Fix bug

dopamine/agents/fqf/configs/fqf.gin

@@ -32,7 +32,7 @@ tf.train.AdamOptimizer.epsilon = 0.0003125
 
 atari_lib.create_atari_environment.game_name = 'GAME'
 atari_lib.create_atari_environment.sticky_actions = False
-create_agent.agent_name = 'implicit_quantile'
+create_agent.agent_name = 'fqf'
 Runner.num_iterations = 200
 Runner.game = 'GAME'
 Runner.runtype = 'RUNTYPE'

dopamine/agents/fqf/fqf_agent.py

@@ -124,7 +124,10 @@ class FQFAgent(rainbow_agent.RainbowAgent):
     # have been run. This matches the Nature DQN behaviour.
     if self._replay.memory.add_count > self.min_replay_history:
       if self.training_steps % self.update_period == 0:
-        _, _, _, loss, loss1, quan_value, quan, vdiff = self._sess.run(self._train_op)
+        if 'sqloss' in self._runtype:
+            _, _, loss, loss1, quan_value, quan, vdiff = self._sess.run(self._train_op)
+        else:
+            _, _, _, loss, loss1, quan_value, quan, vdiff = self._sess.run(self._train_op)
         if self.training_steps % 50000 == 0:
             batchsize = 32
             quan_value = np.reshape(quan_value, [batchsize, self.num_tau_samples])

dopamine/discrete_domains/run_experiment.py

@@ -27,6 +27,7 @@ import csv, json, pickle
 from dopamine.agents.dqn import dqn_agent
 from dopamine.agents.implicit_quantile import implicit_quantile_agent
 from dopamine.agents.rainbow import rainbow_agent
+from dopamine.agents.fqf import fqf_agent
 from dopamine.discrete_domains import atari_lib
 from dopamine.discrete_domains import checkpointer
 from dopamine.discrete_domains import iteration_statistics
@@ -87,6 +88,10 @@ def create_agent(sess, environment, agent_name=None, summary_writer=None,
     return implicit_quantile_agent.ImplicitQuantileAgent(
         sess, num_actions=environment.action_space.n,
         summary_writer=summary_writer)
+  elif agent_name == 'fqf':
+    return fqf_agent.FQFAgent(
+        sess, num_actions=environment.action_space.n,
+        summary_writer=summary_writer)
   else:
     raise ValueError('Unknown agent: {}'.format(agent_name))