Improved readability

README.md

@@ -34,29 +34,25 @@ python3 train.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eva
 
 ## Test
 
-To evaluate the trained model on Pitts30k/Pitts250k:
+To evaluate the trained model:
 
 ```
-python3 eval.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=pitts30k --resume=/path/to/trained/model/CricaVPR.pth --infer_batch_size=8
-```
-
-To evaluate the trained model on other datasets:
-
-```
-python3 eval.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=msls --resume=/path/to/trained/model/CricaVPR.pth
+python3 eval.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=pitts30k --resume=/path/to/trained/model/CricaVPR.pth
 ```
 
 To add PCA:
 
 ```
-python3 eval.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=pitts30k --resume=/path/to/trained/model/CricaVPR.pth --infer_batch_size=8 --pca_dim=4096 --pca_dataset_folder=pitts30k/images/train
+python3 eval.py --eval_datasets_folder=/path/to/your/datasets_vg/datasets --eval_dataset_name=pitts30k --resume=/path/to/trained/model/CricaVPR.pth --pca_dim=4096 --pca_dataset_folder=pitts30k/images/train
 ```
 
 ## Trained Model
 
 You can directly download the trained model [HERE](https://drive.google.com/file/d/171lCcxZFFnvEvo88ntIwELeBegcMTEJs/view?usp=sharing).
 
-**Our another work (two-stage VPR based on DINOv2) SelaVPR achieved SOTA performance on several datasets. The code has been released [HERE](https://github.com/Lu-Feng/SelaVPR).**
+## Related Work
+
+**Our another work (two-stage VPR based on DINOv2) SelaVPR achieved SOTA performance on several datasets. The code is released at [HERE](https://github.com/Lu-Feng/SelaVPR).**
 
 ## Acknowledgements
 

eval.py

@@ -21,6 +21,9 @@
 args.save_dir = join("test", args.save_dir, start_time.strftime('%Y-%m-%d_%H-%M-%S'))
 commons.setup_logging(args.save_dir)
 commons.make_deterministic(args.seed)
+
+if args.eval_dataset_name.startswith("pitts"):     # set infer_batch_size = 8 for pitts30k/pitts250k
+    args.infer_batch_size = args.infer_batch_size // 2
 logging.info(f"Arguments: {args}")
 logging.info(f"The outputs are being saved in {args.save_dir}")
 

network.py

@@ -44,11 +44,14 @@ def __init__(self, args):
         super().__init__()
         self.backbone = get_backbone(args)
         self.aggregation = nn.Sequential(L2Norm(), GeM(work_with_tokens=None), Flatten())
+
+        # In TransformerEncoderLayer, "batch_first=False" means the input tensors should be provided as (seq, batch, feature) to encode on the "seq" dimension.
+        # Our input tensor is provided as (batch, seq, feature), which performs encoding on the "batch" dimension.
         encoder_layer = nn.TransformerEncoderLayer(d_model=768, nhead=16, dim_feedforward=2048, activation="gelu", dropout=0.1, batch_first=False)
         self.encoder = nn.TransformerEncoder(encoder_layer, num_layers=2) # Cross-image encoder
 
         if args.fc_output_dim != None:
-            # Concatenate fully connected layer to the aggregation layer
+            # Concatenate fully connected layer
             self.aggregation = nn.Sequential(self.aggregation,
                                              nn.Linear(args.features_dim, args.fc_output_dim),
                                              L2Norm())