cranial-xix / famo Goto Github PK

I tried using FAMO and it seems to work well, for batches with no labels for a task I am setting the task loss to 0. Any better approach you would reccomend?

Hi there,

I am trying to reproduce the results on QM9. However, the $\Delta_m%$ result I obtained is quite different from the result presented in Table 2 from FAMO paper. I directly ran the scripts provided in this repo, with $\gamma=0.001$ and random seed 42. So, could you kindly inform me if there are any additional modifications required in the scripts to reproduce the result?

Hi, thanks for your time. I'm running FAMO on the MTRL setting according to the installation you provided. Currently I got the following errors

hydra.errors.HydraException: Error calling 'mtrl.experiment.metaworld.Experiment' : Error calling 'metaworld.MT10' : Error loading module 'metaworld.MT10'

Did you have this kind of error before and could you share the solutions? Do I need to change some configs or .yaml files when I run the code? Thanks very much!

Hi, as I noticed in your previous issue, you said "--scale-y=True" would affect FAMO's performance. Will it affect the other methods' performance, like MGDA, LS, or IMTL in your code?

Thanks for the great work and the code. While I was reproducing the results, I found several minor issues.

1. FAMO/experiments/cityscapes/trainer.py

   Line 172 in fe9c98f

   test_data, test_label, test_depth = test_dataset.next()

   
   For the latest version of torch, I got some errors from "test_dataset.next()", and "next(test_dataset)" seemed to work well.

2. FAMO/experiments/cityscapes/trainer.py

   Line 209 in fe9c98f

   f"TEST: {avg_cost[epoch, 7]:.4f} {avg_cost[epoch, 8]:.4f} {avg_cost[epoch, 9]:.4f} | "

   
   If I understand correctly, the indices for this line should be 6, 7, 8, instead of 7, 8, 9?

Hi, Thank you for your excellent work, I am still relatively new to AI and have a question about your famo.py code:
Why should clip_grad_norm_ before loss.backward()?

Hi, could you provide the single-task results on the 40-task CelebA dataset? I am running experiments of an MTL method and want to compare the $\Delta_m$ with FAMO and other baselines. Unlike Cityscapes, NYUv2, and QM9, the single-task results on CelebA seem not to be contained in the codes in this repo. Although I can run single-task experiments, the obtained results may be slightly different from the results in your paper. Then, the comparison of $\Delta_m$ may be unfair. I would greatly appreciate it if you could provide me with more details about the single-task results on CelebA. Thanks in advance!