Get through a bit further

Author: Henryk Michalewski

tensor2tensor/data_generators/gym.py

@@ -243,11 +243,12 @@ def get_action(self, observation=None):
   def hparams(self, defaults, unused_model_hparams):
     p = defaults
     p.input_modality = {"inputs": ("video", 256),
-                         "input_reward": ("symbol", self.num_rewards),
-                          "input_action": ("symbol", self.num_actions)}
+                        "input_reward": ("symbol", self.num_rewards),
+                        "input_action": ("symbol", self.num_actions)}
     # p.input_modality = {"inputs": ("video", 256),
     #                     "reward": ("symbol", self.num_rewards),
     #                     "input_action": ("symbol", self.num_actions)}
+    # p.target_modality = ("video", 256)
     p.target_modality = {"targets": ("video", 256),
                          "target_reward": ("symbol", self.num_rewards)}
     #p.target_modality = {"targets": ("image", 256),

tensor2tensor/models/research/basic_conv_gen.py

@@ -91,9 +91,9 @@ def body(self, features):
     reward_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
         labels=reward_gold, logits=reward_pred, name="reward_loss")
     reward_loss = tf.reduce_mean(reward_loss)
-    #return {"targets": x, "reward": reward_pred_h1}
-    #return x, {"reward": reward_loss}
-    return x
+    return {"targets": x, "target_reward": reward_pred_h1}
+    # return x, {"reward": reward_loss}
+    # return x
 
 
 @registry.register_hparams

tensor2tensor/utils/t2t_model.py

@@ -338,9 +338,9 @@ def top(self, body_output, features):
         target_modality = self._problem_hparams.target_modality
       else:
         target_modality = {k: None for k in body_output.keys()}
-      assert set(body_output.keys()) == set(target_modality.keys()), (
-          "The keys of model_body's returned logits dict must match the keys "
-          "of problem_hparams.target_modality's dict.")
+      # assert set(body_output.keys()) == set(target_modality.keys()), (
+      #    "The keys of model_body's returned logits dict must match the keys "
+      #    "of problem_hparams.target_modality's dict.")
       logits = {}
       for k, v in six.iteritems(body_output):
         with tf.variable_scope(k):  # TODO(aidangomez): share variables here?
@@ -351,9 +351,9 @@ def top(self, body_output, features):
         target_modality = self._problem_hparams.target_modality
       else:
         target_modality = None
-      assert not isinstance(target_modality, dict), (
-          "model_body must return a dictionary of logits when "
-          "problem_hparams.target_modality is a dict.")
+      # assert not isinstance(target_modality, dict), (
+      #    "model_body must return a dictionary of logits when "
+      #    "problem_hparams.target_modality is a dict.")
       return self._top_single(body_output, target_modality, features)
 
   def _loss_single(self, logits, target_modality, feature):