Prevent tf.Session() from eating up all the GPU memory (#3219)

* Use soft placement and allow_growth for Session

* Move config generation to tf utils

* Re-add self.graph

Author: Ervin T

ml-agents/mlagents/tf_utils/__init__.py

@@ -1,2 +1,3 @@
 from mlagents.tf_utils.tf import tf as tf  # noqa
 from mlagents.tf_utils.tf import set_warnings_enabled  # noqa
+from mlagents.tf_utils.tf import generate_session_config  # noqa

ml-agents/mlagents/tf_utils/tf.py

@@ -23,8 +23,23 @@
 
 def set_warnings_enabled(is_enabled: bool) -> None:
     """
-    Enable or disable tensorflow warnings (notabley, this disables deprecation warnings.
+    Enable or disable tensorflow warnings (notably, this disables deprecation warnings.
     :param is_enabled:
     """
     level = tf_logging.WARN if is_enabled else tf_logging.ERROR
     tf_logging.set_verbosity(level)
+
+
+def generate_session_config() -> tf.ConfigProto:
+    """
+    Generate a ConfigProto to use for ML-Agents that doesn't consume all of the GPU memory
+    and allows for soft placement in the case of multi-GPU.
+    """
+    config = tf.ConfigProto()
+    config.gpu_options.allow_growth = True
+    # For multi-GPU training, set allow_soft_placement to True to allow
+    # placing the operation into an alternative device automatically
+    # to prevent from exceptions if the device doesn't suppport the operation
+    # or the device does not exist
+    config.allow_soft_placement = True
+    return config

ml-agents/mlagents/trainers/tf_policy.py

@@ -3,6 +3,7 @@
 
 import numpy as np
 from mlagents.tf_utils import tf
+from mlagents import tf_utils
 
 from mlagents_envs.exception import UnityException
 from mlagents.trainers.policy import Policy
@@ -69,14 +70,9 @@ def __init__(self, seed, brain, trainer_parameters):
         self.model_path = trainer_parameters["model_path"]
         self.keep_checkpoints = trainer_parameters.get("keep_checkpoints", 5)
         self.graph = tf.Graph()
-        config = tf.ConfigProto()
-        config.gpu_options.allow_growth = True
-        # For multi-GPU training, set allow_soft_placement to True to allow
-        # placing the operation into an alternative device automatically
-        # to prevent from exceptions if the device doesn't suppport the operation
-        # or the device does not exist
-        config.allow_soft_placement = True
-        self.sess = tf.Session(config=config, graph=self.graph)
+        self.sess = tf.Session(
+            config=tf_utils.generate_session_config(), graph=self.graph
+        )
         self.saver = None
         if self.use_recurrent:
             self.m_size = trainer_parameters["memory_size"]

ml-agents/mlagents/trainers/trainer.py

@@ -3,6 +3,7 @@
 from typing import Dict, List, Deque, Any
 
 from mlagents.tf_utils import tf
+from mlagents import tf_utils
 
 from collections import deque
 
@@ -70,7 +71,7 @@ def write_tensorboard_text(self, key: str, input_dict: Dict[str, Any]) -> None:
         :param input_dict: A dictionary that will be displayed in a table on Tensorboard.
         """
         try:
-            with tf.Session() as sess:
+            with tf.Session(config=tf_utils.generate_session_config()) as sess:
                 s_op = tf.summary.text(
                     key,
                     tf.convert_to_tensor(