ray/rllib/agents/ppo/tests/test_ppo.py

import copy
import numpy as np
import unittest

import ray
from ray.rllib.agents.callbacks import DefaultCallbacks
import ray.rllib.agents.ppo as ppo
from ray.rllib.agents.ppo.ppo_tf_policy import ppo_surrogate_loss as \
    ppo_surrogate_loss_tf
from ray.rllib.agents.ppo.ppo_torch_policy import ppo_surrogate_loss as \
    ppo_surrogate_loss_torch
from ray.rllib.evaluation.postprocessing import compute_gae_for_sample_batch, \
    Postprocessing
from ray.rllib.models.tf.tf_action_dist import Categorical
from ray.rllib.models.torch.torch_modelv2 import TorchModelV2
from ray.rllib.models.torch.torch_action_dist import TorchCategorical
from ray.rllib.policy.sample_batch import DEFAULT_POLICY_ID, SampleBatch
from ray.rllib.utils.metrics.learner_info import LEARNER_INFO, \
    LEARNER_STATS_KEY
from ray.rllib.utils.numpy import fc
from ray.rllib.utils.test_utils import check, check_compute_single_action, \
    check_train_results, framework_iterator

# Fake CartPole episode of n time steps.
FAKE_BATCH = SampleBatch({
    SampleBatch.OBS: np.array(
        [[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8], [0.9, 1.0, 1.1, 1.2]],
        dtype=np.float32),
    SampleBatch.ACTIONS: np.array([0, 1, 1]),
    SampleBatch.PREV_ACTIONS: np.array([0, 1, 1]),
    SampleBatch.REWARDS: np.array([1.0, -1.0, .5], dtype=np.float32),
    SampleBatch.PREV_REWARDS: np.array([1.0, -1.0, .5], dtype=np.float32),
    SampleBatch.DONES: np.array([False, False, True]),
    SampleBatch.VF_PREDS: np.array([0.5, 0.6, 0.7], dtype=np.float32),
    SampleBatch.ACTION_DIST_INPUTS: np.array(
        [[-2., 0.5], [-3., -0.3], [-0.1, 2.5]], dtype=np.float32),
    SampleBatch.ACTION_LOGP: np.array([-0.5, -0.1, -0.2], dtype=np.float32),
    SampleBatch.EPS_ID: np.array([0, 0, 0]),
    SampleBatch.AGENT_INDEX: np.array([0, 0, 0]),
})


class MyCallbacks(DefaultCallbacks):
    @staticmethod
    def _check_lr_torch(policy, policy_id):
        for j, opt in enumerate(policy._optimizers):
            for p in opt.param_groups:
                assert p["lr"] == policy.cur_lr, "LR scheduling error!"

    @staticmethod
    def _check_lr_tf(policy, policy_id):
        lr = policy.cur_lr
        sess = policy.get_session()
        if sess:
            lr = sess.run(lr)
            optim_lr = sess.run(policy._optimizer._lr)
        else:
            lr = lr.numpy()
            optim_lr = policy._optimizer.lr.numpy()
        assert lr == optim_lr, "LR scheduling error!"

    def on_train_result(self, *, trainer, result: dict, **kwargs):
        stats = result["info"][LEARNER_INFO][DEFAULT_POLICY_ID][
            LEARNER_STATS_KEY]
        # Learning rate should go to 0 after 1 iter.
        check(stats["cur_lr"], 5e-5 if trainer.iteration == 1 else 0.0)
        # Entropy coeff goes to 0.05, then 0.0 (per iter).
        check(stats["entropy_coeff"], 0.1 if trainer.iteration == 1 else 0.05)

        trainer.workers.foreach_policy(self._check_lr_torch if trainer.config[
            "framework"] == "torch" else self._check_lr_tf)


class TestPPO(unittest.TestCase):
    @classmethod
    def setUpClass(cls):
        ray.init()

