Fix dict/tuple hybrid action space for tensorflow eager execution (#8781)

rllib/models/tf/tf_action_dist.py

@@ -230,7 +230,7 @@ class DiagGaussian(TFActionDistribution):
     @override(ActionDistribution)
     def logp(self, x):
         return -0.5 * tf.reduce_sum(
-            tf.square((x - self.mean) / self.std), axis=1) - \
+            tf.square((tf.to_float(x) - self.mean) / self.std), axis=1) - \
             0.5 * np.log(2.0 * np.pi) * tf.to_float(tf.shape(x)[1]) - \
             tf.reduce_sum(self.log_std, axis=1)
 

rllib/policy/eager_tf_policy.py

@@ -5,6 +5,7 @@ It supports both traced and non-traced eager execution modes."""
 import functools
 import logging
 import numpy as np
+from gym.spaces import Tuple, Dict
 
 from ray.util.debug import log_once
 from ray.rllib.models.catalog import ModelCatalog
@@ -586,10 +587,17 @@ def build_eager_tf_policy(name,
                 SampleBatch.NEXT_OBS: np.array(
                     [self.observation_space.sample()]),
                 SampleBatch.DONES: np.array([False], dtype=np.bool),
-                SampleBatch.ACTIONS: tf.nest.map_structure(
-                    lambda c: np.array([c]), self.action_space.sample()),
                 SampleBatch.REWARDS: np.array([0], dtype=np.float32),
             }
+            if isinstance(self.action_space, Tuple) or isinstance(
+                    self.action_space, Dict):
+                dummy_batch[SampleBatch.ACTIONS] = [
+                    flatten_to_single_ndarray(self.action_space.sample())
+                ]
+            else:
+                dummy_batch[SampleBatch.ACTIONS] = tf.nest.map_structure(
+                    lambda c: np.array([c]), self.action_space.sample())
+
             if obs_include_prev_action_reward:
                 dummy_batch.update({
                     SampleBatch.PREV_ACTIONS: dummy_batch[SampleBatch.ACTIONS],

rllib/tests/test_supported_spaces.py

@@ -48,7 +48,7 @@ OBSERVATION_SPACES_TO_TEST = {
 }
 
 
-def check_support(alg, config, check_bounds=False):
+def check_support(alg, config, check_bounds=False, tfe=False):
     config["log_level"] = "ERROR"
 
     def _do_check(alg, config, a_name, o_name):
@@ -95,7 +95,10 @@ def check_support(alg, config, check_bounds=False):
                     pass
         print(stat)
 
-    for _ in framework_iterator(config, frameworks=("tf", "torch")):
+    frameworks = ("tf", "torch")
+    if tfe:
+        frameworks += ("tfe", )
+    for _ in framework_iterator(config, frameworks=frameworks):
         # Check all action spaces.
         for a_name, action_space in ACTION_SPACES_TO_TEST.items():
             _do_check(alg, config, a_name, "discrete")
@@ -141,7 +144,7 @@ class ModelSupportedSpaces(unittest.TestCase):
 
     def test_dqn(self):
         config = {"timesteps_per_iteration": 1}
-        check_support("DQN", config)
+        check_support("DQN", config, tfe=True)
 
     def test_es(self):
         check_support(
@@ -163,11 +166,11 @@ class ModelSupportedSpaces(unittest.TestCase):
             "rollout_fragment_length": 10,
             "sgd_minibatch_size": 1,
         }
-        check_support("PPO", config, check_bounds=True)
+        check_support("PPO", config, check_bounds=True, tfe=True)
 
     def test_pg(self):
         config = {"num_workers": 1, "optimizer": {}}
-        check_support("PG", config, check_bounds=True)
+        check_support("PG", config, check_bounds=True, tfe=True)
 
     def test_sac(self):
         check_support("SAC", {}, check_bounds=True)