question about loading data

[Norooa]

Hello, I attempted to reproduce the code for non_modular_snn and noticed that in train mode, data is loaded only once, after which training is repeated solely with 10 images. Could you please advise if reproducing this requires modifying the method of data loading?

[Norooa]

or I can change the arg-num_query_imgs?

[Somayeh-h]

Hi @Norooa,

Thanks for your question! I set the training to use 10 images (5 per traverse, 2 traverses total) to keep the demo simple. If you check the dataset_imagenames folder, you’ll see files like nordland_imageNames_spring.txt for the Nordland dataset, listing the entire 27k images.

The network is trained using the reference set of a VPR dataset. To train with more reference images, increase the num_labels argument. Just make sure the n_e (number of output neurons) is at least equal to or greater than num_labels for good performance.

Here are some example configurations for Nordland (3,300 images) and Oxford RobotCar (450 images). You can find these in non_modular_snn/single_snn_model_processing.py as well:

Nordland:

python3 modular_snn/modular_snn_processing.py --process_mode="train" --dataset="nordland" --folder_id="NRD_SFS" --skip=8 --update_interval=250 --epochs=30 --seed=0 --num_labels=3300 --num_test_labels=3300 --num_query_imgs=3300 --n_e=3300  
python3 modular_snn/modular_snn_processing.py --process_mode="test" --dataset="nordland" --folder_id="NRD_SFS" --skip=8 --update_interval=250 --epochs=30 --seed=0 --num_labels=3300 --num_test_labels=3300 --num_query_imgs=3300 --n_e=3300  

Oxford RobotCar:

python3 modular_snn/modular_snn_processing.py --process_mode="train" --dataset="ORC" --folder_id="ORC" --skip=8 --update_interval=250 --epochs=30 --seed=0 --num_labels=450 --num_test_labels=450 --num_query_imgs=450 --n_e=450  
python3 modular_snn/modular_snn_processing.py --process_mode="test" --dataset="ORC" --folder_id="ORC" --skip=8 --update_interval=250 --epochs=30 --seed=0 --num_labels=450 --num_test_labels=450 --num_query_imgs=450 --n_e=450  

Let me know if you have more questions!
Best,
Somayeh