    @classmethod
    def tearDownClass(cls):
        ray.shutdown()

    def test_ppo_compilation_and_schedule_mixins(self):
        """Test whether a PPOTrainer can be built with all frameworks."""
        config = copy.deepcopy(ppo.DEFAULT_CONFIG)
        # For checking lr-schedule correctness.
        config["callbacks"] = MyCallbacks

        config["num_workers"] = 1
        config["num_sgd_iter"] = 2
        # Settings in case we use an LSTM.
        config["model"]["lstm_cell_size"] = 10
        config["model"]["max_seq_len"] = 20
        # Use default-native keras models whenever possible.
        # config["model"]["_use_default_native_models"] = True

        # Setup lr- and entropy schedules for testing.
        config["lr_schedule"] = [[0, config["lr"]], [128, 0.0]]
        # Set entropy_coeff to a faulty value to proof that it'll get
        # overridden by the schedule below (which is expected).
        config["entropy_coeff"] = 100.0
        config["entropy_coeff_schedule"] = [[0, 0.1], [256, 0.0]]

        config["train_batch_size"] = 128
        # Test with compression.
        config["compress_observations"] = True
        num_iterations = 2

        for fw in framework_iterator(config, with_eager_tracing=True):
            for env in ["FrozenLake-v1", "MsPacmanNoFrameskip-v4"]:
                print("Env={}".format(env))
                for lstm in [True, False]:
                    print("LSTM={}".format(lstm))
                    config["model"]["use_lstm"] = lstm
                    config["model"]["lstm_use_prev_action"] = lstm
                    config["model"]["lstm_use_prev_reward"] = lstm

                    trainer = ppo.PPOTrainer(config=config, env=env)
                    policy = trainer.get_policy()
                    entropy_coeff = trainer.get_policy().entropy_coeff
                    lr = policy.cur_lr
                    if fw == "tf":
                        entropy_coeff, lr = policy.get_session().run(
                            [entropy_coeff, lr])
                    check(entropy_coeff, 0.1)
                    check(lr, config["lr"])

                    for i in range(num_iterations):
                        results = trainer.train()
                        check_train_results(results)
                        print(results)

                    check_compute_single_action(
                        trainer,
                        include_prev_action_reward=True,
                        include_state=lstm)
                    trainer.stop()

    def test_ppo_exploration_setup(self):
        """Tests, whether PPO runs with different exploration setups."""
        config = copy.deepcopy(ppo.DEFAULT_CONFIG)
        config["num_workers"] = 0  # Run locally.
        config["env_config"] = {"is_slippery": False, "map_name": "4x4"}
        obs = np.array(0)

        # Test against all frameworks.
        for fw in framework_iterator(config):
            # Default Agent should be setup with StochasticSampling.
            trainer = ppo.PPOTrainer(config=config, env="FrozenLake-v1")
            # explore=False, always expect the same (deterministic) action.
            a_ = trainer.compute_single_action(
                obs,
                explore=False,
                prev_action=np.array(2),
                prev_reward=np.array(1.0))
            # Test whether this is really the argmax action over the logits.
            if fw != "tf":
                last_out = trainer.get_policy().model.last_output()
                if fw == "torch":
                    check(a_, np.argmax(last_out.detach().cpu().numpy(), 1)[0])
                else:
                    check(a_, np.argmax(last_out.numpy(), 1)[0])
            for _ in range(50):
                a = trainer.compute_single_action(
                    obs,
                    explore=False,
                    prev_action=np.array(2),
                    prev_reward=np.array(1.0))
                check(a, a_)

            # With explore=True (default), expect stochastic actions.
            actions = []
            for _ in range(300):
                actions.append(
                    trainer.compute_single_action(
                        obs,
                        prev_action=np.array(2),
                        prev_reward=np.array(1.0)))
            check(np.mean(actions), 1.5, atol=0.2)
            trainer.stop()

    def test_ppo_free_log_std(self):
        """Tests the free log std option works."""
        config = copy.deepcopy(ppo.DEFAULT_CONFIG)
        config["num_workers"] = 0  # Run locally.
        config["gamma"] = 0.99
        config["model"]["fcnet_hiddens"] = [10]
        config["model"]["fcnet_activation"] = "linear"
        config["model"]["free_log_std"] = True
        config["model"]["vf_share_layers"] = True

        for fw, sess in framework_iterator(config, session=True):
            trainer = ppo.PPOTrainer(config=config, env="CartPole-v0")
            policy = trainer.get_policy()

