def __init__(self, model_defs, finish_reward=100, dim_size=6,n_action_steps=2, resource_size=2, dataflow="dla"):
super(MaestroEnvironment,self).__init__()
dst_path = "../../cost_model/maestro"
maestro = dst_path
self._executable = "{}".format(maestro)
random_file_name = random.randint(0, 2 ** 31)
self.random_file_name = "{}".format(random_file_name)
global action_space, action_bound, action_bottom
action_space, action_bound, action_bottom = get_action_space()
self.state = np.array([0.5]*8)
self.last_runtime = 2 ** 64
self.last_energy = 2**64
self.last_throughput = 1
self.observation = [0,0, 0,0,0,0]
self.resource_state = [0, 0]
self.consecutive_fail = 0
self.max_fail = 0
self.total_resources = action_bound
self.action_bound = action_bound[:n_action_steps]
self.total_step = len(model_defs)
self.model_defs = model_defs
self.model_defs_saved = copy.deepcopy(model_defs)
model_bound = np.max(model_defs, axis=0, keepdims=True)
self.model_defs_norm = model_defs/model_bound
self.model_defs_norm_saved = copy.deepcopy(self.model_defs_norm)
self.best_reward_whole = float("-inf")
self.best_rewards = []
self.best_sol = None
self.reward = 0
self.mae_reward = 0
self.finish_reward = finish_reward
self.mae_reward_decay = 0.
self.worst_reward = None
self.best_reward = None
self.reward_scale = 0
self.n_action_steps = n_action_steps
self.emptyinfo = [-1] * (len(self.observation) + len(self.resource_state))
self.resource_size = resource_size
self.reward_whole_eps = 0
self.dim_size = dim_size
self.reward_rec = []
self.min_reward = None
self.running_ave_reward = None
self.worst_reward_list = [None for _ in range(self.total_step)]
self.sig = 1
self.mac_rec = []
self.epoch = 0
self.sol_record = []
self.sol_reward_record = []
self.dataflow = dataflow
self.constraint_value = 2**63
self.constraint = "area"
self.prev_reward_whole_eps = 0
self.exp_table = {}
self.draw = np.arange(0, self.total_step)
self.is_gemm = False
os.makedirs(outdir_exp, exist_ok=True)
chkpt_file_t = os.path.join(outdir_exp, "{}".format("result"))
outfile = chkpt_file_t + "_o.plt"
chkpt_file = chkpt_file_t + "_c.plt"
img_file = chkpt_file_t + ".png"
log_file = chkpt_file_t + ".csv"
expLog_file = chkpt_file_t + ".log"
m_file_path = "../../data/model/"
m_file = os.path.join(m_file_path, opt.model + ".csv")
df = pd.read_csv(m_file)
model_defs = df.to_numpy()
_, dim_size = model_defs.shape
# fd = open(expLog_file, "w")
action_space, action_bound, action_bottom = get_action_space()
env = MaestroEnvironment(model_defs=model_defs,
dim_size=dim_size,
dataflow=opt.df)
env.set_fitness(opt.fitness)
env.set_constraint(opt.cstr)
constraint_temp = [
env.get_ref_constraint([action_bound[0], action_bound[1]]),
env.get_ref_constraint([action_bottom[0], action_bottom[1]]),
env.get_ref_constraint([action_bound[0], action_bottom[1]]),
env.get_ref_constraint([action_bottom[0], action_bound[1]])
]
max_constraint, min_constraint = max(constraint_temp), min(constraint_temp)
print("Max constraint: {}".format(max_constraint))
print("Min constraint: {}".format(min_constraint))
set_constraint = min_constraint + (max_constraint - min_constraint) * ratio
def __init__(self,
model_defs,
finish_reward=100,
dim_size=6,
n_action_steps=2,
resource_size=2,
dataflow="dla",
is_discrete=True,
true_continuous=False,
chkpt_file="./chkpt.plt"):
super(MaestroEnvironment, self).__init__()
dst_path = "../../cost_model/maestro"
maestro = dst_path
self._executable = "{}".format(maestro)
random.seed()
random_file_name = random.randint(0, 2**31)
self.random_file_name = "{}".format(random_file_name)
self.chkpt_file = chkpt_file
self.is_gemm = False
global action_space, action_bound, action_bottom
action_space, action_bound, action_bottom = get_action_space()
self.action_bottom = action_bound
self.state = np.array([0.5] * 8)
self.last_runtime = 2**64
self.last_energy = 2**64
self.last_throughput = 1
self.observation = [0, 0, 0, 0, 0, 0]
self.resource_state = [0, 0]
self.consecutive_fail = 0
self.max_fail = 0
self.total_resources = action_bound
self.action_bound = action_bound[:n_action_steps]
self.total_step = len(model_defs)
self.model_defs = model_defs
self.model_defs_saved = copy.deepcopy(model_defs)
model_bound = np.max(model_defs, axis=0, keepdims=True)
self.model_defs_norm = model_defs / model_bound
self.model_defs_norm_saved = copy.deepcopy(self.model_defs_norm)
self.best_reward_whole = float("-inf")
self.best_rewards = []
self.best_sol = None
self.reward = 0
self.mae_reward = 0
self.finish_reward = finish_reward
self.mae_reward_decay = 0.
self.worst_reward = None
self.best_reward = None
self.reward_scale = 0
self.n_action_steps = n_action_steps
self.emptyinfo = [-1] * (len(self.observation) +
len(self.resource_state))
self.resource_size = resource_size
self.reward_whole_eps = 0
self.dim_size = dim_size
self.reward_rec = []
self.min_reward = None
self.running_ave_reward = None
self.worst_reward_list = [None for _ in range(self.total_step)]
self.sig = 1
self.mac_rec = []
self.epoch = 0
self.sol_record = []
self.sol_reward_record = []
self.dataflow = dataflow
self.constraint_value = 2**63
self.constraint = "area"
self.prev_reward_whole_eps = 0
self.exp_table = {}
self.draw = np.arange(0, self.total_step)
self.is_discrete = is_discrete
self.true_contiunous = true_continuous and not is_discrete
self.state_size = len(self.model_defs_norm[0]) + 3
if self.is_discrete:
self.N_DISCRETE_ACTIONS = len(action_space[0])
self.action_space = spaces.MultiDiscrete(
[self.N_DISCRETE_ACTIONS, self.N_DISCRETE_ACTIONS])
else:
self.action_space = spaces.Box(low=0,
high=1,
shape=(self.n_action_steps, ),
dtype=np.float32)
self.observation_space = spaces.Box(low=0,
high=1,
shape=(self.state_size, ),
dtype=np.float32)
self.metadata = {'render.modes': []}
self.reward_range = (-float('inf'), float('inf'))
self.resources_denorm = np.zeros((self.resource_size, ))