updated gifs in README + fix for multiworker crashes + improved Atari DQN and Dueling DDQN presets

README.md

@@ -17,7 +17,12 @@ Training an agent to solve an environment is as easy as running:
 coach -p CartPole_DQN -r
 ```
 
-<img src="img/doom_deathmatch.gif" alt="Doom Deathmatch" width="267" height="200"/> <img src="img/carla.gif" alt="CARLA" width="284" height="200"/> <img src="img/montezuma.gif" alt="MontezumaRevenge" width="152" height="200"/>
+<img src="img/fetch_slide.gif" alt="Fetch Slide"/> <img src="img/pendulum.gif" alt="Pendulum"/> <img src="img/starcraft.gif" alt="Starcraft"/>
+<br>
+<img src="img/doom_deathmatch.gif" alt="Doom Deathmatch"/> <img src="img/carla.gif" alt="CARLA"/> <img src="img/montezuma.gif" alt="MontezumaRevenge"/>
+<br>
+<img src="img/doom_health.gif" alt="Doom Health Gathering"/><img src="img/minitaur.gif" alt="PyBullet Minitaur"/> <img src="img/ant.gif" alt="Gym Extensions Ant"/>
+<br><br>
 
 Blog posts from the Intel® AI website:
 * [Release 0.8.0](https://ai.intel.com/reinforcement-learning-coach-intel/) (initial release)

benchmarks/README.md

@@ -25,10 +25,11 @@ The environments that were used for testing include:
 ![#FF4040](https://placehold.it/15/FF4040/000000?text=+) *Not training*
 
 
+
 |                         |**Status**                                                |**Environments**|**Comments**|
 | ----------------------- |:--------------------------------------------------------:|:--------------:|:--------:|
-|**[DQN](dqn)**                  | ![#ceffad](https://placehold.it/15/ceffad/000000?text=+) |Atari           | Pong is not training |
+|**[DQN](dqn)**                  | ![#2E8B57](https://placehold.it/15/2E8B57/000000?text=+) |Atari           |  |
-|**[Dueling DDQN](dueling_ddqn)**| ![#ceffad](https://placehold.it/15/ceffad/000000?text=+) |Atari           | Pong is not training |
+|**[Dueling DDQN](dueling_ddqn)**| ![#2E8B57](https://placehold.it/15/2E8B57/000000?text=+) |Atari           |  |
 |**[Dueling DDQN with PER](dueling_ddqn_with_per)**| ![#2E8B57](https://placehold.it/15/2E8B57/000000?text=+) |Atari           | |
 |**[Bootstrapped DQN](bootstrapped_dqn)**| ![#2E8B57](https://placehold.it/15/2E8B57/000000?text=+) |Atari           | |
 |**[QR-DQN](qr_dqn)**            | ![#2E8B57](https://placehold.it/15/2E8B57/000000?text=+) |Atari           | |

rl_coach/graph_managers/graph_manager.py

@@ -142,6 +142,9 @@ class GraphManager(object):
         config = tf.ConfigProto()
         config.allow_soft_placement = True  # allow placing ops on cpu if they are not fit for gpu
         config.gpu_options.allow_growth = True  # allow the gpu memory allocated for the worker to grow if needed
+        config.gpu_options.per_process_gpu_memory_fraction = 0.2
+        config.intra_op_parallelism_threads = 1
+        config.inter_op_parallelism_threads = 1
 
         from rl_coach.architectures.tensorflow_components.distributed_tf_utils import create_and_start_parameters_server, \
             create_cluster_spec, create_worker_server_and_device
@@ -169,6 +172,8 @@ class GraphManager(object):
         config.allow_soft_placement = True  # allow placing ops on cpu if they are not fit for gpu
         config.gpu_options.allow_growth = True  # allow the gpu memory allocated for the worker to grow if needed
         # config.gpu_options.per_process_gpu_memory_fraction = 0.2
+        config.intra_op_parallelism_threads = 1
+        config.inter_op_parallelism_threads = 1
 
         if isinstance(task_parameters, DistributedTaskParameters):
             # the distributed tensorflow setting

rl_coach/presets/Atari_DQN.py

@@ -20,7 +20,8 @@ schedule_params.heatup_steps = EnvironmentSteps(50000)
 # Agent #
 #########
 agent_params = DQNAgentParameters()
-agent_params.network_wrappers['main'].learning_rate = 0.00025
+# since we are using Adam instead of RMSProp, we adjust the learning rate as well
+agent_params.network_wrappers['main'].learning_rate = 0.0001
 
 ###############
 # Environment #

rl_coach/presets/Atari_Dueling_DDQN.py

@@ -23,7 +23,9 @@ schedule_params.heatup_steps = EnvironmentSteps(50000)
 # Agent #
 #########
 agent_params = DDQNAgentParameters()
-agent_params.network_wrappers['main'].learning_rate = 0.00025
+
+# since we are using Adam instead of RMSProp, we adjust the learning rate as well
+agent_params.network_wrappers['main'].learning_rate = 0.0001
 agent_params.network_wrappers['main'].middleware_parameters.scheme = MiddlewareScheme.Empty
 agent_params.network_wrappers['main'].heads_parameters = [DuelingQHeadParameters()]
 agent_params.network_wrappers['main'].rescale_gradient_from_head_by_factor = [1/math.sqrt(2)]