def _tasks(config, params):
tasks = params.get('tasks', ['cheetah_run'])
if tasks == 'all':
tasks = [
'cartpole_balance', 'cartpole_swingup', 'finger_spin',
'cheetah_run', 'cup_catch', 'walker_walk'
]
tasks = [getattr(tasks_lib, name)(config, params) for name in tasks]
config.isolate_envs = params.get('isolate_envs', 'thread')
def common_spaces_ctor(task, action_spaces):
env = task.env_ctor()
env = control.wrappers.SelectObservations(env, ['image'])
env = control.wrappers.PadActions(env, action_spaces)
return env
if len(tasks) > 1:
action_spaces = [task.env_ctor().action_space for task in tasks]
for index, task in enumerate(tasks):
env_ctor = functools.partial(common_spaces_ctor, task,
action_spaces)
tasks[index] = tasks_lib.Task(task.name, env_ctor, task.max_length,
['reward'])
for name in tasks[0].state_components:
config.heads[name] = networks.feed_forward
config.zero_step_losses[name] = 1.0
config.tasks = tasks
return config
def _tasks(config, params):
tasks = params.get('tasks', ['cheetah_run']) # tasks = ['breakout']
if tasks == 'all':
tasks = [
'cartpole_balance', 'cartpole_swingup', 'finger_spin',
'cheetah_run', 'cup_catch', 'walker_walk', 'pendulum', 'carla',
'breakout'
]
tasks = [
getattr(tasks_lib, name)(config, params) for name in tasks
] # tasks_lib is tasks.py # config is not used for creating tasks.
def common_spaces_ctor(task, action_spaces):
env = task.env_ctor()
env = control.wrappers.SelectObservations(env, ['image'])
env = control.wrappers.PadActions(env, action_spaces)
return env
if len(tasks) > 1:
action_spaces = [task.env_ctor().action_space for task in tasks]
for index, task in enumerate(tasks):
env_ctor = functools.partial(common_spaces_ctor, task,
action_spaces)
tasks[index] = tasks_lib.Task(task.name, env_ctor, task.max_length,
['reward'])
for name in tasks[
0].state_components: # state_components of task, e.g. ['reward', 'position', 'velocity']
config.heads[name] = networks.feed_forward
config.zero_step_losses[name] = 1.0
config.tasks = tasks
return config
def _tasks(config, params):
tasks = params.get('tasks', ['cheetah_run'])
tasks = [getattr(tasks_lib, name)(config, params) for name in tasks]
config.isolate_envs = params.get('isolate_envs', 'thread')
def common_spaces_ctor(task, action_spaces):
env = task.env_ctor()
env = control.wrappers.SelectObservations(env, ['image'])
env = control.wrappers.PadActions(env, action_spaces)
return env
if len(tasks) > 1:
action_spaces = [task.env_ctor().action_space for task in tasks]
for index, task in enumerate(tasks):
env_ctor = tools.bind(common_spaces_ctor, task, action_spaces)
tasks[index] = tasks_lib.Task(task.name, env_ctor, task.max_length,
['reward'])
for name in tasks[0].state_components:
if name == 'reward' or params.get('state_diagnostics', False):
config.heads[name] = tools.bind(config.head_network,
stop_gradient=name
not in config.gradient_heads)
config.loss_scales[name] = 1.0
config.tasks = tasks
return config
def vizdoom_takecover_tasks(config, params):
tasks = params.get('tasks', ['cheetah_run'])
tasks = [getattr(tasks_lib, name)(config, params) for name in tasks]
# tasks = [getattr(tasks_lib, name)(config, params) for idx, name in enumerate(tasks)]
config.isolate_envs = params.get('isolate_envs', 'thread')
env_ctor_called = 0
