def test_ppo_loss(self):
self.rng_key, key1, key2, key3 = jax_random.split(self.rng_key, num=4)
B, T, A, OBS = 2, 10, 2, (28, 28, 3) # pylint: disable=invalid-name
batch_observation_shape = (-1, -1) + OBS
old_policy_params, _ = ppo.policy_net(key1, batch_observation_shape, A,
[stax.Flatten(2)])
new_policy_params, policy_apply = ppo.policy_net(
key2, batch_observation_shape, A, [stax.Flatten(2)])
value_params, value_apply = ppo.value_net(key3,
batch_observation_shape, A,
[stax.Flatten(2)])
# Generate a batch of observations.
observations = np.random.uniform(size=(B, T + 1) + OBS)
actions = np.random.randint(0, A, size=(B, T))
rewards = np.random.uniform(0, 1, size=(B, T))
mask = np.ones_like(rewards)
# Just test that this computes at all.
_ = ppo.ppo_loss(policy_apply, new_policy_params, old_policy_params,
value_apply, value_params, observations, actions,
rewards, mask)
def test_collect_trajectories(self):
observation_shape = (2, 3, 4)
num_actions = 2
policy_params, policy_apply = ppo.policy_net(
self.rng_key,
(-1, -1) + observation_shape,
num_actions,
# flatten except batch and time
# step dimensions.
[stax.Flatten(2)])
# We'll get done at time-step #10, starting from 0, therefore in 11 steps.
done_time_step = 5
env = fake_env.FakeEnv(observation_shape,
num_actions,
done_time_step=done_time_step)
num_trajectories = 5
trajectories = ppo.collect_trajectories(env,
policy_apply,
policy_params,
num_trajectories,
policy="categorical-sampling")
# Number of trajectories is as expected.
self.assertEqual(num_trajectories, len(trajectories))
# Shapes of observations, actions and rewards are as expected.
for observations, actions, rewards in trajectories:
# observations are one more in number than rewards or actions.
self.assertEqual((done_time_step + 2, ) + observation_shape,
observations.shape)
self.assertEqual((done_time_step + 1, ), actions.shape)
self.assertEqual((done_time_step + 1, ), rewards.shape)
def test_policy_net(self):
observation_shape = (3, 4)
num_actions = 2
policy_params, policy_apply = ppo.policy_net(
self.rng_key,
(-1, -1) + observation_shape,
num_actions,
# flatten except batch and time
# step dimensions.
[stax.Flatten(2)])
# Generate a batch of observations.
batch = 2
time_steps = 10
batch_of_observations = np.random.uniform(size=(batch, time_steps) +
observation_shape)
# Apply the policy net on observations
policy_output = policy_apply(policy_params, batch_of_observations)
# Verify certain expectations on the output.
self.assertEqual((batch, time_steps, num_actions), policy_output.shape)
# Also last axis normalizes to 1, since these are probabilities.
sum_actions = np.sum(policy_output, axis=-1)
self.assertAllClose(np.ones_like(sum_actions), sum_actions)
def test_value_net(self):
observation_shape = (3, 4, 5)
num_actions = 2
value_params, value_apply = ppo.value_net(
self.rng_key, (-1, -1) + observation_shape, num_actions,
[stax.Flatten(num_axis_to_keep=2)])
batch = 2
time_steps = 10
batch_of_observations = np.random.uniform(size=(batch, time_steps) +
observation_shape)
value_output = value_apply(value_params, batch_of_observations)
# NOTE: The extra dimension at the end because of Dense(1).
self.assertEqual((batch, time_steps, 1), value_output.shape)
def test_policy_and_value_net(self):
observation_shape = (3, 4, 5)
batch_observation_shape = (-1, -1) + observation_shape
num_actions = 2
pnv_params, pnv_apply = ppo.policy_and_value_net(
self.rng_key, batch_observation_shape, num_actions,
[stax.Flatten(2)])
batch = 2
time_steps = 10
batch_of_observations = np.random.uniform(size=(batch, time_steps) +
observation_shape)
pnv_output = pnv_apply(pnv_params, batch_of_observations)
# Output is a list, first is probab of actions and the next is value output.
self.assertEqual(2, len(pnv_output))
self.assertEqual((batch, time_steps, num_actions), pnv_output[0].shape)
self.assertEqual((batch, time_steps, 1), pnv_output[1].shape)
