def loadGame(param_dir):
gs = dict()
with open(param_dir, 'rb') as f:
shimon_hero_params = pickle.load(f)
for paramater in shimon_hero_params:
gs[paramater] = shimon_hero_params[paramater]
game = sh.Game(gs)
return game
img_channels = 4 # Stack 4 frames to infer movement
# Initialize instance of Shimon Hero game
gs = dict()
#gs['NUMBER_OF_ARMS'] = 2
gs['NOTE_HEIGHT'] = 20
gs['NOTE_WIDTH'] = 5
gs['ARM_WIDTH'] = 5
gs['ARM_SPEED'] = 2
gs['TOTAL_NUM_NOTES'] = 12
gs['POINT_THRESHOLD'] = -1.
gs['PENALTY_MISSING_NOTE'] = -0.2
gs['PROB_NOTE_SPAWNED'] = 0.006
game = sh.Game(gs) # Instantiate Shimon Hero game
timestr = time.strftime("%m-%d_%H-%M-%S") # save the current time to name the model
shimon_hero_param = game.get_settings()
if SAVE_MODEL:
name = "model_" + timestr + "_param.p"
dir_path = "./saved_models/" + name
with open(dir_path, 'wb') as f:
pickle.dump(shimon_hero_param, f, pickle.HIGHEST_PROTOCOL)
# Note that in order for saving to pick;e to work, I had to comment out sock.socket object from Zach's Shimon Hero
# In the newest version of Keras (February 2017), you need to pass a kwarg called "dim_ordering" when initializing weights
def my_init(shape, name=None, dim_ordering=None):
return initializations.normal(shape, scale=0.01, name=name)
def buildmodel():
action_dict = {0: -1, 1: 0, 2: 1} # Shimon hero will interpret [-1] as one arm left, [1] as one arm right. Or [-1 -1] as two arm left and left.
# So if QNN gives [1 0 0], then in
img_rows , img_cols = 80, 80 # All images are downsampled to 80 x 80
img_channels = 4 # Stack 4 frames to infer movement
# Initialize instance of Shimon Hero game
gs = dict()
#gs['NUMBER_OF_ARMS'] = 4
gs['NOTE_HEIGHT'] = 20
gs['NOTE_WIDTH'] = 10
gs['TOTAL_NUM_NOTES'] = 12
game = sh.Game(gs)
timestr = time.strftime("%m-%d_%H-%M-%S") # save the current time to name the model
# In the newest version of Keras (February 2017), you need to pass a kwarg called "dim_ordering" when initializing weights
def my_init(shape, name=None, dim_ordering=None):
return initializations.normal(shape, scale=0.01, name=name)
def buildmodel():
# Build the model using the same specifications as the DeepMind paper
# Keras is a good library for protrotyping a quick model and will later
# be modified/used in conjunction with pure tensorflow code
# TODO: Need to figure out how to run on GPU
print("Building CNN Model")
model = Sequential()
NUM_ACTIONS = 4 # number of valid actions
GAMMA = 0.99 # decay rate of past observations
OBSERVATION = 3200. # timesteps to observe before training
EXPLORE = 3000000. # frames over which to anneal epsilon
FINAL_EPSILON = 0.0001 # final value of epsilon
INITIAL_EPSILON = 0.1 # starting value of epsilon
REPLAY_MEMORY = 50000 # number of previous transitions to remember
BATCH = 32 # size of minibatch
FRAME_PER_ACTION = 1 # This controls how many frames to wait before deciding on an action. If F_P_A = 1, then Shimon
# chooses a new action every tick, which causes erratic movements with no exploration middle spaces
img_rows, img_cols = 80, 80 # All images are downsampled to 80 x 80
img_channels = 4 # Stack 4 frames to infer movement
# Initialize instance of Shimon Hero game
game = sh.Game(human_playing=False)
timestr = time.strftime(
"%m-%d_%H-%M-%S") # save the current time to name the model
# In the newest version of Keras (February 2017), you need to pass a kwarg called "dim_ordering" when initializing weights
def my_init(shape, name=None, dim_ordering=None):
return initializations.normal(shape, scale=0.01, name=name)
def buildmodel():
# Build the model using the same specifications as the DeepMind paper
# Keras is a good library for protrotyping a quick model and will later
# be modified/used in conjunction with pure tensorflow code
# TODO: Need to figure out how to run on GPU
NUM_ACTIONS = 4 # number of valid actions
GAMMA = 0.99 # decay rate of past observations
OBSERVATION = 3200. # timesteps to observe before training
EXPLORE = 3000000. # frames over which to anneal epsilon
FINAL_EPSILON = 0.0001 # final value of epsilon
INITIAL_EPSILON = 0.1 # starting value of epsilon
REPLAY_MEMORY = 50000 # number of previous transitions to remember
BATCH = 32 # size of minibatch
FRAME_PER_ACTION = 1 # This controls how many frames to wait before deciding on an action. If F_P_A = 1, then Shimon
# chooses a new action every tick, which causes erratic movements with no exploration middle spaces
img_rows, img_cols = 80, 80 # All images are downsampled to 80 x 80
img_channels = 4 # Stack 4 frames to infer movement
# Initialize instance of Shimon Hero game
game = sh.Game()
# In the newest version of Keras (February 2017), you need to pass a kwarg called "dim_ordering" when initializing weights
def my_init(shape, name=None, dim_ordering=None):
return initializations.normal(shape, scale=0.01, name=name)
def buildmodel():
# Build the model using the same specifications as the DeepMind paper
# Keras is a good library for protrotyping a quick model and will later
# be modified/used in conjunction with pure tensorflow code
# TODO: Need to figure out how to run on GPU
print("Building CNN Model")
model = Sequential()