def common_spaces_ctor(task, action_spaces, index=None):
env = task.env_ctor()
env = control.wrappers.SelectObservations(env, ['image'])
env = control.wrappers.PadActions(env, action_spaces)
return env
if len(tasks) > 1:
action_spaces = [task.env_ctor().action_space for task in tasks]
for index, task in enumerate(tasks):
env_ctor = functools.partial(common_spaces_ctor, task,
action_spaces)
# env_ctor = lambda: common_spaces_ctor(task, action_spaces)
tasks[index] = tasks_lib.Task(task.name, env_ctor, task.max_length,
['reward'])
for name in tasks[0].state_components:
config.heads[name] = networks.feed_forward
config.zero_step_losses[name] = 1.0
config.tasks = tasks
config.test_tasks = tasks
config.random_collect_tasks = tasks
return config
def _tasks(config, params):
tasks = params.get('tasks', ['cheetah_run'])
if tasks == 'all':
tasks = [
'cartpole_balance', 'cartpole_swingup', 'finger_spin',
'cheetah_run', 'cup_catch', 'walker_walk', 'vizdoom_basic',
'gym_cheetah', 'gym_breakout', 'gym_seaquest', 'gym_pong',
'gym_vizdoom_takecover'
]
if tasks == ['gym_vizdoom_cig']:
tasks = []
# Multi Planet Train Tasks
tasks.append('gym_vizdoom_cig_0_1')
tasks.append('gym_vizdoom_cig_1_1')
tasks.append('gym_vizdoom_cig_2_1')
tasks.append('gym_vizdoom_cig_3_1')
tasks.append('gym_vizdoom_cig_4_1')
tasks.append('gym_vizdoom_cig_5_1')
tasks.append('gym_vizdoom_cig_6_1')
tasks.append('gym_vizdoom_cig_7_1')
tasks.append('gym_vizdoom_cig_8_1')
tasks.append('gym_vizdoom_cig_9_1')
# Multi Planet Test Tasks
# tasks.append('gym_vizdoom_cig_0_2')
# tasks.append('gym_vizdoom_cig_1_2')
# tasks.append('gym_vizdoom_cig_2_2')
# tasks.append('gym_vizdoom_cig_3_2')
# tasks.append('gym_vizdoom_cig_4_2')
# tasks.append('gym_vizdoom_cig_5_2')
# tasks.append('gym_vizdoom_cig_6_2')
# tasks.append('gym_vizdoom_cig_7_2')
# tasks.append('gym_vizdoom_cig_8_2')
tasks.append('gym_vizdoom_cig_singleplayer_test')
tasks.append('gym_vizdoom_cig_singleplayer')
# tasks = ['gym_vizdoom_cig_0', 'gym_vizdoom_cig_1', 'gym_vizdoom_cig_2', 'gym_vizdoom_cig_singleplayer_test', 'gym_vizdoom_cig_singleplayer']
# tasks = ['gym_vizdoom_cig_multiplayer', 'gym_vizdoom_cig_singleplayer']
tasks = [getattr(tasks_lib, name)(config, params) for name in tasks]
# tasks = [getattr(tasks_lib, name)(config, params) for idx, name in enumerate(tasks)]
config.isolate_envs = params.get('isolate_envs', 'thread')
env_ctor_called = 0
def common_spaces_ctor(task, action_spaces, index=None):
env = task.env_ctor()
env = control.wrappers.SelectObservations(env, ['image'])
env = control.wrappers.PadActions(env, action_spaces)
return env
if len(tasks) > 1:
action_spaces = [task.env_ctor().action_space for task in tasks]
for index, task in enumerate(tasks):
env_ctor = functools.partial(common_spaces_ctor, task,
action_spaces)
# env_ctor = lambda: common_spaces_ctor(task, action_spaces)
tasks[index] = tasks_lib.Task(task.name, env_ctor, task.max_length,
['reward'])
for name in tasks[0].state_components:
config.heads[name] = networks.feed_forward
config.zero_step_losses[name] = 1.0
config.tasks = [tasks[-2]]
config.test_tasks = [tasks[-2]]
config.random_collect_tasks = [tasks[-1]]
assert len(config.tasks) == 1
assert len(config.test_tasks) == 1
assert len(config.random_collect_tasks) == 1
return config