def play(model):
global lastState
global lastaction
car_distance = 0
game_state = carmunk.GameState()
# Do nothing to get initial.
_, state = game_state.frame_step((2))
train_state = np.append(lastState, state[0])
train_state = np.append(train_state, lastaction)
train_state = np.expand_dims(train_state, axis=0)
# Move.
while True:
car_distance += 1
# Choose action.
action = (np.argmax(model.predict(train_state, batch_size=1)))
print(action)
# Take action.
_, state = game_state.frame_step(action)
train_state = np.append(lastState, state[0])
train_state = np.append(train_state, action)
train_state = np.expand_dims(train_state, axis=0)
lastState = state[0]
# Tell us something.
if car_distance % 1000 == 0:
print("Current distance: %d frames." % car_distance)
def play(model):
car_distance = 0
game_state = carmunk.GameState()
# Do nothing to get initial.
_, state = game_state.frame_step((2))
# Move.
while True:
car_distance += 1
# Choose action.
#action = (np.argmax(model.predict(state, batch_size=1)))
# Take action.
if random.random() < 0.4:
action = np.random.randint(0, 3) # random
else:
# Get Q values for each action.
qval = model.predict(state, batch_size=1)
action = (np.argmax(qval)) # best
_, state = game_state.frame_step(action)
# Tell us something.
if car_distance % 1000 == 0:
print("Current distance: %d frames." % car_distance)
def play(model):
"""
DOCSTRING
"""
car_distance = 0
game_state = carmunk.GameState()
_, state = game_state.frame_step((2))
while True:
car_distance += 1
action = (numpy.argmax(model.predict(state, batch_size=1)))
_, state = game_state.frame_step(action)
if car_distance % 1000 == 0:
print("Current distance: %d frames." % car_distance)
def play(model):
car_distance = 0
game_state = carmunk.GameState()
# Do nothing to get initial.
_, state = game_state.frame_step((2))
# Move.
while True:
car_distance += 1
# Choose action.
action = (np.argmax(model.predict(state, batch_size=1)))
# Take action.
_, state = game_state.frame_step(action)
# Tell us something.
if car_distance % 1000 == 0:
print("Current distance: %d frames." % car_distance)
def train_net(model, params):
"""
DOCSTRING
"""
filename = params_to_filename(params)
observe = 1000 # Number of frames to observe before training.
epsilon = 1
train_frames = 1000000 # Number of frames to play.
batchSize = params['batchSize']
buffer = params['buffer']
max_car_distance = 0
car_distance = 0
t = 0
data_collect = []
replay = [] # stores tuples of (S, A, R, S').
loss_log = []
game_state = carmunk.GameState()
_, state = game_state.frame_step((2))
start_time = timeit.default_timer()
while t < train_frames:
t += 1
car_distance += 1
if random.random() < epsilon or t < observe:
action = np.random.randint(0, 3)
else:
qval = model.predict(state, batch_size=1)
action = (np.argmax(qval)) # best
reward, new_state = game_state.frame_step(action)
replay.append((state, action, reward, new_state))
if t > observe:
if len(replay) > buffer:
replay.pop(0)
minibatch = random.sample(replay, batchSize)
X_train, y_train = process_minibatch(minibatch, model)
history = nn.LossHistory()
model.fit(X_train,
y_train,
batch_size=batchSize,
nb_epoch=1,
verbose=0,
callbacks=[history])
loss_log.append(history.losses)
state = new_state
if epsilon > 0.1 and t > observe:
epsilon -= (1 / train_frames)
if reward == -500:
data_collect.append([t, car_distance])
if car_distance > max_car_distance:
max_car_distance = car_distance
tot_time = timeit.default_timer() - start_time
fps = car_distance / tot_time
print("Max: %d at %d\tepsilon %f\t(%d)\t%f fps" %
(max_car_distance, t, epsilon, car_distance, fps))
car_distance = 0
start_time = timeit.default_timer()
if t % 25000 == 0:
model.save_weights('saved-models/' + filename + '-' + str(t) +
'.h5',
overwrite=True)
print("Saving model %s - %d" % (filename, t))
log_results(filename, data_collect, loss_log)
def train_net(model, params):
filename = params_to_filename(params)
observe = 1000 # Number of frames to observe before training.
epsilon = 1
train_frames = 100000 # Number of frames to play.
batchSize = params['batchSize']
buffer = params['buffer']
# Just stuff used below.
max_car_distance = 0
car_distance = 0
t = 0
data_collect = []
replay = [] # stores tuples of (S, A, R, S').
loss_log = []
# Create a new game instance.
game_state = carmunk.GameState()
# Get initial state by doing nothing and getting the state.
_, state = game_state.frame_step((2))
# Let's time it.
start_time = timeit.default_timer()
# Run the frames.
while t < train_frames:
t += 1
car_distance += 1
# Choose an action.
if random.random() < epsilon or t < observe:
action = np.random.randint(0, 3) # random
else:
# Get Q values for each action.
qval = model.predict(state, batch_size=1)
action = (np.argmax(qval)) # best
# Take action, observe new state and get our treat.
reward, new_state = game_state.frame_step(action)
# Experience replay storage.
replay.append((state, action, reward, new_state))
# If we're done observing, start training.
if t > observe:
# If we've stored enough in our buffer, pop the oldest.
if len(replay) > buffer:
replay.pop(0)
# Randomly sample our experience replay memory
minibatch = random.sample(replay, batchSize)
# Get training values.
X_train, y_train = process_minibatch2(minibatch, model)
# Train the model on this batch.
history = LossHistory()
model.fit(X_train,
y_train,
batch_size=batchSize,
nb_epoch=1,
verbose=0,
callbacks=[history])
loss_log.append(history.losses)
# Update the starting state with S'.
state = new_state
# Decrement epsilon over time.
if epsilon > 0.1 and t > observe:
epsilon -= (1.0 / train_frames)
# We died, so update stuff.
if reward == -500:
# Log the car's distance at this T.
data_collect.append([t, car_distance])
# Update max.
if car_distance > max_car_distance:
max_car_distance = car_distance
# Time it.
tot_time = timeit.default_timer() - start_time
fps = car_distance / tot_time
# Output some stuff so we can watch.
print("Max: %d at %d\tepsilon %f\t(%d)\t%f fps" %
(max_car_distance, t, epsilon, car_distance, fps))
# Reset.
car_distance = 0
start_time = timeit.default_timer()
# Save the model every 25,000 frames.
if t % 25000 == 0:
model.save_weights('saved-models/' + filename + '-' + str(t) +
'.h5',
overwrite=True)
print("Saving model %s - %d" % (filename, t))
# Log results after we're done all frames.
log_results(filename, data_collect, loss_log)