def test_discrete_qr_q_function(feature_size, action_size, n_quantiles,
batch_size, gamma):
encoder = DummyEncoder(feature_size)
q_func = DiscreteQRQFunction(encoder, action_size, n_quantiles)
# check output shape
x = torch.rand(batch_size, feature_size)
y = q_func(x)
assert y.shape == (batch_size, action_size)
# check taus
taus = q_func._make_taus(encoder(x))
step = 1 / n_quantiles
for i in range(n_quantiles):
assert np.allclose(taus[0][i].numpy(), i * step + step / 2.0)
# check compute_target
action = torch.randint(high=action_size, size=(batch_size, ))
target = q_func.compute_target(x, action)
assert target.shape == (batch_size, n_quantiles)
# check compute_target with action=None
targets = q_func.compute_target(x)
assert targets.shape == (batch_size, action_size, n_quantiles)
# check quantile huber loss
obs_t = torch.rand(batch_size, feature_size)
act_t = torch.randint(action_size, size=(batch_size, ))
rew_tp1 = torch.rand(batch_size, 1)
q_tp1 = torch.rand(batch_size, n_quantiles)
ter_tp1 = torch.randint(2, size=(batch_size, 1))
# shape check
loss = q_func.compute_error(obs_t,
act_t,
rew_tp1,
q_tp1,
ter_tp1,
reduction="none")
assert loss.shape == (batch_size, 1)
# mean loss
loss = q_func.compute_error(obs_t, act_t, rew_tp1, q_tp1, ter_tp1)
target = rew_tp1.numpy() + gamma * q_tp1.numpy() * (1 - ter_tp1.numpy())
y = _pick_value_by_action(q_func._compute_quantiles(encoder(obs_t), taus),
act_t)
reshaped_target = np.reshape(target, (batch_size, -1, 1))
reshaped_y = np.reshape(y.detach().numpy(), (batch_size, 1, -1))
reshaped_taus = np.reshape(taus, (1, 1, -1))
ref_loss = ref_quantile_huber_loss(reshaped_y, reshaped_target,
reshaped_taus, n_quantiles)
assert np.allclose(loss.cpu().detach(), ref_loss.mean())
# check layer connection
check_parameter_updates(q_func, (obs_t, act_t, rew_tp1, q_tp1, ter_tp1))
def test_discrete_qr_q_function(feature_size, action_size, n_quantiles,
batch_size, gamma):
encoder = DummyEncoder(feature_size)
q_func = DiscreteQRQFunction(encoder, action_size, n_quantiles)
# check output shape
x = torch.rand(batch_size, feature_size)
y = q_func(x)
assert y.shape == (batch_size, action_size)
action = torch.randint(high=action_size, size=(batch_size, ))
target = q_func.compute_target(x, action)
quantiles = q_func(x, as_quantiles=True)
assert target.shape == (batch_size, n_quantiles)
assert (quantiles[torch.arange(batch_size), action] == target).all()
# check quantile huber loss
obs_t = torch.rand(batch_size, feature_size)
act_t = torch.randint(action_size, size=(batch_size, ))
rew_tp1 = torch.rand(batch_size, 1)
q_tp1 = torch.rand(batch_size, n_quantiles)
# shape check
loss = q_func.compute_error(obs_t, act_t, rew_tp1, q_tp1, reduction='none')
assert loss.shape == (batch_size, 1)
# mean loss
loss = q_func.compute_error(obs_t, act_t, rew_tp1, q_tp1)
target = (rew_tp1.numpy() + gamma * q_tp1.numpy())
y = _pick_value_by_action(q_func(obs_t, as_quantiles=True), act_t)
taus = _make_taus_prime(n_quantiles, 'cpu:0').numpy()
reshaped_target = np.reshape(target, (batch_size, -1, 1))
reshaped_y = np.reshape(y.detach().numpy(), (batch_size, 1, -1))
reshaped_taus = np.reshape(taus, (1, 1, -1))
ref_loss = ref_quantile_huber_loss(reshaped_y, reshaped_target,
reshaped_taus, n_quantiles)
assert np.allclose(loss.cpu().detach(), ref_loss.mean())
# check layer connection
check_parameter_updates(q_func, (obs_t, act_t, rew_tp1, q_tp1))
def test_ensemble_discrete_q_function(feature_size, action_size, batch_size,
gamma, ensemble_size, q_func_type,
n_quantiles, embed_size, bootstrap):
q_funcs = []
for _ in range(ensemble_size):
encoder = DummyEncoder(feature_size)
if q_func_type == 'mean':
q_func = DiscreteQFunction(encoder, action_size)
elif q_func_type == 'qr':
q_func = DiscreteQRQFunction(encoder, action_size, n_quantiles)
elif q_func_type == 'iqn':
q_func = DiscreteIQNQFunction(encoder, action_size, n_quantiles,
embed_size)
elif q_func_type == 'fqf':
q_func = DiscreteFQFQFunction(encoder, action_size, n_quantiles,
embed_size)
q_funcs.append(q_func)
q_func = EnsembleDiscreteQFunction(q_funcs, bootstrap)
# check output shape
x = torch.rand(batch_size, feature_size)
values = q_func(x, 'none')
assert values.shape == (ensemble_size, batch_size, action_size)
# check compute_target
action = torch.randint(high=action_size, size=(batch_size, ))
target = q_func.compute_target(x, action)
if q_func_type == 'mean':
assert target.shape == (batch_size, 1)
min_values = values.min(dim=0).values
assert torch.allclose(min_values[torch.arange(batch_size), action],
target.view(-1))
else:
assert target.shape == (batch_size, n_quantiles)
# check reductions
if q_func_type != 'iqn':
assert torch.allclose(values.min(dim=0).values, q_func(x, 'min'))
assert torch.allclose(values.max(dim=0).values, q_func(x, 'max'))
assert torch.allclose(values.mean(dim=0), q_func(x, 'mean'))
# check td computation
obs_t = torch.rand(batch_size, feature_size)
act_t = torch.randint(0,
action_size,
size=(batch_size, 1),
dtype=torch.int64)
rew_tp1 = torch.rand(batch_size, 1)
if q_func_type == 'mean':
q_tp1 = torch.rand(batch_size, 1)
else:
q_tp1 = torch.rand(batch_size, n_quantiles)
ref_td_sum = 0.0
for i in range(ensemble_size):
f = q_func.q_funcs[i]
ref_td_sum += f.compute_error(obs_t, act_t, rew_tp1, q_tp1, gamma)
loss = q_func.compute_error(obs_t, act_t, rew_tp1, q_tp1, gamma)
if bootstrap:
assert not torch.allclose(ref_td_sum, loss)
elif q_func_type != 'iqn':
assert torch.allclose(ref_td_sum, loss)
# check layer connection
check_parameter_updates(q_func, (obs_t, act_t, rew_tp1, q_tp1))
def test_ensemble_discrete_q_function(
feature_size,
action_size,
batch_size,
gamma,
ensemble_size,
q_func_factory,
n_quantiles,
embed_size,
bootstrap,
use_independent_target,
):
q_funcs = []
for _ in range(ensemble_size):
encoder = DummyEncoder(feature_size)
if q_func_factory == "mean":
q_func = DiscreteMeanQFunction(encoder, action_size)
elif q_func_factory == "qr":
q_func = DiscreteQRQFunction(encoder, action_size, n_quantiles)
elif q_func_factory == "iqn":
q_func = DiscreteIQNQFunction(encoder, action_size, n_quantiles,
n_quantiles, embed_size)
elif q_func_factory == "fqf":
q_func = DiscreteFQFQFunction(encoder, action_size, n_quantiles,
embed_size)
q_funcs.append(q_func)
q_func = EnsembleDiscreteQFunction(q_funcs, bootstrap)
# check output shape
x = torch.rand(batch_size, feature_size)
values = q_func(x, "none")
assert values.shape == (ensemble_size, batch_size, action_size)
# check compute_target
action = torch.randint(high=action_size, size=(batch_size, ))
target = q_func.compute_target(x, action)
if q_func_factory == "mean":
assert target.shape == (batch_size, 1)
min_values = values.min(dim=0).values
assert torch.allclose(min_values[torch.arange(batch_size), action],
target.view(-1))
else:
assert target.shape == (batch_size, n_quantiles)
# check compute_target with action=None
targets = q_func.compute_target(x)
if q_func_factory == "mean":
assert targets.shape == (batch_size, action_size)
else:
assert targets.shape == (batch_size, action_size, n_quantiles)
# check reductions
if q_func_factory != "iqn":
assert torch.allclose(values.min(dim=0).values, q_func(x, "min"))
assert torch.allclose(values.max(dim=0).values, q_func(x, "max"))
assert torch.allclose(values.mean(dim=0), q_func(x, "mean"))
# check td computation
obs_t = torch.rand(batch_size, feature_size)
act_t = torch.randint(0,
action_size,
size=(batch_size, 1),
dtype=torch.int64)
rew_tp1 = torch.rand(batch_size, 1)
ter_tp1 = torch.randint(2, size=(batch_size, 1))
if q_func_factory == "mean":
if use_independent_target:
q_tp1 = torch.rand(ensemble_size, batch_size, 1)
else:
q_tp1 = torch.rand(batch_size, 1)
else:
if use_independent_target:
q_tp1 = torch.rand(ensemble_size, batch_size, n_quantiles)
else:
q_tp1 = torch.rand(batch_size, n_quantiles)
ref_td_sum = 0.0
for i in range(ensemble_size):
f = q_func.q_funcs[i]
if use_independent_target:
target = q_tp1[i]
else:
target = q_tp1
ref_td_sum += f.compute_error(obs_t, act_t, rew_tp1, target, ter_tp1,
gamma)
loss = q_func.compute_error(obs_t, act_t, rew_tp1, q_tp1, ter_tp1, gamma,
use_independent_target)
if bootstrap:
assert not torch.allclose(ref_td_sum, loss)
elif q_func_factory != "iqn":
assert torch.allclose(ref_td_sum, loss)
# with mask
mask = torch.rand(ensemble_size, batch_size, 1)
loss = q_func.compute_error(
obs_t,
act_t,
rew_tp1,
q_tp1,
ter_tp1,
gamma,
use_independent_target,
mask,
)
# check layer connection
check_parameter_updates(
q_func,
(obs_t, act_t, rew_tp1, q_tp1, ter_tp1, gamma, use_independent_target),
)