Uncleaned LoRA components cause errors during training

[ShenhanQian]

Hi, thanks for sharing the complete code base!

When I tried to train the model, the following two places caused errors because either args.load_lora or args.lora is undefined for MonSt3R:

https://github.com/Junyi42/croco/blob/6f1362bcd0a76821ca3091dc2d0c3253e95c7794/utils/misc.py#L307-L310

https://github.com/Junyi42/croco/blob/6f1362bcd0a76821ca3091dc2d0c3253e95c7794/utils/misc.py#L341-L344

[lifuguan]

Same problem here. Additionally, it seems the paper doesn’t mention any LoRA configurations or training details. I’m curious to know more about that.

[Junyi42]

Hi @ShenhanQian @lifuguan,

Thanks for pointing out the issue. The LORA configuration is an experimental feature and not applied in the paper. We have updated the code in croco to fix this issue: Junyi42/croco#1.

Best

[ShenhanQian]

Thanks for the update.