[rllib] Use suppress_output instead of run_silent.sh script for tests (#4386)

* fix

* enable custom loss

* Update run_rllib_tests.sh

* enable tests

* fix action prob

* Update suppress_output

* fix example

* fix

python/ray/rllib/examples/custom_loss.py

@@ -51,9 +51,9 @@ class CustomLossModel(Model):
         input_ops = reader.tf_input_ops()
 
         # define a secondary loss by building a graph copy with weight sharing
+        obs = tf.cast(input_ops["obs"], tf.float32)
         logits, _ = self._build_layers_v2({
-            "obs": restore_original_dimensions(input_ops["obs"],
-                                               self.obs_space)
+            "obs": restore_original_dimensions(obs, self.obs_space)
         }, self.num_outputs, self.options)
 
         # You can also add self-supervised losses easily by referencing tensors

python/ray/rllib/offline/off_policy_estimator.py

@@ -92,8 +92,9 @@ class OffPolicyEstimator(object):
             raise ValueError(
                 "Off-policy estimation is not possible unless the inputs "
                 "include action probabilities (i.e., the policy is stochastic "
-                "and emits the 'action_prob' key). You can set "
-                "`input_evaluation: []` to resolve this.")
+                "and emits the 'action_prob' key). For DQN this means using "
+                "`soft_q: True`. You can also set `input_evaluation: []` to "
+                "disable estimation.")
 
     @DeveloperAPI
     def get_metrics(self):

python/ray/rllib/tests/run_silent.sh

@@ -1,21 +0,0 @@
-#!/bin/bash
-
-TMPFILE=`mktemp`
-DIRECTORY=`dirname $0`
-SCRIPT=$1
-shift
-
-if [ -x $DIRECTORY/../$SCRIPT ]; then
-    time $DIRECTORY/../$SCRIPT "$@" >$TMPFILE 2>&1
-else
-    time python $DIRECTORY/../$SCRIPT "$@" >$TMPFILE 2>&1
-fi
-
-CODE=$?
-if [ $CODE != 0 ]; then
-    cat $TMPFILE
-    echo "FAILED $CODE"
-    exit $CODE
-fi
-
-exit 0