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
wob_vnc.configure(env, REMOTE_ADDR)
net = model_vnc.ModelMultimodal(input_shape=wob_vnc.WOB_SHAPE,
n_actions=env.action_space.n)
if args.model:
net.load_state_dict(torch.load(args.model))
preprocessor = model_vnc.MultimodalPreprocessor.load(args.model[:-4] +
".pre")
else:
preprocessor = model_vnc.MultimodalPreprocessor()
env.reset()
for round_idx in range(args.count):
action = env.action_space.sample()
step_idx = 0
while True:
obs, reward, done, info, idle_count = step_env(env, action)
print(step_idx, reward, done, idle_count)
img_name = "%s_r%02d_s%04d_%.3f_i%02d_d%d.png" % (
args.name, round_idx, step_idx, reward, idle_count, int(done))
obs_v = preprocessor([obs])
logits_v = net(obs_v)[0]
policy = F.softmax(logits_v).data.numpy()[0]
action = np.random.choice(len(policy), p=policy)
wob_vnc.save_obs(obs[0], img_name, action=action)
step_idx += 1
if done or reward != 0:
print("Round %d done" % round_idx)
break
pass
env = universe.wrappers.experimental.SoftmaxClickMouse(env)
env = wob_vnc.MiniWoBCropper(env, keep_text=True)
wob_vnc.configure(env, REMOTE_ADDR)
net = model_vnc.ModelMultimodal(input_shape=wob_vnc.WOB_SHAPE, n_actions=env.action_space.n)
if args.model:
net.load_state_dict(torch.load(args.model))
preprocessor = model_vnc.MultimodalPreprocessor.load(args.model[:-4] + ".pre")
else:
preprocessor = model_vnc.MultimodalPreprocessor()
env.reset()
for round_idx in range(args.count):
action = env.action_space.sample()
step_idx = 0
while True:
obs, reward, done, info, idle_count = step_env(env, action)
print(step_idx, reward, done, idle_count)
img_name = "%s_r%02d_s%04d_%.3f_i%02d_d%d.png" % (
args.name, round_idx, step_idx, reward, idle_count, int(done))
obs_v = preprocessor([obs])
logits_v = net(obs_v)[0]
policy = F.softmax(logits_v, dim=1).data.numpy()[0]
action = np.random.choice(len(policy), p=policy)
wob_vnc.save_obs(obs[0], img_name, action=action)
step_idx += 1
if done or reward != 0:
print("Round %d done" % round_idx)
break
pass
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