Add files using upload-large-folder tool

scripts/inference.py

@@ -11,7 +11,6 @@ if _SUBMIT_ROOT not in sys.path:
 
 _LM_PATH = os.path.join(_SUBMIT_ROOT, "model", "LM")
 _VL_B_PATH = os.path.join(_SUBMIT_ROOT, "model", "VL-B")
-_VL_F_PATH = os.path.join(_SUBMIT_ROOT, "model", "VL-F")
 _DIRECTION_MAP_PATH = os.path.join(_SCRIPTS_DIR, "object_direction_map.yaml")
 
 
@@ -33,7 +32,6 @@ def load_robospatialBrain_models(low_memory=False):
             "direction_map": direction_map,
             "low_memory": True,
             "_vl_b_path": _VL_B_PATH,
-            "_vl_f_path": _VL_F_PATH,
             "_current_vl": None,
             "_current_vl_type": None,
         }
@@ -41,13 +39,9 @@ def load_robospatialBrain_models(low_memory=False):
     vl_b = load_robobrain(_VL_B_PATH)
     torch.cuda.empty_cache()
 
-    vl_f = load_robobrain(_VL_F_PATH)
-    torch.cuda.empty_cache()
-
     return {
         "lm_clf": lm_clf,
         "vl_b": vl_b,
-        "vl_f": vl_f,
         "direction_map": direction_map,
         "low_memory": False,
     }
@@ -60,17 +54,16 @@ def _swap_vl(kwargs, needed_type, load_robobrain):
             kwargs["_current_vl"].clear()
             kwargs["_current_vl"] = None
             torch.cuda.empty_cache()
-        path = kwargs["_vl_f_path"] if needed_type == "F" else kwargs["_vl_b_path"]
-        kwargs["_current_vl"] = load_robobrain(path)
+        kwargs["_current_vl"] = load_robobrain(kwargs["_vl_b_path"])
         kwargs["_current_vl_type"] = needed_type
 
 
 def _run_compatibility(question, image_path, kwargs, low_memory, run_robobrain, load_robobrain):
     if low_memory:
-        _swap_vl(kwargs, "F", load_robobrain)
+        _swap_vl(kwargs, "B", load_robobrain)
         return run_robobrain(question, image_path, None, kwargs["_current_vl"], LM_classify="compatibility")
     else:
-        return run_robobrain(question, image_path, None, kwargs["vl_f"], LM_classify="compatibility")
+        return run_robobrain(question, image_path, None, kwargs["vl_b"], LM_classify="compatibility")
 
 
 def inference_single(question, image_path, kwargs):

scripts/robobrain_runner.py

@@ -122,6 +122,11 @@ def run_robobrain(question, image_path, depth_path, kwargs, return_thinking=Fals
         "Answer: yes or no\n\n"
         "Do not include additional text after this line.\n"
     )
+    _COMPATIBILITY_POST_PROMPT = (
+        "\nLook at the spatial positions and relationships of objects in the image."
+        " Based on your visual analysis, answer yes or no.\n"
+        "Your final line must be: Answer: yes or no"
+    )
 
     if LM_classify is not None:
         if LM_classify == "context":
@@ -129,7 +134,7 @@ def run_robobrain(question, image_path, depth_path, kwargs, return_thinking=Fals
             add_think = True
             q_type = "pointing"
         elif LM_classify == "compatibility":
-            post_prompt = _BINARY_POST_PROMPT
+            post_prompt = _COMPATIBILITY_POST_PROMPT
             add_think = True
             q_type = "binary"
         else: