Fix/train limit logging (#140)

* Ignore _do_training_repeats diagnostics when they're None

* Fix max-failures typo in instrument.py

* Add commented gpu dependencies to conda environment.yml

environment.yml

@@ -7,5 +7,7 @@ dependencies:
 - pip>=20.0
 - conda>=4.7
 - pyopengl-accelerate==3.1.3b1
+# - cudatoolkit==10.1.243
+# - cudnn==7.6.5
 - pip:
     - -r ./requirements.txt

examples/instrument.py

@@ -258,7 +258,7 @@ def run_example_debug(example_module_name, example_argv):
     of all cpus once ray local mode supports custom resources.
     """
 
-    debug_example_argv = ['--with-server=False', '--max-failure=0']
+    debug_example_argv = ['--with-server=False', '--max-failures=0']
     for option in example_argv:
         if '--trial-cpus' in option:
             available_cpus = multiprocessing.cpu_count()

softlearning/algorithms/rl_algorithm.py

@@ -184,8 +184,10 @@ def _train(self):
                 gt.stamp('sample')
 
                 if self.ready_to_train:
-                    update_diagnostics.append(self._do_training_repeats(
-                        timestep=self._total_timestep))
+                    repeat_diagnostics = self._do_training_repeats(
+                        timestep=self._total_timestep)
+                    if repeat_diagnostics is not None:
+                        update_diagnostics.append(repeat_diagnostics)
 
                 gt.stamp('train')