Question Regarding Parameter Count of the CViT-S Model #6
Description
Hi @sifanexisted ,
First of all, thank you so much for your excellent work on CVIT: CONTINUOUS VISION TRANSFORMER FOR OPERATOR LEARNING and for making the code publicly available! It's a fantastic resource for the community.
I'm currently working with your implementation and trying to align it with the details reported in the paper. I noticed a small discrepancy regarding the parameter count for the CVIT-S model and was hoping you could provide some clarification.
In Table 1 of the paper, the parameter count for the CViT-S model is listed as 13M.
However, when I traing the model（NS） using the provided codebase ( using the command CUDA_VISIBLE_DEVICES=0 python3 main.py --config=configs/cvit_small_8x8.py，CUDA_VISIBLE_DEVICES=0 python3 main.py --config=configs/cvit_small_8x8.py --config.mode=eval), I calculate a total of approximately 17M parameters.
I0619 10:36:48.839630 139871536145472 standard_logger.py:34] {'step': 199999, 'event_type': 'restore', 'directory': '/home/d/code/cvit/cvit-main/ns/ns_cvit_small_8x8/ckpt', 'checkpointer_start_time': 1750300607.4151492, 'checkpointer_duration_secs': 1.424405813217163, 'checkpoint_manager_start_time': 1750300607.4150863, 'checkpoint_manager_duration_secs': 1.4244709014892578}
Total number of parameters: 17,094,403
My question is: Could you please help clarify how the 13M parameter count reported in the paper was calculated?
Any guidance on this would be greatly appreciated as it would help me ensure I am working with the correct model setup for my experiments.
Thank you again for your time and for your great contribution to the field!
Best regards,