def env_setup(name='MiniGrid-Empty-8x8-v0'):
env_name = name
env = gym.make(env_name)
env = RGBImgObsWrapper(env)
env.max_steps = min(env.max_steps, 200)
env.seed(12345)
env.reset()
return env
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
gin.parse_config_files_and_bindings([
os.path.join(mon_minigrid.GIN_FILES_PREFIX, '{}.gin'.format(
FLAGS.env_name))
],
bindings=FLAGS.gin_bindings,
skip_unknown=False)
env_id = mon_minigrid.register_environment()
env = gym.make(env_id)
env = RGBImgObsWrapper(env) # Get pixel observations
# Get tabular observation and drop the 'mission' field:
env = tabular_wrapper.TabularWrapper(env, get_rgb=True)
env.reset()
num_frames = 0
max_num_frames = 500
if not tf.io.gfile.exists(FLAGS.file_path):
tf.io.gfile.makedirs(FLAGS.file_path)
print('Available actions:')
for a in ACTION_MAPPINGS:
print('\t{}: "{}"'.format(ACTION_MAPPINGS[a], a))
print()
undisc_return = 0
while num_frames < max_num_frames:
draw_ascii_view(env)
a = input('action: ')
if a not in ACTION_MAPPINGS:
print('Unrecognized action.')
continue
action = env.DirectionalActions[ACTION_MAPPINGS[a]].value
obs, reward, done, _ = env.step(action)
undisc_return += reward
num_frames += 1
print('t:', num_frames, ' s:', obs['state'])
# Draw environment frame just for simple visualization
plt.imshow(obs['image'])
path = os.path.join(FLAGS.file_path, 'obs_{}.png'.format(num_frames))
plt.savefig(path)
plt.clf()
if done:
break
print('Undiscounted return: %.2f' % undisc_return)
env.close()
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
gin.parse_config_files_and_bindings(
[os.path.join(mon_minigrid.GIN_FILES_PREFIX, 'classic_fourrooms.gin')],
bindings=FLAGS.gin_bindings,
skip_unknown=False)
env_id = mon_minigrid.register_environment()
env = gym.make(env_id)
env = RGBImgObsWrapper(env) # Get pixel observations
# Get tabular observation and drop the 'mission' field:
env = tabular_wrapper.TabularWrapper(env, get_rgb=True)
env.reset()
num_frames = 0
max_num_frames = 500
if not tf.io.gfile.exists(FLAGS.file_path):
tf.io.gfile.makedirs(FLAGS.file_path)
undisc_return = 0
while num_frames < max_num_frames:
# Act randomly
obs, reward, done, _ = env.step(env.action_space.sample())
undisc_return += reward
num_frames += 1
print('t:', num_frames, ' s:', obs['state'])
# Draw environment frame just for simple visualization
plt.imshow(obs['image'])
path = FLAGS.file_path + str(num_frames) + '.png'
plt.savefig(path)
plt.clf()
if done:
break
print('Undiscounted return: %.2f' % undisc_return)
env.close()
def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
gin.parse_config_files_and_bindings([
os.path.join(mon_minigrid.GIN_FILES_PREFIX, '{}.gin'.format(FLAGS.env))
],
bindings=FLAGS.gin_bindings,
skip_unknown=False)
env_id = mon_minigrid.register_environment()
env = gym.make(env_id)
env = RGBImgObsWrapper(env) # Get pixel observations
# Get tabular observation and drop the 'mission' field:
env = mdp_wrapper.MDPWrapper(env)
env = coloring_wrapper.ColoringWrapper(env)
values = np.zeros(env.num_states)
error = FLAGS.tolerance * 2
i = 0
while error > FLAGS.tolerance:
new_values = np.copy(values)
for s in range(env.num_states):
max_value = 0.
for a in range(env.num_actions):
curr_value = (env.rewards[s, a] + FLAGS.gamma *
np.matmul(env.transition_probs[s, a, :], values))
if curr_value > max_value:
max_value = curr_value
new_values[s] = max_value
error = np.max(abs(new_values - values))
values = new_values
i += 1
if i % 1000 == 0:
print('Error after {} iterations: {}'.format(i, error))
print('Found V* in {} iterations'.format(i))
print(values)
if FLAGS.values_image_file is not None:
cmap = cm.get_cmap('plasma', 256)
norm = colors.Normalize(vmin=min(values), vmax=max(values))
obs_image = env.render_custom_observation(env.reset(),
values,
cmap,
boundary_values=[1.0, 4.5])
m = cm.ScalarMappable(cmap=cmap, norm=norm)
m.set_array(obs_image)
plt.imshow(obs_image)
plt.colorbar(m)
plt.savefig(FLAGS.values_image_file)
plt.clf()
env.close()