb_tracker:
env_name = args.env
if not env_name.startswith('wob.mini.'):
env_name = "wob.mini." + env_name
name = env_name.split('.')[-1] + "_" + args.name
writer = SummaryWriter(comment="-wob_click_mm_" + name)
saves_path = os.path.join(SAVES_DIR, name)
os.makedirs(saves_path, exist_ok=True)
env = gym.make(env_name)
env = universe.wrappers.experimental.SoftmaxClickMouse(env)
env = wob_vnc.MiniWoBCropper(env, keep_text=True)
wob_vnc.configure(
env,
wob_vnc.remotes_url(port_ofs=args.port_ofs,
hostname=REMOTES_HOST,
count=REMOTES_COUNT))
net = model_vnc.ModelMultimodal(input_shape=wob_vnc.WOB_SHAPE,
n_actions=env.action_space.n)
if args.cuda:
net.cuda()
print(net)
optimizer = optim.Adam(net.parameters(), lr=LEARNING_RATE, eps=1e-3)
preprocessor = model_vnc.MultimodalPreprocessor()
agent = ptan.agent.PolicyAgent(lambda x: net(x)[0],
cuda=args.cuda,
apply_softmax=True,
preprocessor=preprocessor)
exp_source = ptan.experience.ExperienceSourceFirstLast(