def initializeALE(romFile, rec_dir):
ale = ALEInterface()
max_frames_per_episode = ale.getInt("max_num_frames_per_episode")
ale.setInt("random_seed", 123)
ale.setFloat("repeat_action_probability", 0.0)
ale.setInt("frame_skip", 5)
# Set record flags
ale.setString(b'record_screen_dir', rec_dir + '/')
ale.setString("record_sound_filename", rec_dir + "/sound.wav")
# We set fragsize to 64 to ensure proper sound sync
ale.setInt("fragsize", 64)
# Set USE_SDL to true to display the screen. ALE must be compilied
# with SDL enabled for this to work. On OSX, pygame init is used to
# proxy-call SDL_main.
USE_SDL = False
if USE_SDL:
if sys.platform == 'darwin':
import pygame
pygame.init()
ale.setBool('sound', False) # Sound doesn't work on OSX
elif sys.platform.startswith('linux'):
ale.setBool('sound', True)
ale.setBool('display_screen', True)
ale.loadROM(romFile)
actionSet = ale.getMinimalActionSet()
return ale, actionSet
def initializeALE(romFile):
ale = ALEInterface()
ale.setInt("max_num_frames_per_episode", 18000)
ale.setInt("random_seed", 123)
ale.setFloat("repeat_action_probability", 0.0)
ale.setInt("frame_skip", 5)
random_seed = ale.getInt("random_seed")
print("random_seed: " + str(random_seed))
# Set USE_SDL to true to display the screen. ALE must be compilied
# with SDL enabled for this to work. On OSX, pygame init is used to
# proxy-call SDL_main.
USE_SDL = False
if USE_SDL:
if sys.platform == 'darwin':
import pygame
pygame.init()
ale.setBool('sound', False) # Sound doesn't work on OSX
elif sys.platform.startswith('linux'):
ale.setBool('sound', True)
ale.setBool('display_screen', True)
ale.loadROM(romFile)
actionSet = ale.getMinimalActionSet()
return ale, actionSet
class pyrlcade_environment(object):
def init(self,rom_file,ale_frame_skip):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt("max_num_frames_per_episode");
self.ale.set("random_seed",123)
self.ale.set("disable_color_averaging",1)
self.ale.set("frame_skip",ale_frame_skip)
self.ale.loadROM(rom_file)
self.legal_actions = self.ale.getMinimalActionSet()
ram_size = self.ale.getRAMSize()
self.ram = np.zeros((ram_size),dtype=np.uint8)
self.ale.getRAM(self.ram)
self.state = self.ale.getRAM(self.ram)
def reset_state(self):
self.ale.reset_game()
def set_action(self,a):
self.action = a
def step(self):
self.reward = self.ale.act(self.action)
is_terminal = self.ale.game_over()
return is_terminal
def get_state(self):
self.ale.getRAM(self.ram)
return self.ram
def get_reward(self):
return self.reward
class emulator:
def __init__(self, rom_name, vis):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt("max_mum_frames_per_episode")
self.ale.setInt("random_seed", 123)
self.ale.setInt("frame_skip", 4)
self.ale.loadROM('roms/' + rom_name)
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
print self.legal_actions
self.screen_width, self.screen_height = self.ale.getScreenDims()
print("width/height: "+ str(self.screen_width) + "/" + str(self.screen_height))
self.vis = vis
if vis:
cv2.startWindowThread()
cv2.namedWindow("preview")
def get_image(self):
# numpy_surface = np.zeros(self.screen_height*self.screen_width*3, dtype=np.uint8)
# self.ale.getScreenRGB(numpy_surface)
# image = np.reshape(numpy_surface, (self.screen_height, self.screen_width, 3))
image = self.ale.getScreenRGB()
image = np.reshape(image, (self.screen_height, self.screen_width, 3))
return image
def newGame(self):
self.ale.reset_game()
return self.get_image(), 0, False
def next(self, action_indx):
reward = self.ale.act(action_indx)
nextstate = self.get_image()
if self.vis:
cv2.imshow('preview', nextstate)
return nextstate, reward, self.ale.game_over()
def train(self):
for episode in range(10):
total_reward = 0
frame_number = 0
while not self.ale.game_over():
a = self.legal_actions[random.randrange(len(self.legal_actions))]
# Apply an action and get the resulting reward
reward = self.ale.act(a);
total_reward += reward
screen = self.ale.getScreenRGB()
screen = np.array(screen).reshape([self.screen_height, self.screen_width, -1])
frame_number = self.ale.getEpisodeFrameNumber()
cv2.imshow("screen", screen/255.0)
cv2.waitKey(0)
self.ale.saveScreenPNG("test_"+str(frame_number)+".png")
print('Episode %d ended with score: %d' % (episode, total_reward))
print('Frame number is : ', frame_number)
self.ale.reset_game()
class emulator:
def __init__(self, rom_name, vis, windowname='preview'):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt(
"max_num_frames_per_episode")
self.ale.setInt("random_seed", 123)
self.ale.setInt("frame_skip", 4)
self.ale.loadROM('roms/' + rom_name)
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
self.windowname = windowname
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
self.init_frame_number = 0
# print(self.legal_actions)
self.screen_width, self.screen_height = self.ale.getScreenDims()
print("width/height: " + str(self.screen_width) + "/" +
str(self.screen_height))
self.vis = vis
if vis:
cv2.startWindowThread()
cv2.namedWindow(self.windowname)
def get_image(self):
numpy_surface = np.zeros(self.screen_height * self.screen_width * 3,
dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface,
(self.screen_height, self.screen_width, 3))
return image
def newGame(self):
