def randomSteps(env,steps,dqn):
t0 = time.time()
env.reset()
i = 0
frame_stack = Pipe(4)
initial_no_op = np.random.randint(4, NO_OP_MAX)
for _ in range(0,steps):
if i < initial_no_op:
# WE PERFORM A RANDOM NUMBER OF NO_OP ACTIONS
action = NO_OP_CODE
state, reward, done, info = env.step(action)
greyObservation = rgb2gray(state)
state = downSample(greyObservation)
frame_stack.push(state)
i += 1
else:
state = np.stack(frame_stack.items, axis=2).reshape((IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
action = np.random.randint(0, len(ACTIONS))
actionH =np.zeros(len(ACTIONS))
actionH[action] = 1
next_state, reward, game_over, info = env.step(action)
greyObservation = rgb2gray(next_state)
next_state = downSample(greyObservation)
frame_stack.push(next_state)
next_state = np.stack(frame_stack.items, axis=2).reshape((IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
dqn.storeExperience(
state.astype(type),
actionH,
reward,
next_state.astype(type),
game_over)
if done:
#print("Episode finished after {} timesteps".format(_ + 1))
env.reset()
i=0
frame_stack=[]
t1 = time.time()
print("Fullfilling replay memory operation took:",t1-t0,)
print('Size of replay memory %s bytes and has %s elements' % ((sys.getsizeof(dqn.replayMemory)),len(dqn.replayMemory)))
print
#Saving and loading networks
saver = tf.train.Saver()
checkpoint = tf.train.get_checkpoint_state("saved_networks")
if checkpoint and checkpoint.model_checkpoint_path and LOAD_NETWORK:
saver.restore(sess, checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
game = int(re.match('.*?([0-9]+)$', checkpoint.model_checkpoint_path).group(1))
else:
print("Could not find old network weights")
sess.run(tf.global_variables_initializer())
game = 0
game_scores =[]
initial_no_op = np.random.randint(4, NO_OP_MAX)
i=0
frame_stack=Pipe(4)
score=0
summary_writer = tf.summary.FileWriter('logs',sess.graph)
summary_writer = tf.summary.FileWriter(LOG_DIRECTORY + RUN_STRING,
sess.graph)
print('Started training')
for step in range(STEPS):
if i < initial_no_op:
# WE PERFORM A RANDOM NUMBER OF NO_OP ACTIONS
action = NO_OP_CODE
state, reward, done, info = env.step(action)
greyObservation = rgb2gray(state)
state = downSample(greyObservation)
frame_stack.push(state)
i+=1
else:
