parser = argparse.ArgumentParser()
parser.add_argument("-d",
"--demo",
required=True,
help="Dir name to scan for demos")
parser.add_argument("-e",
"--env",
default=DEFAULT_ENV,
help="Environment name to load, default=" +
DEFAULT_ENV)
parser.add_argument("-o",
"--output",
required=True,
help="Output prefix to save images")
args = parser.parse_args()
demo = vnc_demo.load_demo(args.demo, args.env, read_text=True)
print("Loaded %d demo samples" % len(demo))
env = gym.make(args.env)
env = universe.wrappers.experimental.SoftmaxClickMouse(env)
for idx, (obs, action) in enumerate(demo):
fname = "%s_%04d.png" % (args.output, idx)
action_coords = env._points[action]
img, text = obs
wob_vnc.save_obs(img, fname, action_coords)
print(fname, text)
pass
help="Host with docker containers")
args = parser.parse_args()
device = torch.device("cuda" if args.cuda else "cpu")
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)
demo_samples = None
if args.demo:
demo_samples = vnc_demo.load_demo(args.demo, env_name, read_text=True)
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, keep_text=True)
wob_vnc.configure(
env,
wob_vnc.remotes_url(port_ofs=args.port_ofs,
hostname=args.host,
count=REMOTES_COUNT))
parser.add_argument("--demo", help="Demo dir to load. Default=No demo")
parser.add_argument("--host", default='localhost', help="Host with docker containers")
args = parser.parse_args()
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_" + 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)
discr = env._action_to_discrete(event)
discr2 = vnc_demo.mouse_to_action(x, y)
assert discr == discr2
pass
if __name__ == "__main__":
# test_mouse_coords()
parser = argparse.ArgumentParser()
parser.add_argument("-d", "--demo", required=True, help="Dir name to scan for demos")
parser.add_argument("-e", "--env", default=DEFAULT_ENV, help="Environment name to load, default=" + DEFAULT_ENV)
parser.add_argument("-o", "--output", required=True, help="Output prefix to save images")
args = parser.parse_args()
demo = vnc_demo.load_demo(args.demo, args.env, read_text=True)
print("Loaded %d demo samples" % len(demo))
env = gym.make(args.env)
env = universe.wrappers.experimental.SoftmaxClickMouse(env)
for idx, (obs, action) in enumerate(demo):
fname = "%s_%04d.png" % (args.output, idx)
action_coords = env._points[action]
img, text = obs
wob_vnc.save_obs(img, fname, action_coords)
print(fname, text)
pass