[rllib] Apex crash when compress_observations: False (#2426)

We shouldn't try to decompress uncompressed data.

Also, fix resource requests for ddpg + GPU.

python/ray/rllib/agents/ddpg/apex.py

@@ -3,6 +3,7 @@ from __future__ import division
 from __future__ import print_function
 
 from ray.rllib.agents.ddpg.ddpg import DDPGAgent, DEFAULT_CONFIG as DDPG_CONFIG
+from ray.tune.trial import Resources
 from ray.utils import merge_dicts
 
 APEX_DDPG_DEFAULT_CONFIG = merge_dicts(
@@ -16,6 +17,7 @@ APEX_DDPG_DEFAULT_CONFIG = merge_dicts(
                 "debug": False
             }),
         "n_step": 3,
+        "gpu": False,
         "num_workers": 32,
         "buffer_size": 2000000,
         "learning_starts": 50000,
@@ -40,6 +42,15 @@ class ApexDDPGAgent(DDPGAgent):
     _agent_name = "APEX_DDPG"
     _default_config = APEX_DDPG_DEFAULT_CONFIG
 
+    @classmethod
+    def default_resource_request(cls, config):
+        cf = dict(cls._default_config, **config)
+        return Resources(
+            cpu=1 + cf["optimizer"]["num_replay_buffer_shards"],
+            gpu=cf["gpu"] and 1 or 0,
+            extra_cpu=cf["num_cpus_per_worker"] * cf["num_workers"],
+            extra_gpu=cf["num_gpus_per_worker"] * cf["num_workers"])
+
     def update_target_if_needed(self):
         # Ape-X updates based on num steps trained, not sampled
         if self.optimizer.num_steps_trained - self.last_target_update_ts > \

python/ray/rllib/agents/ddpg/ddpg.py

@@ -63,6 +63,8 @@ DEFAULT_CONFIG = with_common_config({
     "prioritized_replay_eps": 1e-6,
     # Whether to clip rewards to [-1, 1] prior to adding to the replay buffer.
     "clip_rewards": True,
+    # Whether to LZ4 compress observations
+    "compress_observations": False,
 
     # === Optimization ===
     # Learning rate for adam optimizer