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]
ep_avg = []
for ee in range(episodes):
ob = env.reset()
score_ = 0
steps = 0
t_steps = 0
while True:
steps += 1
action = student.act(ob)
next_ob, reward, done, _ = env.step(action)
score_ += reward
student.build_memory(ob, action, reward, next_ob, done)
ob = next_ob
if done:
env = gym.make('MsPacman-v0')
env = FireResetEnv(env)
env = MapState(env, lambda im: cv2.resize(im, (84, 84)))
env = FrameStack(env, 4)
student = student_dqn(env)
episodes = 10000000
save_every = 1000
record_scores = 50
scores = []
mini_score = []
for ee in range(episodes):
ob_big = env.reset()
ob = ob_big[None, :, :, 6:] / 255.
score_ = 0
while True:
action = student.act(ob)
next_ob_big, reward, done, _ = env.step(action)
next_ob = next_ob_big[None, :, :, 6:] / 255.
score_ += reward
student.build_memory(ob, action, reward, next_ob, done)
ob = next_ob
if done:
mini_score.append(score_)
if (ee + 1) % record_scores == 0 or ee == 0: