Experimental settings

[WoojuLee24]

Hello, thank you for your work. I am reproducing PureACL and I have some questions.

1. GPU memory I experimented on the RTX A5000 (24G) with the default setting: batch size 3, Adam optimizer with a learning rate of 10^(-4) But, the GPU memory is exceeded even with a batch size of 2. The GPU memory fluctuates from 14G to over 24G. I am curious about how the batch size was set. (For example, whether Distributed Data-Parallel (DDP) was used or not, 3 GPUs were used with a batch size of 1 per GPU..)

2. In the paper, The feature extractor weights of PureACL were initialized with the pre-trained weights from SIBCL. Could you provide the experimental results of PureACL without pre-trained weights of SIBCL?

Thank you!

[ShanWang-Shan]

1. During training, I used a single 3090 GPU without DDP. For handling 4 surrounding image inputs, you may need to reduce the batch size.
2. I haven't conducted experiments without pre-trained SIBCL weights, but I have tested across datasets. For example, training PureACL on Ford data using SIBCL Kitti weights resulted in the same level of accuracy.

[tho-kn]

I tried using the SIBCL Kitti weights from SIBCL's repository, copied encoder, decoder, and adaptation weight, but didn't end up with a good result with first epoch. I checked out the provided weight contents, and it seems like it was trained for 1170365 steps, does it take as many steps to converge?

[ShanWang-Shan]

Yes, the training process can be quite slow, and sometimes there is an unexpected performance drop at a particular step. In such cases, I stop the training and restart from the last saved good weights, which helps to avoid this problem.