def launch(self, env_config, use_eval=False):
environment_path = (env_config["environment_path_eval"]
if use_eval else env_config["environment_path"])
port = env_config.get("port", 0)
if use_eval and port:
port += 2
use_visual = env_config.get("use_visual", False)
use_vector = env_config.get("use_vector", True)
multiagent = env_config.get("multiagent", False)
uint8_visual = env_config.get("uint8_visual", True)
flatten_branched = env_config.get("flatten_branched", True)
self.env = UnityEnv(
environment_path,
port,
use_visual=use_visual,
use_vector=use_vector,
uint8_visual=uint8_visual,
multiagent=multiagent,
flatten_branched=flatten_branched,
)
self.action_space = self.env._action_space
self.observation_space = self.env._observation_space
# agent name must be unique among **all** agents
self.agent_name = [
f'{port}_{i}' for i in range(self.env.number_agents)
]
def main():
env = UnityEnv(
"/homes/gkumar/Documents/UnityProjects/maze/Build/mazeBasic_Discrete_imageOnly",
0,
use_visual=True,
uint8_visual=True)
HOSTNAME = os.uname()[1]
logger.configure('./logs/' +
HOSTNAME) # Çhange to log in a different directory
act = deepq.learn(
env,
"cnn", # conv_only is also a good choice for GridWorld
lr=2.5e-4,
total_timesteps=1000000,
buffer_size=50000,
exploration_fraction=0.05,
exploration_final_eps=0.1,
print_freq=20,
train_freq=5,
learning_starts=20000,
target_network_update_freq=50,
gamma=0.99,
prioritized_replay=False,
checkpoint_freq=1000,
checkpoint_path=
'./logs', # Change to save model in a different directory
dueling=True)
print("Saving model to unity_model.pkl")
act.save("unity_model.pkl")
def get_gym_env(self, unity_file):
"""
@param unity_file the full path of Unity environment
Returns an Open AI Gym that wraps given Unity environment
based on selected trainer parameters.
"""
return UnityEnv(unity_file,
self.params.get('worker_id', 0),
use_visual=self.params.get('use_visual', False),
uint8_visual=self.params.get('use_uint8_visual', False),
multiagent=self.params.get('multiagent', False),
flatten_branched=self.params.get('flatten_branched', False),
allow_multiple_visual_obs=self.params.get('allow_multiple_visual_obs', False),
no_graphics=self.params.get('no_graphics', True))
def wrap_unity_env(env_path, frame_skip=0, frame_stack=False, chw_style=False, **unity_config):
worker_id = unity_config.get('port', 9527)
use_visual = unity_config.get('use_visual', True)
uint8_visual = unity_config.get('uint8_visual', True)
flatten_branched = unity_config.get('flatten_branched', True)
multiagent = unity_config.get('multiagent', False)
env = UnityEnv(env_path,
worker_id=worker_id,
use_visual=use_visual,
uint8_visual=uint8_visual,
flatten_branched=flatten_branched,
multiagent=multiagent)
# Be careful with `Decision Interval` in agent script of Unity ml-agent
if frame_skip > 0:
env = MaxAndSkipEnv(env, frame_skip)
if chw_style:
env = CHWStyle(env)
if frame_stack:
env = FrameStack(env, frame_stack, chw_style)
return env
def main():
mask_unused_gpus()
env = UnityEnv("../unity_envs/kais_banana", 0, use_visual=True, uint8_visual=True, flatten_branched=True)
logger.configure('./logs') # Çhange to log in a different directory
act = deepq.learn(
env,
"cnn", # conv_only is also a good choice for GridWorld
lr=2.5e-4,
total_timesteps=100000, #0
buffer_size=50000,
exploration_fraction=0.05,
exploration_final_eps=0.1,
print_freq=20,
train_freq=5,
learning_starts=20000,
target_network_update_freq=50,
gamma=0.99,
prioritized_replay=False,
checkpoint_freq=1000,
checkpoint_path='./logs', # Change to save model in a different directory
dueling=True
)
print("Saving model to unity_model.pkl")
act.save("unity_model.pkl")
start = time.time()
mask_unused_gpus()
# Avoid Tensorflow eats up GPU memory
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
K.set_session(sess)
#Setting up the env
#TODO Worker_id can be changed to run in parallell
#Flatten_branched gives us a onehot encoding of all 54 action combinations.
print("Opening unity env")
env = UnityEnv(
"../unity_envs/kais_banana2",
worker_id=39,
use_visual=True,
flatten_branched=True
) #KOE: Note: If I accept images as uint8_visual=True, I have to convert to float later.
print("Resetting env")
initial_observation = env.reset()
#KOETODO This would have to be manually configured for each environment.
#KOE: What is this misc??
#misc = game_state.game_variables # [Health]
#prev_misc = misc
#KOE: I think this should be the same as my battery measure.
misc = 100 # [Health]
prev_misc = misc
# game.get_available_buttons_size() # [Turn Left, Turn Right, Move Forward]
start = time.time()
mask_unused_gpus()
# Avoid Tensorflow eats up GPU memory
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
K.set_session(sess)
#Setting up the env
#TODO Worker_id can be changed to run in parallell
#Flatten_branched gives us a onehot encoding of all 54 action combinations.
print("Opening unity env")
env = UnityEnv("../unity_envs/kais_banana3",
worker_id=24,
use_visual=True,
uint8_visual=True,
flatten_branched=True)
print("Resetting env")
initial_observation = env.reset()
#KOETODO This would have to be manually configured for each environment.
#KOE: What is this misc??
#misc = game_state.game_variables # [Health]
#prev_misc = misc
#KOE: I think this should be the same as my battery measure.
misc = 100 # [Health]
prev_misc = misc
# game.get_available_buttons_size() # [Turn Left, Turn Right, Move Forward]
class UnityEnvWrapper:
def __init__(self, env_config=None, use_eval=False, rpc_mode=False):
self.env = None
if not rpc_mode:
assert (env_config is not None)
self.launch(env_config, use_eval)
def launch(self, env_config, use_eval=False):
environment_path = (env_config["environment_path_eval"]
if use_eval else env_config["environment_path"])
port = env_config.get("port", 0)
if use_eval and port:
port += 2
use_visual = env_config.get("use_visual", False)
use_vector = env_config.get("use_vector", True)
multiagent = env_config.get("multiagent", False)
uint8_visual = env_config.get("uint8_visual", True)
flatten_branched = env_config.get("flatten_branched", True)
self.env = UnityEnv(
environment_path,
port,
use_visual=use_visual,
use_vector=use_vector,
uint8_visual=uint8_visual,
multiagent=multiagent,
flatten_branched=flatten_branched,
)
self.action_space = self.env._action_space
self.observation_space = self.env._observation_space
# agent name must be unique among **all** agents
self.agent_name = [
f'{port}_{i}' for i in range(self.env.number_agents)
]
def _transform_list_to_dict(self, objs):
return {name: obj for name, obj in zip(self.agent_name, objs)}
def _transform_dict_to_list(self, objs):
return [objs[name] for name in self.agent_name]
def step(self, act, action_settings=None):
action = np.stack(self._transform_dict_to_list(act)).tolist()
observation, reward, done, info = self.env.step(action)
transform = self._transform_list_to_dict
info = list(map(json.loads, info['text_observation']))
for i, x in enumerate(info):
x['done'] = done[i]
done = [False] * 4
done_dict = transform(done)
done_dict['__all__'] = False # no early termination (for logging)
return transform(observation), transform(reward), done_dict, transform(
info)
def reset(self, reset_settings=None):
obs = self.env.reset()
return self._transform_list_to_dict(obs)
def get_env_spaces(self):
spaces = self.action_space, self.observation_space, self.agent_name
p = pickle.dumps(spaces)
z = zlib.compress(p)
return z
def get_action_count(self):
if isinstance(self.env.action_space, gym.spaces.Discrete):
return self.env.action_space.n
elif isinstance(self.env.action_space, gym.spaces.MultiDiscrete):
return self.env.action_space.nvec.tolist()
raise NotImplementedError
def sample(self):
return self.env.action_space.sample()
