async def __call__(self, iput):
sub_name = [None, 'ptr1right', 'ptr2right'][int(
(iput.cnt[0] + 1) % 3)]
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
async def __call__(self, iput):
if (iput.ret_name is None) or (iput.obs[10] !=
1) or (iput.obs[21] != 1):
sub_name = 'lshift'
else:
sub_name = None
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
async def __call__(self, iput):
if (iput.ret_name is None):
sub_name = 'ptr2right'
elif (iput.obs[10] != 1) or (iput.obs[21] != 1) or (
iput.ret_name == 'ptr2right'):
sub_name = 'bstep'
elif (iput.obs[10] == 1) and (iput.obs[21] == 1):
sub_name = None
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
async def __call__(self, iput):
if iput.cnt == 0:
# p = [0.01] * 2
# p[1] = 0.99
# a = np.random.choice(range(2), p=p)
a = 1
sub_name = f'A{a}'
else:
sub_name = None
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
async def __call__(self, iput):
digit = iput.obs.argmax()
if iput.cnt == 0:
if digit % 2 == 0:
sub_name = 'P0'
else:
sub_name = 'P1'
else:
sub_name = None
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
async def __call__(self, iput):
env = iput.env
env.pointer2 = max(env.pointer2 - 1, 1)
num1 = torch_utils.one_hot(env.list_to_sort[env.pointer1 - 1], 10)
num2 = torch_utils.one_hot(env.list_to_sort[env.pointer2 - 1], 10)
env.last_obs = [
num1,
torch.tensor([
float((env.pointer1 == 1) or (env.pointer1 == env.length))
]), num2,
torch.tensor([
float((env.pointer2 == 1) or (env.pointer2 == env.length))
])
]
return torch.empty(0)
async def __call__(self, iput):
if (int(iput.obs[:10].argmax()) > int(iput.obs[11:21].argmax())
) and (iput.ret_name != 'rshift'):
sub_name = 'swap'
elif (int(iput.obs[:10].argmax()) <= int(
iput.obs[11:21].argmax())) and (iput.ret_name !=
'rshift'):
sub_name = 'rshift'
elif (iput.ret_name == 'rshift'):
sub_name = None
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
async def __call__(self, iput):
if (iput.ret_name in [None, 'resetptr'
]) and (iput.cnt[0] !=
(2 * iput.arg - 2)):
sub_name = 'bubble'
elif (iput.ret_name == 'bubble') and (iput.cnt[0] !=
(2 * iput.arg - 2)):
sub_name = 'resetptr'
elif iput.cnt[0] == (2 * iput.arg - 2):
sub_name = None
sub_arg = torch.empty(0)
oput = DictTree(
sub_name=sub_name,
sub_arg=sub_arg,
)
return oput
def rollout(self, env):
self.eval() # set to evaluation mode
with torch.no_grad():
init_arg = env.reset()
memory = self.reset(init_arg)
trace = DictTree(
metadata=DictTree(init_arg=self._get_value(init_arg, teacher=False)),
data=DictTree(steps=[]),
)
ret_name = None
ret_val = torch.empty(0)
done = False
while not done:
obs = env.observe()
iput = DictTree(
mem_in=memory,
ret_name=ret_name,
ret_val=ret_val,
obs=obs,
)
oput = asyncio.get_event_loop().run_until_complete(self(iput))
trace.data.steps.append(DictTree(
mem_in=memory,
ret_name=ret_name,
ret_val=self._get_value(ret_val, teacher=False),
obs=self._get_value(obs, teacher=False),
mem_out=oput.mem_out,
act_name=oput.act_name,
act_arg=oput.act_arg,
))
if oput.act_name is None:
done = True
else:
memory = oput.mem_out
ret_name = oput.act_name
ret_val = asyncio.get_event_loop().run_until_complete(env.step(oput.act_name, oput.act_arg))
self._postprocess(trace)
self.train() # set to training mode
return trace
async def __call__(self, iput):
return torch.empty(0)
def reset(self):
self.digit = None
return DictTree(value=torch.empty(0))
