def make_vec_envs(env_name,
seed,
num_processes,
gamma,
log_dir,
add_timestep,
device,
allow_early_resets,
num_frame_stack=None,
args=None):
if 'golmultienv' in env_name.lower():
num_processes = 1 # smuggle in real num_proc in args so we can run them as one NN
envs = [
make_env(env_name,
seed,
i,
log_dir,
add_timestep,
allow_early_resets,
map_width=args.map_width,
render_gui=args.render,
print_map=args.print_map,
noreward=args.no_reward,
max_step=args.max_step,
args=args) for i in range(num_processes)
]
if 'golmultienv' in env_name.lower():
return envs[0]()
if len(envs) > 1:
print(envs)
envs = SubprocVecEnv(envs)
else:
if sys.version[0] == '2':
envs = DummyVecEnv('DummyVecEnv', (), {1: envs})
else:
envs = DummyVecEnv(envs)
if len(envs.observation_space.shape) == 1:
if gamma is None:
envs = VecNormalize(envs, ret=False)
else:
envs = VecNormalize(envs, gamma=gamma)
envs = VecPyTorch(envs, device)
if num_frame_stack is not None:
print('stacking {} frames'.format(num_frame_stack))
envs = VecPyTorchFrameStack(envs, num_frame_stack, device)
elif len(envs.observation_space.shape) == 3:
envs = VecPyTorchFrameStack(envs, 1, device)
return envs
def view_policy_ddpg():
env = DummyVecEnv([lambda: EnvHandler(make_env())])
view(env,
seed=None,
total_timesteps=10000,
reward_scale=1.0,
render=True,
render_eval=False,
noise_type=None,
normalize_returns=False,
normalize_observations=False,
critic_l2_reg=1e-2,
actor_lr=1e-4,
critic_lr=1e-3,
popart=False,
gamma=0.99,
clip_norm=None,
nb_train_steps=50,
nb_eval_steps=100,
nb_save_epochs=None,
batch_size=64,
tau=0.01,
action_range=(-250.0, 250.0),
observation_range=(-5.0, 5.0),
eval_env=None,
load_path="./checkpoints/00007",
save_dir=None,
param_noise_adaption_interval=50)
def view_policy_ppo():
env = DummyVecEnv([lambda: EnvHandler(make_env())])
view(env=env,
episodes=100,
total_timesteps=1000000,
nsteps=200,
nminibatches=1,
cliprange=0.2,
ent_coef=0.0,
lam=0.95,
gamma=0.99,
noptepochs=4,
save_interval=100,
save_dir=".",
load_path="./checkpoints/00500",
normalize_observations=False,
normalize_returns=False)
def make_vec_envs(env_name, seed, num_processes, gamma, log_dir, add_timestep,
device, allow_early_resets):
envs = [
make_env(env_name, seed, i, log_dir, add_timestep, allow_early_resets)
for i in range(num_processes)
]
if len(envs) > 1:
envs = SubprocVecEnv(envs)
else:
envs = DummyVecEnv(envs)
envs = VecPyTorch(envs, device)
'''
if len(envs.observation_space.shape) == 3:
print('Creating frame stacking wrapper')
envs = VecPyTorchFrameStack(envs, 4, device)
#print(envs.observation_space) '''
return envs
else:
action = 1
obs, reward, done, _ = env.step([action])
episode_reward += reward
if render:
env.render()
if done:
break
eval_reward.append(episode_reward)
return np.mean(eval_reward)
if __name__ == '__main__':
# Create and wrap the environment
env = gym.make('game-stock-exchange-continuous-v0')
env = DummyVecEnv([lambda: env])
action_dim = 2
obs_shape = env.observation_space.shape
rpm = ReplayMemory(MEMORY_SIZE)
model = Model(act_dim = action_dim)
algorithm = DQN(model, act_dim = action_dim, gamma = GAMMA, lr = LEARNING_RATE)
agent = Agent(algorithm, obs_shape[0],obs_shape[1],action_dim)
while len(rpm) < MEMORY_WARMUP_SIZE:
run_episode(env,agent,rpm)
max_episode = 2000
episode = 0
while episode < max_episode: