def state_to_tensor(self, state):
key = str(state)
tensor = self.tensor_cache.get(key)
if tensor is None:
tensor = SIMULATOR.state_to_tensor(state).to(Device.get_device())
self.tensor_cache[key] = tensor
return tensor
def run_exper(model, steps, get_features, pre_proc_features):
r_tup, e_tup = [], []
rover_poss = []
total_stats = {'total': 0, 'good': 0}
from environment import SIMULATOR
# initializing our environment
my_sim = SIMULATOR()
# beginning of an episode
state_temp = my_sim.reset()
observation = my_sim.state_to_tensor(state_temp)
state_obs = observation
total_moves = 0
# main loop
prev_input = None
for i in range(steps):
# preprocess the observation, set input as difference between images
cur_input = observation
x = cur_input.astype(
np.float).ravel() if prev_input is not None else np.zeros(70)
x = x[10:80] if prev_input is not None else x
x = np.array([x[i] for i in range(len(x)) if not (i % 10 == 0)])
x = np.array([x[i] for i in range(len(x)) if not ((i - 8) % 9 == 0)])
prev_input = cur_input
x, rover_pos = get_rover_pos(x, r_tup, e_tup, rover_poss)
rover_poss.append(rover_pos)
x = np.array(x)
"""
x = x[x != 0]
if(len(x) == 1):
x = np.zeros(4)
x = x.tolist()
x.append(-7.)
x = np.array(x)
"""
x_t = pre_proc_features.fit_transform(x.reshape(-1, 1))
x_t = x_t.reshape(1, INPUT_SIZE)[0]
# forward the policy network and sample action according to the proba distribution
proba = model.predict(np.expand_dims(x_t, axis=1).T)
action = proba.argmax()
#run one step
state_temp, reward, done, r_tup, e_tup = my_sim.step(action)
observation = my_sim.state_to_tensor(state_temp)
#my_sim.render()
total_moves += 1
if (total_moves == MAX_STEPS):
done = True
total_moves = 0
# if episode is over, reset to beginning
if done:
total_stats['total'] += 1
so = np.asarray(state_obs).ravel().tolist()
o = np.asarray(observation).ravel().tolist()
#print("state obs ===============")
#print(state_obs)
#print("obs ===============")
#print(observation)
try:
index_obs = so.index(7.0)
except ValueError:
index_obs = -1
try:
index_curr = o.index(7.0)
except ValueError:
index_curr = -1
if (index_obs != -1 and index_curr == -1):
#print("Good Game")
#print(so)
#print(o)
total_stats['good'] += 1
state_temp = my_sim.reset()
observation = my_sim.state_to_tensor(state_temp)
state_obs = observation
rover_poss = []
#my_sim.render()
return total_stats
def run_exper(model, steps, get_features, pre_proc_features):
from environment import SIMULATOR
# initializing our environment
my_sim = SIMULATOR()
# beginning of an episode
state_temp = my_sim.reset()
observation = my_sim.state_to_tensor(state_temp)
r_tup, e_tup, rover_poss = [], [], []
# main loop
prev_input = None
total_moves = 0
MAX_MOVES = 25
for i in range(steps):
total_moves += 1
start = time.perf_counter()
cur_input = observation
x = cur_input.astype(
np.float).ravel() if prev_input is not None else np.zeros(70)
x = x[10:80] if prev_input is not None else x
x = np.array([x[i] for i in range(len(x)) if not (i % 10 == 0)])
x = np.array([x[i] for i in range(len(x)) if not ((i - 8) % 9 == 0)])
x, rov_pos = get_rover_pos(x, r_tup, e_tup, rover_poss)
x = np.array(x)
rover_poss.append(rov_pos)
"""
x = x[x != 0]
if(len(x) == 1):
x = np.zeros(4)
x = x.tolist()
x.append(-7.)
x = np.array(x)
"""
#print_map(x)
x_t = pre_proc_features.fit_transform(x.reshape(-1, 1))
x_t = x_t.reshape(1, INPUT_SIZE)[0]
print("Shape = ", x_t.shape)
prev_input = cur_input
# forward the policy network and sample action according to the proba distribution
#print_map(x)
proba = model.predict(np.expand_dims(x_t, axis=1).T)
end = time.perf_counter()
action = proba[0].argmax()
print("Time taken = ", end - start)
#run one step
state_temp, reward, done, r_tup, e_tup = my_sim.step(action)
observation = my_sim.state_to_tensor(state_temp)
my_sim.render()
time.sleep(1)
if total_moves == MAX_MOVES:
total_moves = 0
done = True
# if episode is over, reset to beginning
if done:
state_temp = my_sim.reset()
observation = my_sim.state_to_tensor(state_temp)
my_sim.render()
rover_poss = []
default='sparse',
type=str,
help='Choose between encoded or sparse')
args = parser.parse_args()
data_type = args.data_type
model = get_model(data_type)
import numpy as np
import gym
# gym initialization
from environment import SIMULATOR
my_sim = SIMULATOR()
state_temp = my_sim.reset()
observation = my_sim.state_to_tensor(state_temp)
prev_input = None
# Hyperparameters to calculate discount rewards
gamma = 0.99
# initialization of variables used in the main loop
x_train, y_train, y_pred, rewards, r_tup, e_tup, rover_poss = [], [], [], [], [], [], []
reward_sum = 0
episode_nb = 0
resume = True
running_reward = None
EPOCHS_BEFORE_SAVING = 50
moves_count = 0
MAX_NEG_REWARD = -100
get_features, pre_proc_features = get_pre_proc_info(data_type)