Fix apparent bugs to improve reproducibility

README.md

@@ -82,8 +82,9 @@ h5py
 scipy
 cython
 plyfile
+numpy==1.26.4
 'trimesh>=2.35.39,<2.35.40'
-'networkx>=2.2,<2.3'
+networkx==3.2.1
 'torch=2.0.1+cu118'
 google-generativeai
 peft>=0.7.0

main.py

@@ -164,6 +164,7 @@ def make_args_parser():
     ##### Distributed #####
     parser.add_argument("--ngpus", default=1, type=int, help='number of gpus')
     parser.add_argument("--dist_url", default='tcp://localhost:12345', type=str)
+    parser.add_argument("--local-rank", type=int, default=0)
 
     args = parser.parse_args()
     args.use_height = not args.no_height
@@ -249,7 +250,7 @@ def build_dataset(args):
                     use_height=args.use_height,
                     augment=True,
                     use_additional_encoders=args.use_additional_encoders,
-                    use_rl_training=args.use_rl_training,  # Pass RL training flag
+                    # use_rl_training=args.use_rl_training,  # Pass RL training flag
                 )
             )
         datasets['test'].append(
@@ -421,4 +422,4 @@ def launch_distributed(args):
         set_start_method("spawn")
     except RuntimeError:
         pass
-    launch_distributed(args)
+    launch_distributed(args)

models/3dr1/captioner.py

@@ -396,7 +396,7 @@ def _get_instruction_response(self,
             point_cloud_color = None
             if 'point_clouds_color' in inputs:
                 point_cloud_color = inputs["point_clouds_color"][0]
-            selected_view_features, view_indices = self.view_selection(point_cloud, instruction_text, point_cloud_color)t)
+            selected_view_features, view_indices = self.view_selection(point_cloud, instruction_text, point_cloud_color)
 
             # Add view selection features to encoder hidden states
             # This is a simplified integration - in practice you might want more sophisticated fusion

script/train.generalist.sh

@@ -13,7 +13,7 @@ python -m torch.distributed.launch \
     --master_port=29500 \
     main.py \
     --checkpoint_dir ./checkpoints/fast_eval \
-    --dataset scenecold \
+    --dataset scenecold_dataset \
     --vocab "qwen/Qwen2.5-7B" \
     --qformer_vocab "bert-base-uncased" \
     --captioner 3dr1 \
@@ -53,4 +53,4 @@ python -m torch.distributed.launch \
     --eval_max_samples 1000 \
     --eval_use_fp16 \
     --eval_skip_metrics \
-    --seed 42
+    --seed 42