# Instead of resetting the game, we load a checkpoint and start from there.
# self.ale.reset_game()
self.ale.restoreState(
self.ale.decodeState(checkpoints[random.randint(
0, 99)].astype('uint8')))
self.init_frame_number = self.ale.getFrameNumber()
#self.ale.restoreState(self.ale.decodeState(np.reshape(checkpoint,(1009,1))))
return self.get_image()
def next(self, action_indx):
reward = self.ale.act(action_indx)
nextstate = self.get_image()
# scipy.misc.imsave('test.png',nextstate)
if self.vis:
cv2.imshow(self.windowname, nextstate)
return nextstate, reward, self.ale.game_over()
def get_frame_number(self):
return self.ale.getFrameNumber() - self.init_frame_number
class emulator(object):
def __init__(self, rom_name, vis, frameskip=1, windowname='preview'):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt(
"max_num_frames_per_episode")
self.ale.setInt("random_seed", 123)
self.ale.setInt("frame_skip", frameskip)
romfile = str(ROM_PATH) + str(rom_name)
if not os.path.exists(romfile):
print('No ROM file found at "' + romfile +
'".\nAdjust ROM_PATH or double-check the filt exists.')
self.ale.loadROM(romfile)
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
self.windowname = windowname
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
# print(self.legal_actions)
self.screen_width, self.screen_height = self.ale.getScreenDims()
print("width/height: " + str(self.screen_width) + "/" +
str(self.screen_height))
self.vis = vis
if vis:
cv2.startWindowThread()
cv2.namedWindow(
self.windowname,
flags=cv2.WINDOW_AUTOSIZE) # permit manual resizing
def get_image(self):
numpy_surface = np.zeros(self.screen_height * self.screen_width * 3,
dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface,
(self.screen_height, self.screen_width, 3))
return image
def newGame(self):
self.ale.reset_game()
return self.get_image()
def next(self, action_indx):
reward = self.ale.act(action_indx)
nextstate = self.get_image()
# scipy.misc.imsave('test.png',nextstate)
if self.vis:
cv2.imshow(self.windowname, nextstate)
if sys.platform == 'darwin':
# if we don't do this, can hang on OS X
cv2.waitKey(2)
return nextstate, reward, self.ale.game_over()
class Atari:
# Constructor
def __init__(self, rom_name):
# 1º Passo: carregamos o jogo e definimos seus parâmetros
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt(
b"max_num_frames_per_episode")
self.ale.setInt(b"random_seed", 123)
self.ale.setInt(b"frame_skip", 4)
self.ale.loadROM(('game/' + rom_name).encode())
self.screen_width, self.screen_height = self.ale.getScreenDims()
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
# 2º Passo: criamos a janela para exibição
self.windowname = rom_name
cv2.startWindowThread()
cv2.namedWindow(rom_name)
# Essa função será utilizada para receber uma imagem do emulador, já em um formato esperado
# por nosso algoritmo de treinamento.
def get_image(self):
numpy_surface = np.zeros(self.screen_height * self.screen_width * 3,
dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface,
(self.screen_height, self.screen_width, 3))
return image
# Simplesmente inicializa o jogo
def newGame(self):
self.ale.reset_game()
return self.get_image()
# Essa função será responsável por retornar as informações da observação do estado após certa ação ser tomada.
def next(self, action):
reward = self.ale.act(self.legal_actions[np.argmax(action)])
nextstate = self.get_image()
cv2.imshow(self.windowname, nextstate)
if self.ale.game_over():
self.newGame()
return nextstate, reward, self.ale.game_over()
class emulator:
def __init__(self, rom_name, vis,windowname='preview'):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt("max_num_frames_per_episode");
self.ale.setInt("random_seed",123)
self.ale.setInt("frame_skip",4)
self.ale.loadROM('roms/' + rom_name )
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
self.windowname = windowname
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
self.init_frame_number = 0
# print(self.legal_actions)
self.screen_width,self.screen_height = self.ale.getScreenDims()
print("width/height: " +str(self.screen_width) + "/" + str(self.screen_height))
self.vis = vis
if vis:
cv2.startWindowThread()
cv2.namedWindow(self.windowname)
def get_image(self):
numpy_surface = np.zeros(self.screen_height*self.screen_width*3, dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface, (self.screen_height, self.screen_width, 3))
return image
def newGame(self):