            # Check the free log std var is created.
            if fw == "torch":
                matching = [
                    v for (n, v) in policy.model.named_parameters()
                    if "log_std" in n
                ]
            else:
                matching = [
                    v for v in policy.model.trainable_variables()
                    if "log_std" in str(v)
                ]
            assert len(matching) == 1, matching
            log_std_var = matching[0]

            def get_value():
                if fw == "tf":
                    return policy.get_session().run(log_std_var)[0]
                elif fw == "torch":
                    return log_std_var.detach().cpu().numpy()[0]
                else:
                    return log_std_var.numpy()[0]

            # Check the variable is initially zero.
            init_std = get_value()
            assert init_std == 0.0, init_std
            batch = compute_gae_for_sample_batch(policy, FAKE_BATCH.copy())
            if fw == "torch":
                batch = policy._lazy_tensor_dict(batch)
            policy.learn_on_batch(batch)

            # Check the variable is updated.
            post_std = get_value()
            assert post_std != 0.0, post_std
            trainer.stop()

    def test_ppo_loss_function(self):
        """Tests the PPO loss function math."""
        config = copy.deepcopy(ppo.DEFAULT_CONFIG)
        config["num_workers"] = 0  # Run locally.
        config["gamma"] = 0.99
        config["model"]["fcnet_hiddens"] = [10]
        config["model"]["fcnet_activation"] = "linear"
        config["model"]["vf_share_layers"] = True

        for fw, sess in framework_iterator(config, session=True):
            trainer = ppo.PPOTrainer(config=config, env="CartPole-v0")
            policy = trainer.get_policy()

            # Check no free log std var by default.
            if fw == "torch":
                matching = [
                    v for (n, v) in policy.model.named_parameters()
                    if "log_std" in n
                ]
            else:
                matching = [
                    v for v in policy.model.trainable_variables()
                    if "log_std" in str(v)
                ]
            assert len(matching) == 0, matching

            # Post-process (calculate simple (non-GAE) advantages) and attach
            # to train_batch dict.
            # A = [0.99^2 * 0.5 + 0.99 * -1.0 + 1.0, 0.99 * 0.5 - 1.0, 0.5] =
            # [0.50005, -0.505, 0.5]
            train_batch = compute_gae_for_sample_batch(policy,
                                                       FAKE_BATCH.copy())
            if fw == "torch":
                train_batch = policy._lazy_tensor_dict(train_batch)

            # Check Advantage values.
            check(train_batch[Postprocessing.VALUE_TARGETS],
                  [0.50005, -0.505, 0.5])

            # Calculate actual PPO loss.
            if fw in ["tf2", "tfe"]:
                ppo_surrogate_loss_tf(policy, policy.model, Categorical,
                                      train_batch)
            elif fw == "torch":
                ppo_surrogate_loss_torch(policy, policy.model,
                                         TorchCategorical, train_batch)

            vars = policy.model.variables() if fw != "torch" else \
                list(policy.model.parameters())
            if fw == "tf":
                vars = policy.get_session().run(vars)
            expected_shared_out = fc(
                train_batch[SampleBatch.CUR_OBS],
                vars[0 if fw != "torch" else 2],
                vars[1 if fw != "torch" else 3],
                framework=fw)
            expected_logits = fc(
                expected_shared_out,
                vars[2 if fw != "torch" else 0],
                vars[3 if fw != "torch" else 1],
                framework=fw)
            expected_value_outs = fc(
                expected_shared_out, vars[4], vars[5], framework=fw)

