def test():
""" test distillation and evaluation """
LEARNING_RATE = 0.0001
GAME = 'BreakoutNoFrameskip-v4'
BATCH_SIZE = 32
EPSILON = 0.05
ADD_MEM_NUM = 3000
UPDATE_NUM = 200
EPOCH = 1
MEM_SIZE = 50000
MODEL_PATH = './model/teacher/breakout-1.h5f'
LOSS_FUC = 'mse'
EVAL_ITERATION = 3000
logger = LogWriter(ROOT_PATH, BATCH_SIZE)
logger.save_setting(args)
env = make_atari(GAME)
env = wrap_deepmind(env, frame_stack=True, scale=True)
teacher = Teacher(MODEL_PATH, env, EPSILON, MEM_SIZE, EVAL_ITERATION)
student = SingleDtStudent(env, LEARNING_RATE, logger, BATCH_SIZE, EPSILON,
teacher, ADD_MEM_NUM, UPDATE_NUM, EPOCH,
LOSS_FUC, TARGET_NET_SIZE)
student.distill()
logger.save_weights(student, 'student_{}'.format(LOSS_FUC))
logger.log_total_time_cost()
# log
root = 'result_EVAL'
if not os.path.exists(root):
os.mkdir(root)
print('*** Create folder: {} ***'.format(root))
now_time = time.strftime('%y%m%d_%H%M%S', time.localtime())
save_path = os.path.join(root, now_time).replace('\\', '/')
if not os.path.exists(save_path):
os.mkdir(save_path)
print('*** Create folder: {} ***'.format(save_path))
# evaluate teacher
teacher.evaluate(save_path)
# evaluate student
for log_path in glob.glob('./result_DT/*'):
Evaluator_deprecate(env,
log_path,
save_path,
eval_iteration=EVAL_ITERATION).evaluate()
def SingleDistillation_main():
logger = LogWriter(ROOT_PATH, BATCH_SIZE)
logger.save_setting(args)
env = make_atari(GAME)
env = wrap_deepmind(env, frame_stack=True, scale=True)
teacher = Teacher(MODEL_PATH, env, EPSILON, MEM_SIZE)
student = SingleDtStudent(env, LEARNING_RATE, logger, BATCH_SIZE, EPSILON,
teacher, ADD_MEM_NUM, UPDATE_NUM, EPOCH,
LOSS_FUC, TARGET_NET_SIZE)
student.distill()
logger.save_weights(student)
logger.save_model_arch(student)
logger.log_total_time_cost()
def Evaluation_deprecate():
"""
evaluation the performance of both teacher and students under Single-target-distillation situation
there should be only one model file and csv file under the directory of './model/teacher'
multiple log directory under path of './result_DT'
"""
# get the game_name from setting.csv
with open(glob.glob('./model/teacher/*.csv')[0]) as f:
reader = csv.reader(f)
settings_dict = {row[0]: row[1] for row in reader}
game_name = settings_dict['game']
print("*** GAME of teacher:{} ***".format(game_name))
# make environment
env = make_atari(GAME)
env = wrap_deepmind(env, frame_stack=True, scale=True)
# log
root = 'result_EVAL'
if not os.path.exists(root):
os.mkdir(root)
print('*** Create folder: {} ***'.format(root))
now_time = time.strftime('%y%m%d_%H%M%S', time.localtime())
save_path = os.path.join(root, now_time).replace('\\', '/')
if not os.path.exists(save_path):
os.mkdir(save_path)
print('*** Create folder: {} ***'.format(save_path))
# evaluate teacher
Teacher(glob.glob('./model/teacher/*.h5f')[0].replace('\\', '/'),
env,
eval_iteration=EVAL_ITERATION,
is_small=True).evaluate(save_path)
# evaluate student
for log_path in glob.glob('./result_DT/*'):
Evaluator_deprecate(env,
log_path.replace('\\', '/'),
save_path,
eval_iteration=EVAL_ITERATION).evaluate()
threads_no = 1
env = EnvWrapper(threads_no, **sim_args)
#%%
env.reset()
ob_space = env.observation_space
ac_space = env.action_space
print("Observation space shape:", ob_space)
print("Action space shape:", ac_space)
assert ob_space is not None
#%%
teacher = Teacher(env, 1, Preprocessor(False))
lr = 4e-4
config = Config(buffer_size=3 * steps_per_ep * threads_no,
batch_size=32,
gamma=0.7,
tau=1e-3,
lr=lr,
update_every=1)
agent = Agent(QNetworkTf, history_length, action_size=7, config=config)
agent.set_epsilon(0.9, 0.001, EPISODE_COUNT - 2)
# Test the model
hyperparams = {**config.__dict__, **sim_args}
tags = [
"Rew: normalized speed", "Final", f"{Agent.NAME}", sim_args['scenario'],
#%%
env.reset()
ob_space = env.observation_space
ac_space = env.action_space
print("Observation space shape:", ob_space)
print("Action space shape:", ac_space)
assert ob_space is not None
#%%
class Agent:
def __init__(self, action_space):
self.action_space = action_space
self.actor_loss = 0
self.critic_loss = 0
def act(self, *args):
return np.random.sample(self.action_space)
def step(self, *args):
pass
def reset(self):
pass
#%%
teacher = Teacher(env, 1, Preprocessor(True))
agent = Agent(env.action_space)
logger = teacher.dry_run(agent, int(simTime/stepTime))
}
rng += 1
if args.beb:
sim_args["dryRun"] = True
print("Steps per episode:", steps_per_ep)
threads_no = 1
env = EnvWrapper(threads_no, **sim_args)
#%%
env.reset()
ob_space = env.observation_space
ac_space = env.action_space
print("Observation space shape:", ob_space)
print("Action space shape:", ac_space)
assert ob_space is not None
tags = [
f"{agent_name}", "Final", sim_args['scenario'],
f"Station count: {sim_args['nWifi']}",
*[f"{key}: {sim_args[key]}" for key in list(sim_args)[:3]]
]
#%%
teacher = Teacher(env, 1, Preprocessor(False))
agent = Agent(env.action_space)
logger = teacher.eval(agent, simTime, stepTime, history_length, tags)