def test_DDPG(self):
config = generate_algorithm_config("DDPG", {})
config["frame_config"]["models"] = [
"DDPGActorDisc",
"DDPGActorDisc",
"DDPGCritic",
"DDPGCritic",
]
config["frame_config"]["model_kwargs"] = [{
"state_dim": 4,
"action_dim": 2
}] * 4
ddpg = init_algorithm_from_config(config)
env = gym.make("CartPole-v0")
dataset = RLGymDiscActDataset(ddpg, env, render_every_episode=1)
self.assert_valid_disc_output(next(dataset))
def test_A2C(self):
config = generate_algorithm_config("A2C", {})
config["frame_config"]["models"] = ["A2CActorDisc", "A2CCritic"]
config["frame_config"]["model_kwargs"] = [
{
"state_dim": 4,
"action_num": 2
},
{
"state_dim": 4
},
]
a2c = init_algorithm_from_config(config)
env = gym.make("CartPole-v0")
dataset = RLGymDiscActDataset(a2c, env, render_every_episode=1)
self.assert_valid_disc_output(next(dataset))
def test_A2C(self):
config = generate_algorithm_config("A2C", {})
config["frame_config"]["models"] = ["A2CActorCont", "A2CCritic"]
config["frame_config"]["model_kwargs"] = [
{
"state_dim": 3,
"action_dim": 1,
"action_range": 2
},
{
"state_dim": 3
},
]
a2c = init_algorithm_from_config(config)
env = gym.make("Pendulum-v0")
dataset = RLGymContActDataset(a2c, env, render_every_episode=1)
self.assert_valid_cont_output(next(dataset))
def test_DQN(self):
config = generate_algorithm_config("DQN", {})
config["frame_config"]["models"] = ["QNet", "QNet"]
config["frame_config"]["model_kwargs"] = [
{
"state_dim": 4,
"action_num": 2
},
{
"state_dim": 4,
"action_num": 2
},
]
dqn = init_algorithm_from_config(config)
env = gym.make("CartPole-v0")
dataset = RLGymDiscActDataset(dqn, env, render_every_episode=1)
self.assert_valid_disc_output(next(dataset))
def test_gym_env_dataset_creator():
# Discrete action environment
config = generate_gym_env_config("CartPole-v0", {})
config = generate_algorithm_config("DDPG", config)
config["frame_config"]["models"] = [
"DDPGActorCont",
"DDPGActorCont",
"DDPGCritic",
"DDPGCritic",
]
config["frame_config"]["model_kwargs"] = [{
"state_dim": 3,
"action_dim": 1,
"action_range": 2
}] * 2 + [{
"state_dim": 3,
"action_dim": 1
}] * 2
ddpg = init_algorithm_from_config(config)
assert isinstance(
gym_env_dataset_creator(ddpg, config["train_env_config"]),
RLGymDiscActDataset)
assert isinstance(gym_env_dataset_creator(ddpg, config["test_env_config"]),
RLGymDiscActDataset)
# Continuous action environment
config = generate_gym_env_config("Pendulum-v0", {})
assert isinstance(
gym_env_dataset_creator(ddpg, config["train_env_config"]),
RLGymContActDataset)
assert isinstance(gym_env_dataset_creator(ddpg, config["test_env_config"]),
RLGymContActDataset)
# Unsupported environment,
# like algorithmic, which uses a tuple action space
# or robotics, which uses the goal action space
config = generate_gym_env_config("Copy-v0", {})
with pytest.raises(ValueError, match="not supported"):
gym_env_dataset_creator(ddpg, config["train_env_config"])
with pytest.raises(ValueError, match="not supported"):
gym_env_dataset_creator(ddpg, config["test_env_config"])
def test_DDPG(self):
config = generate_algorithm_config("DDPG", {})
config["frame_config"]["models"] = [
"DDPGActorCont",
"DDPGActorCont",
"DDPGCritic",
"DDPGCritic",
]
config["frame_config"]["model_kwargs"] = [{
"state_dim": 3,
"action_dim": 1,
"action_range": 2
}] * 2 + [{
"state_dim": 3,
"action_dim": 1
}] * 2
ddpg = init_algorithm_from_config(config)
env = gym.make("Pendulum-v0")
dataset = RLGymContActDataset(ddpg, env, render_every_episode=1)
self.assert_valid_cont_output(next(dataset))