            kl, entropy, pg_loss, vf_loss, overall_loss = \
                self._ppo_loss_helper(
                    policy, policy.model,
                    Categorical if fw != "torch" else TorchCategorical,
                    train_batch,
                    expected_logits, expected_value_outs,
                    sess=sess
                )
            if sess:
                policy_sess = policy.get_session()
                k, e, pl, v, tl = policy_sess.run(
                    [
                        policy._mean_kl_loss,
                        policy._mean_entropy,
                        policy._mean_policy_loss,
                        policy._mean_vf_loss,
                        policy._total_loss,
                    ],
                    feed_dict=policy._get_loss_inputs_dict(
                        train_batch, shuffle=False))
                check(k, kl)
                check(e, entropy)
                check(pl, np.mean(-pg_loss))
                check(v, np.mean(vf_loss), decimals=4)
                check(tl, overall_loss, decimals=4)
            elif fw == "torch":
                check(policy.model.tower_stats["mean_kl_loss"], kl)
                check(policy.model.tower_stats["mean_entropy"], entropy)
                check(policy.model.tower_stats["mean_policy_loss"],
                      np.mean(-pg_loss))
                check(
                    policy.model.tower_stats["mean_vf_loss"],
                    np.mean(vf_loss),
                    decimals=4)
                check(
                    policy.model.tower_stats["total_loss"],
                    overall_loss,
                    decimals=4)
            else:
                check(policy._mean_kl_loss, kl)
                check(policy._mean_entropy, entropy)
                check(policy._mean_policy_loss, np.mean(-pg_loss))
                check(policy._mean_vf_loss, np.mean(vf_loss), decimals=4)
                check(policy._total_loss, overall_loss, decimals=4)
            trainer.stop()

    def _ppo_loss_helper(self,
                         policy,
                         model,
                         dist_class,
                         train_batch,
                         logits,
                         vf_outs,
                         sess=None):
        """
        Calculates the expected PPO loss (components) given Policy,
        Model, distribution, some batch, logits & vf outputs, using numpy.
        """
        # Calculate expected PPO loss results.
        dist = dist_class(logits, policy.model)
        dist_prev = dist_class(train_batch[SampleBatch.ACTION_DIST_INPUTS],
                               policy.model)
        expected_logp = dist.logp(train_batch[SampleBatch.ACTIONS])
        if isinstance(model, TorchModelV2):
            train_batch.set_get_interceptor(None)
            expected_rho = np.exp(expected_logp.detach().cpu().numpy() -
                                  train_batch[SampleBatch.ACTION_LOGP])
            # KL(prev vs current action dist)-loss component.
            kl = np.mean(dist_prev.kl(dist).detach().cpu().numpy())
            # Entropy-loss component.
            entropy = np.mean(dist.entropy().detach().cpu().numpy())
        else:
            if sess:
                expected_logp = sess.run(expected_logp)
            expected_rho = np.exp(expected_logp -
                                  train_batch[SampleBatch.ACTION_LOGP])
            # KL(prev vs current action dist)-loss component.
            kl = dist_prev.kl(dist)
            if sess:
                kl = sess.run(kl)
            kl = np.mean(kl)
            # Entropy-loss component.
            entropy = dist.entropy()
            if sess:
                entropy = sess.run(entropy)
            entropy = np.mean(entropy)

        # Policy loss component.
        pg_loss = np.minimum(
            train_batch[Postprocessing.ADVANTAGES] * expected_rho,
            train_batch[Postprocessing.ADVANTAGES] * np.clip(
                expected_rho, 1 - policy.config["clip_param"],
                1 + policy.config["clip_param"]))

        # Value function loss component.
        vf_loss1 = np.power(
            vf_outs - train_batch[Postprocessing.VALUE_TARGETS], 2.0)
        vf_clipped = train_batch[SampleBatch.VF_PREDS] + np.clip(
            vf_outs - train_batch[SampleBatch.VF_PREDS],
            -policy.config["vf_clip_param"], policy.config["vf_clip_param"])
        vf_loss2 = np.power(
            vf_clipped - train_batch[Postprocessing.VALUE_TARGETS], 2.0)
        vf_loss = np.maximum(vf_loss1, vf_loss2)

        # Overall loss.
        if sess:
            policy_sess = policy.get_session()
            kl_coeff, entropy_coeff = policy_sess.run(
                [policy.kl_coeff, policy.entropy_coeff])
        else:
            kl_coeff, entropy_coeff = policy.kl_coeff, policy.entropy_coeff
        overall_loss = np.mean(-pg_loss + kl_coeff * kl +
                               policy.config["vf_loss_coeff"] * vf_loss -
                               entropy_coeff * entropy)
        return kl, entropy, pg_loss, vf_loss, overall_loss


if __name__ == "__main__":
    import pytest
    import sys
    sys.exit(pytest.main(["-v", __file__]))