mmscan-devkit

OpenRobotLab · Nov 29, 2024 · 11c5c4e · 11c5c4e
1 parent a84d590
commit 11c5c4e
Show file tree

Hide file tree

Showing 3 changed files with 33 additions and 17 deletions.
diff --git a/mmscan/utils/euler_utils.py b/mmscan/utils/euler_utils.py
@@ -4,8 +4,15 @@
 
 import numpy as np
 import torch
-from pytorch3d.ops import box3d_overlap
-from pytorch3d.transforms import euler_angles_to_matrix, matrix_to_euler_angles
+
+try:
+    from pytorch3d.ops import box3d_overlap
+    from pytorch3d.transforms import (euler_angles_to_matrix,
+                                      matrix_to_euler_angles)
+except ImportError:
+    box3d_overlap = None
+    euler_angles_to_matrix = None
+    matrix_to_euler_angles = None
 from torch import Tensor
 
 

diff --git a/models/LEO/model/leo_agent.py b/models/LEO/model/leo_agent.py
@@ -281,19 +281,8 @@ def forward(self, data_dict):
         bs = len(data_dict['prompt_after_obj'])
         if 'obj_tokens' not in data_dict:
 
-            try:
-                data_dict = self.pcd_encoder(data_dict)
-            except:
-                torch.save(
-                    self.state_dict(),
-                    '/mnt/petrelfs/linjingli/tmp/data/big_tmp/model_dict_1028.pth'
-                )
-                torch.save(
-                    data_dict,
-                    '/mnt/petrelfs/linjingli/tmp/data/big_tmp/debug_leo_1028.pt'
-                )
-                import IPython
-                IPython.embed()
+            data_dict = self.pcd_encoder(data_dict)
+
         data_dict['obj_tokens'] = self.pcd_proj(
             data_dict['obj_tokens'].to(device))
         # data_dict['obj_tokens'] = data_dict['obj_tokens'] + self.pcd_type_embed

diff --git a/models/LL3DA/eval_utils/evaluate_mmscan.py b/models/LL3DA/eval_utils/evaluate_mmscan.py
@@ -18,6 +18,11 @@
 
 from mmscan import QA_Evaluator
 
+model_config = {
+    'simcse': '/mnt/petrelfs/linjingli/mmscan_modelzoo-main/evaluation/pc',
+    'sbert': '/mnt/petrelfs/linjingli/mmscan_modelzoo-main/evaluation/st'
+}
+
 
 def to_mmscan_form(raw_input):
     _input = {}
@@ -39,8 +44,9 @@ def evaluate(
     logout=print,
     curr_train_iter=-1,
 ):
-    """define the mmscan_eval here."""
-    model_config = {'simcse': '', 'sbert': ''}
+
+    # prepare ground truth caption labels
+    print('preparing corpus...')
 
     evaluator = QA_Evaluator(model_config)
 
@@ -130,6 +136,20 @@ def evaluate(
             logout(f'Evaluate {epoch_str}; Batch [{curr_iter}/{num_batches}]; '
                    f'Evaluating on iter: {curr_train_iter}; '
                    f'Iter time {time_delta.avg:0.2f}; Mem {mem_mb:0.2f}MB')
+            if curr_iter % 200 == 0:
+                with open(
+                        os.path.join(args.checkpoint_dir,
+                                     f'qa_pred_gt_val_{curr_iter}.json'),
+                        'w') as f:
+                    pred_gt_val = {}
+                    for index_, scene_object_id_key in enumerate(candidates):
+                        pred_gt_val[scene_object_id_key] = {
+                            'instruction': scene_object_id_key.split('@')[1],
+                            'pred': candidates[scene_object_id_key],
+                            'gt': corpus[scene_object_id_key],
+                        }
+                    json.dump(pred_gt_val, f, indent=4)
+                print(f'save pred_gt_val {curr_iter}')
         barrier()
 
     if is_primary():