Report Issues in MOGFN-AL

[badeok0716]

Hi, I'm currently working on multi-objective biological sequence design problems. I appreciate you sharing the implementation of MOGFN-AL!

I found two issues in your code as follows.

1. mogfn_seq.py#L547, mogfn_seq.py#L572
   pos_logits[:, lens+1] = -1000 # can't change last token
pos_logits_temp[:, lens+1] = -1000 # can't change last token
   These lines might prevent the modification of non-last tokens if lens contains more than one distinct value. I think that these lines should be changed into
   pos_logits[np.arange(len(lens)), lens+1] = -1000 # can't change last token
pos_logits_temp[np.arange(len(lens)), lens+1] = -1000 # can't change last token

2. mogfn_seq.py#526-L635
   command:
   python scripts/black_box_opt.py optimizer=mogfn_seq task=regex tokenizer=protein optimizer.encoder_obj=mlm surrogate=multi_task_exact_gp acquisition=nehvi optimizer.use_acqf=True optimizer.pref_cond=False optimizer.num_opt_steps=1500 optimizer.max_len=30 optimizer.sample_beta=32 optimizer.beta_sched=1 optimizer.pref_alpha=1 wandb_mode=disabled
   I slightly changed the code to utilizedtype=torch.double and I ran the command on the machine with RTX3090 GPU.
   Most of the trials had no problem, but some of the trials were stuck at mogfn_seq.py#L585. This error occured becausepos_logits had values near -1000 for the problematic trials. I think that it is better to change -1000 into much larger value like -1e6 for better stability of the code.

• Minor issue about max_len.
  mogfn_seq.py#L94 and
  mogfn_seq.py#L538 seem to do the same thing twice. I think that it is better to remove one of the -2.

I would greatly appreciate your input to validate my concerns. Thanks!

[MJ10]

Hi @badeok0716. Apologies for the late reply, I did not see this issue earlier.
Thanks a lot for the spotting these issues, I agree with all the suggestions you made! Please feel free to raise a PR at https://github.com/MJ10/mogfn-al/ if you would like to, or I can make the changes.