def test_default_dtype(self):
buffer_size = 256
rb = ReplayBuffer(buffer_size,{"done": {}},
default_dtype = np.float32)
rb.add(done=1)
self.assertEqual(rb.sample(1)["done"][0].dtype,np.float32)
def test_multi_processing2(self):
buffer_size = 256
rb = ReplayBuffer(buffer_size,{"done": {}})
self.assertEqual(rb.get_next_index(),0)
self.assertEqual(rb.get_stored_size(),0)
p = Process(target=add,args=[rb])
q = Process(target=add,args=[rb])
p.start()
q.start()
p.join()
q.join()
self.assertEqual(rb.get_next_index() ,200)
self.assertEqual(rb.get_stored_size(),200)
def test_dtypes(self):
buffer_size = 1
for d in [
np.byte, np.ubyte, np.short, np.ushort, np.intc, np.uintc,
np.int_, np.uint, np.longlong, np.ulonglong, np.half,
np.double, np.longdouble, np.int8, np.int16, np.int32,
np.int64, np.uint8, np.uint16, np.uint32, np.uint64, np.intp,
np.float16, np.float32, np.float64
]:
if np.dtype(d).itemsize > 8:
# 128bit is not supported
continue
with self.subTest(dtype=d):
rb = MPReplayBuffer(buffer_size, {"a": {"dtype": d}})
rb.add(a=np.zeros((1, ), dtype=d))
a = rb.get_all_transitions()["a"].ravel()
self.assertEqual(a.dtype, d)
np.testing.assert_allclose(a, np.zeros((1, ), dtype=d))
def test_multi_processing(self):
buffer_size = 256
rb = ReplayBuffer(buffer_size,{"obs": {"dtype": int}})
self.assertEqual(rb.get_next_index(),0)
self.assertEqual(rb.get_stored_size(),0)
p = Process(target=add_args,args=[rb, [{"obs": i} for i in range(100)]])
p.start()
p.join()
self.assertEqual(rb.get_next_index(),100)
self.assertEqual(rb.get_stored_size(),100)
s = rb.get_all_transitions()
np.testing.assert_allclose(s["obs"].ravel(),np.arange(100,dtype=int))
def test_np_float16(self):
buffer_size = 256
rb = MPReplayBuffer(buffer_size, {"done": {"dtype": np.float16}})
self.assertEqual(rb.get_next_index(), 0)
self.assertEqual(rb.get_stored_size(), 0)
p = Process(target=add_args,
args=[rb, [{
"done": i
} for i in range(100)]])
p.start()
p.join()
self.assertEqual(rb.get_next_index(), 100)
self.assertEqual(rb.get_stored_size(), 100)
done = rb.get_all_transitions()["done"]
np.testing.assert_allclose(done.ravel(),
np.arange(100, dtype=np.float16))
self.assertEqual(done.dtype, np.float16)
def test_add(self):
buffer_size = 256
obs_shape = (15,15)
act_dim = 5
rb = ReplayBuffer(buffer_size,{"obs":{"shape": obs_shape},
"act":{"shape": act_dim},
"rew":{},
"next_obs": {"shape": obs_shape},
"done": {}})
self.assertEqual(rb.get_next_index(),0)
self.assertEqual(rb.get_stored_size(),0)
obs = np.zeros(obs_shape)
act = np.ones(act_dim)
rew = 1
next_obs = obs + 1
done = 0
rb.add(obs=obs,act=act,rew=rew,next_obs=next_obs,done=done)
self.assertEqual(rb.get_next_index(),1)
self.assertEqual(rb.get_stored_size(),1)
with self.assertRaises(KeyError):
rb.add(obs=obs)
self.assertEqual(rb.get_next_index(),1)
self.assertEqual(rb.get_stored_size(),1)
obs = np.stack((obs,obs))
act = np.stack((act,act))
rew = (1,0)
next_obs = np.stack((next_obs,next_obs))
done = (0.0,1.0)
rb.add(obs=obs,act=act,rew=rew,next_obs=next_obs,done=done)
self.assertEqual(rb.get_next_index(),3)
self.assertEqual(rb.get_stored_size(),3)
def test_buffer(self):
buffer_size = 256
obs_shape = (15,15)
act_dim = 5
N = 512
erb = ReplayBuffer(buffer_size,{"obs":{"shape": obs_shape},
"act":{"shape": act_dim},
"rew":{},
"next_obs":{"shape": obs_shape},
"done":{}})
for i in range(N):
obs = np.full(obs_shape,i,dtype=np.double)
act = np.full(act_dim,i,dtype=np.double)
rew = i
next_obs = obs + 1
done = 0
erb.add(obs=obs,act=act,rew=rew,next_obs=next_obs,done=done)
erb._encode_sample(range(buffer_size))
erb.sample(32)
erb.clear()
self.assertEqual(erb.get_next_index(),0)
self.assertEqual(erb.get_stored_size(),0)