def number_agents(self):
return self.env.number_agents
def env_close(self):
if self.env:
self.env.close()
self.env = None
def close(self):
self.env_close()
def hello(self):
print('Hello World')
import cv2
np.set_printoptions(threshold=np.inf)
# create grid and write to file
maze_trials, max_episode, max_steps = 1, 100000, 900
action_repeat = 30
buffer = ReplayBuffer(10000)
learner = Learner(buffer)
# assume that unity reads that file and generates maze dynamically
env = UnityEnv(
"/home/gaurav/MySharedRepository/mazeContinuousTarget_fixed_camera/Build/mazeContinuousTarget_fixed_camera",
0,
use_visual=True,
uint8_visual=True)
def drawTrajectory():
action_repeat = 150
# size of experience
# 11 velocities on x axis
# 11 velocities on y axis
# 11 actions on x axis
# 11 actions on y axis
# 150 action repeats
# 2 position values
def sampleTrajectory():
action_repeat = 300
action_range_around_zero = 20 # should be even
#action_range_around_zero = [-9, -8, -7, -6, -5, 0, 5, 6, 7, 8, 9]
period = 1
if action_range_around_zero % 2 != 0:
return False
env = UnityEnv(
"/homes/gkumar/Documents/UnityProjects/mazeContinuousTarget_fixed_camera_data_collection/Build/mazeContinuousTarget_fixed_camera_data_collection",
0,
use_visual=True,
uint8_visual=True)
list_of_data = []
for i in range(int(-1 * action_range_around_zero / 2),
int(action_range_around_zero / 2 + 1),
period): # velocity X [-5, -3, -1, 1, 3, 5]
for j in range(int(-1 * action_range_around_zero / 2),
int(action_range_around_zero / 2 + 1),
period): # velocity Y [-5, -3, -1, 1, 3, 5]
print(i, j)
for k in range(int(-1 * action_range_around_zero / 2),
int(action_range_around_zero / 2 + 1),
period): # action X [-5, -3, -1, 1, 3, 5]
for l in range(int(-1 * action_range_around_zero / 2),
int(action_range_around_zero / 2 + 1),
period): # action Y [-5, -3, -1, 1, 3, 5]
single_tuple = np.zeros(4 + 2 * action_repeat)
obs_fovea = env.reset()
obs_fovea_next, reward, done, info = env.step([[i], [j],
[k], [l]])
# action
single_tuple[0] = i
single_tuple[1] = j
# velocity
single_tuple[2] = k
single_tuple[3] = l
for m in range(0, action_repeat):
single_tuple[
3 + m * 2 +
1] = info["brain_info"].vector_observations[0][2]
single_tuple[
3 + m * 2 +
2] = info["brain_info"].vector_observations[0][3]
x_vel_new = info["brain_info"].vector_observations[0][
6]
y_vel_new = info["brain_info"].vector_observations[0][
7]
if math.sqrt(
math.pow((single_tuple[3 + m * 2 + 1] -
single_tuple[4]), 2) +
math.pow((single_tuple[3 + m * 2 + 2] -
single_tuple[5]), 2)) < 6:
obs_fovea_next, reward, done, info = env.step(
[[i], [j], [x_vel_new], [y_vel_new]])
else:
for n in range(m, action_repeat):
single_tuple[3 + n * 2 +
1] = single_tuple[3 +
(m - 1) * 2 + 1]
single_tuple[3 + n * 2 +
2] = single_tuple[3 +
(m - 1) * 2 + 2]
break
list_of_data.append(single_tuple)
h5f = h5py.File('data.h5', 'w')
h5f.create_dataset('dataset_1', data=list_of_data)
h5f.close()
import gym
from baselines import deepq
from baselines import logger
import time
from gym_unity.envs.unity_env import UnityEnv
import subprocess as sp
import os
env = UnityEnv("../unity_envs/kais_banana",
0,
use_visual=True,
uint8_visual=True,
flatten_branched=True)
act = deepq.learn(env,
network='cnn',
total_timesteps=0,
load_path="logs_backup/model") #"unity_model.pkl")
#Visualizing
#TODO Maybe slow down the simulation by inserting some delays here.
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
