def get_player(viz=False, train=False):
env = AtariPlayer(ROM_FILE, frame_skip=ACTION_REPEAT, viz=viz,
live_lost_as_eoe=train, max_num_frames=30000)
env = FireResetEnv(env)
env = WarpFrame(env, IMAGE_SIZE)
if not train:
# in training, history is taken care of in expreplay buffer
env = FrameStack(env, FRAME_HISTORY)
return env
def get_player(train=False, dumpdir=None):
env = gym.make(ENV_NAME)
if dumpdir:
env = gym.wrappers.Monitor(env, dumpdir)
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, IMAGE_SIZE))
env = FrameStack(env, 4)
if train:
env = LimitLength(env, 60000)
return env
def get_player(train=False, dumpdir=None):
env = gym.make(ENV_NAME)
if dumpdir:
env = gym.wrappers.Monitor(env, dumpdir)
env = FireResetEnv(env)
env = WarpFrame(env, IMAGE_SIZE)
env = FrameStack(env, 4)
if train:
env = LimitLength(env, 60000)
return env
def get_player(rom, viz=False, train=False):
env = AtariPlayer(rom,
frame_skip=ACTION_REPEAT,
viz=viz,
live_lost_as_eoe=train,
max_num_frames=60000)
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, IMAGE_SIZE))
if not train:
# in training, context is taken care of in expreplay buffer
env = FrameStack(env, CONTEXT_LEN)
return env
def get_player(viz=False, train=False):
env = AtariPlayer(ROM_FILE,
frame_skip=ACTION_REPEAT,
viz=viz,
live_lost_as_eoe=train,
max_num_frames=60000)
env = FireResetEnv(env)
env = MapState(env,
lambda im: cv2.resize(im, IMAGE_SIZE)[:, :, np.newaxis])
if not train:
# in training, history is taken care of in expreplay buffer
env = FrameStack(env, FRAME_HISTORY)
return env
def get_player(viz=False, train=False):
if USE_GYM:
env = gym.make(ENV_NAME)
else:
from atari import AtariPlayer
env = AtariPlayer(ENV_NAME, frame_skip=4, viz=viz,
live_lost_as_eoe=train, max_num_frames=60000)
env = FireResetEnv(env)
env = MapState(env, lambda im: resize_keepdims(im, IMAGE_SIZE))
if not train:
# in training, history is taken care of in expreplay buffer
env = FrameStack(env, FRAME_HISTORY)
if train and USE_GYM:
env = LimitLength(env, 60000)
return env
def get_player(rom,
image_size,
viz=False,
train=False,
frame_skip=1,
context_len=1):
env = AtariPlayer(rom,
frame_skip=frame_skip,
viz=viz,
live_lost_as_eoe=train,
max_num_frames=60000)
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, image_size))
if not train:
# in training, context is taken care of in expreplay buffer
env = FrameStack(env, context_len)
return env
def get_player(train=False, dumpdir=None):
use_gym = not ENV_NAME.endswith(".bin")
if use_gym:
env = gym.make(ENV_NAME)
else:
from atari import AtariPlayer
env = AtariPlayer(ENV_NAME,
frame_skip=4,
viz=False,
live_lost_as_eoe=train,
max_num_frames=60000,
grayscale=False)
if dumpdir:
env = gym.wrappers.Monitor(env, dumpdir, video_callable=lambda _: True)
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, IMAGE_SIZE))
env = FrameStack(env, 4)
if train and use_gym:
env = LimitLength(env, 60000)
return env
target_f[0][action] = target
self.model.fit(ob_train, target_f, epochs=1, verbose=0)
self.teaching = False
self.frustration = self.frustration_max
self.has_taught = 0
if self.epsilon > self.min_epsilon:
self.epsilon *= self.epsilon_decay
if __name__ == "__main__":
ENV_NAME = 'MsPacman-v0'
NUM_ACTIONS = get_player().action_space.n
env = gym.make(ENV_NAME)
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, (84,84)))
env = FrameStack(env, 4)
pred = OfflinePredictor(PredictConfig(
model=Model(),
session_init=get_model_loader("models/MsPacman-v0.tfmodel"),
input_names=['state'],
output_names=['policy']))
student = student_dqn(env, teacher=pred)
episodes = 1000000
scores = []
teacher_step_nums = []
step_nums = []
ep_avgs = [0]
os.environ['TENSORPACK_TRAIN_API'] = 'v2' # will become default soon
from tensorpack import *
from tensorpack.utils.concurrency import ensure_proc_terminate, start_proc_mask_signal
from tensorpack.utils.serialize import dumps
from tensorpack.tfutils import symbolic_functions as symbf
from tensorpack.tfutils.gradproc import MapGradient, SummaryGradient
from tensorpack.utils.gpu import get_nr_gpu
import gym
from simulator import SimulatorProcess, SimulatorMaster, TransitionExperience
from common import Evaluator, eval_model_multithread, play_n_episodes, play_n_episodes_recording
from atari_wrapper import MapState, FrameStack, FireResetEnv, LimitLength
env = gym.make('MsPacman-v0')
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, (84, 84)))
env = FrameStack(env, 4)
class Model(ModelDesc):
def _get_inputs(self):
assert NUM_ACTIONS is not None
return [
InputDesc(tf.uint8, (None, ) + IMAGE_SHAPE3, 'state'),
InputDesc(tf.int64, (None, ), 'action'),
InputDesc(tf.float32, (None, ), 'futurereward'),
InputDesc(tf.float32, (None, ), 'action_prob'),
]
def _get_NN_prediction(self, image):