found bug with interpret and fixed it!

predicators/datasets/generate_atom_trajs_with_vlm.py

@@ -831,7 +831,6 @@ def _generate_ground_atoms_with_vlm_oo_code_gen(
             ground_atoms = utils.abstract(state, candidates | known_predicates)
             ground_atoms_traj.append(ground_atoms)
         ground_atoms_trajs.append(ground_atoms_traj)
-    import ipdb; ipdb.set_trace()
     return ground_atoms_trajs
 
 
@@ -870,12 +869,18 @@ def _create_prompt_from_image_option_traj(
     for i, a in enumerate(image_option_traj.actions):
         state = image_option_traj.states[i]
         demo_str.append(f"state {i}:")
-        demo_str.append(state.dict_str(indent=2, object_features=True))
+        # NOTE: it's important to set the round_feat_vals argument to False
+        # here. If we set it to True, then the VLM might mistakenly propose
+        # predicates that work given rounding, but fail otherwise.
+        # So for instance, a predicate classifier that does `== 0` would
+        # work for a value 0.00123 rounded to a single decimal place,
+        # but wouldn't actually work when deployed on the number 0.00123!
+        demo_str.append(state.dict_str(indent=2, object_features=True, round_feat_vals=False))
         demo_str.append(f"action {i}: {a.name}")
     num_states = len(image_option_traj.states)
     state = image_option_traj.states[-1]
     demo_str.append(f"state {num_states}:")
-    demo_str.append(state.dict_str(indent=2, object_features=True))
+    demo_str.append(state.dict_str(indent=2, object_features=True, round_feat_vals=False))
     demo_str_ = '\n'.join(demo_str)
     template = template.replace("[DEMO_TRAJECTORY]", demo_str_)
 

predicators/structs.py

@@ -215,14 +215,14 @@ def pretty_str(self) -> str:
         suffix = "\n" + "#" * ll + "\n"
         return prefix + "\n\n".join(table_strs) + suffix
 
-    def dict_str(self, indent: int = 0, object_features: bool = True) -> str:
+    def dict_str(self, indent: int = 0, object_features: bool = True, round_feat_vals: bool = True) -> str:
         """Return a dictionary representation of the state."""
         state_dict = {}
         for obj in self:
             obj_dict = {}
             if obj.type.name == "robot" or object_features:
                 for attribute, value in zip(obj.type.feature_names, self[obj]):
-                    if isinstance(value, (float, int, np.float32)):
+                    if isinstance(value, (float, int, np.float32)) and round_feat_vals:
                         value = round(float(value), 1)
                     obj_dict[attribute] = value
             obj_name = obj.name

setup.py

@@ -38,6 +38,8 @@
         "ImageHash",
         "google-generativeai",
         "tenacity",
+        "opencv-python",
+        "torchvision"
     ],
     include_package_data=True,
     extras_require={