# Instead of resetting the game, we load a checkpoint and start from there.
# self.ale.reset_game()
self.ale.restoreState(self.ale.decodeState(checkpoints[random.randint(0,99)].astype('uint8')))
self.init_frame_number = self.ale.getFrameNumber()
#self.ale.restoreState(self.ale.decodeState(np.reshape(checkpoint,(1009,1))))
return self.get_image()
def next(self, action_indx):
reward = self.ale.act(action_indx)
nextstate = self.get_image()
# scipy.misc.imsave('test.png',nextstate)
if self.vis:
cv2.imshow(self.windowname,nextstate)
return nextstate, reward, self.ale.game_over()
def get_frame_number(self):
return self.ale.getFrameNumber() - self.init_frame_number
class Emulator:
def __init__(self, rom_name, vis):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt(
"max_num_frames_per_episode")
self.ale.setInt("random_seed", 123)
self.ale.setInt("frame_skip", 4)
self.ale.loadROM('roms/' + rom_name)
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
#print(self.legal_actions)
self.screen_width, self.screen_height = self.ale.getScreenDims()
print("width/height: " + str(self.screen_width) + "/" +
str(self.screen_height))
self.vis = vis
if vis:
cv2.startWindowThread()
cv2.namedWindow("preview")
def get_image(self):
numpy_surface = np.zeros(self.screen_height * self.screen_width * 3,
dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface,
(self.screen_height, self.screen_width, 3))
#added by ben may 2016
print image
print '&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&&& printing'
return image
def newGame(self):
self.ale.reset_game()
return self.get_image()
def next(self, action_indx):
reward = self.ale.act(action_indx)
nextstate = self.get_image()
# scipy.misc.imsave('test.png',nextstate)
# scipy.misc.imsave('test.png',nextstate)
if self.vis:
cv2.imshow('preview', nextstate)
return nextstate, reward, self.ale.game_over()
class Atari:
def __init__(self, rom_name):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt(
"max_num_frames_per_episode")
self.ale.setInt("random_seed", 123)
self.ale.setInt("frame_skip", 4)
self.ale.loadROM('game/' + rom_name)
self.screen_width, self.screen_height = self.ale.getScreenDims()
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
# print len(self.legal_actions)
self.windowname = rom_name
cv2.startWindowThread()
cv2.namedWindow(rom_name)
def get_image(self):
numpy_surface = np.zeros(self.screen_height * self.screen_width * 3,
dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface,
(self.screen_height, self.screen_width, 3))
return image
def newGame(self):
self.ale.reset_game()
return self.get_image()
def next(self, action):
reward = self.ale.act(self.legal_actions[np.argmax(action)])
nextstate = self.get_image()
cv2.imshow(self.windowname, nextstate)
if self.ale.game_over():
self.newGame()
# print "reward %d" % reward
return nextstate, reward, self.ale.game_over()
class Atari:
def __init__(self,rom_name):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt("max_num_frames_per_episode")
self.ale.setInt("random_seed",123)
self.ale.setInt("frame_skip",4)
self.ale.loadROM(rom_name)
self.screen_width,self.screen_height = self.ale.getScreenDims()
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
print len(self.legal_actions)
self.windowname = rom_name
cv2.startWindowThread()
cv2.namedWindow(rom_name)
def preprocess(self, image):
image = cv2.cvtColor(cv2.resize(image, (84, 110)), cv2.COLOR_BGR2GRAY)
image = image[26:110,:]
ret, image = cv2.threshold(image,1,255,cv2.THRESH_BINARY)
return np.reshape(image,(84,84, 1))
def get_image(self):
numpy_surface = np.zeros(self.screen_height*self.screen_width*3, dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface, (self.screen_height, self.screen_width, 3))
return self.preprocess(image)
def newGame(self):
self.ale.reset_game()
return self.get_image()
def next(self, action):
reward = self.ale.act(self.legal_actions[np.argmax(action)])
nextstate = self.get_image()
cv2.imshow(self.windowname,nextstate)
if self.ale.game_over():
self.newGame()
#print "reward %d" % reward
return nextstate, reward, self.ale.game_over()
class emulator:
def __init__(self, rom_name, vis):
if vis:
import cv2
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt("max_num_frames_per_episode");
self.ale.setInt("random_seed",123)
self.ale.setInt("frame_skip",4)
self.ale.loadROM('roms/' + rom_name )
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
# print(self.legal_actions)
self.screen_width,self.screen_height = self.ale.getScreenDims()
print("width/height: " +str(self.screen_width) + "/" + str(self.screen_height))
self.vis = vis
if vis:
cv2.startWindowThread()
cv2.namedWindow("preview")
def get_image(self):
numpy_surface = np.zeros(self.screen_height*self.screen_width*3, dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface, (self.screen_height, self.screen_width, 3))
return image
def newGame(self):
self.ale.reset_game()
return self.get_image()
def next(self, action_indx):
reward = self.ale.act(action_indx)
nextstate = self.get_image()
# scipy.misc.imsave('test.png',nextstate)
if self.vis:
cv2.imshow('preview',nextstate)
return nextstate, reward, self.ale.game_over()
class pyrlcade_environment(object):
def init(self, rom_file, ale_frame_skip):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt(
"max_num_frames_per_episode")
self.ale.set("random_seed", 123)
self.ale.set("disable_color_averaging", 1)
self.ale.set("frame_skip", ale_frame_skip)
self.ale.loadROM(rom_file)
self.legal_actions = self.ale.getMinimalActionSet()
ram_size = self.ale.getRAMSize()
self.ram = np.zeros((ram_size), dtype=np.uint8)
self.ale.getRAM(self.ram)
self.state = self.ale.getRAM(self.ram)
def reset_state(self):
self.ale.reset_game()
def set_action(self, a):
self.action = a
def step(self):
self.reward = self.ale.act(self.action)
is_terminal = self.ale.game_over()
return is_terminal
def get_state(self):
self.ale.getRAM(self.ram)
return self.ram
def get_reward(self):
return self.reward
class Atari:
def __init__(self,rom_name):
self.ale = ALEInterface()
self.max_frames_per_episode = self.ale.getInt("max_num_frames_per_episode")
self.ale.setInt("random_seed",123)
self.ale.setInt("frame_skip",4)
self.ale.loadROM('./' +rom_name)
self.screen_width,self.screen_height = self.ale.getScreenDims()
self.legal_actions = self.ale.getMinimalActionSet()
self.action_map = dict()
for i in range(len(self.legal_actions)):
self.action_map[self.legal_actions[i]] = i
#print len(self.legal_actions)
self.windowname = rom_name
#cv2.startWindowThread()
#cv2.namedWindow(rom_name)
def get_image(self):
numpy_surface = np.zeros(self.screen_height*self.screen_width*3, dtype=np.uint8)
self.ale.getScreenRGB(numpy_surface)
image = np.reshape(numpy_surface, (self.screen_height, self.screen_width, 3))
return image
def newGame(self):
self.ale.reset_game()
return self.get_image()
def next(self, action):
reward = self.ale.act(self.legal_actions[np.argmax(action)])
nextstate = self.get_image()
#cv2.imshow(self.windowname,nextstate)
if self.ale.game_over():
self.newGame()
#print "reward %d" % reward
return nextstate, reward, self.ale.game_over()
# Author: Ben Goodrich
#
# This is a direct port to python of the shared library example from ALE provided in
# doc/examples/sharedLibraryInterfaceExample.cpp
import sys
from ale_python_interface import ALEInterface
import numpy as np
if (len(sys.argv) < 2):
print("Usage ./ale_python_test1.py <ROM_FILE_NAME>")
sys.exit()
ale = ALEInterface()
max_frames_per_episode = ale.getInt("max_num_frames_per_episode")
ale.set("random_seed", 123)
random_seed = ale.getInt("random_seed")
print("random_seed: " + str(random_seed))
ale.loadROM(sys.argv[1])
legal_actions = ale.getLegalActionSet()
for episode in range(10):
total_reward = 0.0
while not ale.game_over():
a = legal_actions[np.random.randint(legal_actions.size)]
reward = ale.act(a)
total_reward += reward
print("Episode " + str(episode) + " ended with score: " +
14, #11001 fire up/right
16, #11010 fire down/right
14, #11011 fire up/down/right (invalid)
11, #11100 fire left/right (invalid)
14, #11101 fire left/right/up (invalid)
16, #11110 fire left/right/down (invalid)
14 #11111 fire up/down/left/right (invalid)
)
if(len(sys.argv) < 3):
print("Usage ./ale_logger.py <ROM_FILE_NAME> <LOG_OUTPUT_FILE>")
sys.exit()
ale = ALEInterface()
max_frames_per_episode = ale.getInt("max_num_frames_per_episode");
ale.setInt("random_seed",123)
random_seed = ale.getInt("random_seed")
print("random_seed: " + str(random_seed))
ale.loadROM(sys.argv[1])
legal_actions = ale.getMinimalActionSet()
print legal_actions
(screen_width,screen_height) = ale.getScreenDims()
print("width/height: " +str(screen_width) + "/" + str(screen_height))
(display_width,display_height) = (1024,420)
#init pygame
